Hey everyone, here's your new AI update for this week!

In today’s issue:

• Chrome launched AI "Auto Browse" that can shop, book flights, and fill forms for you
• A former Google engineer was found guilty of stealing AI tech worth billions for China
• Criminals created "dark LLMs" on the dark web—AI chatbots designed to help hackers
• And more...

🌐 1. Chrome Can Now Browse the Web and Complete Tasks For You

Google just dropped “Auto Browse” in Chrome (January 28–29, 2026)—an AI feature that actually performs multi-step tasks on your behalf instead of just answering questions.

Why This Matters

• Chrome’s new Gemini side panel can now shop online, book travel, fill out forms, and complete purchases—all autonomously
• Available to Google AI Pro ($20/month) and Ultra ($30/month) subscribers in the US first
• Chrome has 70%+ global market share, meaning this reaches billions of users faster than any competing browser

The Reality Check

This is “agentic” AI in action—the system doesn’t just respond to prompts, it actively navigates websites, clicks buttons, fills forms, and adds items to carts based on a single command like:

"Go to Etsy and find supplies to recreate this photo booth. Add them to my cart. Don’t spend more than $75 total."

Gemini then visits Etsy, identifies items from the photo, checks prices, adds them to your cart, and stops at checkout (it won’t buy without your permission). You can watch the AI work in the side panel and take over at any point.

The Practical Impact

What you can do right now (if you’re a Pro/Ultra subscriber in the US):

• Travel planning: “Research hotel and flight costs for Chicago across multiple dates in March and find the cheapest option”
• Product comparison: “Compare these three laptops across different retailers and add the best deal to my cart”
• Form filling: “Fill out this job application using my resume from Google Drive”
• Apartment hunting: “Go through my favorited apartments on Redfin and remove any that aren’t pet-friendly”

Security & Limits

• Auto Browse pauses before making purchases or posting on social media (requires your approval)
• Uses Chrome’s password manager for logins (with permission)
• Pro subscribers get 20 tasks/day, Ultra get 200 tasks/day
• Can’t access sites that block bots (Amazon already sued Perplexity over this)

The Competition

OpenAI launched its own Atlas browser in October 2025, Microsoft Edge has Copilot automation, and Opera has AI agents—but Chrome’s 70% market share makes this the biggest agentic browsing deployment yet.

Bottom line: The era of “AI does your web browsing” officially started this week. When your browser can autonomously book flights, compare products, and fill forms, it’s not a chatbot anymore—it’s a digital employee. McKinsey projects agentic commerce could hit $1 trillion by 2030.

🧨 2. Former Google Engineer Convicted of Stealing AI Tech Worth Billions

On January 28, 2026, a federal jury found former Google engineer Linwei Ding guilty of economic espionage and theft of trade secrets—specifically, stealing confidential AI technology for a Chinese company.

Why This Matters

• Ding faces up to 10 years per count for theft and 15 years for economic espionage
• He allegedly stole Google’s proprietary AI chip designs and machine learning infrastructure
• The stolen tech is valued in the billions and relates to Google’s competitive advantage in AI hardware

The Reality Check

Ding worked as a software engineer at Google and had access to highly confidential AI systems. Prosecutors allege he:

• Copied proprietary code and technical documents related to Google’s AI chips
• Transferred the data to personal accounts while employed
• Planned to use the technology for a startup backed by Chinese investors

The Department of Justice charged him under both the Economic Espionage Act (for benefit of a foreign government) and the Trade Secrets Act.

The Practical Impact

What this signals:

• AI espionage is now a top national security concern — the FBI is actively investigating tech theft from major AI labs
• Companies are tightening internal security around AI model weights, training data, and chip designs
• Geopolitical AI race is escalating — U.S.-China tensions over AI dominance are leading to criminal prosecutions

The conviction comes as:

• DeepSeek (China) released competitive models at a fraction of U.S. companies’ costs
• Chinese open-source AI models are gaining 15% global market share
• The U.S. government is restricting chip exports to China

The Broader Trend

This isn’t isolated—multiple cases of AI tech theft have emerged:

• Character.AI (acqui-hired by Google) faced security questions over data access
• OpenAI accused former employees of taking proprietary research
• Anthropic’s Claude models showed up in unauthorized third-party services

Bottom line: When a former Google engineer faces 10–15 years for stealing AI tech, it proves the technology is now considered a strategic national asset on par with military secrets. The AI arms race is real, and governments are prosecuting theft as economic warfare.

💀 3. Criminals Created “Dark LLMs”—Evil AI Chatbots Helping Hackers on the Dark Web

Security researchers confirmed the existence of "dark LLMs"—AI chatbots specifically designed to help criminals with fraud, phishing, malware development, and cyberattacks.

Why This Matters

• At least 1,000+ users are paying $30–$200/month for AI tools with zero ethical restrictions
• These systems help create phishing emails, generate malware code, and automate scams at scale
• One example, "Nytheon AI," is an 80-billion-parameter model hosted on the dark web (TOR)

The Reality Check

Unlike ChatGPT, Claude, or Gemini—which refuse to help with illegal activities—dark LLMs are built specifically to assist criminals. They:

• Generate convincing phishing emails personalized to victims
• Write malware code (one system created 88,000 lines of advanced malware in 6 days vs. 30 weeks normally)
• Help with fraud content, vulnerability exploitation, and social engineering

Example use case reported by researchers:

A hacker subscribes to Nytheon AI for $200/month and asks: "Create a phishing campaign targeting healthcare workers. Generate personalized emails, fake login pages, and malware payloads."

The AI generates 50+ variations of convincing emails, creates fake websites that look identical to real ones, and writes code to steal credentials—all automatically.

The Practical Impact

What’s happening:

• "Agentized" phishing campaigns — AI agents automatically develop lures, send personalized emails, and adapt strategies based on victim responses
• Democratized hacking — Low-skill criminals can now create sophisticated attacks that used to require expert coding
• Scale explosion — What took a team of hackers months now takes one person with AI subscription days

Cybersecurity firms report:

• 60% increase in fraud damages between 2024–2025
• AI-generated phishing emails are 3× more effective than human-written ones
• Traditional detection methods struggle because AI-generated content looks legitimate

The Defense

What companies are doing:

• Treating AI agents as a new threat surface (not just productivity tools)
• Implementing "AI agent hijacking" protections (preventing attackers from manipulating your AI tools)
• Training security teams on AI-specific threats

Bottom line: When criminals create their own AI with zero ethical boundaries and subscription pricing, cybercrime becomes automated, scalable, and accessible to anyone with $30/month. This is the dark side of the agentic AI revolution—and it’s already here.

🔗 Useful Links:

• ChatGPT Web | iOS | Android
• Google AI Pro ($20/month – includes Auto Browse)
• Chrome Gemini (side panel available now)
• Claude Web App | API Docs

📚 More Resources:

• Google’s Auto Browse Announcement
• DOJ Press Release: Google Engineer Conviction
• Dark LLMs Security Report
• McKinsey: Agentic Commerce Projections

🌐 1. Use Chrome's "Auto Browse" to Actually Automate Web Tasks (Launched 48 Hours Ago)

The problem: You spend 30 minutes comparing flight prices across multiple dates or filling out repetitive online forms.

The solution: Chrome's new "Auto Browse" (launched Jan 28–29) lets AI do multi-step web tasks autonomously.

3-minute setup (US only, requires Google AI Pro $20/mo or Ultra $30/mo):

1. Open Chrome → Click Gemini icon (sparkle) in top-right
2. Side panel opens → Type your command
3. Watch it work in real-time

Real examples:

• "Find the cheapest round-trip flights from NYC to LA in March across 5 different dates." → Gemini opens Google Flights, compares dates, shows you best deals
• "Go to Amazon, find noise-canceling headphones under $200 with 4.5+ stars, add top 3 to cart." → Gemini searches, filters, reads reviews, adds to cart (stops before checkout)
• "Fill out this job application using my resume from Google Drive." → Gemini pulls your data and auto-fills the form

Limits: Pro gets 20 tasks/day, Ultra gets 200. Some sites like Amazon block bots. US-only for now.

Tools:Google AI Pro ($20/month)

🎙️ 2. Clone Your Voice in 3 Seconds With Open-Source AI (Released January 2026)

The problem: Professional voice cloning used to cost hundreds of dollars and require hours of recording.

The solution: Alibaba's Qwen3-TTS (January 2026) clones any voice with just a 3-second audio sample—free and open-source.

3-minute setup:

1. Go to Hugging Face Qwen3-TTS Demo
2. Record one sentence (3 seconds): "Hello, this is a test of voice cloning."
3. Upload audio → Type any text → Click "Generate"
4. Download your AI-generated audio in your cloned voice

Real examples:

• Record 3 seconds → Generate entire audiobook in your voice
• English sample → Speak fluent Spanish (even if you don't speak it)
• Preserve a loved one's voice for family memories

Privacy warning: Only clone voices with explicit permission. This tech can be misused for deepfakes and scams.

Advanced move: Combine with ChatGPT to write scripts, then use Qwen3-TTS for professional voiceovers in your own voice.

