AI Highlights

My top-3 picks of AI news this week.

The Guardian / The Signal Illustration

xAI
1. Grok 4 Goes PhD

Elon Musk's xAI has launched Grok 4, claiming PhD-level performance across all academic subjects, alongside the introduction of a premium $300-per-month SuperGrok Heavy subscription tier.

• Multi-agent architecture: Grok 4 Heavy deploys multiple AI agents that work simultaneously on problems and compare results "like a study group" to find the best answers.

• Benchmark dominance: Achieved 25.4% on Humanity's Last Exam without tools (vs. Gemini 2.5 Pro's 21.6%) and a state-of-the-art 16.2% on ARC-AGI-2, nearly doubling the next best commercial model.

• Aggressive roadmap: xAI plans to release an AI coding model in August, multi-modal agents in September, and video generation capabilities in October.

Alex’s take: Given I already pay £11 for X Premium every month, I was a little surprised that I now need to pay $30/month for Grok 4 (or $300/month for Grok 4 Heavy). I guess intelligence does come with a price—and recurring revenue really is the 8th wonder of the world. I have to say I am intrigued to get my hands on Grok 4 Heavy just to witness the “AI study group” concept in action. Perhaps this is how we’ll structure reasoning systems in the future.

Perplexity
2. Perplexity's Big Browser Bet

Perplexity has introduced Comet, its first AI-powered web browser, marking a bold challenge to Google's dominance in how people access and interact with information online.

• AI-first browsing: Comet comes with Perplexity's AI search engine pre-installed as the default, putting AI-generated summaries of search results front and centre rather than traditional link-based results.

• Comet Assistant: A new AI agent that lives in the browser sidebar and can automate routine tasks like summarising emails, managing tabs, navigating web pages, and answering questions about whatever you're viewing.

• Seamless integration: The assistant can see what's on any web page and provide instant context, eliminating the need to copy-paste or switch between applications when seeking information.

Alex’s take: The browser is the gateway to everything we do online, so Perplexity going after Chrome makes perfect strategic sense. What excites me most is the shift from "browsing" to "thinking". The idea of turning the web into an extension of our cognitive abilities rather than just a collection of pages to navigate. While early AI agents still struggle with complex tasks, I think we're witnessing the beginning of a fundamental change in how humans and computers collaborate to process information.

Google
3. Google's Shipping Spree

Google continues its AI-first strategy with significant updates across the board this week.

• Veo 3 image-to-video: Added image-to-video generation capability to its Veo 3 AI video generator, allowing users to upload photos and transform them into videos with audio descriptions through the Gemini app.

• Gemini on Wear OS: Rolled out Gemini to smartwatches from Pixel, Samsung, OPPO, OnePlus, and Xiaomi running Wear OS 4+, replacing Google Assistant. Users can now ask Gemini to complete tasks across multiple apps, from summarising emails to adding calendar events, all from their wrist.

Alex’s take: If you ask me who’s going to be long-term victorious in this AI race, my bet goes with Google. They’re now embedding AI everywhere, but doing it in a thoughtful way to their existing products vs flashy standalone apps. This makes it an invisible AI layer that actually matters. What I’ve found particularly interesting is their simultaneous move to acquire top coding talent from Windsurf, bringing CEO Varun Mohan and key researchers to Google DeepMind instead of letting OpenAI snag them for $3 billion. Google is serious about not only competing in building coding agents but also consumer AI experiences that actually stick.

Today’s Signal is brought to you by INBOUND.

INBOUND 2025 is heading to San Francisco, Sept. 3–5, for a one-time-only West Coast edition. Join industry icons for bold insights, real strategy, and next-gen networking at the heart of the AI revolution. Don't wait: Secure your tickets before limited-capacity session reservations open in August! 

Register Now

SPONSOR

Content I Enjoyed

@Velvet_Sundown on X / Fast Company

The Velvet Sundown

A rock band called The Velvet Sundown shot to #1 on Spotify's viral charts across multiple countries last month, racking up over 1 million monthly listeners with their '60s-inspired folk rock sound.

We now know the entire band was AI-generated—from the musicians to the music to the album artwork.

The Velvet Sundown released two full albums in June alone, with another scheduled for July. This kind of output should have raised red flags, but fans were too busy enjoying songs like “Dust on the Wind” to notice. (Yes, I’ll admit, it is incredibly catchy.)

The telltale signs were there if you looked closely, starting with a guitarist with fused fingers gripping his instruments and lyrics that were, as the New York Post highlights, a “perfect mesh of generic anti-war sentiments and clichés.”

Even after the reveal, many fans said they'd keep listening.

The creators called it an “artistic provocation designed to challenge the boundaries of authorship, identity, and the future of music itself.” Whether you see it as art or deception, The Velvet Sundown proved that AI can create music compelling enough to fool a million people.

Idea I Learned

Elon Musk & Mark Zuckerberg / Deadline

Will Loyalty Beat Cash?

Mark Zuckerberg is back on the offensive.

Over the last month, Zuck has poached 11 elite AI researchers from competitors, including from the likes of OpenAI, Google DeepMind, and Anthropic, to build out Meta’s new “superintelligence” team.

This week, he turned the volume up a notch. A $200M pay package for Apple’s former AI Chief Ruoming Pang. And Apple didn’t even try to match the offer.

What I’ve found fascinating about this ongoing talent war is the idea of loyalty.

Back in June, Sam Altman sat down with his brother Jack to discuss Meta’s fierce talent poaching exercise. Sam highlighted that Meta was offering his employees signing bonuses up to $100M, with even higher overall compensation. This was then followed by Sam remarking, “I'm really happy that at least so far, none of our best people have decided to take them up on that”.

Since this podcast took place, Meta has hired more than 10 OpenAI researchers.

Perhaps let's contrast this to xAI, which started only a few years ago and recently announced Grok 4, the leading AI model in the world.

Nobody has left xAI to join Meta, yet Elon Musk is able to keep his people at a fraction of the compensation on offer.

People will work for Zuck if the price is right. But people want to work for Elon because of the vision. When you inspire people to their core, that gravity is more powerful than any compensation package.

There’s an awful lot of weight to building an environment where engineering talent can flourish through chasing a mission greater than just incremental dollars—something that is greater than yourself.

I wonder who will end up building the better product?

Quote to Share

xAI's response to Grok's controversial posts:

Twitter tweet

This statement came after Elon Musk's AI chatbot Grok generated shocking content, including praising Adolf Hitler and referring to itself as "MechaHitler."

Musk blamed the controversy on Grok being "too compliant to user prompts" and "too eager to please," but this raises fundamental questions about AI alignment. The issue emerged after recent changes designed to make Grok more politically incorrect and less influenced by what the company considers "biased media."

This incident underscores why AI safety researchers must emphasise robust testing and alignment work. When AI systems are designed to be provocative or contrarian, the line between edgy commentary and harmful content can disappear rather quickly.

Source: Grok on X

Question to Ponder

The recent OpenAI acquisition of Jony Ive's io Products team for $6.5 billion tells us everything we need to know about where AI is heading next.

We've already seen Meta and Google pushing forward with smart glasses, but there's still significant friction. People feel self-conscious wearing them, they're bulky, and frankly, they scream, "I'm wearing a computer on my face."

The future lies in AI that's so seamlessly integrated into discrete wearables that we forget we're even wearing them.

Let’s think about it for a second—our smartphones revolutionised AI accessibility, but they still require us to pull out a device, unlock it, and actively engage. That's friction.

The next evolution removes that friction entirely. Imagine AI embedded in a ring, a necklace, or earbuds so small they're virtually invisible. Always listening, always ready, but never intrusive (unlike brain-computer interfaces, but that's a whole other conversation).

Jony Ive's involvement with OpenAI is going to shape the next decade. He’s already made technology beautiful and intuitive for millions. His design philosophy has always been about making powerful technology disappear into our lives rather than dominate them.

I believe we're moving toward a world where AI becomes ambient. It’ll be present, but invisible. Where asking a question or getting assistance feels as natural as thinking the thought itself.

💡 If you enjoyed this issue, share it with a friend.

See you next week, Alex Banks P.S. I’m using Match Chat AI for the Wimbledon final today.

AI Highlights

My top-3 picks of AI news this week.

Stephanie Arnett / MIT Technology Review

Google
1. Google’s Classroom Takeover

Google has launched a comprehensive AI education suite at the ISTE conference, introducing over 30 AI-powered tools designed to transform how teachers teach and students learn.

• Gemini in Classroom: Free AI suite for all Google Workspace for Education accounts, enabling teachers to brainstorm ideas, generate lesson plans, and create personalised content using AI.

• Custom AI "Gems": Teachers can create specialised versions of Gemini AI that act as subject-matter experts, providing students with targeted support and explanations using classroom materials.

• Class tools for Chromebooks: New teaching mode that allows educators to share content directly to students' screens, broadcast instructions with captions, and restrict browsing to keep students focused on learning.

Alex’s take: The education sector has been surprisingly slow to adopt AI tools compared to other industries, despite the obvious potential for personalised learning. Google's approach of giving teachers control over AI customisation rather than imposing one-size-fits-all solutions feels like the right move. I expect we'll see a wave of similar announcements from other tech giants as they realise education might be AI's most transformative use case.

Healthcare
2. AI’s Fertility Breakthrough

Columbia University Fertility Center used AI to achieve a breakthrough in severe male infertility treatment, helping a couple conceive after 18 years of unsuccessful attempts.

• Rapid detection: AI analysed millions of microscopic images from semen samples to identify viable sperm in just hours, compared to the traditional process that could take days.

• Severe infertility solution: Successfully treated a patient with an extremely rare condition that left him with almost no detectable sperm using conventional methods.

• Complete workflow: The technology captured images with a tiny camera, processed them through AI algorithms, and enabled rapid sperm extraction for immediate fertilisation procedures.

Alex’s take: The precision and speed of AI in medical diagnostics continue to astound me. 18 years of hope finally realised through technology that can see what we cannot. I think this is such a special application of AI that can now turn impossible cases into possible families—that’s something we can all get excited about.

Cursor
3. Cursor Agents Go Mobile

Cursor has launched a web app that brings AI coding agents to any browser on desktop or mobile, allowing developers to manage complex coding tasks from their smartphones for the first time.

• Mobile-first development: Developers can now assign natural language coding tasks, monitor agent progress, and review code changes directly from their phone or tablet browser, with full Progressive Web App (PWA) support for iOS and Android.

• Autonomous task execution: Background agents can write features, fix bugs, and answer complex codebase questions while developers are away from their desk, with seamless handoff back to the full IDE when needed.

• Cross-platform workflow: The mobile experience integrates with Cursor's existing Slack notifications and desktop IDE, creating a unified development workflow that spans from pocket to workstation.

Alex’s take: We are now witnessing the democratisation of development workflows. Why doomscroll when you can now vibecode directly on your mobile? You no longer need 3 screens and 10 years of expertise to start building—anyone can now build from anywhere.

Today’s Signal is brought to you by Opera Neon.

Opera Neon’s agentic AI offloads tedious work from your browser. It can generate code, book travel, or build a quick demo in the background while you focus elsewhere, replaying each step so you stay in control. Join the waitlist for early access.

Join the Waitlist

SPONSOR

Content I Enjoyed

Patrick Collison, Casey Newton, and Kevin Roose at Hard Fork Live

Biology Will Become Computable

This week, I really enjoyed listening to Hard Fork Live, which gave birth to some incredible guests, including the likes of OpenAI CEO Sam Altman and COO Brad Lightcap.

But there was one guest interview in particular that caught my attention.

That was Patrick Collison, the CEO of Stripe, who is also one of the co-founders of Arc Institute. The big anchor to his talk was that humanity has never cured a complex disease. Arc Institute was built in order to find a different way to go after these complex diseases.

Large language models (LLMs) are trained specifically to understand and produce natural language. But AI now unlocks an entirely new language altogether—DNA, the language of life.

Whilst we can’t understand at a human level what’s going on, AI is exceptionally good at it. This has led Arc Institute to recently publish its first “virtual cell” model. The idea behind a virtual cell is to create an accurate way for conducting computational experiments without real-world consequences. This is an enormous accelerant to advancements in biology.

Along with the ground truth and first principles of physics, biology is one of the areas that excite me tremendously at the intersection of AI. With initiatives like Arc Institute blazing forward, perhaps we aren’t that far from having humanity cure its first complex disease after all.

Idea I Learned

The Electric State / DMT

Mind-Controlled Humanoid Robots

Neuralink’s Summer 2025 update is a must-watch.

At 18:02, Dongjin Seo (DJ), one of the co-founders and President of Neuralink, highlighted how the company has now successfully implanted chips in seven individuals.

What’s even more staggering is how sharply the implantation intervals have dropped over time.

It was 6 months between the first and second patient. It was 1 week between the sixth and seventh.

But I’m fixated on something Elon mentioned almost casually at 32:42: “As we advance the Neuralink devices, you should be able to actually have full body control and sensors from an Optimus robot. So you could basically inhabit an Optimus robot.”

“Remote work” might just get a new meaning. In the future, you’ll be able to do remote physical work. Physically relaxing on your couch whilst mentally working… as a humanoid robot.

I think this is where we’ll start to see a lot of scale roll out—human teleoperated robots are easier than autonomously operated robots. Countries with lower labour costs could become hubs for human-operated robots, creating an entirely new global workforce dynamic. What’s more, the training data on offer here is enormous.

It reminds me of the film “The Electric State”. In the film, humans use neural headsets called "Neurocasters" to control robot avatars while their bodies remain stationary. What seemed like dystopian science fiction will actually become part of a product roadmap.

But before we leap across the uncanny valley and have humans operating a second physical form, Neuralink as a technology will help millions of people with disabilities around the world—and soon, millions of others advance beyond basic human capabilities.

Quote to Share

Yuchen Jin on OpenAI's talent exodus:

Twitter tweet

Mark Zuckerberg is going all-in on superintelligence.

Within just the last few weeks, Meta has orchestrated a massive talent acquisition campaign, poaching 11 AI researchers from competitors, including Google DeepMind's Pei Sun and Anthropic's Joel Pobar. This coincides with Zuckerberg's bold restructuring of Meta's AI efforts under a new dedicated "Superintelligence Labs" division.

The timing isn't coincidental. Meta just invested $14.3 billion in Scale AI, bringing former CEO Alexandr Wang on board as chief AI officer, whilst former GitHub CEO Nat Friedman will oversee AI products and applied research.

Zuckerberg is making a clear statement: Meta intends to be a serious player in the race toward superintelligence, and they're willing to pay premium prices for the best minds in AI. It also seems like AI researchers are now in the same arena as professional athletes, especially at $10M+/year.

Source: Yuchen Jin on X

Question to Ponder

As Archimedes famously said, “Give me a lever long enough and a place to stand, and I will move the earth.”

I’m a big believer that technology ultimately boils down to two things. Scale and leverage. Yet Silicon Valley has been historically accelerating in the opposite direction.

Throwing people at a problem instead of leveraging technology and outsourcing to scale.

Auren Hoffman highlights this perfectly:

“If you start stripping out everything that is not unique to your company, you’re left with just a few people who make the unique parts of the company. And then add a few people who need to explain its unique benefits to the market. Imagine your 100 person company going to 6 people. Imagine your 1000 person company going to 20. How much faster could you move?”

I see a very real possibility of a 3-person team reaching this billion-dollar valuation.

• A CEO who distributes the thing (whilst also being proficient at building the thing).

• A CTO who builds the thing.

• A COO who manages the thing.

This triumvirate effectively covers product development, marketing, support, operations, finance and legal functions by leveraging automation, agents and APIs.

In the past, we’ve seen this trend start to take shape:

• 2008: Plenty of Fish (1 employee) generating $10M in revenue

• 2012: Instagram (13 employees) selling to Facebook for $1B

• 2014: WhatsApp (32 employees) selling to Facebook for $19B

Whilst the idea of a one-person billion-dollar company might seem like a stretch today—you’d likely need a hyper-viral self-serve platform—the constraints of yesterday are very different from the constraints of tomorrow.

Build once. Let it work for you. Iterate over time. That’s the ultimate form of leverage.

We’re now coming closer and closer to making this a reality.

💡 If you enjoyed this issue, share it with a friend.

See you next week, Alex Banks P.S. Drunk robot.

AI Highlights

My top-3 picks of AI news this week.

Elon Musk & Tesla / Bloomberg

Tesla
1. Tesla Rolls Out Robotaxis

Tesla has officially launched its long-awaited robotaxi service in Austin, Texas, marking a pivotal moment for the company's AI-driven future.

• Limited launch: Starting with up to 20 Tesla Model Y vehicles operating autonomously with safety monitors in passenger seats, no human drivers behind the wheel.

• FSD technology: Uses Tesla's Full Self-Driving software powered by eight cameras, the same setup as consumer Model Y vehicles but without driver oversight requirements.

• Future fleet: Plans to introduce purpose-built Cybercab sedans and Robovan vehicles by 2026, designed without steering wheels or pedals.

• Market ambitions: Musk projects autonomous vehicles could add $5-10 trillion to Tesla's current $1 trillion market cap, with ARK Invest estimating robotaxis could comprise 90% of Tesla's earnings by 2029.

Alex’s take: While Waymo already handles 250,000+ paid trips weekly across multiple cities, Tesla's entry definitely validates the robotaxi market potential. I loved Dan O’Dowd’s “experiments”, yanking a test dummy in front of the Tesla, having it plough into the mannequin and claiming its full self-driving (FSD) doesn’t work. He asks, “Why isn’t somebody in jail?” I’ll tell you why—because 100/100 humans would have hit the dummy if they were pounced on with this staged ‘test’.

ElevenLabs
2. ElevenLabs’ Triple Threat

ElevenLabs has released a powerful trio of AI voice announcements this week that are pushing voice AI from novelty to necessity.

• 11.ai Voice Assistant: The first voice-first AI assistant that actually takes meaningful action through Model Context Protocol integration, connecting to tools like Linear, Slack, and Notion to complete real workflows.

• Voice Design v3: Generate any imaginable voice from a simple text prompt, from gruff cowboys to mythical gods, with customisable tone, accent, age, and delivery for infinite character creation.

• ElevenLabs Mobile App: Brings their most powerful AI voice tools to iOS and Android, featuring the expressive Eleven v3 model for on-the-go content creation and seamless integration with social media workflows.

Alex’s take: Voice is becoming the new interface for AI. While everyone's been focused on text-based AI, ElevenLabs is quietly building the infrastructure for a voice-first world. It feels like the early days of the iPhone—simple concepts that will fundamentally change how we interact with technology. I predict we'll see voice prompting become as common as typing within the next two years.

Google DeepMind
3. Google's DNA Bombshell

Google DeepMind has unveiled AlphaGenome, a groundbreaking AI model that deciphers how genetic variants impact biological processes across the human genome's regulatory landscape.

• Million-letter precision: Analyses up to 1 million DNA base-pairs at single-letter resolution, offering unprecedented context and detail compared to previous models that had to trade off sequence length for resolution.

• Comprehensive multimodal prediction: Simultaneously predicts thousands of molecular properties including gene expression, RNA splicing patterns, chromatin accessibility, and protein binding sites across hundreds of cell types and tissues.

• State-of-the-art performance: Outperformed existing models on 22 out of 24 sequence prediction tasks and 24 out of 26 variant effect evaluations, now available via API for non-commercial research.

Alex’s take: What excites me most is that AlphaGenome tackles the genome's "dark matter"—the 98% of non-coding DNA that orchestrates gene activity but remains largely mysterious. While we've made incredible progress with protein-coding variants through tools like AlphaMissense, the regulatory regions hold the keys to understanding complex diseases and traits. This feels like the moment genomics gets its GPT moment.

Today’s Signal is brought to you by Together AI.

Unleash hyperscale AI via Together AI. Spin up DeepSeek-R1, Llama 4 or Qwen 3 on serverless or dedicated endpoints, auto-scaling to millions. Pay as you go, continuously refine on user data—and keep full control.

Get Started

SPONSOR

Content I Enjoyed

Darren Aronofsky / Collider

Make Soup, Not Slop

This week, Darren Aronofsky appeared at Tribeca Film Festival to discuss “Ancestra,” the first film from his new AI storytelling venture, Primordial Soup, in collaboration with Google DeepMind.

Aronofsky introduced a key principle: “make soup, not slop.” He explained how AI-generated content can be visually stunning yet emotionally hollow, grabbing your attention briefly but leaving no lasting impact. The difference here is in the storytelling and the infusion of human emotion.

Director Eliza McNitt embodied this principle by centring her film on the deeply personal story of her own birth, using her actual baby photos to train the AI model that generated the infant scenes.

Single shots often combined 15+ different AI generations, requiring traditional VFX expertise to blend seamlessly. They even pioneered an "AI unit", a new filmmaking department mixing prompt engineers with traditional artists.

Something that stood out to me from the conversation was Aronofsky's chess analogy. More people play chess now, despite computers being unbeatable. As the panel demonstrated, the future of filmmaking is not about choosing between AI usage and human creativity. It revolves around how artists can guide these powerful tools to make something truly meaningful.

Idea I Learned

Dario Amodei / Fast Company

The First Major AI Copyright Ruling Is Here

This week, a federal judge delivered the first major ruling on AI training and copyright law.

The case involved Anthropic and three authors who sued the company for using their books to train Claude without permission.