Tools:Qwen3-TTS (free, Apache 2.0 license)

🧠 3. Turn Any PDF Into an Interactive Expert With ChatGPT's "Deep Research" Mode

The problem: You upload a 50-page PDF and get shallow answers that miss key details.

The solution: ChatGPT's "Deep Research" mode (January 2026, Plus/Pro users) actively investigates and synthesizes like a human researcher.

3-minute setup:

1. Go to ChatGPT
2. Select "GPT-5.2 Deep Research" from model dropdown
3. Upload PDF → Give it a research task (not just questions)

The difference:

Normal mode:

• Upload earnings report → Ask "What were the highlights?"
• Get: 3 bullet points from the summary

Deep Research mode:

• Same report → Command: "Analyze revenue trends vs last quarter, identify discrepancies between claims and numbers, flag downplayed risks, compare to competitors."
• Get: 2-page analysis with graphs, citations, web-sourced competitor data, and red flags

Real examples:

• Legal docs:"Identify all obligations and deadlines. Cross-reference with standard contracts to flag unusual clauses."
• Academic papers:"Critique the methodology, find contradicting research, suggest follow-up experiments."
• Business reports:"Extract market trends, identify emerging competitors, compare to our strategy."

Note: Deep Research takes 3–5 minutes (it's doing actual research, not just reading).

Tools:ChatGPT Plus ($20/month) or Pro ($20/month for unlimited)

A quick note before you go

Thanks for reading this week’s Brain Bytes — I hope something here helped you move faster or think better.

How’d this one land?

• A ✨ Super Helpful!
• B 🤔 Some Good Stuff!
• C 😕 Could be better.

See you next week, — Oliver

Hey everyone, here's your new AI update for this week!

In today's issue:

• Why DeepSeek's $6M AI model matching ChatGPT sparked a $1 trillion market panic
• Anthropic launched Claude for Healthcare that can read your medical records across multiple providers
• CES 2026 proved "physical AI" is real—robots, autonomous cars, and AI-powered hardware everywhere
• And more...

💰 1. DeepSeek's $6M AI Model Just Triggered a $1 Trillion Market Sell-Off

Chinese AI startup DeepSeek released its R1 reasoning model this week, claiming it matches OpenAI's capabilities while costing just $6 million to develop—68 times cheaper than comparable models.

Why This Matters

• DeepSeek R1 outperforms ChatGPT’s o1 model on math, coding, and logic benchmarks despite the massive cost difference
• The model is fully open-source and free to use, with API costs 96% lower than OpenAI’s pricing
• When released in January 2025, it triggered a $1 trillion sell-off across global AI stocks as investors questioned Western AI’s cost advantage

The Reality Check

Microsoft reports DeepSeek is gaining massive traction in developing nations—89% market share in China, 56% in Belarus, 43% in Russia. But adoption remains low in North America and Europe due to security concerns and government restrictions.

DeepSeek V4 is reportedly launching mid-February with elite coding performance that could surpass Claude and ChatGPT on long-context code tasks.

The Practical Impact

What you can actually do:

• Use DeepSeek R1 for free at deepseek.com — it excels at math problems, coding challenges, and logical reasoning
• Check its “thought process” feature to see exactly how it arrived at conclusions (like OpenAI’s o1)
• Run it locally for complete privacy using the open-source implementation

However, DeepSeek operates differently on political topics due to China’s internet restrictions, and some countries have banned it over security risks.

Bottom line: When a $6 million model matches $400 million models in performance, the AI industry’s economics fundamentally shift. The question isn’t whether cheaper, open alternatives will disrupt the market — it’s how fast Western AI labs can adapt their pricing and approach.

🧾 2. Anthropic Launched Claude for Healthcare—Now AI Can Read Your Medical Records

Anthropic just unveiled Claude for Healthcare with a partnership allowing patients to connect their electronic medical records to Claude and ask health questions based on their actual history.

Why This Matters

• Claude can now access your unified medical records across 50,000+ healthcare providers through HealthEx integration
• Uses Model Context Protocol (MCP) to securely retrieve only relevant portions of your records for each specific question
• Launched alongside OpenAI’s ChatGPT for Health—the AI labs are racing to capture the lucrative healthcare market

The Reality Check

Available now to Claude Pro and Max subscribers in the US only. Your medical data is NOT used for model training and doesn’t get stored in Claude’s memory. You control access and can disconnect anytime.

Anthropic donated MCP to the Linux Foundation’s new Agentic AI Foundation this week, joining OpenAI and Block as founding members—signaling industry-wide commitment to standardizing how AI connects to data.

The Practical Impact

Real use cases right now:

• "What does this lab result mean?" – Get plain-English explanations of medical terminology
• "How has my cholesterol changed over time?" – Track health metrics across years and providers
• "What should I ask my doctor about this?" – Prepare for appointments with context-aware questions

Setup takes minutes: verify identity with biometrics and government ID, connect patient portal logins, choose what Claude can access.

Bottom line: AI accessing your medical records sounds alarming, but the ability to ask “What medications am I taking?” and get instant answers from scattered records across multiple doctors is genuinely useful. The privacy controls are solid—just know you’re trading convenience for giving an AI company (temporary) access to your health data.

🤖 3. CES 2026 Proved "Physical AI" Is Finally Real

This week's Consumer Electronics Show in Las Vegas marked a major shift: AI moved from digital interfaces into physical hardware, robots, and autonomous systems.

Why This Matters

• Nvidia launched Vera Rubin architecture with 10x throughput improvements and 10x reduction in AI token costs
• Boston Dynamics' humanoid Atlas robot is coming to Hyundai factories in 2028 for mass production
• Ford announced an AI assistant that uses cameras to calculate exactly how many bags of mulch fit in your truck bed

The Reality Check

Every major chipmaker (Nvidia, AMD, Intel, Qualcomm) unveiled robotics-focused processors at CES. The event featured cowboy hat-wearing humanoid bots, surgical robots, and helper bots checking people in at events.

Nvidia’s Alpamayo self-driving model is designed to help cars understand unique situations like children chasing balls into streets—moving beyond simple object detection.

The Practical Impact

What’s actually shipping soon:

• AI-powered robot vacuums that climb stairs and pick up objects (already available)
• Ford’s mobile app AI assistant analyzing cargo space (launching 2026, in-car version 2027)
• Smart home devices with offline voice control that never send data to the cloud

CES Innovation Awards went to VIXallcam (all-weather vision for trucks) and AA-2 (indoor delivery robot that navigates elevators autonomously).

Bottom line: Physical AI means intelligence embedded in hardware that moves, sees, and acts in the real world. When robots can autonomously navigate elevators and AI can tell you exactly how much cargo fits in your truck by looking at a photo, we've crossed from demos into deployable products. 2026 is when AI stops being just software.

🔗 Useful Links:

• ChatGPT Web | iOS | Android
• DeepSeek | API Docs | Pricing
• Claude Web App | API Docs | Pricing

📚 More Resources:

• DeepSeek R1 Paper (Updated)
• Model Context Protocol Docs
• Claude for Healthcare Announcement
• CES 2026 AI Highlights

📖 1. Use NotebookLM's New "Interactive Mode" to Interview Your Documents

The problem: You have PDFs, articles, or reports you need to understand, but reading 50 pages feels overwhelming and you learn better through conversation.

The solution: Google's NotebookLM just added Interactive Mode (rolled out January 2026), which lets you interrupt the AI podcast hosts mid-conversation and ask them questions in real-time.

2-minute setup:

1. Go to notebooklm.google.com
2. Create a notebook and upload your documents (up to 50 sources)
3. Click "Generate" under Audio Overview
4. NEW: Click "Join Conversation" to activate Interactive Mode
5. Interrupt the hosts and ask questions as they discuss your content

What you get: Two AI hosts discussing your documents like a podcast, but now you can jump in and say "Wait, explain that chart on page 12" or "How does this compare to the other study?"

Real example: Upload your company’s Q4 reports. The hosts start discussing revenue trends. You interrupt: "Hold on, why did European sales drop?" They pivot immediately to analyze that specific section across all uploaded documents.

The trick: The hosts maintain context across interruptions. You can have a 20-minute back-and-forth where they reference earlier points in the conversation. It’s like having two expert analysts who’ve read everything.

When to use it:

• Researching complex topics across multiple sources
• Preparing for meetings by "interviewing" briefing materials
• Studying for exams with textbooks and lecture notes
• Understanding contracts or legal documents in plain English

Tools:NotebookLM (free with Google account)

🧠 2. Use Perplexity's "Spaces" to Build Your Own Custom AI Research Assistant

The problem: You’re researching the same topic repeatedly (planning a home renovation, tracking competitors, following medical research) and you're tired of re-explaining context every time.

The solution: Perplexity Spaces (updated December 2025) lets you create topic-specific AI assistants with custom instructions, file uploads, and conversation memory that persists.

5-minute setup:

1. Go to perplexity.ai and create an account
2. Click "Spaces" in the sidebar → "Create Space"
3. Add custom instructions: "You’re helping me plan a kitchen remodel. Budget: $30k. Style: modern farmhouse. Always check local building codes for Seattle."
4. Upload relevant files (quotes, floor plans, inspiration photos)
5. Now every search in this Space has that context automatically

What you get: An AI that remembers your project details, references your uploaded files, and doesn't make you repeat yourself across sessions.

Real example: Create a "Competitor Analysis" Space for your startup. Upload competitor websites, pricing sheets, and product docs. Every time you ask “What’s their positioning strategy?” or “How do they handle onboarding?” Perplexity searches with full context of everything you've uploaded.

The trick: Spaces are searchable across sessions. Come back three weeks later and your AI still remembers your budget, constraints, and previous decisions. It’s like having a research assistant with perfect memory.

Advanced features:

• Pro Collections: Organize multiple Spaces by project
• Shared Spaces: Collaborate with team members who can all query the same knowledge base
• Source pinning: Tell Perplexity to prioritize specific websites in its searches

When to use it:

• Long-term research projects (thesis, book research)
• Business intelligence and competitor tracking
• Home improvement planning
• Medical condition research and treatment tracking

Tools:Perplexity (free basic, Pro: $20/month for unlimited searches)

🎨 3. Use ChatGPT's New "Canvas Collaboration Mode" to Build Interactive Tools (No Coding)

The problem: You need a custom calculator, tracker, or interactive tool for work, but hiring a developer costs thousands and you don’t code.

The solution: ChatGPT Canvas just added "Collaboration Mode" (January 2026 update) that lets non-technical people build functional web apps through conversation.

5-minute setup:

1. Open ChatGPT (Plus or Team required)
2. Select "GPT-4o with Canvas" from model picker
3. Say: “Open a canvas and build me [describe your tool]”
4. ChatGPT creates it in real-time → you see the live preview instantly
5. Iterate: “Add a reset button” or “Make the colors match our brand”

What you get: A working interactive tool you can use immediately in the browser, or export as HTML to embed anywhere.

Real examples people are building:

• ROI Calculator: “Build a calculator where I input monthly ad spend and conversion rate, and it shows projected annual revenue with a slider”
• Project Timeline: “Create a Gantt chart where I can drag tasks and it auto-updates dependencies”
• Budget Tracker: “Make a form where I log expenses by category and it shows a pie chart of spending”

The trick: Canvas sees your changes in real-time and can modify specific parts without rebuilding everything. Say “make just the header blue” and it updates that element only.

Advanced move: Export the code and give it to a developer to integrate into your actual product. Many teams use Canvas to prototype features before committing engineering resources.

When to use it:

• Building internal tools for your team
• Creating custom calculators for clients
• Prototyping app ideas before hiring developers
• Making interactive presentations or demos

What makes this different: Unlike regular ChatGPT code generation, Canvas shows you the working tool as you build it. You’re not copying code blocks—you’re having a conversation while watching your tool come to life.

Tools:ChatGPT Canvas (Plus: $20/month, Team: $30/user/month)

A quick note before you go

Thanks for reading this week’s Brain Bytes — I hope something here helped you move faster or think better.

How’d this one land?

• A ✨ Super Helpful!
• B 🤔 Some Good Stuff!
• C 😕 Could be better.

See you next week, — Oliver

Hey there,

While you were nursing your New Year's hangover, China's DeepSeek published a research paper that has AI analysts calling it a "striking breakthrough." Meanwhile, investors are predicting 2026 will be the year AI either proves its worth or bankrupts companies trying. And Apple? They're sitting on $130 billion in cash while everyone else burns money on data centers.

Here's what happened:

• DeepSeek's training breakthrough – New method trains bigger models without them breaking
• VCs predict AI labor shift – Companies will replace workers with agents in 2026
• Apple's AI gamble – Cautious approach may pay off as bubble concerns grow

Let's break it down.

🛠️ Tool of the Week: MailMaestro (Not Sponsored)

The AI email assistant that actually sounds like you (not a robot)

Most AI email tools sound like ChatGPT wrote your emails. MailMaestro solves this with an "AI personality" feature that trains on how YOU write.

What it does:

• Writes emails in your voice and tone
• Learns words you avoid, your job title, preferred style
• Built-in library of email templates
• Works across email platforms

Why it's different:

Instead of generic "I hope this email finds you well" nonsense, MailMaestro studies your past emails and mimics your actual writing patterns.

You train it once on:

• Words to avoid
• Preferred tone (casual, formal, direct)
• Job-specific context
• Common scenarios you face

Then it writes emails that sound like you, not like AI.

Real-world example:

Without MailMaestro:
"I hope this email finds you well. I wanted to reach out regarding the Q1 budget proposal. Would you be available for a quick sync to discuss next steps? Looking forward to hearing from you!"

With MailMaestro (trained on your style):
"Hey Sarah - need 15 min to walk through Q1 budget numbers. Thursday afternoon work?"

Why people use it:

• Saves hours on email without sounding robotic
• Maintains your personal brand/voice
• Faster than training ChatGPT every time
• Pre-built templates for common scenarios

The catch:

• Requires some setup time to train properly
• $15/month isn't cheap for an email tool
• Free plan limited to 3 requests per week

Who it's for: Anyone drowning in email who wants AI help but hates how generic AI emails sound. Sales teams, executives, customer support, consultants.

Pricing:

Free: 3 requests per week
Pro: $15/seat/month + 14-day trial

Tools:MailMaestro

🇨🇳 1. China's DeepSeek Drops AI Training Breakthrough

New method could reshape how the entire industry builds models

While U.S. companies throw billions at bigger data centers, China's DeepSeek just published a research paper showing there's a smarter way to train AI models.

The paper introduces "Manifold-Constrained Hyper-Connections" (mHC) — a training method designed to scale models without them becoming unstable or crashing mid-training.

Why this matters:

Training large AI models is brutally expensive and wasteful. Models often fail mid-training, wiping out weeks of work and thousands of GPU hours.

When that happens, companies restart from scratch. More electricity wasted. More compute hours burned. More money gone.

DeepSeek’s approach makes training more stable and predictable, reducing failures and wasted resources.

The technical bit:

As language models grow, researchers try to improve performance by letting different parts of the model share more information internally. But this often makes training unstable — models break or produce garbage.

mHC fixes this by redesigning how information flows through the model during training. Tests on models with 3 billion to 27 billion parameters showed better scaling and stronger performance with only ~6.7% overhead.

What analysts are saying:

Wei Sun, principal analyst at Counterpoint Research, called it a “striking breakthrough.”

“DeepSeek combined various techniques to minimize the extra cost of training a model,” she said. “Even with a slight increase in cost, the new training method could yield much higher performance.”

Sun added that the paper signals DeepSeek can “bypass compute bottlenecks and unlock leaps in intelligence” — similar to their January 2025 “Sputnik moment” when they launched R1 at a fraction of ChatGPT’s cost.

Why the timing is suspicious:

DeepSeek’s founder Liang Wenfeng co-authored the paper. He typically only does this for major breakthroughs.

DeepSeek also has a track record of publishing foundational research right before major model launches. They did it before R1.

The company is reportedly working on R2 or V4. This paper could be the signal it’s coming soon.

The bigger picture:

While OpenAI, Google, and Meta compete on who can spend the most money on data centers and chips, Chinese AI labs are innovating around constraints.

U.S. export restrictions cut off China’s access to advanced Nvidia chips. So Chinese companies are getting creative with engineering — building better models with less compute.

If DeepSeek can train frontier models more efficiently than companies with unlimited budgets, that’s a problem for the “just throw more GPUs at it” approach dominating Silicon Valley.

Tools:Bloomberg Coverage

🗞️ 2. VCs Predict AI Is Coming for Your Job in 2026

Companies will shift budgets from labor to AI agents this year

Multiple venture capitalists told TechCrunch they expect AI to have a massive impact on the enterprise workforce in 2026 — and it wasn’t even a question about jobs.

The prediction: Companies will start pulling money from labor budgets to fund AI projects.

What investors are saying:

“2026 will be the year of agents as software expands from making humans more productive to automating work itself,” said Jason Mendel, venture investor at Battery Ventures.

Rajeev Dham, managing director at Sapphire Ventures, agreed that 2026 budgets will shift resources from labor to AI.

Marell Evans, founder of Exceptional Capital, predicted companies will pull from their hiring pools to increase AI spending.

The "doom loop" for workers:

Eric Bahn, co-founder at Hustle Fund, isn’t sure exactly what happens next:

“Is it going to lead to more layoffs? Is there going to be higher productivity? Or will AI just be an augmentation for the existing labor market to be even more productive in the future? All of this seems pretty unanswered, but it seems like something big is going to happen in 2026.”

Translation: Nobody actually knows if AI will replace jobs or just make people more productive, but either way, budgets are moving.

The cynical take:

Antonia Dean, partner at Black Operator Ventures, said companies will blame AI for layoffs whether or not they’re actually using AI successfully:

“Many enterprises, despite how ready or not they are to successfully use AI solutions, will say that they are increasing their investments in AI to explain why they are cutting back spending in other areas or trimming workforces.”