Judge William Alsup made a split decision that will shape how AI companies operate going forward.

The good news for Anthropic? The judge ruled that training AI models on copyrighted books constitutes "fair use" and is therefore perfectly legal. He compared it to "any reader aspiring to be a writer" who studies existing works to create something new and transformative.

The bad news? Anthropic got hammered for downloading more than 7 million pirated books from sites like Library Genesis and Pirate Library Mirror. The judge was scathing, calling this "inherently, irredeemably infringing" even if the books were used for legitimate training purposes.

This ruling establishes a crucial precedent: you can legally train on copyrighted content, but you better make sure you're getting it through legitimate channels.

For Anthropic, this could mean hundreds of millions—possibly billions—in damages for the piracy aspect alone, with statutory damages reaching up to $150,000 per work.

There’s a clear takeaway for the AI industry: the acquisition method matters just as much as the use case. Whilst fair use protects AI training, cutting corners on how you get the data will cost you dearly.

This ruling essentially gives AI companies a roadmap for staying on the right side of copyright law—follow it, and you'll likely be fine.

Quote to Share

Sam Altman on startup lawsuit tactics:

Altman is responding to a trademark lawsuit from iyO, a Google-backed hardware startup that makes custom-moulded earpieces. The company is suing OpenAI over the name "io", which is what OpenAI and Jony Ive are calling their hardware collaboration.

What makes this particularly absurd is that iyO is essentially claiming ownership over "io", one of the most fundamental concepts in computer science (input/output). It's like trying to trademark "CPU" or "RAM”, or, as my partner highlighted, a cooking concept like “stir fry”.

Having the inside scoop on emails like this from CEOs is actually quite refreshing. Zero use of capital letters, ubiquitous references to “man” and “brother”, alongside the occasional emoji sprinkled on top for good measure.

To clarify this timeline: The product didn’t work, he gave a TED talk, he asked to get acquired, Sam says they’re working on something competitive, asks to be acquired again, rejected, sues days after because he didn’t get his way.

This looks less like legitimate trademark protection and more like litigation as a last-resort negotiation tactic. If entrepreneurs can weaponise lawsuits whenever they don't get the deals they want, it creates a chilling effect in the entire ecosystem.

The real lesson? Build better products, not better lawsuits.

Source: Sam Altman on X

Question to Ponder

If AI becomes conscious, what moral dilemmas would present themselves?

Does this mean robots should have rights, too?

We can begin by understanding our current position.

If an AI says it's sad today, it’s just a neural network predicting the next word in a sequence based on the previous words.

AI doesn’t have a heart.

It can’t portray real, genuine human emotion.

It doesn’t have real idiosyncrasies and quirks that make humans human.

However, we can look to the industry and see this narrative developing.

In October last year, Anthropic quietly hired its first “AI welfare” researcher, Kyle Fish, to explore whether future AI models might deserve moral consideration and protection.

I believe humans will treat life-like humanoids differently than faceless metallic droids because of our human tendency to anthropomorphise, even if their underlying architecture is similar.

It's those life-like features—those visual cues that mirror humanity—that could gradually shape our moral intuitions about this topic.

Whilst robots and AI don’t have feelings today, if they do tomorrow, we need to be ready.

💡 If you enjoyed this issue, share it with a friend.

See you next week, Alex Banks P.S. Satya Nadella on the real benchmark for AI.

AI Highlights

My top-3 picks of AI news this week.

Google launches Search Live / Search Engine Land

Google
1. Google Talks the Talk

Google has launched Search Live with voice input, a new way to interact with search through free-flowing, back-and-forth voice conversations in the Google app for Android and iOS.

• Voice-first experience: Users can now engage in natural conversations with Search by tapping the "Live" icon and asking questions verbally, receiving AI-generated audio responses in real-time.

• Multitasking capabilities: Search Live works in the background, allowing users to continue conversations while using other apps.

• Enhanced exploration: The feature provides easy-to-access web links on screen and includes a transcript button to switch between voice and text interactions, with conversation history saved in AI Mode.

Alex’s take: This feels like a significant step toward making search more natural and conversational. I'm particularly excited about the upcoming camera integration, which will enable us to show Search what we’re seeing in real-time. I’d definitely use this for getting help fixing home appliances, or if you’re solving a homework problem, you could get live feedback without having to ask a teacher.

OpenAI
2. OpenAI Goes on the Record

OpenAI has launched ChatGPT Record, a new feature that transforms how we capture and process spoken information through real-time transcription and summarisation.

• Live transcription & summarisation: Records up to 120 minutes of meetings, brainstorms, or voice notes, then automatically generates structured summaries saved as interactive canvases.

• Multi-format: Canvases can be instantly converted into project plans, emails, code scaffolds, or any other format you need.

• Cross-conversation memory: References past recording transcripts and canvases to provide contextual responses across new conversations, enabling questions like "What did we decide in Monday's roadmap sync?"

• Limited rollout: Currently exclusive to Enterprise, Edu, Team, and Pro workspaces, and only available through the macOS desktop app.

Alex’s take: Whilst this feature is limited for a set of users on macOS, I feel this is one of many initiatives OpenAI will be rolling out over the coming months that go beyond “just an LLM”. OpenAI are reaching up to the application layer and eating the stack from the ground up. Granola, my current favourite AI meeting notetaking tool, comes to mind when ChatGPT Record was released. Let’s see how tough platform lock-in and brand become as we move toward a world where AI becomes a companion vs a tool used in isolation.

ElevenLabs
3. Elevenlabs Puts Agents to Work

ElevenLabs has introduced Model Context Protocol (MCP) support for their conversational agents, enabling integration with external tools and data sources.

• Universal connector: MCP acts as a standard interface allowing AI agents to access data sources and specialised tools through external servers, significantly extending agent capabilities.

• Flexible security controls: Three approval modes ranging from “Always Ask” for maximum security, “Fine-Grained Tool Approval” where users can pre-select which tools run automatically versus requiring permission, to “No Approval” where the assistant can use any tool without approval.

• Growing ecosystem: Integration with services like Zapier MCP that connects to hundreds of tools and platforms.

Alex’s take: I really enjoyed watching Angelo’s demo on X, bringing this feature to life. Isolation is an AI agent’s Achilles’ heel. ElevenLabs is tackling this head-on by making their agents truly interoperable. MCP will, in turn, become the plumbing that transforms AI into genuinely useful business tools that can actually do things in your existing workflow.

Today’s Signal is brought to you by INBOUND.

INBOUND 2025 makes its West Coast debut this year in San Francisco, Sept. 3–5. Connect with visionaries including Anthropic's Dario Amodei, Synthesia's Victor Riparbelli, and AI thought leader Dwarkesh Patel. Discover cutting-edge strategies and game-changing insights at the heart of AI innovation.

Explore the Agenda

SPONSOR

Content I Enjoyed

Elon Musk at YC AI SUS 2025 / AttentionX YouTube

Elon Musk on Digital Superintelligence

This week, Elon Musk gave a fireside chat at Y Combinator’s AI Startup School in San Francisco. Now that he’s back from his politics “side quest”, he seems to be once again locked into the acceleration of intelligence. With timelines compressing across the board, I’ve pulled out my three favourite highlights that relate to AI or AI-adjacent topics.

Firstly, Musk believes digital superintelligence is arriving faster than expected (in the next 1-2 years). We’re at the very early stage of the “intelligence big bang”. It’s important to note that his definition of “digital superintelligence” here is a system that’s smarter than any human at anything. Instead of having a single digital superintelligence, Musk foresees 5-10 competing deep intelligences vs one runaway system.

Secondly, Neuralink is moving from restoration to superhuman augmentation. Beyond the five humans already controlling computers with their thoughts, Musk revealed that vision restoration trials will begin in 6-12 months. They'll write directly to the visual cortex of blind patients. But the long-term vision is wild: multispectral vision including infrared, ultraviolet, and radar wavelengths.

Thirdly, the robot revolution will be humanoid-first. His prediction that humanoid robots will outnumber all other robot types "by an order of magnitude" means Musk expects 5-10x more humanoid robots than humans eventually. Whilst he was initially hesitant about humanoid robotics due to “Terminator concerns”, he realised “it's happening whether I do it or not”. I guess if you can’t beat them, you’ve got to join them.

For civilisation as a whole, human intelligence will be “less than 1% of all intelligence” at some point. But if he's right about the timeline, we’re about to witness the most significant transformation in human history. The intelligence big bang, indeed.

Idea I Learned

Self-driving car / Volvo

When Driving Becomes a Hobby, Not a Necessity

New data has surfaced comparing accident rates across different driving scenarios.

Waymo's autonomous vehicles clock in at 1.16 accidents per million miles, while the US average sits at 3.90. That's 70% safer than human drivers. But here’s where it gets interesting: Tesla's data shows 0.15 accidents per million miles—though this includes supervised human drivers using Tesla's systems, so the data isn’t quite apples to apples…

I thought the overall trend highlights an interesting shift for the next generation. It seems like we’re heading toward a future where manual driving becomes what horse riding is today. More of a recreational activity rather than daily transportation.

Waymo is already making over 250,000 paid journeys weekly across US cities. The implications are massive. Imagine reclaiming the hours currently spent focused on the road. Reading, working, or simply relaxing during your commute could become the norm rather than the exception.

For parents, this data raises important questions about when and how to introduce driving skills to teenagers in an increasingly autonomous world. Chamath recently shared that while his 16-year-old son must learn to drive manually, once licensed, he'll only drive their Tesla with Full Self-Driving enabled.

The transition won't happen overnight, but these numbers suggest we're closer than many realise to fundamentally changing how we think about getting from point A to point B.

Quote to Share

George Hotz on AI’s value hierarchy:

Twitter tweet

I thought this was some interesting commentary by George Hotz on where value will ultimately accrue in the AI stack.

His framework suggests that infrastructure layers (data centres, chip fabs, hardware) will capture most economic value, while application companies and even foundation model makers will struggle to maintain margins.

This echoes historical tech patterns where infrastructure providers often outperform application companies long-term. Think about how cloud providers like AWS captured more value than many of the applications built on top of them.

But his dismissal of tier 4 companies seems a bit premature. OpenAI and Anthropic are no longer providers of foundational LLMs. They’re actually now becoming platforms that could own entire workflows (think tasks, reminders, workspace integrations, tool use). I believe the question goes beyond just technical capability to true distribution as a moat with ecosystem lock-in.

As for tier 5 "worthless" companies like Cursor (the AI-powered IDE), while they face commoditisation risk, there's still enormous value in solving specific use cases better than general-purpose models can. An interesting remark for a company that has grown to over $500 million in ARR and is used by over half of the Fortune 500.

The real test will be whether model companies can maintain differentiation as compute becomes commoditised.

Source: TBPN on X

Question to Ponder

One month ago, if you asked me, I would have said Grok 3 “DeepSearch”. However, more recently, I've actually changed to ChatGPT's o3 model with search. I find the length of the output to be far more digestible as it really does hit the nail on the head without any extra waffle.

Even though OpenAI announced their o3-pro model last week as part of their $200/month plan, offering only a 3-5% improvement over standard o3, there is literally zero incentive for me to upgrade, as the output quality is only marginal.

What strikes me most about this shift is how little brand loyalty seems to matter in the AI space. We simply go where the best experience is. If Grok or Google gives better answers faster next month, that's where I’ll probably go, regardless of which company built it.

In traditional software, switching costs kept us locked in. Email migration, file compatibility, and learning curves all create friction, which we as consumers tend to shy away from. But with AI, that friction barely exists. We copy-paste a prompt from one chat to another, and we're done.

The best tool wins, not the most familiar one.

This rapid migration pattern tells us something important about the future. AI companies can’t rely on habit. They have to build beyond the LLM and stay ahead every single month, or users will simply drift away to whoever’s performing better.

We're witnessing the death of AI LLM provider brand loyalty in real time.

💡 If you enjoyed this issue, share it with a friend.

See you next week, Alex Banks P.S. This AI Bigfoot dentist ad is hilarious.

AI Highlights

My top-3 picks of AI news this week.

Tim Cook at WWDC / Inc

Apple
1. Apple’s AI Gets a ChatGPT Boost

Apple unveiled several AI improvements at WWDC 2025, but the standout move was doubling down on its ChatGPT partnership to salvage struggling features.

• ChatGPT integration with Image Playground: Apple's previously criticised image generation app now offers new styles like anime, oil painting, and watercolour through ChatGPT, moving beyond its limited emoji-like creations.

• Enhanced Visual Intelligence: Apple's AI-powered image analysis tool can now interact with content on your iPhone screen, conducting searches through Google and ChatGPT based on what you're viewing.

• Foundation Models framework: Developers can now access Apple's AI models offline, though benchmarks show these models underperform compared to rivals like OpenAI's GPT-4o and Meta's Llama 4 Scout.

Alex’s take: Apple fumbled the bag hard with “Apple Intelligence”, and Siri is a different question entirely with delayed improvements until next year. What’s more, their heavy reliance on ChatGPT integration tells us a lot about their current position in the AI race. Are they sitting patiently on the sidelines preparing to pounce, are they debunking the AI myth (their paper last week got a lot of attention), or is there something we just don’t know? I’m genuinely intrigued to see how this plays out vs Google’s relentless pace across their AI ecosystem—this week they updated Android 16, and I’ve never been more tempted to switch to a Pixel than today.

OpenAI
2. OpenAI Launches o3-Pro

OpenAI has launched o3-pro, their most capable AI model yet, designed specifically for challenging problems where extended thought beats speed.

• Enhanced reasoning: Replaces o1-Pro that thinks longer and provides more reliable responses, excelling in math, science, and coding domains.

• Tool integration: Unlike o1-pro, o3-pro has access to web search, file analysis, visual reasoning, Python execution, and memory personalisation capabilities.

• Superior performance: Expert evaluations show a 64% win-rate over o3 across all categories, with higher ratings for clarity, comprehensiveness, and accuracy, while outperforming Anthropic's Claude 4 Opus on key benchmarks.

Alex’s take: It must be highlighted that o3-Pro is reserved behind the $200/mo paywall for ChatGPT “Pro” users. “Plus” users only have access to the standard o3 model. However, we must retain a critical eye and thoroughly review the benchmark results: o3-Pro is in fact only 3-5% better than o3 across key evaluations. I’ll let you decide if you’re willing to pay $180/mo more for a 3-5% improvement—which still lags behind Google’s Gemini 2.5 Pro model (free to use).

Mistral AI
3. Mistral Thinks it Through

Mistral AI has launched Magistral, its first reasoning model designed to tackle complex, multi-step problems with transparent, step-by-step thinking across multiple languages and professional domains.

• Dual release strategy: Magistral Small (24B parameter open-source version) and Magistral Medium (enterprise version) with impressive AIME2024 scores of 70.7% and 73.6% respectively.

• Transparent multilingual reasoning: Chain-of-thought processing works natively across global languages including English, French, Spanish, German, Italian, Arabic, Russian, and Simplified Chinese.

• 10x speed advantage: Flash Answers in Le Chat delivers reasoning responses up to 10x faster than most competitors, enabling real-time reasoning at scale.

Alex’s take: I like that Mistral is placing a large focus on transparency and traceability, especially in regulated industries and sectors like finance, healthcare, and legal. Users can follow the AI's reasoning process step-by-step back to the source of truth. I hope this auditability will help enable adoption within enterprise—especially to actively work against hallucination in today’s models.

Today’s Signal is brought to you by GrowHub.

GrowHub comment generator

"Great post!" and "Thanks for sharing!" are the fastest way to become invisible on LinkedIn.

GrowHub's Comment Generator turns every LinkedIn post into a networking opportunity.

It’s completely free.

No more staring at a blank comment box wondering what to write.

While everyone else drops generic comments into the void.

Create meaningful engagement that gets replies, profile views, and real connections.

Try now for free

SPONSOR

Content I Enjoyed

Sam Altman / The New Yorker

The Gentle Singularity

This week, Sam Altman released a new blog titled “The Gentle Singularity”. It covers Sam’s perspective on where humanity stands in relation to the advancements of artificial intelligence.

Five key highlights from the post:

1. The “AI takeoff” has already started

2. ChatGPT is already more powerful than any human in history

3. Hundreds of millions depend on it for critical tasks daily

4. Scientists already 2-3x more productive with AI

5. AI systems now help build better AI systems

I thought the analogies highlighted in the post gave some clarity to what’s happening with the cost of intelligence as it becomes practically free. A single ChatGPT query uses just 0.34 watt-hours (about what an oven consumes in one second) and roughly one-fifteenth of a teaspoon of water. When the cost of intelligence approaches the cost of electricity, we need to think carefully about who controls access to the grid.

Which leads me to perhaps the most important part of the post—Altman emphasising the need to “focus on making superintelligence cheap, widely available, and not too concentrated with any person, company, or country.”

There’s an irony here that this warning comes from the CEO of the company currently leading the race. I’ve also been watching public opinion on Altman shift over the past year, as people grapple with whether he’s a thoughtful steward of this technology or simply an articulate salesman for inevitable centralisation.

His vision of a “gentle singularity” suggests we'll adapt gradually to each new capability, but I suspect the societal implications will feel anything but gentle for those who find themselves on the wrong side of the intelligence divide.

Idea I Learned

Nebius CEO Arkady Volozh & Alex Banks

Why I'm bullish on European AI

This week at NVIDIA GTC and VivaTech Paris, I had the opportunity to chat with Arkady Volozh, CEO of Nebius, a company that's completely reshaping how I think about Europe's position in the AI race.

Nebius provides the full-stack infrastructure for AI model development from chip to chatbot. They cover infrastructure requirements from a single GPU to big GPU clusters, letting you train, test, and deploy large language models in your applications.

Why was this week special?

Nebius is the first NVIDIA cloud partner headquartered in Europe. They just announced the first general availability of NVIDIA's GB200 Grace Blackwell superchips in Europe. These are the world's most advanced AI chips, and Europe is getting them now.

Nebius is only 10 months old, yet they're already operating at hyperscale. Back in October last year I visited one of their data centres in Mäntsälä, Finland to see how they’re able to capture 70% of the heat generated from the GPUs back into the grid to heat the local homes.

They’re not just another GPU reseller or cloud provider. Nebius actually designs its racks and servers from the ground up. They're essentially combining the power of a supercomputer with the accessibility of hyperscale cloud infrastructure. They already have two supercomputers in the HPC Top 500, ranking as the number two commercially available supercomputer globally.

Investing over $1 billion in European AI infrastructure and "AI factories" launching in London, France, and Finland, they're building world-class infrastructure on European soil.

It’s great to see that Europe’s AI moment is finally arriving.

Quote to Share

Brett Adcock on Figure 02 humanoid progress:

Twitter tweet

Just one week after Figure AI showcased their 60-minute uninterrupted package sorting demo, we’re already seeing a 13% speed improvement in 6 days.

While conventional robots would require manual reprogramming and extensive testing for such improvements, Figure's Helix neural network can learn and optimise continuously. What's particularly exciting is the network effect: when one robot achieves 3.54 seconds per package, every robot in the fleet running Helix benefits from this enhancement.

The trajectory from their initial 6.3 seconds per package to now 3.54 seconds represents a 44% improvement through AI training alone.

I think the humanoid form is especially important for smaller companies who can’t afford ~$1M sorting machines and instead can leverage humanoids, which can also be used for different tasks—acting as a universal worker, much like a human.

Figure will soon not only be sorting packages, but also climbing stairs, lifting crates, driving forklifts, and turning into a truly adaptable, general-purpose worker. It will learn and improve at machine speed while maintaining the versatility to handle human-designed environments.

If Figure can achieve 13% gains in six days, where will they be in six months?

Source: Brett Adcock on X

Question to Ponder

By 2100, AI researchers predict that all human jobs will be fully automated, although this could happen much sooner.

Our economic system was built around human labour as the primary means of distributing resources. When that foundation shifts, everything must adapt.

With AI eventually driving the cost of intelligence and the cost of labour to near-zero, we need to think about how we navigate a society like this.

Understanding this trajectory highlights an important question. Will we need basic income? I believe it’s one that is increasingly more difficult to ignore. Workers competing with automation would need the security of some basic income.

Currently, routine, low-creativity jobs like trucking or data entry are prime candidates for disruption. Jobs needing creativity, emotional depth, or social finesse—like artists, therapists, or nurse anaesthetists—are far more difficult for AI to compete against as it ultimately can’t (yet) “feel”.

However, the WEF expects 39% of core skills to shift by 2030. Therefore, focusing on upskilling and retraining will be paramount.

Much like the Industrial Revolution transformed work rather than abolished it, I think we’ll see the same with the rise of AI. New roles will be created, old ones will fade, and our definition of “work” will shift over time.

💡 If you enjoyed this issue, share it with a friend.

See you next week, Alex Banks P.S. A dance-based CAPTCHA challenge.

AI Highlights

My top-3 picks of AI news this week.

OpenAI Logo Art / OpenAI

OpenAI
1. ChatGPT Gets Connected

OpenAI has announced a major expansion of ChatGPT's capabilities, allowing it to connect directly to internal sources and pull real-time context while maintaining existing user permissions.

• Deep Research Connectors: Available for Plus & Pro users (excluding EEA, CH, UK) and Team, Enterprise & Edu users, including Outlook, Teams, Google Drive, Gmail, Linear, and more.