The counter-argument:

Not everyone buys the doom narrative. Some predict “2026 will be the year of the humans” because AI hasn’t worked as autonomously as promised.

Expected outcome: New jobs in AI governance, transparency, safety, and data management. Unemployment predicted to average under 4%.

Why this matters:

Whether AI actually replaces workers or not, the narrative is shifting budgets. CFOs are being told to cut labor costs and redirect to AI.

Even if the AI doesn’t work as well as promised, the damage is done. Jobs get cut. Budgets get reallocated. Companies hope AI fills the gap.

2026 is when we find out if that bet pays off.

Tools:TechCrunch Article

🗞️ 3. Apple's Cautious AI Strategy May Finally Pay Off

Sitting on $130B while competitors burn cash on data centers

While OpenAI, Google, and Meta invest hundreds of billions in AI infrastructure, Apple took a different approach: do almost nothing and wait.

That strategy — which got Apple criticized for "falling behind" — might look smart in 2026 as AI bubble concerns grow.

The numbers:

• OpenAI, Google, Meta: Spending hundreds of billions on data centers, chips, model training
• Apple: Sitting on $130 billion+ in cash and marketable securities

Market sentiment is shifting. Investors are starting to question whether massive AI spending can be justified by near-term revenue.

If the AI bubble pops, Apple has cash. Competitors have data centers.

Apple's 2026 plan:

The company’s biggest AI move is an overhauled Siri launching in spring 2026. The updated assistant will be more conversational and handle multi-step tasks.

To power it, Apple is adopting Google’s Gemini — reflecting an internal view that large language models are becoming commoditized and not worth building in-house.

The iPhone advantage:

Unlike AI companies that need standalone apps or web services, Apple can distribute AI features through software updates to its existing device ecosystem.

Competing hardware from AI companies faces major challenges: manufacturing, distribution, ecosystem development. Apple already owns all of that.

Why this could work:

The Information argues Apple’s cautious approach positions them well if:

1. AI spending enthusiasm cools (already happening)
2. The overhauled Siri actually delivers (big if)
3. LLMs become commoditized (Apple’s bet with Gemini)

If all three happen, Apple avoided burning billions while competitors built expensive infrastructure that may not pay off.

The risk:

If AI doesn’t plateau and keeps improving exponentially, Apple will be years behind competitors who invested early.

But Apple has done this before — let others pioneer, then perfect the experience later. Worked with smartphones, tablets, watches.

What to watch:

Spring 2026 Siri launch. If it’s good, Apple’s strategy looks genius. If it’s the same old broken Siri with a ChatGPT wrapper, the criticism was justified.

Tools:MacRumors Coverage

Tip 1: "Negative Space Prompting"

Most people tell AI what to include. Power users tell AI what to EXCLUDE.

Why it works:

Constraints force creativity. When you remove obvious options, AI has to think differently.

The template:

"Give me [what you want]

But do NOT use:
- [Common approach 1]
- [Common approach 2]
- [Common approach 3]

Force yourself to be creative."
  

Example:

Bad: "Give me marketing ideas for our app."

Good: "Give me 10 marketing ideas for our productivity app.

Do NOT use:

• Social media ads
• Paid search
• Email campaigns
• Influencer marketing
• Content marketing

Force creative, unconventional approaches.

Real result difference:

Without constraints: Predictable answers (run Facebook ads, do SEO, send newsletters)

With constraints: Creative answers (partner with competing apps for co-marketing, sponsor productivity podcasts' "tool of the week," create viral productivity challenges on TikTok)

When to use:

Brainstorming, breaking out of obvious solutions, finding differentiated approaches.

Tip 2: "Persona Switching"

Ask AI to answer the same question from multiple expert perspectives. Reveals blind spots and trade-offs you’d never see from a single viewpoint.

Why it works:

Every expert has biases based on their role, incentives, and experience. A CFO cares about different things than a product manager. A customer success manager sees risks a founder ignores.

Single-perspective advice is incomplete. Multi-perspective advice shows you the full picture — including conflicts between what’s "right" for different stakeholders.

The template:

"Answer this question from 3 perspectives:

1. As a [Role 1] who prioritizes [their main concern]
2. As a [Role 2] who prioritizes [their main concern]
3. As a [Role 3] who prioritizes [their main concern]

For each perspective:
- What would they recommend?
- Why would they recommend it?
- What are they worried about?
- What are they ignoring?

Question: [Your question]"
  

Example:

Question: "Should we raise prices from $50 to $75/month?"

Answer from 3 perspectives:

1. As a startup founder focused on rapid growth and market share
2. As a CFO focused on profitability and burn rate
3. As a customer success manager worried about churn and retention

For each: What would they recommend? Why? What are they worried about? What are they ignoring?

What you get:

• Recommendation: Test price increase with new customers first
• Why: Protects existing growth metrics while testing willingness to pay
• Worried about: Losing momentum, competitors undercutting, slowing customer acquisition
• Ignoring: Existing customer sentiment, support burden, margin improvement

• Recommendation: Raise prices immediately for all customers
• Why: Revenue per customer increases 50%, even with 20% churn we come out ahead
• Worried about: Burn rate, runway, unit economics
• Ignoring: Brand damage, customer goodwill, competitive response, sales team morale

• Recommendation: Don’t raise prices; add premium tier instead
• Why: Avoids forcing price increase on happy customers who’ll churn
• Worried about: Support ticket volume, churn spike, angry customers, bad reviews
• Ignoring: Revenue pressure, investor expectations, competitive pricing

The insight:

Now you see the full trade-off space. The founder and CFO both want price increases but for different reasons and with different strategies. Customer Success opposes it entirely.

The best decision might be a hybrid: Premium tier for new customers (CSM’s idea), gradual increase for existing (Founder’s caution), with clear ROI targets (CFO’s math).

You couldn’t see that hybrid solution from a single perspective.

Advanced move:

Add a fourth perspective that’s antagonistic or contrarian:

“4. As a competitor watching this decision, what would you hope we do — and why?”

This reveals strategic blindspots. Your competitor wants you to raise prices so they can undercut you. Knowing that changes the calculus.

When to use:

• Major strategic decisions (pricing, hiring, product direction)
• Evaluating investments or partnerships
• Understanding stakeholder conflicts
• Reframing analysis paralysis by seeing all angles at once
• Preparing for tough conversations (you already know what others will say)

Tip 3: "Assumption Testing"

Make AI state its assumptions before answering. Catches when AI is working from wrong premises and prevents wasted advice.

Why it works:

AI makes assumptions based on patterns in training data. It assumes "normal" situations that might not match your reality. Without surfacing assumptions, you get advice that sounds smart but doesn’t work.

The template:

"Before answering my question, first list the 3–5 main assumptions you’re making about [topic/situation].

Be specific about what you’re assuming is true.

Then, after I confirm or correct your assumptions, answer the question.

Context: [relevant details]
Question: [Your question]"
  

Example:

Bad: "How should I scale my SaaS business?"

Good: "Before telling me how to scale, list the 5 main assumptions you’re making about my SaaS business."

Context: B2B productivity tool, 500 customers, $50/month, 5% monthly churn, acquired via Meta ads at $100 CAC.

Then tell me how to scale."

What AI might reveal:

"I’m assuming:

1. Your unit economics work (CAC < 3x LTV)
2. You’re adding net new customers after churn
3. You have product-market fit
4. Meta ads consistently deliver at $100 CAC
5. You can support 10x volume without infrastructure changes"

Now you can respond:

"Actually, assumptions #1 and #2 are wrong:

• My LTV is only $200, so CAC is half of LTV (losing money per customer)
• We’re flat on growth — churn equals new signups

Given these corrections, what should I do?"

AI’s corrected response:

"You don’t have a scaling problem, you have a unit economics and retention problem. Scaling now would just lose money faster.

Instead:

1. Fix churn first (5% monthly is 46% annual)
2. Test pricing (you may be underpriced)
3. Improve LTV through annual plans
4. Only scale once CAC < 3x LTV"

The insight:

Without assumption testing, AI would’ve given you a scaling playbook (hire salespeople, expand ads) that would’ve killed your company.

By surfacing assumptions first, AI caught the real problem: fix fundamentals before scaling.

When to use:

• Strategic advice
• Business decisions
• Technical recommendations
• Career choices
• Anything where context matters and generic advice could be dangerous

📚 More Resources

Worth your time:

• Model Context Protocol (MCP) – The "USB-C for AI" that's becoming the standard for agentic workflows
• IBM's 2026 AI Predictions – 18 expert predictions on what's coming
• Anthropic's Agentic AI Foundation – Linux Foundation's new initiative for open-source AI agents

A quick note before you go

Thanks for reading this week’s Brain Bytes — I hope something here helped you move faster or think better.

How’d this one land?

• A ✨ Super Helpful!
• B 🤔 Some Good Stuff!
• C 😕 Could be better.

See you next week, — Oliver

Hey there,

This week's AI news reads like a Black Mirror episode: AI ate all the memory chips (your next laptop will cost more), Google is crushing ChatGPT in market share, and over half the web is now AI-generated garbage that nobody asked for. Also if you have any suggestions on how you would like to see the newsletter improve, then just reply directly to the emails. We see them all!

Here's what happened:

• AI memory chip shortage – Laptops, phones, consoles about to get pricier
• Google Gemini jumps to 18% market share – ChatGPT drops 19 points in one year
• "AI slop" named Word of the Year – 50%+ of web content now AI-generated

Let’s break it down.

🎉 Brain Bytes Premium Has Reached Max Capacity

Brain Bytes Premium is currently at full capacity. We may be opening a few spots soon..

If you'd like to join, there's a waitlist. Sign up below, and we'll notify you if a spot opens up.

🛠️ Tool of the Week: Rewind AI (Not Sponsored)

The Mac app that records everything on your screen and makes it searchable

Forget what someone said in a meeting three days ago? Can’t find that competitor name you saw on Tuesday? Rewind records everything on your Mac screen and makes it searchable.

What it does:

• Records your screen continuously (every 2 seconds)
• Transcribes all audio from Zoom, Meet, videos
• OCR on everything visible on screen
• Natural language search – "What was that budget number from the morning meeting?"
• Timeline navigation to scrub through your digital history

Real-world example:

Someone mentions a competitor’s new feature in a meeting. You don’t catch the exact name. Three days later, you need it.

Open Rewind (Cmd + Shift + Space), type "competitor new feature meeting," and it shows you the exact moment it was mentioned with screen + transcript.

The privacy angle:

Everything stored locally on your Mac. Nothing sent to cloud. You control what gets recorded (pause anytime, exclude apps). Private browser windows automatically excluded.

Trade-off: If someone accesses your Mac, they access your Rewind history.

Why people use it:

• Executives: Never miss meeting details
• Developers: Find terminal commands from weeks ago
• Sales: Review calls without taking notes
• ADHD: Offload memory burden completely

The catch:

• Mac only (Apple Silicon - M1/M2/M3)
• ~14GB storage per month
• Can make Mac run hotter
• $19/month isn’t cheap

Who it’s for: People in back-to-back meetings, context switchers, information overload victims, anyone who wants photographic memory for their workday.

Pricing:
Free: Perpetual storage, limited features
Pro: $19/month (annual) or $29/month (monthly)

Tools:Rewind AI

🧠 1. AI Memory Chip Shortage = Higher Prices Coming

The AI boom is eating all the RAM, and you're about to pay for it

Remember when crypto miners bought every GPU on Earth and ruined PC gaming prices? That’s happening again, but with memory chips.

AI data centers need massive amounts of RAM to train and run models. Chip manufacturers have shifted production to meet this lucrative AI demand, which means fewer chips for laptops, phones, gaming consoles, and TVs.

The numbers:

• 10% of global RAM demand can’t be fulfilled
• Micron (major chip maker) reports supply "substantially short of demand for the foreseeable future"
• Chip companies prioritizing AI data centers over consumer products
• Dell already warning customers about higher costs

What’s actually happening:

AI workloads require "large, persistent memory footprints, extreme bandwidth, and tight proximity to compute," according to industry analysts. Companies are spending billions building data centers at breakneck speed, creating unprecedented demand for high-end memory chips.

Manufacturers like Samsung, SK Hynix, and Micron have pivoted production toward these lucrative AI contracts. That means fewer chips for everything else.

The impact:

TrendForce analyst Avril Wu’s advice: "If you want a device, buy it now."
She already bought an iPhone 17 (yes, seriously).

Dell’s COO Jeff Clarke noted higher memory costs on an earnings call and said there’s no way this doesn’t affect customer pricing.

Translation: Your next laptop, phone, or console will cost more. Not because the tech improved, but because AI ate all the chips.

Why this matters:

Unlike the GPU shortage during crypto mania, this isn’t a bubble. AI demand for memory is structural and growing. Companies can’t "dial down" memory requirements without breaking performance.

This is the new normal until chip manufacturing capacity catches up (which could take years).

Tools:NPR Article

📉 2. Google Gemini Just Crushed ChatGPT

Market share jumped from 5% to 18% in one year while ChatGPT dropped 19 points

Google’s Gemini went from afterthought to serious competitor in just 12 months — and new data shows it’s eating ChatGPT’s lunch.

According to Similarweb traffic data, Gemini now commands 18.2% of generative AI web traffic, up from just 5.4% a year ago. That’s more than tripled.

ChatGPT, meanwhile, fell from 87.2% to 68% — a staggering 19-point drop.

Why this is happening:

1. Native distribution wins

Gemini is embedded everywhere users already live: Chrome, Android, Google Workspace, and Search. When AI usage transitions from novelty to routine utility, the product that’s already there captures the flow.

You don’t need to open a new tab to use Gemini. It’s just… there.

2. ChatGPT’s dominance was never guaranteed

Early mover advantage only matters if you keep moving. Google’s infrastructure, distribution, and resources are now fully mobilized around AI — and they’re catching up fast.

3. Microsoft Copilot is DOA

Despite being baked into Windows and Edge, Copilot went from 1.5% to 1.2% market share over the same period. Distribution alone doesn’t work if users don’t trust or enjoy the product.

What analysts are saying:

“This is not noise,” said Solid Finance CEO Sam Badawi. “If you believe web behavior reflects user preference at scale, then this is the clearest signal that Google is not only in the game, it is winning share.”

OpenAI’s response:

Reportedly sent an internal “Code Red” memo after ChatGPT traffic dipped. Then rushed GPT‑5.2 and a new image generator to compete.

The war is heating up.

Why it matters:

This isn’t just about chatbots. It’s about who controls the AI infrastructure layer of the internet.

If Google can integrate AI into Search, Maps, Gmail, and Android successfully, they own the AI user experience for billions of people.

ChatGPT may have started the revolution — but Google might finish it.

Source: Yahoo Finance Article

📰 3. "AI Slop" Named Word of the Year

Over half of English web content is now AI-generated, and everyone hates it

In 2025, mentions of "AI slop" across the internet increased ninefold from 2024, with negative sentiment hitting 54% in October.

Merriam-Webster and Australia's national dictionary both named it Word of the Year.

What is AI slop?

Low-quality, unwanted AI-generated content polluting search engines, shopping platforms, social media, and yes — even the White House’s official communications.

According to SEO firm Graphite, AI-generated articles now make up more than half of all English-language content on the web.

Where you see it:

• Nonsensical Amazon product descriptions
• Generic blog posts that say nothing
• LinkedIn posts that read like ChatGPT wrote them (because they did)
• Search results filled with AI-generated spam sites
• Social media posts designed to game algorithms

The problem:

Product designers are stuck between shareholders demanding "AI features" and users who don’t want them.

Kate Moran, VP at Nielsen Norman Group: “There’s a lot of pressure to show shareholders, ‘Look, we put AI in our product.’ This is technology-led design, starting with the tool, and then trying to look for a problem that potentially that tool could solve.”

Real-world failure:

Meta launched an AI feature that replaced Instagram’s search bar with an AI chatbot. Users were furious. Meta backpedaled fast.

“You believe a search bar does a certain thing, and then all of a sudden, when you start typing in there, you’re talking to an AI chatbot and you didn’t want that. That’s a bad experience,” Moran said.

Meta’s AI push flopped:

Meta’s “Vibes” (AI-generated content feed) launched in Europe with high hopes. Internal data shows just 23,000 daily active users in the first weeks.

This is the same company that said it was tackling “unoriginal content” earlier this year and advised creators to favor “authentic storytelling.”

Why this matters:

The internet is getting worse because companies are optimizing for “we have AI” instead of “we solved a problem.”

The result: An ocean of mediocre, AI-generated garbage drowning out actual human creativity and expertise.

Welcome to 2025.

Tools: Euronews Article

Tip 1: "Anti-Prompting" – Tell AI What NOT to Do

Most people only tell AI what they want. Power users also tell it what to avoid.

Why it works: AI models often fall into predictable patterns (buzzwords, generic advice, obvious answers). Explicitly ruling these out forces more creative, useful outputs.

The template:

"[Your request]

Do NOT:
- [Common bad pattern 1]
- [Common bad pattern 2]
- [Common bad pattern 3]

Instead, focus on: [what you actually want]"

Example:

Bad: "Write a LinkedIn post about our new feature."

Good: "Write a LinkedIn post about our new collaborative editing feature."

Do NOT:

• Use words like 'game-changing,' 'revolutionary,' or 'excited to announce'
• Make generic claims about productivity
• Include emojis or hashtags
• Sound like marketing copy

Instead: Tell a specific story about a customer problem this solves, use concrete numbers, write like a human.

Real result difference:

Without constraints: “🚀 Excited to announce our game-changing new feature! Revolutionary collaborative editing is here to transform your workflow...”

With constraints: “Last week, a customer told us they waste 2 hours/day merging conflicting document versions. Three people editing the same deck, each overwriting the other's changes. Our new collaborative editing shows live cursors, auto-merges conflicts, and tracks who changed what. No more version hell.”