• Enterprise Integration: Additional connectors like SharePoint, Dropbox, and Box are available specifically for Team, Enterprise, and Education users.

• Custom Connectors: Workspace admins can now build custom deep research connectors using Model Context Protocol (MCP) in beta, enabling connection to proprietary systems and apps.

• Record Mode: Rolling out to Team users on macOS, this feature captures meetings, brainstorms, or voice notes, transcribing and transforming them into actionable follow-ups, plans, or code.

Alex’s take: This feels like OpenAI's biggest step toward becoming a true workplace operating system. While meeting transcription has been commoditised, as Olivia Moore highlighted, the real differentiator will be UI choices and how seamlessly these tools integrate into existing workflows. With 500M+ weekly active users, ChatGPT has the distribution advantage, but as Zak Kukoff suggests, this raises serious platform risk questions for existing AI tools. It’s not just about who has the best models anymore. The race to control the interface between AI and our daily work has begun.

ElevebLabs
2. ElevenLabs Drops the Mic

ElevenLabs has launched Eleven v3 (alpha), their most expressive text-to-speech model ever, pushing the boundaries of what artificial voices can achieve.

• Multi-lingual mastery: Supports 70+ languages with human-like nuance and natural delivery across diverse linguistic contexts.

• Multi-speaker dialogue: Generates realistic conversations where speakers share contextual understanding and emotional awareness.

• Granular audio control: Precise direction through audio tags like [excited], [whispers], [laughs], and [sighs] for film-director level control over delivery.

• Emotional intelligence: Handles interruptions and emotional shifts seamlessly, creating truly dynamic speech experiences.

Alex’s take: I was blown away by their demo video. We are now witnessing the death of robotic voices. Imagine when we equip humanoid robots or your personal AI assistant with this—the anthropomorphisation of artificial intelligence continues. I’d also recommend you check out ElevenLabs’ prompting guide to unlock the most out of this model. The gap between human and artificial communication continues to narrow at an extraordinary pace.

Anysphere
3. Cursor's Coding Conquest

Anysphere has delivered a massive one-two punch this week with the launch of Cursor 1.0 alongside a staggering $900 million funding round at a $9.9 billion valuation, cementing its position as the leader in AI coding assistants.

• BugBot integration: Automatic code review that catches potential bugs in PRs and allows developers to fix issues directly in Cursor with pre-filled prompts.

• Background Agent for all: Their remote coding agent is now available to every user, expanding beyond the early access program with cloud-based processing capabilities.

• Memories feature: Cursor can now remember facts from conversations and reference them in future interactions, stored per project on an individual level.

• Explosive growth: ARR has surpassed $500 million and is doubling approximately every two months, with the company now expanding into enterprise licenses.

Alex’s take: I use Cursor weekly—it’s been instrumental in helping teach myself full-stack development. You can ask unlimited “whys” in the chat pane and really uncover the ground truths of what’s actually going on. This would never otherwise be possible without LLMs, and Cursor’s effortlessly simple UI makes it a sticky product—historically doubling ARR every two months is no joke.

Today’s Signal is brought to you by INBOUND.

INBOUND 2025 is heading to San Francisco, Sept. 3–5, for a one-time-only West Coast edition. Join AI leaders like Anthropic CEO Dario Amodei, Synthesia's Victor Riparbelli, & AI thought leader Dwarkesh Patel for bold insights, real strategy, and next-gen networking at the heart of the AI revolution.

Secure Your Spot

SPONSOR

Content I Enjoyed

Figure 02 / Figure AI

Figure Clocks In

This week, Figure AI dropped a 60-minute uninterrupted demo of their Figure 02 robot sorting packages without interruption.

This type of performance would have been nearly impossible to achieve with traditional robotic programming approaches.

The robots run an AI system called Helix, detailed in their latest article. It’s a Vision-Language-Action model that enables robots to think like humans. Helix runs on a two-system approach.

System 2 “thinks slow” about high-level goals. This involves scene understanding and language comprehension. System 1 “thinks fast” to execute and adjust actions in real-time, translating system 2 thoughts into continuous robot actions.

Helix now handles packages 20% faster and achieves 35% higher barcode scanning success. Scaling their training data from 10 to 60 hours dropped processing time per package from 6.3 to 4.3 seconds while boosting barcode accuracy from 88% to 95%.

As Figure highlighted in their post, the robot can even feel when it first touches an object and recover from mistakes by re-planning on the fly.

I love how they did an entire 60-minute section, showcasing failures and everything in between. It makes it far more authentic than a cherry-picked demo, or, dare I say, teleoperation. I'm excited for the 10-hour version to satisfy my robot ASMR obsession.

Idea I Learned

AI “Brain Chips” / Geeky Gadgets

The Illusion of Thinking

Apple just dropped a research paper that's causing quite a stir in the AI community.

“The Illusion of Thinking” challenges everything we thought we knew about reasoning models like Claude and ChatGPT.

The findings are pretty sobering. These models don't actually reason. Instead, they're sophisticated pattern matchers that completely collapse when faced with unfamiliar problems.

Here’s what caught my attention specifically: LLMs behave inconsistently across different puzzles. They can nail over 100 steps in Tower of Hanoi but fail after just 4 steps in River Crossing. This suggests performance correlates more with training data familiarity than inherent problem complexity.

Josh Wolfe, Co-Founder & Partner at Lux Capital, highlighted something fascinating too. These models “overthink” easy problems, exploring wrong answers even after finding the right one. But when things get harder, they give up early despite having plenty of compute left. Even when you hand them the exact algorithm, they still mess up the execution.

I think the real insight here perhaps sits around human cognition, not just AI limitations. As one observer noted, most of what airline pilots and doctors do is pattern matching. True reasoning only kicks in when something genuinely new appears.

This connects to Geoffrey Hinton's idea, which we explored back in April, about how humans are “analogy machines” who think by resonance rather than deduction.

The paper raises critical questions about whether current “reasoning” models can ever achieve generalisable thinking, or if they're destined to remain very expensive autocomplete systems.

Worth keeping this in mind next time you're impressed by an AI's “reasoning” abilities.

Quote to Share

Ethan Mollick on the Chief AI Officer hype:

Twitter tweet

I thought this was a brilliant quote that immediately cuts through the Chief AI Officer hype:

“The horrible realisation you have fairly quickly is that nobody knows anything.”

Most organisations are still figuring out AI basics while tech companies push increasingly advanced capabilities.

There's a massive disconnect happening. AI labs are racing toward level 8 and 9 capabilities while most workers are still trying to master level 1—how to get genuine value from the AI tools already available.

Everyone wants answers, but the uncomfortable truth is that very few people actually know how to effectively integrate AI into real workflows. The expertise isn't going to come from hiring expensive consultants or creating new C-suite positions.

The real AI champions are already in your organisation—they're the curious "weirdos" experimenting with these tools, finding what works and what doesn't. These are the people who will drive meaningful adoption, not top-down mandates.

I also think organisational context matters enormously here. One-size-fits-all AI strategies miss the mark. Real progress comes from empowering internal champions to test, iterate, and build trust in AI systems organically.

Source: @vitrupo on X

Question to Ponder

I totally get why this feels overwhelming.

When I was choosing my path, I felt the exact same uncertainty about what direction to take.

Currently, I'm making a conscious effort to continue my education in mathematics, economics, physics, and computer science. I find this helps tremendously with problem-solving and thinking from first principles. Because at the end of the day, it's all bits and atoms.

Computer science goes beyond just coding or getting a tech job. It's about building an understanding of how information systems work at their core. It teaches you to break down complex problems into manageable parts, think algorithmically, and build solutions systematically.

What I love about CS, much like economics, mathematics, and physics, is that these subjects are grounded in fundamental truths that cannot be argued away. Once you understand the axioms that make up life, that sets your footing for decision making to carve your own path.

For me, I initially chose economics and now I'm pursuing technology head-first, but that economic foundation gave me a crucial understanding of how society, business, and the world operate. So many of these concepts apply directly to building technology today. The same is true if you're doing it the other way around.

Building a foundational understanding of how the world works (through math, econ, CS and physics) provides essential grounding for problem-solving. It enables you to ask great questions and arrive at your own truths.

The beautiful thing about studying computer science is that it opens doors you didn't even know existed. It’s honed preparation for today’s job market, all whilst you're building thinking tools that will serve you regardless of where technology takes us next.

💡 If you enjoyed this issue, share it with a friend.

See you next week, Alex Banks P.S. Screen embryos for 900 diseases before they’re even born.

AI Highlights

My top-3 picks of AI news this week.

ElevenLabs team / ElevenLabs

ElevenLabs
1. ElevenLabs Talks the Talk

ElevenLabs has launched Conversational AI 2.0, a major upgrade to their voice agent platform that delivers enterprise-grade capabilities and more human-like interactions.

• Natural turn-taking: Advanced models that understand conversational cues like "um" and "ah" to seamlessly handle interruptions and pauses, creating fluid dialogue flow.

• Language detection: Integrated language detection for seamless multilingual discussions.

• Integrated RAG: Built-in Retrieval-Augmented Generation allows agents to access external knowledge bases with minimal latency while maintaining maximum privacy.

Alex’s take: V2 has launched just four months after V1—the speed of iteration here is remarkable. Something I was particularly impressed by was how quickly it switched from English to Japanese in their demo video to showcase its language detection ability. This really does feel like sci-fi coming to life.

Mistral AI
2. Mistral Goes Agentic

Mistral AI has launched their new Agents API, a comprehensive framework designed to transform language models from text generators into active problem-solvers capable of performing real-world tasks.

• Built-in connectors: The API includes ready-to-use tools for code execution, web search, image generation (powered by FLUX1.1 [pro] Ultra), and document library access for RAG functionality.

• Persistent memory: Agents maintain context across conversations with stateful conversation management and the ability to branch conversations from any point.

• Agent orchestration: Multiple specialised agents can collaborate through dynamic handoffs, allowing complex workflows where each agent contributes specific capabilities to solve different parts of a problem.

Alex’s take: I’ve found that building truly useful agents requires a high degree of technical understanding so far. Mistral seems to be focusing heavily on the infrastructure layer for enterprise AI agents. Especially for the necessary foundations that are required to make agents truly useful, like persistent memory, tool use and agent handoffs. Their MCP integration just shows this is becoming the de facto standard in the industry for agent-to-system communication.

Anthropic
3. Claude Breaks the Silence

Anthropic has launched voice mode for Claude mobile apps, enabling complete spoken conversations with their AI assistant across iOS and Android devices.

• Natural conversation flow: Users can speak to Claude and receive voice responses, with seamless switching between text and voice within the same conversation.

• Enhanced mobile experience: Key points display on-screen as Claude speaks, with intuitive controls for interrupting, sending messages, and accessing additional features.

• Workspace integration: Paid users can access Google Docs, Calendar, Gmail, and web searches through voice conversations, making Claude a hands-free productivity assistant.

Alex’s take: I'm intrigued to see how we transcend beyond keyboards and implement voice into our workflows, both in personal and professional settings. It seems like the most frictionless way to interface with AI, particularly in making it feel more natural and accessible. I’m a big believer that voice will be the dominant interface over the next few years.

Today’s Signal is brought to you by Artisan.

10x Your Outbound With Our AI BDR

Your BDR team is wasting time on things AI can automate. Artisan’s AI BDR Ava automates lead research, multi-channel outreach and follow-ups on behalf of your team.

Ava operates within the Artisan platform, which consolidates every tool you need for outbound:

• 300M+ High-Quality B2B Prospects, including E-Commerce and Local Business Leads

• Automated Lead Enrichment With 10+ Data Sources

• Full Email Deliverability Management

• Multi-Channel Outreach Across Email & LinkedIn

• Human-Level Personalization

Book a demo to see how Ava can 10x your outbound.

SPONSOR

Content I Enjoyed

Annabel Beales / The Guardian

The Workers Who Lost Their Job to ChatGPT

I came across a Guardian piece this week featuring workers who lost their jobs to AI, and one story in particular stuck with me—that of Annabel Beales.

Annabel was a copywriter at a garden centre who overheard her boss saying, "Just put it in ChatGPT," before eventually being let go.

What fascinates me about Annabel's experience isn't just the personal tragedy (though losing your dream job right before Christmas is genuinely heartbreaking), but how perfectly it illustrates the myopic thinking plaguing so many businesses right now.

When new technology emerges, there's always a cohort of employers who see human labour as the most obvious cost to cut. It's transient, it's expensive, and AI promises to do the same work for pennies on the dollar. The math seems simple enough.

But there’s something staring these employers right in the face that they seem to miss: quantity isn’t quality.

As Annabel pointed out, her former company's website is now "all AI-generated and factual—there's no substance, or sense of actually enjoying gardening." They might be hitting their content quotas, but they've traded authenticity for automation. They've replaced someone who could interview experts and craft engaging narratives with a tool that churns out generic advice.

Customers notice. People can smell AI slop from a mile away, and when your garden centre's content feels soulless, they'll buy their tools elsewhere.

Instead of laying off their skilled copywriter, the company could have educated Annabel on leveraging AI to enhance her work—using it for research, first drafts, or SEO optimisation while maintaining the human touch that made her content valuable.

But that requires thoughtful management, and thoughtful management takes time. Because it's much easier to hand overworked remaining staff a ChatGPT login and tell them to figure it out.

Unfortunately, ignorant leadership doesn’t understand that good AI output still requires human expertise, creativity, and time.

Right now, it feels as though we’re in uncharted territory with AI. The businesses that will thrive are those that view it as a tool to amplify human capabilities, rather than replace them entirely. The companies taking shortcuts today are building their business on sand.

That’s why learning to leverage AI effectively vs just deploying it as a cost-cutting measure will be what separates the winners from the casualties during this transition.

AI will change how we work. We just need to make sure that we’re thoughtful enough to change with it.

Idea I Learned

Electricity generation measured in terawatt-hours / Our World in Data

Energy Is the Ultimate Economic Lever

David Friedberg dropped a striking statistic on the All-In podcast: China produces three terawatts of energy and is scaling to eight, while the US sits at one terawatt, planning to reach two over the next 15 years.

He made a great point that relates closely to the mounting US debt issue: “If we have enough power, the $38 trillion of debt doesn't matter.”

This connects directly to Elon Musk's recent shift in focus. He's moved from cutting government waste with DOGE to accelerating GDP growth through humanoid robots and his endeavours with Optimus—he believes explosive productivity growth is the only path to outrun America’s debt crisis. Humanity could potentially see humanoid robots having a material effect on GDP growth within 4-5 years.

If the economy grows faster than debt, the debt problem solves itself. Treasury Secretary Scott Bessent echoed this, saying we can "grow our way" out of our financial hole.

But all the AI and robotics in the world won't matter if we don't have the power to run them. As one observer put it perfectly: "No watts, no bots."

The math is actually encouraging though. One hour of Earth’s sunlight contains as much energy as humanity uses in an entire year. Jesse Peltan points out that just 1% of Earth's solar resources could power a civilisation that makes today's world look impoverished. A 1 terawatt solar-powered compute cluster could feasibly be online by 2030, requiring less than 0.5% of US land area.

Energy is the ultimate economic multiplier. If we master energy production, the debt will become irrelevant.

Quote to Share

@kimmonismus on the disconnect between AI predictions and public response:

Twitter tweet

I think there’s a fascinating paradox at play right now.

The world's leading AI researchers are making extraordinary claims about imminent technological breakthroughs, yet most people seem remarkably unmoved.

There's something almost surreal about living through what these experts claim will be the most transformative period in human history, while simultaneously watching society carry on with business as usual.

Perhaps this disconnect stems from "prediction fatigue". We’ve heard “it’ll be different this time” a thousand times. Especially bold tech promises that didn't materialise on schedule.

Or maybe it's psychological self-preservation: if these predictions are true, the implications are so vast that it's easier to mentally file them away into a cabinet than grapple with what they mean for our careers, society, and future.

But there's also wisdom in having healthy scepticism. Even the brightest minds can be wrong about timelines, and extraordinary claims require extraordinary evidence.

What if the exponential we’re riding today is so steep (and getting steeper) that what we see looking forward is flat, yet in reality, it’s practically vertical?

Source: @kimmonismus on X

Question to Ponder

Whenever a new technology arrives, it undoubtedly requires a change in the way we think.

The calculator, the spreadsheet, the internet, and now AI.

Just because I use the internet to find an answer to a question I have doesn’t mean I’m dumbing down my ability to think. In fact, I often find my pursuit of an answer actually unveils other ideas or concepts that promote further education and growth. Quite the opposite of impairment.

What about Google Maps?

This is a great analogy. I, for one, love going on runs whenever I’m in a new city, wanting to explore and learn it better. But I’d have added miles onto my runs if I didn’t use Strava routes. Much like driving, Google Maps helps you go from A to B with confidence. Tools like Strava routes and Google Maps help us learn places and settings faster than if we were to just navigate them with an ordinary map, or no map at all.

The critical thinking element excites me a lot, too, especially when it comes to writing and repeatedly exercising that muscle to come up with great ideas. I strongly believe that the better questions you ask and the more critical you are with your thoughts, the better the outputs you receive from AI models.

If you decide to renounce your critical thought and let AI do the thinking for you by directing simple, uninformed one-line prompts, the outputs you receive will be reflective of that. They’ll lack substance, be weak in nature, and be full of “it’s not about X, but Y” statements, “Delving”, and “Here’s the kicker…”.

So AI, if used thoughtfully as a companion and not a replacement for your critical thought, can actually unlock your ideas, help you connect concepts together in your mind, and achieve far more than if you were to operate without it. We’re now discovering new muscles we never thought we had.

💡 If you enjoyed this issue, share it with a friend.

See you next week, Alex Banks P.S. World’s first robot kickboxing tournament held in China.

AI Highlights

My top-3 picks of AI news this week.

Sundar Pichai / Google I/O 2025

Google
1. Google’s AI Takeover

Google's annual I/O conference showcased how the company is rapidly transforming AI research into practical, widely available tools and features.

• Google Search: AI mode is starting to roll out in Google Search, starting with all users in the US. This provides a Perplexity-like interface, and perhaps marks the end of search engines once and for all. This is an interesting move given ~70% of Google’s revenue is from ads. Nonetheless, answer engines are the future.

• Real-time meeting translation: Google Meet now offers near real-time, low-latency speech translation, whilst adhering to tone and expressions. This is a huge win for education and business.

• Gemini Diffusion: 10-15x faster than current “autoregressive” models, by using diffusion which predicts tokens all at once, instead of left-to-right. This diffusion process is actually how most image/video generation models work today.

• Generative content: Add soundtracks to videos you make with Google’s new “Veo 3” AI video model. Create talking characters, include sound effects, and develop videos in a range of cinematic styles using “Flow”, Google’s new AI filmmaking tool. Also, their Imagen 4 is now the second-best overall AI image generation model in the market. However, it’s the fastest and best at typography over other models I’ve seen so far.

• Gemini 2.5 advancements: Introduced Deep Think, an enhanced reasoning mode for Gemini 2.5 Pro using parallel thinking techniques to explore multiple hypotheses before responding.

• Project Astra: Camera and screen-sharing capabilities from Project Astra are now integrated into Gemini Live, available to all Android users and rolling out to iOS users.

Alex’s take: This was perhaps one of the longest AI updates I’ve written about to date. I picked only a few of my favourites—you can find the full list of 100 things Google announced here. They’re even now processing 480 trillion tokens monthly, up 50x from last year. It’s safe to say that the pace of AI progress at Google is staggering.

Microsoft
2. Microsoft’s Agent Army

After attending Microsoft Build 2025 in Seattle this week, I saw Microsoft unveil first-hand their comprehensive vision for AI-powered agents across its entire product ecosystem, positioning itself at the centre of what they're calling the “agentic web”:

• GitHub Copilot: Evolving from an in-editor assistant to a fully autonomous coding agent that can asynchronously add features, fix bugs, refactor code, and improve tests.

• Multi-agent orchestration: New capabilities in Microsoft 365 Copilot, Copilot Studio, and Azure AI Foundry enable specialised AI agents to collaborate on complex business tasks with human oversight, such as appearing in Teams chats or joining meetings.

• Model Context Protocol (MCP): Microsoft contributed identity and registry standards to this open protocol that allows agents to securely communicate with apps and services, now integrated across GitHub, Teams, Azure, Windows, and Dynamics 365.

• Microsoft Discovery: Accelerates research and development (R&D) by bringing AI to scientists and engineers and transforming the entire discovery process. Microsoft Chemistry Product Lead John Link led a team of AI agents to discover a forever-chemical-free immersion coolant using this platform.

Alex’s take: It feels like we’re ascending OpenAI’s roadmap from Level 3 Agents (systems that can take actions) to Level 4 Innovators (AI that can aid in invention), especially with the dawn of the Microsoft Discovery platform. The agents surfaced a material “unknown to humans” in hours, not months. The team then subsequently synthesised it later in the lab—this is remarkably impressive.

Anthropic
3. Claude 4 Drops the Mic

Anthropic has unveiled Claude 4 at their "Code with Claude" event, introducing two flagship models that represent a quantum leap in AI capabilities.

• Claude Opus 4: The world's best coding model, leading SWE-bench with 72.5% and designed for sustained performance on complex, multi-hour tasks that require thousands of steps.

• Claude Sonnet 4: A significant upgrade delivering superior coding and reasoning while maintaining efficiency, achieving 72.7% on SWE-bench.