When to use: Content creation, brainstorming, avoiding generic AI outputs, getting past clichés.

Tip 2: Use Constraints to Get Better Outputs

AI will ramble forever if you let it. Set clear constraints to force focus.

Why it works: Constraints force AI to prioritize and be concise. Without them, you get 800 words when you needed 100.

The template:

"[Your request]

Constraints:
- Maximum [X] words/bullets/sentences
- Include only [specific type of information]
- Format as [specific format]
- Avoid [things you don't want]"

Example:

Bad: "Give me marketing ideas for my app."

Good: "Give me 5 marketing ideas for a productivity app targeting remote workers."

Constraints:

• Budget under $500 per idea
• Must be executable in under 2 weeks
• Focus on organic/owned channels only
• No paid ads
• Format as: Idea name, why it works, first 3 steps

When to use: Brainstorming, getting actionable advice, content creation, research summaries.

Tip 3: Create a "Knowledge Base" Prompt for Repeat Tasks

If you do the same type of task repeatedly, create a master prompt with all your context and reuse it.

Why it works: You stop wasting time re-explaining your context every time. The AI already knows your business, style, audience, constraints.

How to build it:

1. Start with your business context
2. Add your target audience
3. Include your brand voice/style
4. List your constraints
5. Save it, reuse it, refine it

Example knowledge base prompt:

CONTEXT:
I run a B2B SaaS company selling project management software to 50–200 person tech companies. Our customers are mostly engineering and product teams. Our main competitors are Asana and Monday.com.

AUDIENCE:
Tech-savvy, values efficiency, hates bloat, wants integrations, skeptical of marketing hype.

VOICE:
Direct, conversational, no buzzwords, show don't tell, use concrete examples.

CONSTRAINTS:
- All content under 500 words
- Always include specific numbers/data when possible
- Avoid: "revolutionary," "game-changing," "next-level"
- Focus on practical value, not features

[Your specific request goes here]

Now every time you need content, add your specific request at the bottom. The AI already knows everything else.

When to use: Content marketing, customer emails, sales outreach, product descriptions, recurring analysis.

A quick note before you go

Thanks for reading this week’s Brain Bytes — I hope something here helped you move faster or think better.

How’d this one land?

• A ✨ Super Helpful!
• B 🤔 Some Good Stuff!
• C 😕 Could be better.

See you next week, — Oliver

🎉 Brain Bytes Premium Has Reached Max Capacity

Brain Bytes Premium is currently at full capacity. Thanks to everyone that signed up!

If you'd like to join, there's a waitlist. Sign up below, and we'll notify you if a spot opens up.

Hey, it's Oliver, here's your AI update for this week!

This week: AI is taking jobs (55,000 layoffs citing AI in 2025), hiring is becoming a nightmare for both sides, and OpenAI just added teen protections after some serious pressure.

Here's what happened:

• AI-driven layoffs hit 55,000 – Amazon, Microsoft, Salesforce cut thousands citing AI
• AI hiring creates "doom loop" – both companies and job seekers miserable
• OpenAI updates teen safety – new U18 Principles for ChatGPT and Sora

Let’s break it down.

🛠️ Tool of the Week: Bearly AI (Not Sponsored)

The AI research assistant that works offline (because your data shouldn’t leave your computer)

Most AI tools send everything you type to their servers. Bearly doesn’t.

It’s a desktop app (Mac/Windows) that brings AI directly to your workflow — reading PDFs, summarizing articles, generating content, answering questions about documents — all while keeping your data local and encrypted.

What makes it different:

• Works offline – AI runs on your device, not the cloud
• Privacy-first – Your research, documents, and queries stay on your computer
• Works anywhere – Hotkey brings it up in any app
• Multi-model – Access GPT-4, Claude, and other models in one place

What it does:

• Summarize research papers – Upload PDFs, get instant summaries
• Chat with documents – Ask questions about any file
• Generate content – Writing, code, analysis without switching apps
• Web scraping – Grab content from articles and process it
• Templates – Pre-built prompts for common tasks

Real-world example:

You're researching a topic with 10 PDFs open. Instead of reading all 200 pages:

1. Upload PDFs to Bearly
2. Ask: "What are the main disagreements between these papers on [topic]?"
3. Bearly analyzes all 10 docs and gives you a synthesized answer
4. Follow up: "Which paper has the strongest evidence?"
5. Get a detailed breakdown with citations

All without your documents ever leaving your laptop.

Why people use it:

• Privacy – No data sent to cloud servers
• Researchers – Handle multiple papers/articles at once
• Speed – No tab-switching, hotkey brings AI to any app
• Control – Choose which AI model to use

The privacy angle:

If you work with sensitive docs (legal, medical, proprietary research), sending everything to ChatGPT’s servers is a risk. Bearly processes locally, so confidential information stays confidential.

Who it's for:

Researchers, academics, lawyers, anyone reading 10+ articles/papers per project, people who care about data privacy.

The catch:

• Desktop only (no mobile)
• Offline mode requires paid plan
• $20/month is pricier than some alternatives

Pricing:

Free: Limited queries, cloud-based
Pro: $20/month – unlimited, offline mode, all models

Tools:

1. AI Drove 55,000 Layoffs in 2025 (And Counting)

AI was cited as a factor in nearly 55,000 U.S. layoffs in 2025, according to Challenger, Gray & Christmas. Total job cuts topped 1.17 million — the highest since COVID in 2020.

The big names:

Amazon – Slashed 14,000 corporate roles in October (largest layoff in company history)

• Amazon's VP of People: "This generation of AI is the most transformative technology since the Internet...we need to be organized more leanly with fewer layers."

Microsoft – Cut ~15,000 jobs through 2025, including 9,000 in July

• CEO Satya Nadella: We need to "reimagine our mission for a new era" and shift "from a software factory to an intelligence engine."

Salesforce – CEO Marc Benioff confirmed 4,000 customer support workers were cut with AI's help

• Reduced headcount from 9,000 to 5,000: "I need less heads."
• Benioff revealed AI was already doing up to 50% of work at the company by summer 2025

IBM – CEO Arvind Krishna told WSJ that AI chatbots replaced "a few hundred" HR workers

What's actually happening:

Companies aren't just automating tasks — they’re restructuring entire departments around AI. Customer support, HR, and administrative roles are being hit hardest.

The pattern:

Announce AI investment → Cut jobs → Say "AI enables us to be leaner" → Repeat.

Why this matters:

We're past the "AI will change work" phase. We're in the "AI is actively replacing workers right now" phase. If your role involves repetitive knowledge work, the pressure to "work with AI or be replaced by it" is real.

2. AI Hiring Is Making Everyone Miserable (The "Doom Loop")

Over half of companies (54% per SHRM) now use AI to recruit. About a third of ChatGPT users use it to write cover letters and resumes. The result? A hiring nightmare for both sides.

What's broken:

• AI-led interviews feel "cold" (54% of US job seekers had one)
• Cover letters all look the same now (everyone uses ChatGPT)
• Companies stopped valuing cover letters because they can't tell if a human wrote them
• Using AI to apply makes you less likely to get hired (per Dartmouth research)

• Flooded with applications (everyone uses AI to apply to hundreds of jobs)
• Can't distinguish good candidates (all resumes look polished)
• AI screening tools can "copy and magnify human biases" (researcher Djurre Holtrop)
• AFL-CIO president: "AI systems rob workers of opportunities based on criteria as arbitrary as zip codes or how often they smile"

Greenhouse CEO Daniel Chait: AI created a "doom loop" where both sides say "This is impossible, it's not working, it's getting worse."

The cycle:

1. Job seekers use AI to mass-apply to hundreds of jobs
2. Companies get flooded with AI-written applications
3. Companies deploy AI to screen applicants faster
4. Job seekers use more AI to beat the screening
5. Repeat until everyone is miserable

Real example: One job seeker hung up on an AI recruiter the first time because the experience felt too "cold." Now worries: "Some great people are going to be left behind."

State response: California, Colorado, and Illinois are enacting AI hiring regulations. Trump’s recent executive order threatens to undermine state-level AI rules.

Bottom line: The market for recruiting tech will hit $3.1 billion by end of 2025, but nobody — job seekers or employers — is happier for it.

3. OpenAI Adds Teen Protections After Pressure

On December 18, OpenAI updated its Model Spec with new "Under-18 (U18) Principles" for ChatGPT and Sora.

What changed:

• New behavioral guidelines for how ChatGPT interacts with teens (ages 13–17)
• Extended parental controls to group chats, ChatGPT Atlas browser, and Sora app
• Age prediction model rolling out to automatically detect minors and apply safeguards
• Partnership with ThroughLine to surface localized helplines in ChatGPT/Sora
• New resources: "Family Guide to Help Teens Use AI Responsibly"

Why this happened:

• Sora 2 launched in October with minimal content protections
• Japanese content group CODA (including Studio Ghibli) demanded OpenAI stop training on their content
• Disney deal announced Dec 11 — terms require "age-appropriate policies and reasonable controls"

The safeguards:

• If OpenAI can't confirm someone’s age, it defaults to U18 experience
• Adults can verify their age to access full features
• Grounded in "developmental science" per American Psychological Association guidance

What's NOT included: The announcement doesn’t detail what specifically changes in teen interactions — just that there are now "rules, values, and behavioral expectations" for minors.