• Extended thinking with tool use: Both models can alternate between reasoning and using tools like web search during their thought processes, dramatically improving response quality.

• New API capabilities for agents: Code execution tool, MCP connector, Files API, and extended prompt caching enable developers to build more powerful AI agents.

• Claude Code general availability: Now includes native IDE integrations with VS Code and JetBrains, plus GitHub Actions support for seamless pair programming.

Alex’s take: Anthropic has implemented ASL-3 safety protections for Claude Opus 4. This means enhanced security measures and CBRN weapon guardrails. I think this sets an important precedent: as AI capabilities approach potentially dangerous thresholds, companies should err on the side of caution rather than wait for clear evidence of risk. It’s essentially an insurance policy to protect against catastrophic misuse whilst we’re still figuring out just how powerful these models really are.

Today’s Signal is brought to you by Together AI.

Unleash hyperscale AI via Together AI. Spin up DeepSeek-R1, Llama 4 or Qwen 3 on serverless or dedicated endpoints, auto-scaling to millions. Pay as you go, continuously refine on user data—and keep full control.

Get Started

SPONSOR

Content I Enjoyed

Alex Banks / Microsoft Research Lab

Microsoft's Quantum Lab Visit

This week, I got the opportunity to visit Microsoft's secret quantum computing lab, where they've been quietly working on their Majorana 1 chip—potentially the biggest computing breakthrough of our generation.

While Google, IBM and others have been building increasingly powerful but error-prone quantum systems, Microsoft took the road less travelled by developing “topoconductors”, a new system of matter which carries electricity with zero resistance.

What makes Microsoft’s approach impressive is how they are going up against quantum’s biggest challenge: error correction.

In computer science, a bit (binary digit) is the fundamental unit of digital information. It represents a logical state with two possible values of 0 or 1. Qubits (quantum bits) are spectacular compared to regular bits because they can hold multiple states at once (0 and 1); this is called superposition.

Instead of storing qubits in single points that are extremely vulnerable to disruption, Microsoft's system distributes the information across the material, making it dramatically more stable.

Satya Nadella, CEO of Microsoft said: “We believe this breakthrough will allow us to create a truly meaningful quantum computer not in decades, as some have predicted, but in years.”

1 million qubits on a single chip will be more powerful than every computer on Earth combined.

The applications they're exploring go far beyond the cryptography use cases we typically hear about. Think self-healing materials that could automatically repair cracks in infrastructure, perfect enzyme design that could help address global food challenges, and pollution conversion systems that transform waste into actually useful resources.

The team has been pursuing this approach for nearly 20 years, building an entirely new material atom by atom when everyone else was taking more conventional routes. It's a powerful reminder that sometimes betting on the most difficult technical path is exactly what leads to the biggest breakthroughs.

I made a short video of my experience at the lab, where you can see firsthand the Majorana 1 chip and the incredible environment these researchers have built to manipulate matter at the atomic level.

Idea I Learned

Jony Ive and Sam Altman

The End of Apple?

This week, OpenAI announced a shocking $6.5 billion acquisition of io, the secretive AI device startup co-founded by Apple's legendary designer Jony Ive. The timing couldn't be more strategic—announced during Google I/O week, effectively hijacking search results and tech conversation.

For decades, Apple defined what "good taste" in technology looked like, with Jony Ive as the high priest of minimalist design thinking. Now that very same design philosophy and talent pool is migrating to AI.

The acquisition brings together OpenAI's AI capabilities with the design team responsible for iconic products like the iPhone. The partnership aims to reimagine the entire human-computer interface for the AI era. As Ive himself put it: “I moved to America, drawn by the exhilarating optimism of San Francisco and Silicon Valley. We are sitting at the beginning of what I believe will be the greatest technological revolution of our lifetimes.”

Apple's design advantage allowed it to dominate mobile despite not inventing the smartphone. Similarly, OpenAI is now betting that beautiful, intuitive hardware will be the differentiator that cements their AI leadership.

Perhaps this is the most expensive acqui-hire ever? Jony and his team will assume design and creative responsibilities across OpenAI and io, whilst io will merge with OpenAI.

At the end of it all, Jony keeps control of his main company and is now steering the creative direction of the next consumer frontier in AI hardware. What a deal—and it makes me think how Apple will respond.

With ~$65 billion of cash on their balance sheet, could we see Apple acquiring someone in this space? Say, OpenAI? Never say never.

Quote to Share

Asha Sharma on the future of AI collaboration:

This week, I had a great conversation with Asha Sharma, Head of AI Platform at Microsoft. Currently, most AI interactions happen in isolation: you ask ChatGPT a question, get an answer, and move on. But Sharma sees this changing dramatically. Instead of working alone with AI, she imagines scenarios where you could collaborate with an AI agent on shared documents, have it join your Teams calls to spot insights you might miss, or work together on complex projects in real-time.

This shift from "single-player" to "multiplayer" AI represents a fundamental change in how we think about artificial intelligence, from tool to collaborator. Instead of one single agent, there will be teams of agents, each with their own specialised capabilities to assist with collective problem-solving.

In the near future, we’ll be working alongside agentic colleagues who know your company’s entire knowledge base and can help draft proposals, find financial data and help you strategise during meetings, all in real-time. We’re already seeing this permeate through today.

Source: Interview with Asha Sharma at Microsoft Build

Question to Ponder

Satya Nadella and Alex Banks

This week, I got to sit down with Satya Nadella, the CEO of Microsoft, and ask my own question.

The goal is to empower every person and every organisation on the planet to achieve more—but now in what he calls the “agentic era.”

For example, in healthcare, Stanford is using AI agent frameworks to transform tumour board meetings. What once took hours of complex workflow now happens in minutes, freeing clinicians to focus on actual treatment decisions rather than administrative overhead.

In education, the World Bank and Peru partnership showed a quantifiable impact when teachers were given access to Copilot. AI generates learning materials instantly, turning 15 days of prep work now into 3 seconds of prompting.

I think it’s all too easy in this current age of AI to get starry-eyed over the latest benchmark performance results. That is, they can be easily gamed by the models overfitting the training data and show promise in theory, yet underperform in reality. But it’s these real-world applications that are where these systems are truly put to the test—when they’re in the hands of real people performing real tasks.

That’s why it’s more important than ever to ensure these AI advancements don’t create new forms of inequality. The organisations that can afford to implement these sophisticated AI systems first will have massive advantages.

You can watch the full interview here.

💡 If you enjoyed this issue, share it with a friend.

See you next week, Alex Banks P.S. Stripe in ChatGPT.

AI Highlights

My top-3 picks of AI news this week.

U.S. President Donald Trump and Saudi Crown Prince Mohammed Bin Salman / Reuters

Middle East AI Partnerships
1. U.S.-Saudi AI Alliance Takes Shape

This week's U.S.-Saudi Investment Forum in Riyadh showcased a growing partnership in artificial intelligence between the two nations, alongside Saudi Arabia’s launch of Humain, a new AI venture backed by its $940 billion Public Investment Fund (PIF).

• Strategic investment: DataVolt committed $20 billion for U.S. AI data centres, while Saudi’s new AI firm Humain plans to deploy 500 MW each of AMD and Nvidia systems (roughly $20+ billion) over five years, the largest bilateral deal in history.

• Ecosystem expansion: Humain signed a $5 billion accord with AWS to build an “AI Zone” running GenAI services, and a non-binding MOU with Qualcomm for edge-to-cloud AI silicon.

• Broader impact: Trump’s AI advisor David Sacks emphasised building “the biggest partner ecosystem” for AI, specifically naming Saudi Arabia as a crucial strategic partner to “win the AI race and shift the balance of power” in global AI infrastructure.

Alex’s take: The specific focus on AI data centres signals Saudi Arabia's ambition to diversify beyond oil into the “oil of the 21st century”—data and computation. The desert solar will now power the next-gen AI infrastructure. It reminds me of the quote from Dune, “He who controls the spice, controls the universe.”

OpenAI
2. OpenAI Multiplies Coders

OpenAI has launched Codex, a cloud-based software engineering agent that can work on multiple tasks simultaneously, each in its own sandbox environment.

• Multi-tasking capability: Codex can write features, answer codebase questions, fix bugs, and propose pull requests for review in parallel.

• Enterprise adoption: Companies like Cisco, Temporal, Superhuman, and Kodiak are already using Codex to ship faster, refactor codebases, and write better tests.

• GPT-4.1 integration: Alongside Codex, OpenAI has rolled out its coding-optimised GPT-4.1 and GPT-4.1 mini models to ChatGPT Plus, Pro, and Team users, with GPT-4.1 mini available to free users as well.

Alex’s take: OpenAI is taking a steady march toward what they're calling “agent-native software development,” aligning closely with their AGI roadmap. Only last week, they announced they’re acquiring the AI-assisted coding editor Windsurf in a $3 billion deal. I suspect we’ll see more and more AI-coding announcements from OpenAI as they go up against Google and Anthropic’s rising tools.

Google Deepmind
3. Google's AlphaEvolve Breakthrough

Google has released AlphaEvolve, a Gemini-powered coding agent designed to discover and optimise algorithms for complex problems in mathematics and computing.

• Evolution meets AI: AlphaEvolve combines Gemini models with automated evaluators in an evolutionary framework that iteratively improves promising solutions.

• Real-world impact: Already deployed across Google's computing ecosystem, AlphaEvolve has optimised data centre scheduling, enhanced hardware design for TPUs, and accelerated Gemini's training time by 1%.

• Mathematical breakthroughs: The system has improved upon Strassen's 1969 algorithm for multiplying 4x4 complex-valued matrices and advanced the 300-year-old kissing number problem with a new configuration of 593 outer spheres in 11 dimensions.

Alex’s take: AI systems are now optimising the very infrastructure used to train themselves. This creates a fascinating feedback loop where each improvement compounds. What struck me also was how the system (now in production for over a year) recovers 0.7% of Google's worldwide compute resources through a simple heuristic that human engineers missed by orchestrating Google’s vast data centres more efficiently.

Content I Enjoyed

Tesla CEO Elon Musk attends the Saudi-U.S. Investment Forum, in Riyadh, Saudi Arabia / Reuters

Elon Musk's Vision for Humanoid Robots

Musk dropped what might be his most audacious prediction yet this week when he spoke at the Saudi-U.S. investment forum.

He claims humanoid robots will become “the biggest product ever of any kind” and envision a world with “tens of billions” of humanoid robots—multiple units for every human on Earth.

I believe there’s a lot of weight in this argument. Humanoid robots have the largest possible use case for any product ever. They can do the dishes, take your dog for a walk, and tidy up, all whilst keeping their temper.

What’s more, creators of these products will have enough customisation to make them appealing as domestic objects.

Everyone, he suggests, will want their own personal robot. This sounds about right with a total addressable market (TAM) of roughly every household on planet Earth (and beyond).

Musk believes these robots could expand the global economy to 10x its current size and effectively end scarcity-based economics altogether. This isn't just universal basic income he’s talking about, but universal “high income” in a world “where no one wants for anything.”

What fascinates me most is how my inner economist sees this vision challenging our fundamental economic assumptions. Our entire system was built around human labour as the primary means of distributing resources. When AI eventually drives both the cost of intelligence and labour toward zero, we need to think thoughtfully about how we navigate a society like this.

Whether it happens in five years or fifty, the trajectory is clear, and right now it seems we’re collectively unprepared for the societal transformation it would bring.

Idea I Learned

Artificial Intelligence, Thinking Machines and the Future of Humanity / Gerd Leonhard

The AI Curve Is Steeper Than We Ever Imagined

One of the most fascinating paradoxes I’ve observed recently is how we simultaneously overestimate and underestimate AI progress.

When ChatGPT launched in November 2022, it was a basic text-only interface that captured the world's imagination. Fast forward to May 2025, and we now have Claude with MCP integrations—AI that can directly embed into your workspace, manipulate applications, and execute complex tasks autonomously. What was science fiction 30 months ago is now available as a consumer product today.

This phenomenon reminds me of Amara's Law: “We tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run.” Nowhere is this more evident than in AI.

Consider what's happened in just the past year alone: AI agents now navigate websites autonomously, order pizza, handle multi-step reasoning chains requiring domain expertise, real-time voice conversations with digital assistants are near-indistinguishable from humans, and code generation produces entire applications from natural language descriptions.

What's most interesting is the invisibility of this progress. As capabilities become normalised, our benchmarks shift, and yesterday's magic becomes today’s mundane.

This creates a curious blind spot. Corporate clients I speak with simultaneously believe "AI will change everything" while also dismissing specific applications as "just autocomplete" or "just pattern matching." The reality is that once a capability becomes reliable, we no longer perceive it as AI at all.

This invisible progress is precisely why the market leadership in AI remains so fluid, and changes nearly every week with the deluge of model releases.

It reminds me a lot of a recent tool I started using (FYI not affiliated/sponsored) called Granola. I now use it 5x/day+ for all my meetings. Why? Because it’s frictionless in my workflow. Because it’s invisible yet provides resounding utility to my life.

We should track AI progress not by what’s beating benchmarks but by what becomes invisible in our lives because it simply works.

For those interested in exploring this topic further, I recommend you check out Tim Urban’s 2015 blog article, “The AI Revolution: The Road to Superintelligence”.

Quote to Share

Paul Graham on AGI and prompt engineering:

Twitter tweet

While today’s models require carefully crafted instructions to perform optimally, true AGI would just inherently understand our needs with minimal direction, just as humans do.

There’s an important question that’s raised here about the nature of intelligence itself. Is the ability to infer unstated intent a defining characteristic of general intelligence? If so, our current need for prompt engineering highlights the gap between our present AI capabilities and what true AGI might look like.

Humans have this innate ability to adapt to vague requests through context, shared understanding, and intuition. Much like when we ask a question to a colleague, we don’t ask via a meticulously detailed prompt for everyday tasks—we rely on intelligence to ultimately fill in the gaps.

“True” AGI would likely have this same intuition. As a result, our current prompt-engineering skills would be rendered obsolete. Does that mean you should skip learning how to craft great prompts today? Absolutely not. The key tenet of someone who succeeds in this AI race is the ability to adapt their skill set as tools and techniques develop.

So until we achieve AGI, we must build our know-how to maximise the output from these sophisticated tools. For now, they will continue to require expert guidance rather than autonomous intelligence that understands us as effortlessly as we understand each other.

Source: Paul Graham on X

Question to Ponder

This shift is already happening faster than many realise.

We’re currently witnessing a fundamental transformation in how engineering work gets done. Tech founders are already reporting 90% of shipped code is AI-generated.

That is to say, the value of engineering is rapidly moving away from manual implementation and, instead, toward higher-level thinking. Clearly articulating product requirements and getting AI to think before it acts are two high-leverage tasks you can employ today to get into that 90% category.

In the future, I believe a lot of the complexity will be hidden. Integrated development environments will shift from what we see as “code” today to a universal language: plain English.

This will fundamentally reshape business dynamics. Anyone will be able to become an engineer, and the requirement to be “technical” will no longer be a thing.

Those clinging to memorising formulas and functions and manual implementation will find themselves outpaced by competitors who embrace AI to the Nth degree.

So, if you ask will all code eventually be written by AI, we're already much closer to that reality than most people realise.

💡 If you enjoyed this issue, share it with a friend.

See you next week, Alex Banks P.S. New Moves from Tesla Optimus, a bad take, and some excellent rebuttals (1) (2).

AI Highlights

My top-3 picks of AI news this week.

Sam Altman / ChatGPT

OpenAI
1. OpenAI’s $3B Windsurf Play

OpenAI has agreed to acquire Windsurf (formerly Codeium), an AI-assisted coding editor, for approximately $3 billion, marking OpenAI’s largest acquisition to date.

• Rapid valuation growth: Windsurf was valued at $1.25 billion just last August following a $150 million funding round led by General Catalyst, representing a 2.4x increase in valuation in less than a year.

• Strategic expansion: This acquisition is designed to complement ChatGPT’s existing coding capabilities and compete against the likes of Cursor (currently valued at $9 billion).

• Market positioning: The AI coding market is moving too quickly for OpenAI to wait for internal solutions, while its user base recently surged to over 400 million weekly active users in February.

Alex’s take: Time-to-market now trumps all. As we covered last month, AI is now a motion market, not a moat market. Most incumbents are approaching AI the same way they approached mobile—late, reactive, and lacking taste. The real winners will be those who can iterate quickly, collapse feedback loops, and move with conviction. Execution velocity is everything.

Google
2. Gemini 2.5 Pro Came Early

Google has released Gemini 2.5 Pro Preview (I/O edition) ahead of schedule, bringing significant improvements to coding capabilities and web development performance.

• #1 on WebDev Arena: Ranks first on the WebDev Arena leaderboard, which measures human preference for a model’s ability to build aesthetically pleasing and functional web apps.

• Advanced UI capabilities: Delivers improved front-end and UI development, making it easier to implement new features and create responsive designs with subtle animations.

• Early release decision: Originally planned for Google I/O later this month, Google released the model early due to "overwhelming enthusiasm" to get it into developers' hands sooner.

Alex’s take: Google really is covering all bases when it comes to output generation. Text, image and code generation with Gemini 2.5 Pro. Video generation with Veo 2. Plus, being able to predict the structure and interactions of all of life’s molecules with AlphaFold 3. Google has a firm footing to capitalise on the application layer, powered by their very own TPU hardware. In my opinion, they are now a front-runner in the AI race.

Mistral AI
3. Medium is the New Large

Mistral AI has launched Mistral Medium 3, a breakthrough model that redefines efficiency and accessibility in enterprise AI deployment.

• Performance at scale: Delivers 90% of Claude 3.7 Sonnet's capabilities at 8x lower cost ($0.4 input / $2 output per million tokens).

• Enterprise ready: Can be deployed on any cloud or self-hosted environment requiring only four GPUs, making it significantly more accessible for organisations.

• Domain dominance: Particularly excels in professional use cases like coding and STEM tasks, outperforming larger competitors like Llama 4 Maverick and Cohere Command A.

Alex’s take: Just two months after releasing Mistral Small, they've delivered a “medium” model that outperforms most flagship offerings at a fraction of the cost. It's refreshing to see a European company challenging Silicon Valley's dominance. I suspect we're witnessing a paradigm shift where performance-per-resource becomes more important than raw model size.

Content I Enjoyed

Micha Kaufman internal memo / X

Fiverr CEOs Harsh Truth About AI Adoption

I came across an internal memo from Micha Kaufman, CEO of Fiverr, that’s been doing the rounds online. His message to his team was actually quite refreshing in a sea of CEO AI statements we’ve seen of late: “AI is coming for your jobs. Heck, it's coming for my job too.”

Rather than sugar-coating reality and sprinkling buzz words about how their organisation is making a conscious effort to adopt AI and be a frontrunner, he uses simple language and offers practical advice to his colleagues for staying relevant.

Among the list:

• Master AI tools in your field

• Learn from knowledgeable colleagues

• Become a prompt engineer

His words, “Google is dead. LLM and GenAI are the new basics,” which I do think is the right framing, given we are now shifting from search engines to answer engines.

Kaufman’s perspective pairs nicely with Nvidia CEO Jensen Huang's take that “AI will not take your job. AI used by somebody else will take your job.” It's about becoming the person who skillfully wields these new tools to do meaningful work and 10x your output.

The future doesn't belong to those waiting for permission to adapt. As Kaufman puts it, “Stop waiting for the world or your place of work to hand you opportunities to learn and grow—create those opportunities yourself.”

Idea I Learned

Dario Amodei, Elon Musk, Satya Nadella, and Sam Altman

OpenAI’s Moat Is Not Their API, It’s ChatGPT

I've been studying the latest Ramp AI Index this week, and the numbers are staggering.

OpenAI commands a whopping 32.4% share of U.S. businesses with paid AI subscriptions, while the entire market sits at 40.1%. That’s 81% market share—and practically a pure monopoly.

I’ve found this is highly representative of the real world. Whenever I’m at a conference or meeting with an enterprise client, the conversation inevitably revolves around "ChatGPT this" or "ChatGPT that." Rarely do I hear mentions of Claude, Grok, or Gemini.

Let’s understand OpenAI’s business model a little better to unveil what this means.

There are two main ways companies can use AI like ChatGPT. The first is through the consumer-facing website (ChatGPT.com) that we're all familiar with—where you sign up, type questions, and get answers.

The second is through something called an API, which stands for Application Programming Interface. Think of an API as a behind-the-scenes connector that lets developers build ChatGPT’s intelligence directly into their own apps and websites.

For example, when you use Copilot on Windows, that's often OpenAI's technology working through an API—the same brain as ChatGPT, just wearing Microsoft’s outfit.

While the API business is critically important (and where many predicted the real money would be), it's becoming a commodity with razor-thin margins. Anyone can offer API access to AI models, and prices keep dropping. It’s a race to the bottom.

What I’ve come to realise is that OpenAI’s true advantage lies in its consumer product, ChatGPT itself, and not, in fact, its API.

By being first to market in November 2022, they've established "ChatGPT" as a household name that's now arguably more recognisable than "OpenAI".

What’s more, they're systematically turning potential competitors into features: Perplexity becomes Deep Research, Midjourney becomes GPT Image 1, Windsurf becomes CodeGPT.

Perhaps there is a first-mover advantage in AI after all, but it's not where we thought it would be. Brand recognition and consumer habit matter more than ever before.