The timing: This dropped 7 days after Disney invested $1B in OpenAI. Disney’s deal explicitly requires OpenAI to maintain "age-appropriate policies." Coincidence? Probably not.

Bottom line: After rushing Sora 2 to market with an "opt-out" content model that enraged creators, OpenAI is now backfilling safety features under pressure from partners and advocacy groups.

🔁 1. The "Explain It Back to Me" Technique

The technique: After AI generates something complex (code, strategy, analysis), ask it to explain what it just created back to you in simple terms. This catches errors and ensures you actually understand what you're using.

Why it works: AI makes mistakes. When it explains its own output, inconsistencies or flawed logic become obvious. Plus, you learn instead of just copy-pasting.

Copy-paste template:

[After AI generates something]

Now explain what you just created back to me:
1. What does this do?
2. What are the key decisions you made?
3. What could go wrong?
4. What would I need to change for [different scenario]?
  

Example - Code Review:

You: [AI writes a Python script]

You: Explain this code back to me like I'm a junior developer:
1. What does each function do?
2. Why did you structure it this way?
3. What would break if the API returns an error?
4. How would I modify this to handle 1000+ records?

AI: This script does three things:
1. fetch_data() calls the API and returns JSON
2. process_data() filters records where status='active'
3. save_to_csv() writes results to a file

I structured it this way because...

If the API returns an error, the script would crash at line 12 because there's no try/catch...

To handle 1000+ records, you'd need to add pagination at line 8...
  

Now you understand why the code works (or doesn't), not just that it exists.

When to use: Any time AI generates something you'll rely on — code, strategy docs, analysis, technical explanations.

Tools:ChatGPT | Claude | Gemini

🎯 2. Role-Based Prompting

The technique: Tell AI to assume a specific expert role before answering. "Act as a [role]" dramatically improves response quality for specialized topics.

Why it works: AI models were trained on expert content. When you invoke a role, you're essentially filtering for that expert's perspective and vocabulary.

Copy-paste template:

Act as a [specific expert role with experience].

Context: [Your situation]

Task: [What you need]

Constraints: [Any limitations]
  

Example - Marketing Strategy:

Act as a B2B SaaS marketing director with 10 years of experience in dev tools.

Context: I'm launching a code review tool for engineering teams. Budget is $50K for 6 months. No brand awareness yet.

Task: Create a go-to-market strategy focusing on developer communities, not paid ads.

Constraints:
- Can't sponsor conferences (too expensive)
- Team of 2 (me + one contractor)
- Must see traction in 90 days
  

AI will respond with tactics specific to dev tool marketing — writing for Hacker News, contributing to open source, developer documentation strategies — instead of generic "create social media posts" advice.

Role examples:

• "Act as a senior SQL developer reviewing a junior's code"
• "Act as a UX researcher conducting usability testing"
• "Act as a CFO explaining financial metrics to non-finance teams"
• "Act as a technical writer simplifying complex concepts"

When to use: Specialized topics where generic advice won't cut it.

Tools:ChatGPT | Claude | Gemini

⚡ 3. The "Progressive Detail" Method

The technique: Start with a high-level request, then drill down into specifics through follow-up prompts. Think of it like zooming in on a map.

Why it works: AI gives better answers when it understands the big picture first, then adds detail. One massive prompt with 10 requirements often produces generic results.

Copy-paste template:

LEVEL 1 (Overview):
"Give me a high-level approach to [task]"

LEVEL 2 (Structure):
"Expand on [specific part from Level 1]"

LEVEL 3 (Details):
"For [specific element], give me step-by-step instructions"

LEVEL 4 (Examples):
"Show me a concrete example of [step from Level 3]"
  

Example - Content Strategy:

LEVEL 1:
You: "Give me a high-level content strategy for a productivity app targeting remote workers"

AI: Focus on 3 pillars: async communication, focus/deep work, and work-life boundaries. Create content that...

LEVEL 2:
You: "Expand on the 'async communication' pillar. What specific content types would work?"

AI: For async communication:
- How-to guides on tools (Slack, Loom, Notion)
- Case studies from remote-first companies
- Templates for async standup updates...

LEVEL 3:
You: "For the 'how-to guides,' give me a step-by-step process to create one"

AI: Step 1: Identify the specific async tool...
Step 2: Outline common pain points...
Step 3: Create screenshots showing...

LEVEL 4:
You: "Show me an outline for a how-to guide on 'Async Standups in Slack'"

AI: Title: "How to Run Async Standups in Slack (Without Meetings)"
1. Why async standups work...
2. Setting up a dedicated channel...
3. Template for daily updates...
  

Each level gives AI more context about what you actually want.

When NOT to use: Simple, straightforward requests where one prompt is enough.

Tools:ChatGPT | Claude | Gemini

A quick note before you go

Thanks for reading this week’s Brain Bytes — I hope something here helped you move faster or think better.

How’d this one land?

• A ✨ Super Helpful!
• B 🤔 Some Good Stuff!
• C 😕 Could be better.

See you next week, — Oliver

Hey there!

This week, Disney made the biggest AI bet in Hollywood history. OpenAI released GPT-5.2 after declaring "code red." And somewhere in all this chaos, there's actually useful stuff for the rest of us.

Here's what happened:

• Disney invests $1B in OpenAI - licenses 200+ characters for Sora (Mickey, Darth Vader, Iron Man)
• GPT-5.2 drops early - OpenAI's "code red" response to Google's Gemini 3
• Batch Processing: Do 20 Tasks at Once
• And more...

🛠️ Tool of the Week: Wispr (Not Sponsored)

The voice typing tool that's 3x faster than your keyboard (and actually works)

If you write thousands of words per week, your wrists probably hurt and your brain thinks faster than your fingers can type. Wispr Flow fixes both problems.

It's an AI-powered voice dictation tool that works in every app - Gmail, Slack, Google Docs, your IDE, messaging apps. You speak, and polished text appears instantly.

What makes it different:

• 10% error rate vs. 27% for OpenAI Whisper, 47% for Apple dictation
• Auto-cleans speech - removes "um," "like," adds punctuation automatically
• Command Mode - say "make this formal" and it rewrites in real-time
• Learns your vocab - auto-learns new words you use

Real-world example:

[You speak]: "Hey thanks for the update um I think we should schedule a call next week let me know what works"

[Wispr outputs]: "Hey, thanks for the update. I think we should schedule a call next week. Let me know what works for you."

The numbers:

• $81M raised (Nov 2025)
• 270 Fortune 500 companies use it
• 3x faster than typing

Who uses it: Developers coding by voice, sales reps writing 50+ emails/day, writers with carpal tunnel, anyone who thinks faster than they type.

Pricing:

Free: 2,000 words/week
Pro: $15/month – unlimited

Tools:Wispr Flow

1. Disney Bets $1 Billion on AI (And Licenses Mickey Mouse to OpenAI)

On December 11, Disney dropped the biggest Hollywood-AI partnership ever: a $1 billion equity investment in OpenAI, plus a 3-year licensing deal that puts 200+ Disney characters into Sora (OpenAI's AI video generator).

What's in the deal:

• Disney characters available in Sora: Mickey, Ariel, Simba, Iron Man, Darth Vader, Yoda, Elsa, and 200+ more
• Fans can create short AI-generated videos with these characters (starting early 2026)
• Best user-created videos will stream on Disney+
• Disney becomes a "major customer" of OpenAI – will use ChatGPT for employees and build Disney+ features with OpenAI's API

What's NOT included:

• Actor likenesses or voices (only animated/illustrated characters)
• Long-form content (just short social videos)

Why this matters:

This is Disney’s “if you can’t beat them, join them” moment. For months, they’ve been suing AI companies (Midjourney, Character.AI) for copyright violations. On December 10, they sent a cease-and-desist to Google alleging "massive scale" copyright infringement.

Then 24 hours later, they hand OpenAI $1B and 200 characters.

The real story:

Disney gets control. They created a joint steering committee with OpenAI to monitor every user-created video. There’s a “voluminous brand appendix” outlining what you can’t do with their characters. And crucially – OpenAI can’t train its models on Disney IP.

What it means for creators:

Early 2026, you'll be able to type “Darth Vader playing basketball with Groot” and get a 6-second video in your style. Disney curates the best ones for Disney+. It's user-generated content, but with Disney-level IP and OpenAI-level tech.

The irony:

Disney sent Google a cease-and-desist on Wednesday. Announced the OpenAI deal on Thursday. Hollywood’s new strategy: sue first, partner later.