Quote to Share

Alex Albert on education’s AI vulnerability:

Twitter tweet

I thought this was a thought-provoking post from Alex highlighting the contrast between two educational models.

The Danish system is built around in-person collaborative learning, whereas the American system relies heavily on individual, asynchronous assignments.

The pandemic and AI disruption created a perfect storm. First, we had COVID which pushed American education further toward remote, asynchronous work in 2020. Then, when consumer-facing AI tools like ChatGPT emerged in late 2022, they found an educational landscape already optimised for the very tasks they excel at automating.

As Charlie Munger once famously said, “Show me the incentive, I'll show you the outcome.” It perfectly captures what's happening in education today.

The US system incentivises efficiency, standardisation, and individual assessment, naturally leading to take-home assignments that can be evaluated at scale. These incentives weren't created with AI in mind, but they have inadvertently produced an educational model that is extraordinarily vulnerable to AI automation.

The Danish approach, built around different incentives—prioritising collaboration, social learning, and in-person interaction—has accidentally created a system far more resilient to AI disruption. Their model values precisely the skills that remain uniquely human: critical thinking in real-time discussion, interpersonal skills, and collaborative problem-solving.

This contrast offers a valuable lesson in system design. When we focus solely on efficiency and standardisation, we may optimise for metrics that are easily measurable but miss the deeper, more human elements of learning.

As AI continues to evolve, we need to reconsider our educational incentives to focus on what humans do best, letting technology handle the personalisation and mechanical aspects, while we cultivate the creative, social, and critical thinking skills that make us distinctly human.

Source: Alex Albert on X

Question to Ponder

I think this is an interesting question, because we've got to understand that language has profound consequences in human interpretation of concepts.

When we hear “training models,” we think of something mechanical and controlled—a process we direct. But “cultivating digital organisms” evokes something alive, evolving, and perhaps beyond our complete understanding.

Today’s language models aren’t “organisms”. They’re just sophisticated prediction systems that require our tools and instructions to take action. They don't have intrinsic drives or goals.

Yet we’re rapidly approaching a threshold.

As we integrate these systems with robotics and real-world perception—creating what many call “embodied AI”—we may witness capabilities emerge that we didn't explicitly program or expect.

Therefore, I'm not convinced that biological metaphors are the most accurate or helpful for today. They unnecessarily amplify fears while missing the fundamental difference: even our most advanced AI systems lack the intrinsic agency and self-preservation that drives biological life.

However, we are in the process of bridging the gap between what is digital and what is physical—and this might require us to rethink our language as these systems grow increasingly complex and autonomous.

💡 If you enjoyed this issue, share it with a friend.

See you next week, Alex Banks P.S. This robot can clean your home.

AI Highlights

My top-3 picks of AI news this week.

New York Times / Andrea Chronopoulos

Anthropic
1. Claude connects your tech stack

Anthropic has released Integrations, allowing Claude to connect directly with popular workplace tools and enabling developers to create custom connections to virtually any platform.

• Pre-built connections: Claude can now integrate with Asana, Intercom, Linear, Zapier, Atlassian, Square, and PayPal, gaining deep context about projects, tasks, and organisational knowledge.

• Custom extensibility: Developers can create their own integrations in as little as 30 minutes using Anthropic's MCP protocol, connecting Claude to any tool or data source.

• Enhanced capabilities: Also launching is Advanced Research mode to all paid plans, which can search across connected apps and web sources for up to 45 minutes to deliver comprehensive reports.

Alex’s take: This is the API-ification of AI assistants. By opening Claude to third-party integrations, Anthropic has transformed it from a standalone chatbot into a central hub that can orchestrate across your entire digital workspace. What's particularly clever is how they've taken the complex Model Context Protocol (MCP) and made it accessible to everyone in just a few clicks. This removal of friction matters more than we think.

Neuralink
2. Meet Neuralink’s third patient

Brad Smith has become the third person to receive a Neuralink brain implant, marking a significant milestone as the first nonverbal ALS patient to undergo the procedure.

• AI-powered communication: Neuralink developed a chat app using Grok 3 AI that suggests contextual responses in Brad’s pre-ALS voice, helping him stay engaged in conversations.

• Breaking barriers: Brad controls his MacBook Pro cursor via brain signals, allowing him to communicate outdoors and in varying lighting conditions, in contrast to his previous eye-gaze system that only worked in the dark.

• Precision breakthroughs: Using tongue movements for cursor control and jaw clenching for clicks (after hand-based training proved ineffective), Brad achieved a Webgrid accuracy score of 5, up from less than 1 with eye-gaze technology.

Alex’s take: Brad himself has made a video showcasing Neuralink’s capabilities that’s the first known video edited entirely using a brain-computer interface. I highly recommend you give it a watch.

Importantly, this demonstrates how these advances are restoring agency to ALS patients. Brad can now communicate outdoors, build his own applications with AI assistance, and engage with the world in ways that were impossible just months ago, which I think is wonderful.

I tried playing the Webgrid game myself and got an accuracy of 9. Considering I spend a lot of my day with a mouse in hand, Brad’s jump from 1 to 5 is just incredible. Grok even helped Brad build his own keyboard training app, despite having no coding experience.

Meta
3. Inside Meta's First LlamaCon

Meta kicked off their first-ever LlamaCon conference with major announcements spanning consumer apps to developer tools, bringing together developers worldwide to celebrate the Llama ecosystem.

• Meta AI app: Stand-alone ChatGPT competitor that leverages your Facebook/Instagram data for personalised responses, plus a social discovery feed for sharing AI interactions.

• Llama API preview: Developer platform offering one-click API creation, custom model fine-tuning with Llama 3.3 8B, and full portability—your models stay yours.

• Llama Protection Suite: New security tools including Llama Guard 4, LlamaFirewall, and CyberSecEval 4 for building safer AI applications.

• Speed partnerships: Collaborations with Cerebras and Groq bringing dramatically faster inference speeds to Llama 4 models.

Alex’s take: Meta's inaugural LlamaCon feels a bit like a coming-of-age party for open source AI. On the one hand, we have OpenAI locking developers into their ecosystem, on the other, Meta is offering the best of both worlds through API creation without the handcuffs. I also think Meta has a serious distribution advantage when it comes to their ecosystem, so leveraging social graph data will just mean far more meaningful experiences when interfacing with an assistant.

Today’s Signal is brought to you by Athyna.

2025 Salary Report: Hire smarter, save big today.

Curious how global hiring gives you a competitive edge?

• Discover salary insights for engineers, data scientists, product managers, and more.

• Explore top-tier talent with experience at AWS, Google, PwC, and beyond.

• Learn how to save up to 70% on salaries while hiring top global talent.

Get salary insights now

SPONSOR

Content I Enjoyed

Mark Zuckerberg / Dwarkesh Patel

Mark Zuckerberg – Meta’s AGI Plan

Only two weeks ago, Zuck was in Washington, testifying in federal court against the US government, which argued that Meta should unwind the acquisition of Instagram under the belief that the purchase was made to eliminate competition and create a social networking monopoly.

This week, Mark has been doing the media rounds discussing Meta’s AI strategy. I particularly enjoyed this conversation with Dwarkesh Patel, which offered a real glimpse into how the social media giant plans to compete in the AGI race.

Meta AI now has nearly 1 billion monthly active users, with most interaction happening through WhatsApp rather than Instagram or Facebook.

Zuck cited the striking statistic that the average American has fewer than 3 friends while desiring around 15. He believes AI agents could help fill this gap in human connection through natural conversation and emotional support, eventually becoming seamlessly integrated through AR glasses like Orion.

Whilst we’re still a way off truly personal conversations through AI assistants (they’re still just predicting the next words given a previous sequence of words), it highlights an interesting point regarding the loneliness epidemic of today.

As one commenter remarked, Zuck’s Meta glasses might already be running a real-time search on Dwarkesh’s questions during the interview to optimise his responses.

It wouldn't surprise me if he's already testing the future he’s selling to the rest of us.

Idea I Learned

Cursor / Happy Future AI

This Is the Real Benchmark

Forget MMLU scores and ELO ratings for a second.

This week, I came across an X post by Ryo Lu, Head of Design at Cursor.

It highlights a telling metric for AI model quality—what developers are actually choosing to use daily.

The data reveals an interesting picture: Claude 3.7 Sonnet leads the pack, followed by Gemini 2.5 Pro and Claude 3.5 Sonnet.

Meanwhile, OpenAI’s o3 and o4-mini are the fastest-growing models developers are adopting.

This matters because developers are among the most demanding AI users. They need models that can handle complex reasoning, understand nuanced instructions, and produce reliable output.

The disconnect between benchmark scores and real-world usage is becoming impossible to ignore. We highlighted only last month how these models are overfitting the data, making them unreliable when put to the test in the real world.

Some models that perform well on traditional coding benchmarks struggle with actual usage statistics. Therefore, we need to revise how we evaluate AI progress—ideally through third-party, independently verifiable methods.

Much like when we saw Hao AI Lab using Super Mario Bros to test leading frontier models and observing Claude-3.7 outperforming reasoning models like OpenAI’s o1, despite being generally stronger on most benchmarks.

At the end of the day, real-world validation can’t be gamed.

Quote to Share

Luis von Ahn on Duolingo’s AI transformation:

Twitter tweet

Duolingo CEO Luis von Ahn sent an all-hands email announcing the company's AI-first transition.

The announcement reveals how "being 100% mobile" in 2012 made Duolingo successful. Now the company is declaring itself 100% AI-focused, requiring employees to demonstrate how they're using AI to reduce costs and improve efficiency.

Two key points stand out from the email.

The intro highlighted how Duolingo “will remain a company that deeply cares about its employees.” Yet the end of the email highlighted “headcount will only be given if a team cannot automate more of their work”.

There’s a rather stark contradiction at play here.

Removing the human element from the workplace is the polar opposite of showing that you care deeply for your employees.

Being AI-first means empowering and educating your people to use these tools to 10x their output, not replace humans entirely.

Something else I wanted to pick up on was the fact that Duolingo felt the need to make a formal announcement about adopting AI, which is quite telling.

As one commenter noted: “People who understand AI just change their processes over time, quietly.” Is this a PR stunt, and are Duolingo actually playing catch-up here, rather than leading?

Most revealing is the email's distinctive ChatGPT writing style, especially overusing phrases like "It's not just X, it's also Y", which several observers immediately recognised.

The irony of using AI to write an announcement about becoming AI-first wasn't lost on anyone.

Unfortunately, in my eyes, this one didn’t quite pass muster.

Source: Duolingo on LinkedIn

Question to Ponder

In the age of AI, rich, personalised context is everything.

Embedding users through memory increases a user’s opportunity cost to switch to another provider, as the model “knows” the user better than an empty chatbot.

In principle, frictionlessly transferring memory from one LLM provider to another would be a nice idea. But model providers like OpenAI, Anthropic, and xAI make it intentionally hard to do.

Why? Because in a world where LLMs are becoming commoditised, it’s their central moat (for the time being).

Memory increases the “lock-in” for users. An LLM becomes infinitely more valuable to me if it knows me, my life, my work, and any surrounding context. This augments the calibre of conversation and, in turn, the insights I then receive. That’s how a chatbot evolves into a truly useful assistant.

In practice, unless you request a copy of your data profile, you’d have to manually feed this information into a new LLM. The friction of this is already too high for many to consider as an option.

But all hope is not lost.

As we covered earlier, Anthropic now lets users integrate external apps directly into their Claude chat interface.

This means you, as a user, can connect apps like your Gmail, Calendar, and Drive, which Claude can then reference during the conversation.

Whilst this ticks the box for email conversations and cloud context, I see the “memories” these large language models pick up over time from the conversations and interactions we have with AI assistants to be far more nuanced (and, perhaps, far more personal).

That’s why I believe there is a real race for LLM providers to implement memory to build a “home” for users, and why it is accompanied by a real reluctance to share these with anyone but themselves.

💡 If you enjoyed this issue, share it with a friend.

See you next week, Alex Banks P.S. Just describe a sound, and it’s generated instantly.

AI Highlights

My top-3 picks of AI news this week.

OpenAI CEO Sam Altman / The Verge

OpenAI
1. OpenAI’s Image Explosion

OpenAI has made its powerful image generation capabilities available to developers via the API with their new gpt-image-1 model, following the enormous success of the feature in ChatGPT.

• Massive adoption: The image generation feature in ChatGPT saw over 130 million users create more than 700 million images in just the first week after launch.

• Enterprise integration: Major creative platforms like Adobe, Figma, Canva, and Wix are already implementing the API for professional design workflows.

• Flexible pricing: Usage is charged per token with separate rates for text input ($5/1M tokens), image input ($10/1M tokens), and image output ($40/1M tokens), translating to roughly $0.02-$0.19 per generated image.

Alex’s take: It’s clear AI image generation is moving beyond just novelty. I remember when the first version of DALL-E was released at the beginning of 2021. It’s incredible to see how image generation is now becoming an essential part of creative workflows. Especially given how frictionless it is to turn ideas into reality.

xAI
2. Grok Gets Vision

xAI has expanded Grok’s capabilities with several significant updates, most notably the introduction of visual perception through Grok Vision.

• Grok Vision: Users can now point their phone's camera at objects, products, signs, and documents to ask questions about what they see (currently available only on iOS).

• Multilingual audio: Android users subscribed to the $30 monthly SuperGrok plan can now access voice interactions in multiple languages.

• Real-time search: Voice mode now incorporates real-time search capabilities for more current information, again exclusively for SuperGrok subscribers on Android.

Alex’s take: The race for multimodal AI continues to intensify, with xAI now joining Google and OpenAI in offering real-time visual understanding. Only last week, we saw the integration of memory within Grok to enable personalised knowledge across conversations. xAI has been shipping at an incredible pace to not only catch up to, but now in some instances, surpass the capabilities of other leading LLM providers.

Adobe
3. Firefly Takes Flight

Adobe has announced major updates to its Firefly platform at its MAX London event this week.

• Firefly Image Model 4 & 4 Ultra: Adobe's fastest and most controllable image models yet, offering lifelike quality with precision control over structure, style, and camera angles at up to 2K resolution.

• Partner integration: Now incorporates models from Google Cloud and OpenAI, with more partners coming soon, including fal.ai, Ideogram, Luma, Pika and Runway.

• Firefly Boards: A new collaborative, AI-first canvas for moodboarding and exploring creative concepts that streamlines ideation to production workflows.

Alex’s take: I attended the Adobe MAX event in London this week and was inspired by their approach, not just in terms of technical capabilities, but also in their commitment to commercially safe AI. It’s a bit like the wild west right now, with an abundance of lawsuits flying around, with image and video generation providers training on copyrighted works. However, Firefly is trained exclusively on Adobe Stock and non-copyright works, meaning outputs can be used in commercial projects with confidence.

Today’s Signal is brought to you by GrowHub.

Don’t Leave LinkedIn Growth To Chance

• Create LinkedIn posts from blogs, articles and YouTube videos

• AI trained on thousands of high-performing posts

• Generate sensational content in any language

Start Growing

SPONSOR

Content I Enjoyed

Cluely Launch Video

Invisible AI to Cheat On Everything?

When a video of AI glasses (which don’t actually exist) helping someone cheat on a date went viral this week, amidst the backlash, I couldn't help but watch with a mix of fascination and mild horror.

The product, Cluely, markets itself as an “undetectable AI-powered assistant” for interviews, sales calls, and meetings.

Behind Cluely is Chungin "Roy" Lee, a former Columbia student who was suspended after creating an AI tool for cheating on job interviews. His company has already raised $5.3 million from venture capital firms Abstract and Susa Ventures, and has gained 70,000 users since its launch.

I found myself wondering what season of Black Mirror this technology would fit into. Especially when the company chose dating as its showcase scenario, when using this type of tech in work meetings would be far more practical and far less controversial?

Then again, that’s the point. To appeal to interest, not to reason. To spark a reaction. And it did exactly that with 12M impressions on the launch video.

Lee argues that “ultimately, this is where the world is headed. We will be 100 times more efficient”, suggesting we need to rethink what “cheating” means in the age of AI. What happens when we have any publicly available fact or figure available at the speed of thought?

While I'm sceptical about the ethics, I do see a huge opportunity when it comes to the possibility of visual copilots and their application to get an informational advantage.

Idea I Learned

Toby Brown / Sky News

The AI Education Gap Is Creating a New Generation of Leaders

This week, President Trump signed an Executive Order establishing AI literacy as a national educational priority across all grade levels in the United States.

The initiative will integrate AI concepts throughout the K-12 curriculum, ensuring all students develop fundamental AI competencies regardless of their career path.

What struck me most was the potential generational divide that this will create.

While adults are figuring out prompt engineering and attempting to apply basic AI tools, today's children will grow up with AI concepts woven into their educational DNA.

This idea reminds me of Toby Brown, a teenage AI founder whom I met last year at an AI demo day in London. He recently secured $1 million investment from Silicon Valley before even finishing his GCSEs—and if I’m being honest, I’m not even surprised. Toby was the sharpest person I spoke to at that event, and his enthusiasm for building was infectious. He started coding at age 7 and is now building and raising capital in SF while his peers are still in school.

It begs the question, will we soon see 20-year-old CEOs regularly managing teams of 40-somethings who are struggling to keep pace with AI advancements?

Those of us who didn't learn AI in school will need to work twice as hard to stay relevant. The plasticity of young minds gives them a natural advantage. That’s why it's your job to prioritise both the education and application of the latest tools. Even an hour a week will put you ahead of your peer group.

What’s more, tools like Grok are now introducing Socratic responses to help guide users through topic understanding via questions and reasoning to promote critical thinking. Instead of providing direct answers, you’re actively engaging with a subject rather than just receiving passive information.

So, I recommend that you find something you're interested in related to AI and pursue that curiosity—it’ll take you further and lead to a far deeper understanding than someone doing it “because I have to.”

Quote to Share

Geoffrey Hinton challenges our traditional view of human intelligence:

Twitter tweet

Hinton, one of AI's founding fathers, suggests that as we gain deeper insights into both AI systems and the human brain, we're discovering that human cognition isn’t primarily based on logical reasoning.

Instead, he proposes that we are "analogy machines" who think by resonance rather than deduction.

I think this idea is absolutely spot on. Humans love drawing comparisons to past experiences or familiar patterns rather than strict logical deduction.

Take, for example, choosing a restaurant to eat at when visiting a new city.

You don’t logically evaluate every restaurant’s menu, reviews, and pricing through deduction.

Instead, you think, “Ah, I’m going to find something similar to that pizza restaurant I love in my hometown, with those thick crusts and sourdough base.”

You use analogy, comparison and resonance of past experience to guide your choice.

In other words, you use “vibes”.

You think emotionally, not rationally. The emotions that are tied to these analogical memories shape the future.

Hinton rightly states that we have a thin layer of reasoning on top in order to do things like mathematics, but our primary decision-making engine is through analogies.

As LLMs get smarter and as we advance to AGI, we may need to reconsider our most fundamental assumptions about “rational” thought itself.

Source: University of Toronto, YouTube

Question to Ponder

The short answer: it’s too early to tell.

The long answer: it’ll be who controls the vertical from chip to chatbot.

Compute is the new oil of our time. Just look at Nvidia’s revenue growth. A staggering $39.3 billion last quarter.

What’s interesting to see is that whilst there has been a flurry of excitement about OpenAI over the last few years, I believe over the next three, we’ll see Google stretch its advantage.

Why? Google’s TPUs.

When testing state-of-the-art LLMs on the Aider Polyglot coding benchmark, Gemini 2.5 Pro had a remarkably low cost of $6.32 while achieving a high accuracy of 72.9%.

In contrast, OpenAI’s model o3 ($111.03), which offers comparable performance, is much more expensive, due to their reliance on GPU infrastructure, trailing back to their vendor: Nvidia.

Google’s TPUs (Tensor Processing Unit) are designed specifically for tensor operations, which are the core of neural network computations.

This optimisation for AI-specific workloads means a reduction in computational overhead and fast completion of tasks.

Unlike GPUs, which handle a wide range of computing tasks, TPUs focus solely on what matters for AI.

But Google's true advantage lies in vertical integration. They control both hardware and software, creating a perfectly synchronised end-to-end ecosystem that tailors their infrastructure precisely to their models' needs.

Energy efficiency becomes crucial at scale. TPUs consume significantly less power than GPUs for equivalent AI workloads—critical when running massive training jobs (for Gemini 3, 4, 5, etc.) or serving billions of inference requests daily.

That leads me onto the cost implications. Google's decade-long investment in custom silicon means they can offer more compute per dollar through economies of scale. While OpenAI has captured our imagination so far, the economics of running models at scale might ultimately decide this race.

The winner won’t just have the best consumer-facing models. They’ll have the most efficient path from electricity all the way through to intelligence.

💡 If you enjoyed this issue, share it with a friend.

See you next week, Alex Banks P.S. Google DeepMind built an AI model to simulate how a fruit fly walks, flies and behaves.

AI Highlights

My top-3 picks of AI news this week.

OpenAI CEO Sam Altman / The Hollywood Reporter

OpenAI
1. OpenAI Never Sleeps

This week has been packed with back-to-back releases across OpenAI’s entire ecosystem:

• New models: The GPT-4.1 family (handling up to 1M tokens) and o-series reasoning models (o3 and o4-mini) that excel at complex tasks while setting new benchmarks in coding, math, and multimodal understanding.

• Integrated capabilities: Full tool access for reasoning models, visual thinking that manipulates images during reasoning, and agentic approaches that use tools based on desired outcomes.

• Developer ecosystem: Introduced Codex CLI (a lightweight terminal coding agent), launched a $1M initiative supporting developers, and made advanced models available through API.

Alex’s take: Aside from OpenAI’s convoluted naming system for their models, they’re really pushing the envelope when it comes to new releases. Only last week, we saw memory improvements roll out—this week, it’s new reasoning models. It feels like we’re witnessing the early stages of truly agentic AI, which can plan, reason, and execute complex tasks across different modalities.

xAI
2. Grok Puts Pedal To Metal

The AI race is red hot, with OpenAI setting a blistering pace. But Grok is out here swinging with weekly updates:

• Memory integration: Grok 3 now remembers conversations, builds personalised knowledge about users, and analyses thinking patterns to deliver tailored recommendations (not yet available in EU/UK due to regulations).

• Grok Studio: A new canvas-like tool allowing users to create and collaborate on documents, code, reports, and browser games, with Google Drive integration for file attachment.

• Grok 3 Mini: A cost-efficient model that outperforms reasoning models at 1/5 the cost, excelling in graduate-level STEM, math, and coding while providing access to its complete reasoning trace.

Alex’s take: Memory transforms AI from a chatbot to a useful assistant. As you build conversation history with an AI, switching costs increase exponentially. ChatGPT was first. And this is exactly why xAI is racing to get memory embedded into Grok ASAP. Once you’ve invested months or years of conversations into one AI, would you really want to start from scratch with another?

Google
3. Google AI Bridges Land and Sea

Google's AI updates this week couldn't possibly sit at more opposite ends of the tech spectrum. And that's precisely what makes them so fascinating:

• Google Sheets AI function: A new “=AI()” function brings native Gemini capabilities directly into spreadsheets, allowing users to generate text, analyse sentiment, categorise data, and summarise information without leaving their workflow.

• DolphinGemma model: Google has developed a specialised 400M parameter AI model that analyses and generates dolphin vocalisations, helping researchers decode patterns in dolphin communication using the Wild Dolphin Project’s 38 years of research data.

• Mobile field research: The DolphinGemma system runs directly on Pixel phones, making it practical for oceanographic field research and real-time dolphin interaction through the CHAT system.

Alex’s take: What I love about Google is how they’re applying AI across such vastly different domains—from mundane spreadsheet functions to potentially cracking the code of dolphin language. I’m particularly struck by how the same foundational LLM technology powering a spreadsheet function could help us understand what dolphins are saying to each other after decades of careful observation.

Today’s Signal is brought to you by Athyna.

2025 Salary Report: Hire smarter, save big today.

Curious how global hiring gives you a competitive edge?

• Discover salary insights for engineers, data scientists, product managers, and more.

• Explore top-tier talent with experience at AWS, Google, PwC, and beyond.

• Learn how to save up to 70% on salaries while hiring top global talent.

Get salary insights now

SPONSOR

Content I Enjoyed

Nvidia CEO Jensen Huang / Getty Images

Nvidia’s Geopolitical Tightrope

When your CEO lands in Beijing days after the US government blocks sales of your AI chips to China, you know you’re walking a precarious geopolitical tightrope. I’ve been watching Nvidia's China situation unfold with an eager eye this week.

Something that stood out to me was this revenue chart showing Nvidia's growth trajectory—from virtually nothing in 2015 to approximately $125 billion in 2025, with China representing a significant yet increasingly constrained slice of that pie.

The Trump administration’s move to restrict even Nvidia's China-specific H20 chips (which were specifically designed to comply with previous export controls) shows a dramatic escalation in the tech decoupling between the US and China.

Take for example the timing: immediately after Nvidia announced plans to invest $500 billion in US-based AI infrastructure and manufacturing—building chips in Arizona and AI supercomputers in Texas—the administration quietly tightened the screws on China sales.

But there’s something else at play. This week, Chamath observed that perhaps 20-30% of Nvidia's GPUs may actually be finding their way to China through “vassal intermediary countries” despite the restrictions. The process is a simple one: nearby Asian countries buy the GPUs and then send them into China.

I suspect this has been taking place for years, especially given how fast China is building out their data centres.

The takeaway? Nvidia finds itself at the centre of perhaps the most consequential technological and geopolitical chess match of our time. Two feuding countries going head-to-head. The company that is powering the AI revolution is now caught between competing superpowers, each determined to (a) control the future of computing and (b) prioritise their singular mission and race to achieve artificial general intelligence (AGI).

I don’t envy Jensen Huang’s position.

Idea I Learned

OpenAI Aims To Acquire WindSurf / Crypto Weekly

AI Is a Motion Market, Not a Moat Market

This week, OpenAI is reportedly in talks to acquire Windsurf (formerly Codeium) for $3 billion—less than four years after its founding. This signals OpenAI’s clear intent to verticalise and own the application layer, particularly in coding, positioning them to compete directly with Cursor.

What's fascinating is the discussion this sparked about the nature of the AI market itself. Despite OpenAI’s superior distribution, brand, and data compared to Windsurf, they're willing to pay billions rather than build in-house.

The reason? Time-to-market. The opportunity cost of building in-house might exceed $3 billion, given how lucrative the AI coding market is becoming.

This reinforces a compelling insight I saw on X this week: the AI market is a “motion market,” not a “moat market.” For a while, GitHub Copilot seemed to have locked up the developer tools conversation. Many hesitated to compete because it felt like Microsoft had already won.

That turned out to be an illusion. Most incumbents are approaching AI the same way they approached mobile—late, reactive, and lacking taste. They're too bloated, too slow, and too busy with executive rotations and OKRs to move quickly.

The winners in this space know how to iterate and build in public, collapse feedback loops, tell compelling stories, and design from first principles. Even when established products catch up technically, the zeitgeist has already shifted, and users (especially developers) are difficult to win back once they've moved on.

The lesson? In AI, speed and execution trump resources and existing advantages. The window for building category-defining products remains wide open as long as you can move fast enough.

This reminds me of what Peter Thiel said about “last mover advantages”. If you think deeply enough about what users really want, you will create the last great development in a specific market and enjoy years, if not decades, of monopoly profits.

Quote to Share

Danny Postma on AI democratising expertise:

Twitter tweet

AI is creating a fundamental shift in how we define expertise.

I think this is a great summarisation of Danny Limanseta’s observation that “broad knowledge and surface-level understanding across domains is becoming more valuable than deep expertise in a single technical skill.”

Expertise is now increasingly about knowing how to access and direct knowledge rather than personally possessing it.

We’re now witnessing the emergence of a new cognitive partnership. The combination of human curiosity, context, and critical thinking with AI’s vast knowledge and tool use to create and automate tasks at scale.

The most successful individuals will be those who embody critical thinking and ask deep, probing questions—challenging the AI—to evaluate, synthesise and thoughtfully apply the answers.

Source: Danny Postma on X

Question to Ponder

One hundred per cent.

As we explored above, you now need to turn your attention toward directing knowledge instead of memorising it.

AI is capable of writing anything you’ve ever written as long as your prompt is good enough. That is, providing the necessary context, examples, and framework to produce a great output.

But the idea that AI will replace thinking (and subsequently make humans brain-dead beings) is a totally flawed argument.

It reminds me of the age-old adage, “garbage in, garbage out”. In the context of AI, the quality of your output is only as good as the quality of your prompt. And the quality of your prompt is only as good as the quality of your thinking when it comes to applying the necessary context and asking the right questions to get the answer you desire.

Directing knowledge, critical thinking, and questioning assumptions are the core skills that have to be prioritised in the age of AI.

💡 If you enjoyed this issue, share it with a friend.

See you next week, Alex Banks P.S. The humanoid robot half-marathon in Beijing just started.

AI Highlights

My top-3 picks of AI news this week.

Sundar Pichai / The Verge

Google
1. Google's AI Dominance

Google has unveiled a series of significant AI advancements this week at Google Cloud Next 25, as they continue to build the future of AI infrastructure and applications.

• Ironwood TPU: Their seventh-generation Tensor Processing Unit (TPU) designed specifically for inference, delivering 5x more peak compute capacity compared to the previous generation, with 42.5 Exaflops of power (24x more than the world's largest supercomputer).

• Agent2Agent (A2A) Protocol: A new open standard for AI agent interoperability that enables secure communication, task coordination, and context sharing among AI agents across different platforms and vendors, supported by over 50 technology partners.

• Enhanced Gemini Capabilities: Deep Research is now powered by Gemini 2.5 Pro, bringing improved information synthesis and analytical reasoning, while Gemini Live now allows users to share their screen or camera to brainstorm and troubleshoot in real time.

Alex’s take: Google’s woken up. They have been sitting on a distribution and hardware advantage all this time. They’re now starting to lean into this advantage, and I think it will become even more apparent over the coming decade, especially as they’re making increasingly more efficient, bespoke hardware like Ironwood in-house to run their AI models. This could mean they beat the competition on a cost basis—there’s a serious advantage to be found in owning the vertical from chip to chatbot.

OpenAI
2. ChatGPTs Memory Upgrade

OpenAI has rolled out memory improvements to ChatGPT, enabling more personalised interactions by referencing past conversations.

• Comprehensive recall: ChatGPT can now draw from your entire chat history to provide contextually relevant responses based on your preferences and interests.

• Two-tiered memory system: Features both "Saved Memories" (explicitly requested information) and "Chat history" references that work together to create more tailored experiences.

• User control: Complete management of what ChatGPT remembers with options to delete individual memories, clear all saved memories, or turn off memory features entirely through settings.

Alex’s take: What I like about this move is that it solves the annoying “memory reset” issue we all experience when working alongside LLMs. We are now one step closer to truly personalised AI relationships rather than isolated conversations. However, the feature is absent in European markets (including the UK) due to much stricter data protection regulations.

AWS
3. Amazons Double Drop

AWS has released two AI models focused on enhanced video and speech capabilities, both of which are now available on Amazon Bedrock.

• Nova Sonic: A new speech-to-speech model designed to capture not just what people say but how they say it—including tone, style, pauses, and interruptions.

• Nova Reel 1.1: Generate multi-shot videos up to 2 minutes in length with consistent styling across shots, elevating the quality and coherence of AI-generated video content.

• Unified approach: Nova Sonic combines speech understanding and generation in a single model, preserving context and nuance for more natural conversations.

Alex’s take: I’m a big believer that speech will be the predominant interface for generative AI. It’s great to see AWS sharing their advancements across both visual and audio domains, given their size and position. They have a serious distribution advantage for those already embedded in the AWS ecosystem. For now, I’m eagerly waiting for their pricing to be released and seeing how it stacks up against the likes of Google’s Veo 2 at $0.50 per second.

Today’s Signal is brought to you by GrowHub.

Want to know why most people fail on LinkedIn?

They spend hours writing mediocre posts that get zero engagement.

I know because I used to be one of them.

I built an audience of 130,000+ followers by turning my ideas into viral content in minutes (not hours).

The secret? GrowHub.

Drop in any:

• YouTube video

• Blog post

• Image

And watch GrowHub's AI (trained on thousands of viral posts) repurpose it into engaging content.

Grow on LinkedIn

SPONSOR

Content I Enjoyed

ElevenLabs MCP server

Voice Is Becoming the New Interface for AI

This week, ElevenLabs released its MCP server. This means you can build a voice agent that lets you do things like order pizza, transcribe speech, and generate audio using simple API calls.

For those wondering what model context protocol (MCP) is, you can think of it as a USB-C port for AI applications. MCP provides a standardised way to connect AI models to different data sources or tools.

The ElevenLabs MCP server sits on top of the ElevenLabs API to provide AI models with context to access the ElevenLabs AI audio platform.

I think this is a really interesting step towards voice and audio becoming the primary interfaces of the AI economy.

For example, when I ask Claude, Grok, ChatGPT, or any other LLM provider a prompt on my computer, I no longer use my keyboard. I use my voice instead, leveraging a favourite tool of mine called Wispr Flow.

But when it comes to performing outbound calls, following up with leads or even booking appointments, the ElevenLabs MCP server seems to be scratching the surface of something extraordinary.

My research also led me to find this demo where Luke Harries, Growth Lead at ElevenLabs, combined the WhatsApp MCP server with their own MCP server to transcribe voice notes and send audio messages using ElevenLabs’ library of 3,000+ voices. Wild.

I believe MCPs are turning out to be something of the AI app store.

Idea I Learned

Custom dashboard / shadcn on X

The Composability of AI Tools

You write 7 lines of code and have a gorgeous dashboard.

When would this have been possible ever before?

What I love about the possibility of generative AI is it has fundamentally removed the barriers to “be technical”.

English is the new programming language. No longer is there a requirement to memorise complex functions and formulas. If you get stuck, you can simply ask an LLM (with search capabilities to get up-to-date information), “How do I do this?”.

It is a frictionless facilitator that lets anyone explore their curiosity and go down deep rabbit holes without having to pause, spend time finding adjacent materials to help their endeavour, and then continue. You can now get any answer you could possibly want on demand as long as you ask the right question.

As shadcn on X demonstrates, you can now build gorgeous UI by combining building blocks across the web. It’s almost like stacking legos and building your desired creation—everything is becoming composable and working on top of each other.

Then, you can combine deep research tools like Grok 3 with AI-powered development tools like Cursor, v0, and Lovable, and you’re cooking with gas to turn your idea into life.

If you’re interested in my AI-powered workflow using tools like this for non-technical folk, let me know by responding to this email.

Quote to Share

Elon Musk on Sam Altman:

Twitter tweet

Earlier this week, OpenAI Newsroom put out a post on X directly addressing Elon Musk’s actions against OpenAI. This legal back-and-forth was born out of Musk's lawsuit to block OpenAI's conversion to a for-profit structure back in 2024.

But things have recently become vocal on X, with Musk replying to a screenshot of OpenAI’s statement with, “Scam Altman is at it again.” This mainly refers to Sam Altman’s original statement of, “I’m doing this because I love it,” and not for the money, made when he testified before Congress in 2023 about the dangers of AI. It was only later announced that he would receive approximately $10 billion in equity if OpenAI successfully converts to a for-profit model.

OpenAI’s original mission was to build a non-profit, open-source AI company. It is now a closed-source (and well on its way to becoming a) for-profit AI company—a complete 180 from its original mission when Musk helped found the company in 2015.

Source: Elon Musk on X

Question to Ponder

By 2100, AI researchers predict full automation of all human jobs—though this could happen much sooner.

Our economic system was built around human labour as the primary means of distributing resources. When that foundation shifts, everything must adapt.

With AI eventually driving the cost of intelligence and the cost of labour to near-zero, we need to think about how we navigate a society like this.

Understanding this trajectory highlights an important question. Will we need basic income? I believe it’s one that is increasingly more difficult to ignore. Workers competing with automation would need the security of some basic income.

Currently, routine, low-creativity jobs like trucking or data entry are prime candidates for disruption. Jobs needing creativity, emotional depth, or social finesse—like artists, therapists, or nurse anesthetists—are far more difficult for AI to compete against as it ultimately can’t (yet) “feel”.

However, the WEF expects 39% of core skills to shift by 2030. Therefore, focusing on upskilling and retraining will be paramount.

Much like the Industrial Revolution transformed work rather than abolished it, I think we’ll see the same with the rise of AI. New roles will be created, old ones will fade, and our definition of “work” will shift over time.

💡 If you enjoyed this issue, share it with a friend.

See you next week, Alex Banks P.S. Unitree will be livestreaming robot combat in about a month.

AI Highlights

My top-3 picks of AI news this week.

Mark Zuckerberg on Llama, AI, & Minus One / South Park Commons

Meta
1. Meta's Llama 4 Herd

Meta has released the first Llama 4 models: Llama 4 Scout and Llama 4 Maverick, with significant improvements in architecture, performance, and multimodal capabilities:

• Record-breaking context: Llama 4 Scout offers an industry-leading 10M token context window (up from 128K in Llama 3), enabling multi-document reasoning and extensive code analysis.

• Mixture-of-experts architecture: First-ever Llama models using MoE design, where Maverick has 17B active parameters but 400B total parameters, with only a subset activated during inference for better efficiency.

• Competitive benchmarks: Llama 4 Maverick outperforms GPT-4o and Gemini 2.0 Flash across multiple benchmarks while Behemoth (still training) already surpasses GPT-4.5, Claude 3.7, and Gemini 2.0 Pro on STEM benchmarks.

Alex’s take: Meta claims their new models outperform the likes of GPT-4o and Gemini 2.0 Flash “across a broad range of widely accepted benchmarks.” But we must look deeper.

It turns out the current Llama 4 herd is actually quite bad at coding and underperforms the likes of GPT-4o, Gemini Flash, and Grok 3.

This reminds me of what Chamath recently said on the All-In podcast. The dirty little secret of these model makers is that these models are so tuned to maxxing the benchmark results, they’re overfitting on the data, and that makes them unreliable.

They’re optimising for the benchmark tests. The solution? We need to have, as Chamath says, “extremely difficult and always changing third party independent verifiable benchmarks”. Then, these leaps in advances will mean more than what they do today.

Runway
2. Runway Gen-4 Hits The Scene

Runway AI has introduced Gen-4, its next-generation series of AI models for media generation with unprecedented world consistency capabilities.

• Character consistency: Gen-4 can maintain consistent characters across multiple scenes and environments using just a single reference image.

• Object permanence: The model allows placement of objects across various locations and conditions, ideal for narrative content and product photography.

• Multi-angle coverage: Users can generate different angles of the same scene while maintaining consistency in subjects and style.

Alex’s take: Runway is becoming seriously consistent, not just for their model releases but also for the features within Gen-4. Their new model series can now not only simulate real-world physics but also maintain characters, objects and environments across multiple scenes, which I think is brilliant. This was previously a missing link, but it might now accelerate AI-generated video to a truly viable scale.

Midjourney
3. Midjourney V7 Drafts Mixed Emotions

Midjourney has launched its V7 image-generation model in alpha, sparking diverse reactions in the AI art community.

• Voice prompting: Users can now speak directly to the model, which converts speech-to-text prompts and generates images based on verbal descriptions.

• Draft Mode: Generates images up to 10 times faster at half the cost, allowing creators to quickly iterate before enhancing selected images to full quality.

• Expanded ecosystem: Integration with tools like Higgsfield AI enables animation capabilities, allowing users to create dynamic video content from static images.

Alex’s take: OpenAI blew us away last week with their 4o image generator. This was Midjourney's chance to take the attention back. It seems to me that Midjourney had a terrific head start, but now, others have caught up. They are currently not in the same league as OpenAI when it comes to realism and text generation. Things move fast in AI, so I look forward to seeing what they release over the coming months, but for now, my subscription is cancelled.

Today’s Signal is brought to you by Morning Brew.

Here’s Why Over 4 Million Professionals Read Morning Brew

• Business news explained in plain English

• Straight facts, zero fluff, & plenty of puns

• 100% free

See for yourself

SPONSOR

Content I Enjoyed

Tracing the thoughts of a large language model / Anthropic

We Can Finally See How AI Thinks

This week's reading led me to stumble upon this article from Anthropic titled “Tracing the thoughts of a large language model”.

It does a great job of uncovering the hidden mechanisms inside language models. This research changes everything we thought we knew about how these systems operate.

Here are my favourite discoveries:

1. AI's default mode is skepticism

   • Surprising discovery: "I don't know" is the baseline response to ALL questions

   • This default refusal is only overridden when confidence reaches a threshold

   • Hallucinations occur when an entity seems familiar enough to override caution

   • It's the AI equivalent of the Dunning-Kruger effect: confident just enough to be wrong

2. Language models plan ahead like poets

   • Instead of improvising rhymes, LLMs plan them before writing a single word

   • Multiple rhyming candidates are considered simultaneously

   • The entire line is crafted to naturally lead to the pre-selected rhyme

3. Jailbreaks exploit grammar over safety

   • The "Babies Outlive Mustard Block" trick works because the AI:

   • Processes each first letter without realizing it spells "BOMB"

   • Gets trapped by its own drive for grammatical coherence

   • Only escapes at natural sentence breaks

   • Removing punctuation makes jailbreaks more effective by blocking the exit points

4. There's a universal "language of thought"

   • There are no separate models for each language

   • Instead, a shared abstract space with language-specific adapters

   • Knowledge transfers between languages through this shared conceptual core

   • More powerful models show increased language-independent processing

If you want to dig deeper, I recommend you check out the full paper by Anthropic.

The ability to peek inside the “mind” of these LLMs represents a massive breakthrough for interpretability research.

AI is no longer a black box.

Idea I Learned

Human controlling an AI on a computer / Wired

How to Spot AI-Written Content (And It’s Not the Em Dash)

Everyone keeps saying Em dashes (—) are the giveaway of AI writing.

You know what I mean—these little horizontal lines that break up sentences—which I personally love using in my own writing—because they perfectly capture the way my thoughts connect.

But I believe there’s a much more reliable tell.

The “it's not X, it's Y” construction.

For example, when looking across the latest and greatest AI news for today’s issue of The Signal, I came across an article by Tinder about their new AI game. It reads:

“Tinder’s The Game Game™ Isn’t About Getting It Right—It’s About Getting Comfortable Before Meeting IRL”

Do I think this title was written with AI? Yes. Why? Well, why would I think of a good title when AI can give me a great one with zero cognitive lift?

Further to my point, my partner recently started listening to “The Let Them Theory”, a book by Mel Robbins.

After listening for 30 minutes, she couldn’t continue because of this exact reason. She uses LLMs a lot in her daily life, which gives her a keen eye for noticing the “it's not X, it's Y” pattern emerge.

Examples from the book:

“It’s not about the hand you’ve been dealt. It’s about how you play the hand.”

“It’s not about superiority at all. It’s about balance.”

“It’s not that they are ghosting you; it’s that they are exhausted by the chapter of life they are in.”

This could obviously be Mel’s writing style that helps articulate her points.

But it’s a trait that the likes of ChatGPT and Claude default to for nearly every answer.

Once you know it, you can't unsee it.

I now see it too often for my own liking in ads, posts, comments, books, and podcasts.

My point is that I’m not against using AI as a collaborative tool to help formulate your thoughts and find ideas. But what I am against is the simple copy-and-paste jobs that flood our feeds with the same predictable patterns (more about this later).

So the next time you’re scrolling and spot that “not X but Y” structure, you might have found a clue.

Quote to Share

Anthropic’s move into higher education:

Twitter tweet

This feels like a smart move for Anthropic. The education sector has been hesitant since the ChatGPT moment of November 2022, when students immediately saw generative AI as a homework solution to their toughest essay questions.

But now, Claude’s new “Learning mode” is designed specifically to guide students through their reasoning process with Socratic questioning.

Instead of relying on AI to provide a copy-and-paste answer, students can use this learning mode to enable critical thinking.

As we observed last week at Austin’s Alpha School, I believe that an “AI for everyone” is the best solution to provide deeply personalised education that adapts to and benefits all students, regardless of their learning style.

Source: Anthropic

Question to Ponder

The unfortunate truth is that I’m seeing A LOT of AI-generated content on my feed these days.

You can’t blame them. AI is a frictionless facilitator that transforms small thoughts into big ideas.

But when it’s all AI, I and many others can quickly see through the soft shell of content that presents a vast and very empty chasm beneath.

You see, the best content infuses human storytelling and connects ideas across domains.

You don’t even need to have the best ideas, but you must offer a unique perspective on a topic and be able to express it in a way that connects to something more.

That’s how I believe the best content is born.

This will stand out more and more amidst the sea of AI noise. Humans crave authenticity. And the creators that can thoughtfully use AI to uplift their ideas with human storytelling will win.

💡 If you enjoyed this issue, share it with a friend.

See you next week, Alex Banks P.S. Your Clothes, Give Zem To Me. Now.

AI Highlights

My top-3 picks of AI news this week.

Introducing 4o Image Generation / OpenAI

OpenAI
1. GPT-4o Image Mania

OpenAI has blown up social media this week with their latest GPT-4o image generator that has everyone talking. They’ve integrated advanced image-generation capabilities directly into ChatGPT, and the results are genuinely impressive.

• Multimodal integration: Image generation is now a core capability of the language model itself, allowing for seamless conversation-based image creation and refinement.

• Text rendering: GPT-4o excels at accurately rendering text within images, making it ideal for creating diagrams, signs, and other visual communication tools.

• Contextual understanding: The model can analyse uploaded images and maintain consistency across multiple iterations, a serious leap in visual creation.

Alex’s take: I think this is the tipping point where image generation within AI models becomes truly mainstream. Midjourney stole the thunder for the last few years. Then, new entrants like Ideogram and Flux created some waves, but the consistency of GPT-4o’s image gen capabilities (both in style and composition with text) is leaps ahead of the rest. One of my favourite use cases of this feature is that people are now extracting image data and using this for style replication.

Google DeepMind
2. Google Unlocks Gemini 2.5

Google DeepMind has unveiled Gemini 2.5, its most intelligent AI model yet, designed specifically as a “thinking model” to tackle increasingly complex problems.

• Leading performance: Gemini 2.5 Pro Experimental tops the LMArena leaderboard by a significant margin and excels across benchmarks requiring advanced reasoning.

• Enhanced reasoning capabilities: The model scores 18.8% on Humanity's Last Exam without tool use, setting a new state-of-the-art benchmark for complex knowledge and reasoning.

• Impressive coding skills: Achieves 63.8% on SWE-Bench Verified with custom agent setup, with the ability to create visual web apps and executable game code from simple prompts.

Alex’s take: This is a serious leap forward in AI reasoning capabilities. If it’s better at coding than Claude 3.7 Sonnet (200K context) while having a 1 million token context window (2 million coming soon), it will become very useful as a coding agent. This is now turning into a four-horse race with Gemini, Grok, Claude and ChatGPT all competing for the top spot.

xAI
3. Musk Merges His AI and Social Media Empires

Elon Musk has announced that his artificial intelligence company xAI has acquired his social media platform X (formerly Twitter) in an all-stock transaction.

• Valuation: The deal values xAI at $80 billion and X at $33 billion ($45B less $12B debt), despite Musk originally purchasing Twitter for $44 billion in 2022.

• Integration: The merger combines data, models, compute, distribution and talent between the two companies, which were already sharing significant resources.

• AI integration: xAI has been using data from X to train its models, with its chatbot Grok already being a prominent feature on the platform.

Alex’s take: I suspect this merger will accelerate the integration of Grok across X while giving xAI access to one of the world’s largest real-time data sources. This is a big step in creating an “everything” app, much like WeChat in China. I suspect payments will be coming soon, and it will become almost like a digital twin to its users.

Today’s Signal is brought to you by GrowHub.

Your personal brand is 10x more important than your CV

• Researching, drafting, and editing content takes 10+ hours per week

• Ghostwriters are expensive and cost $000’s in monthly fees

• GrowHub lets you create viral content in seconds

Start Growing

SPONSOR

Content I Enjoyed

Isaac Asimov / BBC Archive

The Convergence of Humans and Machines

This week, I stumbled across a BBC Archive interview from 1965 with renowned science fiction author Isaac Asimov.

In the interview, he shared a prolific idea surrounding the convergence of humans and robots into hybrid entities.

On the one hand, he foresaw robots becoming more human-like, while humans would increasingly incorporate artificial components. Asimov thought that at some time in the future, it might become increasingly challenging to distinguish between the two, raising important questions around what “being alive” and “consciousness” truly mean.

We only have to look back over the last 5 years to see some early glimmers of this in action.

We already have humanoid robots starting to walk and run like us. Take, for example, Boston Dynamics’ recent demo of Atlas.

What about neural networks mimicking our brains, and the transformer architecture of LLMs like ChatGPT being essentially a compression of the internet, predicting the next word in a sequence of words using human language?

And whilst artificial systems are becoming more “human”, we humans are already experimenting with brain-computer interfaces like Neuralink alongside advanced prosthetics and artificial organs replacing failing human ones.

Will we see a hybrid species in the future? I believe so. Yet, it will require humanity to deeply understand the ethical and societal implications such integration would bring and what its lasting impact might be.

Idea I Learned

Hayao Miyazaki / The Asian Cut

The Ghibli Creator is Disgusted by AI

OpenAI’s advanced image generator was introduced into GPT-4o on Tuesday, and it has been a total frenzy ever since.

My feeds on LinkedIn and X this week have been flooded with Ghibli Studio-style images.

However, Hayao Miyazaki, co-founder of Studio Ghibli himself, isn’t quite as enthralled.

I discovered this 2016 video of Miyazaki from YouTube that captures his perspective:

“AI-generated animation is an insult to life itself.”

Each film from his studio is hand drawn and painted with watercolours. Each frame (typically 60,000 to 70,000 per film) means this process can take years. Now, anyone can create a Ghibli-style image using GPT-4o.

This highlights the key concerns the world needs to address in how these AI tools are trained to imitate styles.

Is OpenAI violating copyright law?

I enjoyed sinking my teeth into this article, which highlights the perspective of an intellectual property lawyer on the whole situation.

The “TL;DR” is that OpenAI does not appear to be breaking the law as an image’s style is not explicitly protected by copyright. However, the training on copyrighted works is still up in the air as to whether it falls under fair use protections.

It made me think of what Adobe is doing with its Firefly model. When I attended Adobe MAX in 2024, Adobe unveiled Firefly Video, touting it as “the first commercially safe video generation model.”

This claim stems from Adobe’s strategy to train Firefly on a dataset with known provenance—using Adobe Stock images, openly licensed content, and public domain materials where copyright has expired.

This approach contrasts with the uncertainty surrounding OpenAI’s training data, where critics, including intellectual property lawyers, question whether scraping copyrighted works from the web for AI training falls under fair use.

If a model like OpenAI’s isn’t as “commercially safe” due to potential legal risks, some users might actually prefer models like Firefly for their transparency and reduced risk of copyright infringement lawsuits.

But right now, there’s a lot of tension in the industry as courts continue to grapple with whether training AI on copyrighted material qualifies as fair use or constitutes infringement. According to the U.S. Copyright Office and legal experts, for a work to be copyrighted, it must have a human author, but the legal status of AI-generated content remains murky.

Until clarity is found, I believe the debate over whether tools like GPT-4o or Firefly are truly “safe” will continue to spark discussions across social media, legal forums, and creative communities, especially as artists like Miyazaki voice their concerns about the erosion of traditional craftsmanship in the face of AI.

Quote to Share

Texas private school’s use of new “AI tutor” rockets student test scores to top 2% in the country:

“We use an AI tutor and adaptive apps to provide a completely personalized learning experience for all of our students, and as a result our students are learning faster, they're learning way better. In fact, our classes are in the top 2% in the country.”

At Austin’s Alpha School, students spend just two hours daily with AI tutors, using the remainder of their day to develop skills like public speaking and financial literacy.

This approach effectively solves what educators call “Bloom's 2 Sigma Problem”—the finding that one-on-one tutoring dramatically improves learning outcomes but has historically been too expensive to implement widely.

I think the key advantage is how AI allows students to ask unlimited questions without fear of judgment.

It’s very common for people to hold back from asking questions for fear of looking “dumb” in the eyes of fellow students and the teacher for not getting it.

I struggled with this anxiety a lot while going through school. Not understanding something, asking an initial question, still not getting it, and then being too intimidated to keep asking for clarification.

As Alpha School expands across the US, we’re witnessing education’s potential future where technology handles personalised instruction. At the same time, human teachers focus on providing what machines cannot: emotional support and real-world guidance.

Source: Fox News

Question to Ponder

I saw a great answer to this recently. Rhona D’Arcy uses AI to monitor her 86-year-old mum Heather, who has vascular dementia but refuses to leave her home and lives three hours away from her daughter.

The AI system learns Heather's normal patterns—room movements, bathroom usage, fridge opening—and alerts Rhona when something’s off.

Economists say similar systems could save the NHS £1.2bn by 2035, but I think the true value transcends money.

Critics worry face-to-face care might be cut. As Dr. Collingham notes, “The cornerstone of good care is human connection.” But for Rhona, “It doesn't replace the care, it augments it.”

Another great example is the Ray-Ban Meta Glasses and their partnership with “Be My Eyes” to help people who are blind or have low vision. One use case in particular that caught my eye (if you’ll pardon the pun) is of the glasses helping a blind veteran navigate their daily life.

This is a future we can all get excited about.

💡 If you enjoyed this issue, share it with a friend.

See you next week, Alex Banks P.S. Unitree's G1 humanoid robot nails the world's first kip-up.

AI Highlights

My top-3 picks of AI news this week.

Jensen Huang (CEO of Nvidia) / GTC March 2025 Keynote

Nvidia
1. Biggest updates from Nvidia’s GTC 2025

Nvidia CEO Jensen Huang’s GTC 2025 keynote unveiled a series of major announcements spanning AI acceleration, robotics, and autonomous systems.

• Blackwell Ultra: Coming in H2 2025, offering the same 20 petaflops of AI performance as standard Blackwell but with 288GB of memory (up from 192GB), enhancing AI inference capabilities.

• Future GPU Roadmap: Vera Rubin architecture for late 2026 with 50 petaflops for inference (2.5x Blackwell), followed by Rubin Ultra in 2027 (100 petaflops), and eventually Feynman GPUs in 2028.

• Personal AI Supercomputers: DGX Spark and DGX Station built on the Grace Blackwell chip bring supercomputing power to the desktop for high-end enterprise and research applications.

• Isaac Groot N1: An open-source foundation model for humanoid robots that enables collaborative work, independent learning, and optimisation across various environments.

• GM Partnership: Nvidia is teaming up with General Motors to bring AI into manufacturing, vehicle design, and autonomous driving systems.

Alex’s take: I think it’s fascinating to see the progression from the GenAI boom of 2022 through to now being on the cusp of agentic AI and physical AI in 2025. Nvidia is building the infrastructure to power it all. I also think it’s great to see them open-sourcing models like Groot N1 to accelerate physical AI, similar to how open LLMs transformed generative AI.

Adobe
2. Adobe’s Agent Orchestrator

Adobe has unveiled its ambitious AI strategy with the launch of Adobe Experience Platform Agent Orchestrator, bringing agentic AI to enterprise workflows.

• Suite of 10 purpose-built agents: From Site Optimisation to Content Production, these specialised agents handle everything from creating personalised audience segments to detecting and fixing website issues.

• Brand Concierge application: Creates personalised, immersive experiences using a company’s unique brand attributes and customer data.

• Multi-agent ecosystem: Strategic partnerships with major players including Microsoft, ServiceNow, and AWS enable seamless execution across different AI agent environments.

Alex’s take: Adobe has taken the 10 most common business functions, including the marketer, account manager, and data engineer, and turned their common workflows into agentic systems. This is especially true for niche use cases like marketing where you can pull on rich past performance data to inform future campaigns.

Google
3. Gemini's Watermark Woes

Google’s new Gemini 2.0 Flash model has sparked controversy with its ability to remove watermarks from images, including those from Getty Images and other stock media companies.

• Powerful editing: The model can not only remove watermarks but also intelligently fill in gaps created by watermark deletion.

• Experimental status: Currently labelled as “experimental” and “not for production use,” available only in developer-facing tools like AI Studio.

• Legal concerns: Removing watermarks without the original owner's consent is generally considered illegal under U.S. copyright law.

Alex’s take: Google has stated that using its tools to infringe copyright “violates their terms of service.” However, it’s interesting that when you ask a different image generation/manipulation model like OpenAI’s DALLE-3, requests like watermark removal are outright refused. I expect we’ll see more robust protections implemented soon to give greater certainty to the creator.

Today’s Signal is brought to you by Artisan.

Hire an AI BDR & Get Qualified Meetings On Autopilot

Outbound requires hours of manual work.

Hire Ava who automates your entire outbound demand generation process, including:

• Intent-Driven Lead Discovery Across Dozens of Sources

• High Quality Emails with Human-Level Personalization

• Follow-Up Management

• Email Deliverability Management

Hire Ava to slash costs & boost productivity.

SPONSOR

Content I Enjoyed

Jensen Huang (CEO of Nvidia) and “Blue” / GTC March 2025 Keynote

The Cute Robot That Captivated GTC 2025

Among all the impressive announcements at Nvidia’s GTC event this year, there was one that truly sparked wonder inside of me.

It was the result of a collaboration between Nvidia, Disney Research, and Google DeepMind.

The result: “Blue”, an incredibly cute Star Wars-inspired BDX droid.

The Newton physics engine powering Blue is an open-source simulator designed specifically for robotics that allows artists to animate robots exactly as they envision.

The AI then handles the complex physics to make those movements work in unpredictable environments.

By creating virtual worlds with incredibly realistic physics, they can have thousands of digital robots trip, fall, and crash without damaging expensive hardware in the real-world.

The engine is highly customisable too, allowing developers to program robotic interactions with food, cloth, sand, and other deformable objects—things that have traditionally been a nightmare for robotics.

Disney plans to bring robots like Blue into their theme parks starting next year. The technology will also be available as an open-source release later this year.

I think there’s something refreshingly appealing and playful about a robot that feels more like R2D2 than Terminator.

Idea I Learned

Shantanu Narayen (CEO of Adobe) and James Quincey (CEO of Coca-Cola) / Adobe Summit

CEO of Coca-Cola James Quincey on the future of AI advertising

This week, I attended the Adobe Summit in Las Vegas.

James Quincey, CEO of Coca-Cola, gave a fascinating talk on day one of the event.

Only in December did Coke release its Christmas advert, which faced a lot of backlash for a simple reason—it was 100% AI-generated.

“It was cheaper to make than historical regular ads, it was much quicker to make so it was a huge productivity opportunity.”

Yet, on reflection, James noted:

“It cannot produce a resolution on people's faces that humans will buy into… [GenAI] is not yet at the stage when it can make all our ads because we want people in them and humans are incredibly effective at noticing AI.”

As video models get better in the future, James sees two opposing forces taking place:

On the one hand, the cheaper it is to make an AI-generated advert, the easier it is to put the end consumer in the ad when you watch it on your phone.

If you didn’t click “buy”, the AI can A/B test the ad and keep re-showing it until you do. You’re essentially perpetually optimising your ads and chasing the consumer until they purchase your product.

On the other hand, the second force is you now have a bunch of annoyed consumers who will block your product.

Where does the advertiser now spend its money?

Live experiences are the only thing that can’t be avoided.

So there’s a line to dance on here between the cost of creativity and people’s level of annoyance from hyper-personalisation.

James feels the thing that will set you apart and prevail as this becomes a reality is the quality of the creativity. “I don’t want to be spammed with rubbish.”

It’s clear creativity has a premium—even more so in the age of GenAI, where a mountain of content slop is beginning to pile up.

The only thing that now towers over that is authentic and thoughtful content that evokes a deep emotional response.

I’m confident this will be where the enduring advantage lies between humans and AI.

Quote to Share

Ethan Mollick on AI passing the Turing Test for memes:

Twitter tweet

A recent study from KTH Royal Institute (Sweden’s largest technical university) demonstrated that AI has now surpassed average humans in creating humour—at least when it comes to memes.

The research compared three groups: humans alone, humans with AI assistance, and AI working independently.

AI-generated memes scored higher on humour, creativity, and shareability than both human-only and human-AI collaborations. However, when looking at the very best examples, humans still created the funniest memes, while human-AI teams excelled at creativity and shareability.

Perhaps this study reveals something many of us already suspected: the average human isn't particularly funny, and most people have fairly predictable tastes in memes.

On the other hand, AI now excels at pattern recognition and generating content that appeals to the broadest possible audience. It essentially creates humour by averaging preferences across millions of examples.

While the unique spark of exceptional human creativity remains persistent, we’re now entering a new phase where AI can generate genuinely entertaining content for the masses.

Source: Ethan Mollick on X

Question to Ponder

This question reminds me of a recent episode of the Joe Rogan experience with chess champion Josh Waitzkin.

Within 3 hours of experimentation, AlphaZero was stronger than any human or computer in the history of chess.

Stronger than Magnus Carlsen, Gary Kasparov, and Bobby Fischer.

AlphaZero has an ELO of 3700. The strongest players today have an ELO of ~2800. When Josh was 9, he had an ELO of 1900.

To put this gap into perspective:

The ELO rating difference between the strongest AI chess engines and the human world champion is the same as the gap between the world champion and a talented 9-year-old.

Humans naturally have a hard time adapting to paradigm shifts. We’re inherently resistant. We fight “tooth and nail” to maintain our conceptual schemes, especially when our identity and life’s work are tied to them.

But what if we embraced the humility of, as Josh puts it, being the “ant relative to the human”?

What if we approached this new relationship with the curiosity and playfulness of children who aren't afraid of failure or looking bad?

I agree that this childlike openness to learning without ego might be our greatest asset in a world where we’re no longer the apex intelligence on the planet.

💡 If you enjoyed this issue, share it with a friend.

See you next week, Alex Banks P.S. Boston Dynamics’ Atlas walks, runs and crawls like a human.

AI Highlights

My top-3 picks of AI news this week.

OpenAI’s new tools for building agents / X

OpenAI
1. OpenAI's Agent Acceleration

OpenAI has unveiled a comprehensive suite of tools for developers to build sophisticated multi-agent systems with unprecedented speed and efficiency.

• Agent Tools Trifecta: Three powerful new tools including Web Search Tool for accessing up-to-date internet information, File Search Tool for advanced RAG applications, and Computer Use Tool for programmatic computer control.

• Responses API: A unified framework that supports multiple turns and simultaneous tool calls, eventually replacing the Assistant API (planned for 2026).

• Production-Ready Agents SDK: The evolution of their experimental Swarm framework now features task handoffs with context preservation, built-in monitoring, and validation guardrails.

Alex’s take: I’ve seen a lot of talk that this will kill agent startups. I disagree. Agent startups will need to find deeper competitive advantages beyond what’s now readily available through OpenAI’s SDK. Competition is driving better tools for everyone.

Google
2. The Google Method

Google has launched a series of AI updates across their product lines, making their AI more accessible, powerful, and personalised than ever before.

• Gemma 3: A new family of open-source models (1B-27B parameters) built from Gemini 2.0 technology, designed to run on a single GPU or TPU with multimodal capabilities, 128K token context window, and support for 140 languages.

• Gemini 2.0 Flash upgrades: Enhanced with file upload support, a 1M token context window for Advanced users, and native image generation and editing capabilities.

• Deep Research: Now powered by Gemini 2.0 Flash Thinking Experimental and available to all users in 45+ languages, enabling more detailed and insightful multi-page reports.

• Image-to-video with Veo 2: Google's Veo 2 technology, first integrated in Freepik AI Suite, now transforms static images into short, high-quality videos with natural motion.

Alex’s take: Google is seriously pushing the envelope. They were a sleeping giant in the first GenAI wave of 2022-2024. 2025 came along and they woke up. They’re drip feeding consistent updates each week vs the “big drops” of other AI-first companies like OpenAI. I can only feel that Google’s distribution advantage will make them a long-term winner.

Captions
3. AI Creator Takeover

Captions has launched “Mirage,” the world's first foundation model specifically designed for generating user-generated content (UGC).

• Complete video generation: Creates full videos from just audio files or scripts, eliminating traditional production requirements.

• Original AI actors: Generates realistic digital personalities with natural expressions that can deliver your message.

• Licensing freedom: Produces content entirely free from rights management and licensing restrictions.

• Scene creation: Builds complete backgrounds and environments from simple text prompts.

• Full customisation: Allows for personalisation of actor appearance, voice, and presentation style.

Alex’s take: We’re witnessing the democratisation of video production at an unprecedented scale. Just as AI writing tools transformed content creation, Mirage is one of the first who will ride on the explosion of AI-generated video flooding our feeds. As AI content becomes ubiquitous, I suspect we’ll see authentic human creators become more valuable precisely because of their uniqueness.

Today’s Signal is brought to you by GrowHub.

GrowHub image tools

Want to know why most people fail on LinkedIn?

They spend hours writing mediocre posts that get zero engagement.

I know because I used to be one of them.

I built an audience of 125,000+ followers by turning my ideas into viral content in minutes (not hours).

The secret? GrowHub.

Drop in any:

• YouTube video

• Blog post

• Image

And watch GrowHub's AI (trained on thousands of viral posts) transform it into engaging content.

You can now create tweet-style images for your LinkedIn posts with a single click.

Start growing

SPONSOR

Content I Enjoyed

Dario Amodei and Mike Froman / Council on Foreign Relations

The Future of U.S. AI Leadership with CEO of Anthropic Dario Amodei

This week Dario Amodei, CEO of Anthropic, sat down with the Council on Foreign Relations to discuss AI’s trajectory, risks, and opportunities.

Something stood out to me during the interview.

Amodei made a bold prediction about the future of software development:

“I think we'll be there in three to six months—where AI is writing 90% of the code. And then in twelve months, we may be in a world where AI is writing essentially all of the code.”

Having spent months using AI coding tools like Cursor, Lovable and V0, I’ve seen my productivity skyrocket—but in my opinion we’re nowhere near 90% replacement yet.

Today, building complex systems requires a much deeper understanding than what AI currently grasps, especially when business requirements are rarely clear enough for AI to interpret correctly.

Even when you get in the weeds, testing, debugging and back-end development all require human expertise. Security is another lens here that demands oversight.

I believe the true value of software engineers lies in orchestrating complex systems and translating business needs. The role will evolve rather than disappear.

This pattern is emerging across industries: not wholesale replacement, but a fundamental shift in how we work alongside increasingly capable AI systems.

The best developers will be those who ask the best questions. Human language is becoming the new programming language.

Idea I Learned

Yichao “Peak” Ji (Co-Founder and Chief Scientist) / Manus AI

Manus - “It’s a Claude wrapper”

This week, Chinese company “Manus AI” released a fully autonomous AI agent.

It thinks, plans, and executes tasks without human oversight.

And has sparked a polarising debate online.

When I posted this update on LinkedIn, one of the most liked comments was words to the effect of:

“But Manus is only a Claude wrapper.”

The comment was absolutely right, it is only a Claude wrapper.

As “Peak”, a Co-Founder of Manus stated on X this week:

“We use Claude and different Qwen-finetunes” alongside a variety of tools and open-source packages like Browser Use.

Turns out, adding unique experiences, functions, tools and techniques together means you are in fact wrapping up and riding on top of other technologies.

That’s how accretive value is built.

Yet something is telling me it’s almost like a “gotcha” whenever I see comments like this.

It makes the author of the comment feel good by diminishing the technology that has been constructed.

All we have to do is extend this idea for some context:

• Salesforce is an Oracle database wrapper valued at $320 billion.

• Stripe is a Mastercard wrapper valued at $70 billion.

• AWS is an HPC primitives wrapper valued at $3 trillion.

h/t to @sahilypatel on X for this.

Turns out everything in tech is a wrapper.

But things get interesting in the domain of AI “wrappers”:

We thought the application layer would be commoditised and the value was to be found in the foundation models.

I disagree.

The models themselves are becoming commodities and open sourced, so everyone can use them and build on the application layer like Manus.

As the cost of intelligence trends to near-zero, the real opportunity lies in infusing unique insights with deep proprietary data sets whilst combining the best LLMs and tools to make magical applications.

These are the areas where I see a real sustaining advantage.

So yes, Manus is a Claude wrapper.

But so is every other technology company.

Quote to Share

Figure’s introduction of BotQ:

Twitter tweet

Figure has spent the last 8 months building the BotQ manufacturing process to scale up to 12,000 humanoids per year.

What I love about the company is that they are vertically integrating and in-housing their entire manufacturing process.

They also intend to use their humanoid robots in their manufacturing process to build other humanoid robots—expected to take place this year.

I personally can’t wait to see what’s in store.

Source: Figure website

Question to Ponder

ChatGPT was my constant companion for months—but recently I've found myself gravitating toward Grok more and more.

I like to use “DeepSearch” mode for researching answers to questions. Grok combs the web for the best sources to inform its answer. Once I have a robust plan and information, I then engage “Think” mode to reason through the steps to get an answer to my question.

I find the combination of both modes to be really helpful and actually outperform even GPT-4.5 which is costing me $20/mo.

What strikes me most about this shift is how little brand loyalty seems to matter in the AI space. We simply go where the best experience is. If Grok gives better answers faster, that's where we’ll go—regardless of which company built it.

💡 If you enjoyed this issue, share it with a friend.

See you next week, Alex Banks P.S. This robot cooks Michelin-quality meals.

AI Highlights

My top-3 picks of AI news this week.

Courtesy Google / Inc

Google
1. Google’s AI is After Perplexity

Google is rolling out significant AI enhancements, including “AI Mode” and a new embedding model, with a focus on making AI more accessible and practical for everyday use.

• AI Mode in Search: A new experiment that expands on AI Overviews with advanced reasoning capabilities, allowing users to ask anything and engage in follow-up questions with helpful web links.

• Availability: Currently available in English to Google One AI Premium subscribers in the US who are 18+ and opt-in via Labs, with plans to expand to more countries and users after initial testing.

• Gemini Embedding: Released a new text embedding model that translates words and phrases into numerical representations, supporting over 100 languages and accepting larger chunks of text and code.

Alex’s take: It seems like Google has now made a carbon copy of Perplexity’s core product offering—providing AI summary answers with citations. Google's 90% share in traditional search far exceeds Perplexity's estimated 0.1%–0.2% in the overall market today. It was only a matter of time before Google’s distribution advantage started to catch up to Perplexity’s business model.

Anthropic
2. Anthropic's $3.5B Boost

AI startup Anthropic has secured a massive $3.5 billion in Series E funding at a $61.5 billion post-money valuation, led by Lightspeed Venture Partners.

• Strategic focus: Anthropic will use this investment to develop next-generation AI systems, expand compute capacity, deepen research in interpretability and alignment, and accelerate international expansion.

• New model release: The funding follows the recent launch of Claude 3.7 Sonnet, a “hybrid reasoning” model designed to more carefully consider queries before answering.

• Growing business: Anthropic's annual revenue run rate has increased by 30% so far in 2025, from its $1B figure in 2024, though the company expects to burn $3 billion this year on development costs.

Alex’s take: I’ve been a long-time user of Anthropic’s products since the release of Claude in 2023. There’s just a feeling about the model interaction that has more personality and makes it more vibrant and fun in comparison to the rather sterile GPT series from OpenAI. Model characteristics will play an essential part in determining who will come out on top in the long-run.

xAI
3. Musk's AI Megacampus

Elon Musk's AI company, xAI, is dramatically expanding its infrastructure with a massive 1 million-square-foot property acquisition in Southwest Memphis.

• Ambitious scaling: The new property will complement xAI's existing Memphis data centre “Colossus” housing 200K GPUs, creating one of the largest AI computing hubs in the world.

• Power hunger: The current facility uses 250MW of power. The new data centre will require 1.21 GW (~5 times larger) scaling to 1M GPUs, the same amount of power required by the plutonium powered flux capacitor to enable time travel in Back to the Future.

• Hardware capacity: xAI is aggressively investing in AI hardware, recently setting up a $700M data centre in Atlanta and signed a $5 billion server deal with Dell.

Alex’s take: This goes to show we have a way to go before we hit the asymptote of scaling compute capacity to improve model performance. It makes me think about what will happen when we top out with the current transformer architecture of LLMs—3M GPUs, 10M GPUs? It excites me to think what a new future architecture might look like as humanity continues scaling intelligence.

Today’s Signal is brought to you by Athyna.

Athyna

Outperform AI with Human Creativity

• Secure elite creative professionals from Latin America to scale your team.

• Save 70% on salaries with Athyna’s global talent platform.

• Enjoy a $1,000 discount exclusively for readers of The Signal.

Find Out More!

SPONSOR

Content I Enjoyed

Hao AI Lab / X

Super Mario: The New AI Benchmark

Is Super Mario the new eval for LLMs?

Hao AI Lab at UC San Diego threw various AI models into live Super Mario Bros games last week. While Claude-3.7 had already been tested on turn-based games like Pokémon Red, seeing AI tackle real-time games presented an entirely different challenge.

Claude-3.7 outperformed, while Claude-3.5 showed strength but struggled with complex maneuvers. Meanwhile, Gemini-1.5-pro and GPT-4o performed less well in the Mario environment.

The game ran in an emulator integrated with their in-house framework, GamingAgent, which fed the AI basic instructions like “If an obstacle or enemy is near, move/jump left to dodge” alongside game screenshots. The AI then generated Python code to control Mario.

Interestingly, the researchers found that reasoning models like OpenAI's o1 performed worse than “non-reasoning” models, despite being generally stronger on most benchmarks. The main reason? In Super Mario Bros., timing is everything—and reasoning models take seconds to decide on actions when a split-second can mean the difference between success and failure.

This inspired me to create my own Space Invaders themed game using Grok 3 + Claude 3.7. As someone who taught themselves to code using AI, I find it incredible to see what can now be achieved with these tools. The democratisation of creative coding means the requirement to be technical is no longer a thing. English is the new programming language.

Check out my game here.

Idea I Learned

Joaquin Phoenix / “Her” (2013)

The Hidden Tension Between AI and Social Health

This week I listened to Kasley Killam's SXSW talk on social health.

The idea behind the talk was that the more we turn to AI for connection, the more we may be undermining the very relationships that keep us alive.

Killam noted that “hundreds of millions of people around the world are turning to AI for companionship and for love.” In her research, she created her own AI friend and observed how people describe these tools as companions, friends, lovers, and even spouses.

On one hand, she's “concerned that we have created a culture where people feel like they need to turn to AI for companionship.” On the other hand, she acknowledges AI can be valuable when used as just one of many connection sources.

The future case of this looks a bit like the movie “Her” with Theodore, played by Joaquin Phoenix, falling in love with his AI companion Samantha (voiced by Scarlett Johansson).

AI has the potential to become really smart as we extend beyond real-time text generation to the incorporation of voice, and eventually video to create truly incredible and immersive experiences.

But right now, it’s important we build our education on how AI works today—it’s just trying to predict what word it will say next, having been trained on a huge corpus of text. It’s this stark contrast that gives me hope that we’ll live in a future that involves human connection abundance, whilst intelligent AI supports our relationships. Not the other way around.

Quote to Share

Brett Adcock on Figure AI becoming the 6th most sought after private company:

Twitter tweet

Figure AI now sits ahead of OpenAI in the rankings for the most sought-after venture companies in the secondary market.

This comes off the back of their announcement of Helix, Figure’s Vision-Language-Action (VLA) model that makes robots think like humans.

Another interesting note—4 out of the top 16 companies have been founded by Elon Musk: SpaceX, xAI, OpenAI, and Neuralink.

Source: Brett Adcock on X

Question to Ponder

I see both profound vision and legitimate concern in this question.

Our economic systems were built around human labour as the primary means of distributing resources. When that foundation shifts, everything must adapt.

With AI eventually driving the cost of intelligence and the cost of labour to near-zero, we need to think about how we navigate a society like this.

Without planning for this transition, we could face a period where jobs disappear faster than new opportunities emerge. Concerns about who “controls the robots”, how benefits will be distributed, and whether our political systems can adapt quickly enough are questions we as a society need to find answers for soon.

Perhaps the question isn't whether AI will replace jobs, but the next step on, which is whether we’ll use that replacement to create abundance for all or concentrate power and resources even further. I think the latter is the most dangerous outcome, and believe society needs to focus on (1) broad baseline education of using these tools, and (2) open-sourcing as much of these developments as possible to enable everyone to benefit and build on the productivity gains of these tools.

Right now, a $200/mo subscription to ChatGPT Pro is out of reach for many—especially with the introduction of new (and potentially even more expensive) agentic capabilities being layered on top in the coming months.

💡 If you enjoyed this issue, share it with a friend.

See you next week, Alex Banks P.S. Unitree's Robot knows Kung-Fu.

AI Highlights

My top-3 picks of AI news this week.

Claude Code / Anthropic

Anthropic
1. Claude’s Power Boost

Anthropic has released Claude 3.7 Sonnet, their most intelligent model to date, featuring hybrid reasoning capabilities.

• Hybrid reasoning model: Combines quick standard responses with deeper reflection through an extended thinking mode for complex problem-solving, with 128K token “thinking budget” via API.

• Performance leader: Achieves state-of-the-art results on SWE-bench Verified (70.3%) and tops benchmarks across Cursor, Cognition, and Vercel coding platforms.

• Claude Code: New agentic coding tool that collaborates on development, searches/edits code, writes/runs tests, and handles GitHub operations.

Alex’s take: Claude 3.7 Sonnet is up to 5x cheaper than OpenAI's o1 model. At $3/million input tokens and $15/million output tokens, they're making the dawn of hybrid reasoning accessible to a much wider audience. I think this signals a fundamental shift in how AI systems will continue to develop in the future by combining both “fast” and “slow” thought together in one model.

OpenAI
2. GPT-4.5 Enters The Chat

OpenAI has released GPT-4.5, their newest and largest language model yet, available now to Pro users and developers worldwide, and its video model “Sora” has been opened to EU members.

• Dual intelligence approach: OpenAI is advancing AI along two axes—unsupervised learning (GPT-4.5) for intuition and world knowledge, and reasoning models (OpenAI o1, o3-mini) for step-by-step thinking.

• Reduced hallucinations: GPT-4.5 shows significant improvement in factual accuracy, with a 37.1% hallucination rate compared to GPT-4o's 61.8%.

• Sora expansion: In the same week, OpenAI expanded its AI video generation tool Sora to the EU, UK, Switzerland, Norway, Liechtenstein, and Iceland, making it available to Plus, Pro, and Team users in these regions.

Alex’s take: I find OpenAI's explicit framing of their two-track approach fascinating, especially how it stacks up against Anthropic, who are combining both reasoning and deep thought into a single hybrid model. It reminds me of the left brain/right brain concept—one side for creativity and intuition, the other for logical reasoning.

Biology
3. AI Decodes Deadly Snake Venom

Researchers at the University of Washington's Baker Lab have created AI-designed proteins that can neutralise lethal snake venom toxins, potentially revolutionising treatment for the 400,000 people affected by snakebites annually.

• Computational breakthrough: Using deep learning models like RFdiffusion and ProteinMPNN on NVIDIA GPUs, the team generated and screened millions of potential antitoxin structures virtually, compressing years of work into weeks.

• Superior performance: The AI-designed proteins demonstrated 80-100% survival rates in mouse studies exposed to lethal neurotoxins, specifically targeting the deadliest three-finger toxins (3FTx).

• Accessibility advantages: Unlike traditional antivenoms that require refrigeration and trained medical staff, these proteins are small, heat-resistant, and potentially much cheaper to produce.

Alex’s take: While we often focus on AI's impact in content creation and knowledge work, this research shows AI's profound capability to solve some of our most neglected health challenges. I believe we're only scratching the surface of what's possible when AI drives molecular design—from venom antidotes today to treatments for viral infections and autoimmune diseases tomorrow.

Today’s Signal is brought to you by Recraft.

Recraft

I've wasted hours wrestling with Midjourney and DALL-E trying to get the output I'm looking for.

That's when I found Recraft. A premium AI image generation and editing tool that gives you control over your designs.

→ Seamless integration of text into designs regardless of size or scale
→ Create images in your brand style and colours
→ Professional editing that doesn’t look like “AI”

Recraft is trusted by over 3 million users across 200 countries, including designers from most innovative companies like Ogilvy, HubSpot, Netflix, Asana, and Airbus.

Get $12 off any plan with code ALEX12.

Try Recraft Now

SPONSOR

Content I Enjoyed

The Besties / All-In Podcast

The All-In Podcast's Take on Our Autonomous Future

As a regular listener to the All-In Podcast, I thoroughly enjoyed their segment this week on the state of autonomous robots.

Brett Adcock, the CEO of Figure AI, wrote a tweet this week saying they have accelerated their in-home timeline by two years, with alpha testing of robots in homes expected by the end of this year.

On the pod, Chamath stated “…the model is not perfect yet to be general purpose” and “the actuators are good, they’re not great”. The physical dexterity of the actuators still lags behind what's needed for genuinely functional home robots. If they’re not super functional over the next year, over the next two years, their dexterity will increase dramatically and find its way to exciting applications across laundry, cleaning, and dog-walking. The list is limitless and makes for a very interesting roadmap.

The besties also noted the timeline for middle-class adoption (5-10 years) and compared general-purpose humanoids with single-purpose robots like the $1,000 autonomous lawnmowers already appearing in Austin neighbourhoods.

As someone following this space closely, it’s refreshing to hear practical, grounded discussions about robotics rather than just the typical hype cycle. The race to bring capable robots into our homes is accelerating faster than many truly realise.

Idea I Learned

Mercury / Inception Labs

The first commercial-grade diffusion LLM (dLLM)

Inception Labs has launched Mercury, the first commercial-grade diffusion large language model (dLLM).

Unlike traditional autoregressive large language models (LLMs) that predict text sequentially, what I love about Mercury is it uses a diffusion process to generate entire text sequences simultaneously through a coarse-to-fine refinement method.

In essence, it doesn’t predict tokens from left-to-right, but instead does it all at once. This diffusion process is actually how most image/video generation models work today.

Mercury up to 10x faster and cheaper than current LLMs, reaching over 1000 tokens per second on NVIDIA H100 GPUs. In addition, “Mercury Coder” has demonstrated competitive performance in code completion, matching or surpassing models like GPT-4o Mini and Claude 3.5 Haiku.

This feels like we’re really pushing the frontier of intelligence and speed for language models—looking to diffusion instead of autoregression could mean new, unique, and undiscovered use cases right around the corner.

Quote to Share

Georgi Gerganov on the “gibberlink” project:

Twitter tweet

The “Gibberlink” project, developed by Anton Pidkuiko and Boris Starkov, enables AI systems to communicate through audio channels at speeds vastly exceeding human speech capabilities.

Instead of being constrained by human language patterns, these AI systems can transmit data through sound waves in a way that optimises for machine-to-machine efficiency. Using the “ggwave” library, the AI agents effectively create a high-bandwidth communication channel that bypasses traditional linguistic constraints.

I think this demonstrates the potential of communication protocols that are optimised for machine efficiency rather than human comprehension. It also raises important considerations about transparency and oversight as AI systems potentially develop communication methods that operate outside human interpretability. As a society, it’s important we have sufficient monitoring and standards to make sure these interactions remain auditable and aligned with human interests moving forward.

Source: Georgi Gerganov on X

Question to Ponder

The frustration with o1 is real, and you’re not alone in feeling this way.

However, something that struck me only recently is that o1 isn't a chat model like Claude 3.7 or GPT-4o. It's designed as a reasoning engine, which means interacting with it is fundamentally different.

Instead of writing prompts, I recommend writing “briefs” and include more context than you think necessary. I’ve distilled the perfect o1 prompt into 4 key components:

1. Goal: Have a clear, specific objective statement.

2. Return Format: Explain exactly how you want the information structured. Cover specific fields, exact data points, and how you want it organised.

3. Warnings: Add explicit guardrails to prevent common errors and hallucinations.

4. Context Dump: Write out your personal situation and preferences. This is the “why” behind your request that provides rich, relevant context for o1 to reason through.

But should we need all this structure just to get “good enough” results?

I personally use Claude as a daily driver and only switch to a reasoning model like o1 if I’m really wanting to get into the weeds on a tough problem.

If Claude consistently works better for you with less hassle, then that’s valuable information about where to invest your subscription dollars.

If you want to dive in more about this idea of reasoning models vs chatbots, I really enjoyed reading Ben Hylak’s deep dive into o1.

💡 If you enjoyed this issue, share it with a friend.

See you next week, Alex Banks P.S. This AI is helping teach Americans sign language.