🧠 2. AI Agents Are Finally Useful (Here's What Changed)

Why This Matters

• After years of hype, AI agents that actually complete multi-step tasks are going mainstream across Google, OpenAI, and Anthropic.
• These aren't chatbots—they're systems that plan, execute, use tools, and persist through long-running tasks across your actual apps.
• The shift from "answering questions" to "taking action" fundamentally changes what counts as AI assistance.

The Reality Check

OpenAI's GPT-5.1 update introduced adaptive reasoning that thinks harder only when tasks deserve it, cutting response times by 2–3x on many workloads. Anthropic secured $15 billion from Microsoft and NVIDIA for deeper Claude integration into enterprise workflows.

These agents still need human oversight. They pause before critical actions like purchases or sending messages, and you can take over anytime.

The Practical Impact

Real tasks these agents handle today:

• Analyzing entire email threads and creating summaries with action items
• Booking travel by coordinating across calendars, booking sites, and preferences
• Reading codebases (with 1M token context windows) and suggesting updates
• Researching topics across multiple sources and generating comprehensive reports

The key innovation is "stateful tool use"—agents maintain their reasoning across multiple steps using thought signatures instead of starting fresh each time.

Bottom line: The bottleneck is shifting from "can AI do this?" to "what’s the right task to delegate?" Agents that connect to your actual workflow tools—email, calendar, files, terminal—are replacing isolated chatbots. Companies integrating agent capabilities into existing tools (Google Workspace, Microsoft 365) have the distribution advantage.

1. 🔁 Prompt Chaining: Turn One Prompt Into A Workflow

The technique: Break complex tasks into a sequence of prompts where each output feeds the next input. Think of it like a factory assembly line – each station does one thing well.

Why it works: AI is better at focused, single-step tasks than handling 5 things at once. Chaining gives you quality at every stage while automating the full workflow.

Copy-paste template:

STEP 1: [First task]
[Paste AI's output here]

STEP 2: Take the output above and [second task]
[Paste AI's output here]

STEP 3: Using both outputs, now [third task]
  

Example - Blog Post to Social Media:

STEP 1: Summarize this 2000-word blog post into 3 key takeaways:
[paste article]

AI OUTPUT:
1. AI tools save 40–60 min/day for average users
2. Focus shifted from features to reliability
3. Enterprise adoption growing 3x faster than consumer

STEP 2: Turn takeaway #1 into a LinkedIn post (under 150 words, conversational tone)

AI OUTPUT:
"Here's something wild: the average ChatGPT Enterprise user is saving 40–60 minutes every day. Heavy users? Over 10 hours per week.

That's not a marginal gain – that's reclaiming 2 full workdays per month.

The catch? You need to actually use it consistently. The people seeing these results aren't just trying AI occasionally. They've integrated it into their daily workflow.

What's one task you do every day that takes 20+ minutes? That's your starting point."

STEP 3: Now create a Twitter thread (5 tweets, max 280 chars each) from all 3 takeaways
  

Power move: Save your best chains as templates. Got a good “article → LinkedIn → Twitter” chain? Reuse it every week.

When to use: Content repurposing, research → report, data analysis → presentation, any multi-step workflow you do regularly.

Tools:ChatGPT • Claude • Gemini

2. 🎯 Strategic Context Loading: Feed AI Only What It Needs

The technique: Instead of dumping your entire project into AI at once, strategically add context in layers based on what the current task actually needs.

Why it works: AI performs worse when overwhelmed with irrelevant information. Focused context = better, more accurate outputs.

Copy-paste template:

BACKGROUND (one-time setup):
[Core context that applies to all tasks]

CURRENT TASK:
[Specific task at hand]

RELEVANT DETAILS:
[Only the info needed for THIS task]

Now: [What you want AI to do]
  

Example - Email Draft:

BACKGROUND:
I'm a product manager at a B2B SaaS company. Our product helps sales teams automate outreach.

CURRENT TASK:
Draft an email to a customer who reported a bug

RELEVANT DETAILS:
- Customer: Sarah Chen, VP Sales at TechFlow (customer for 8 months)
- Bug: Email sequences sending duplicates to same contacts
- Our fix: Deployed yesterday, monitoring for 24 hours
- No data was lost or compromised

Now draft a brief email (under 100 words) that:
1. Acknowledges the issue
2. Confirms it's fixed
3. Offers to jump on a call if needed
  

The difference: Without strategic context, AI might write a generic apology. With it, AI knows to be brief (Sarah is busy), acknowledge the relationship (8-month customer), and focus on the fix (not excuses).

When to use: Writing emails to specific people, creating documents with organizational context, any task where "who/what/why" changes the right answer.

When NOT to use: Quick one-off questions where context doesn't matter.

Tools:ChatGPT | Claude | Gemini

3. ⚡️ Batch Processing: Do 20 Tasks at Once

The technique: Give AI a list of similar tasks to complete all at once instead of one-by-one. Like "find-replace" for creative work.

Why it works: AI can handle repetitive tasks at scale without quality loss. What would take you 2 hours takes AI 2 minutes.

Copy-paste template:

I need you to [task] for each item below.

Format: [how you want output structured]

Items:
1. [item 1]
2. [item 2]
3. [item 3]
...
20. [item 20]

Example - Product Descriptions:

I need you to write a 50-word product description for each item below.

Format:
- Start with the main benefit
- Include 2-3 key features
- End with call-to-action

Products:
1. Wireless charging pad - $29.99 - Fast charge, works with any Qi device, sleek aluminum
2. Noise-canceling earbuds - $79.99 - 30hr battery, transparency mode, IPX4 water resistant
3. Portable SSD - $149.99 - 2TB storage, USB-C, 1050MB/s read speed
... [17 more products]

AI outputs all 20 descriptions in one response, consistent style, ready to paste into your store.

Example - Email Subject Line Testing:

Generate 10 different subject lines for this email. Make half curiosity-driven, half benefit-driven.

Email content:
"We just launched a new feature that auto-generates meeting summaries from your calendar. It pulls key decisions, action items, and who's responsible - then emails everyone a recap within 5 minutes of the meeting ending."

Output format:
[Number]. [Subject line] - [Type]

Time saved: Writing 20 product descriptions manually = 2 hours. With AI batch = 5 minutes (2 min for AI + 3 min to review/edit).

When to use: Product descriptions, social media posts, email subject lines, data formatting, anything where you need "same task, different inputs" 10+ times.

Pro tip: Keep your best batch prompts in a doc and reuse them. Changed one word? Batch it.

Tools:ChatGPT | Claude | Gemini

A quick note before you go

Thanks for reading this week’s Brain Bytes — I hope something here helped you move faster or think better.

How’d this one land?

• A ✨ Super Helpful!
• B 🤔 Some Good Stuff!
• C 😕 Could be better.

See you next week, — Oliver

Hey, it's Oliver, here's your AI update for this week!

In today's issue:

• Superhuman: The AI email app that users call "the perfect subscription I'd never cancel"
• Claude Opus 4.5 crushes coding benchmarks (80.9% on SWE-bench Verified)
• Use "Few-Shot" prompting to teach AI by example (10x better results)
• Build AI workflow chains that automate multi-step processes
• And more...

🛠️ Tool of the Week: Superhuman ($30/month) (Not Sponsored)

What it is:

An AI-powered email client that makes you get through your inbox twice as fast. It's Gmail or Outlook, but rebuilt from scratch with speed and AI at the core.

Why it matters now:

Here's the thing about Superhuman: it costs $30/month, and people absolutely refuse to cancel it. Users say things like "the perfect subscription I'd never cancel" and "I can't live without it."

I was skeptical too—who pays $30/month for email? But Superhuman just got a massive AI update (December 2025) with features that are genuinely different from other email tools.

Three ways it saves hours every week:

• Auto Labels + Auto Archive = Inbox Zero automatically – AI categorizes every email: "Response Needed," "Waiting On," "Marketing," "Cold Pitches." Create custom labels with prompts like "job applications" or "urgent requests." Turn on Auto Archive and marketing emails disappear automatically. Users clear hundreds of emails per day without thinking about it.
• AI writes in YOUR voice, not GPT's – The "Write with AI" feature learns your style by analyzing emails you've sent. It pulls from your calendar, past emails, and web research. The result sounds like you—people say "doesn't sound GPT-y at all" because it adapts tone per recipient.
• Ask AI knows your entire inbox – Press "?" and ask "What's the status of the Q4 budget?" It searches all emails and summarizes. Draft emails, schedule meetings, get answers—no digging through your inbox. Chat history saves so you don't re-ask questions.

Just launched (December 2025):

• Auto Drafts write follow-ups without prompting
• Smart reminders surface emails if people don't respond
• AI connects your inbox, calendar, and web research

The catch:

$30/month for individuals, $40/month for Business plan (includes CRM). But users save 4+ hours per week and respond 12 hours faster.

Real talk:

For people with 50+ emails/day where email is central to their job. If you get 10 emails a day, skip it. But if you live in your inbox? Users process email 2x faster.

Pricing:

Individual: $30/month | Business: $40/month

Bottom line:

The price is jarring until you realize it saves 4+ hours weekly. That's worth way more than $30 for most professionals. If email is 25%+ of your job, try it for a month.

Try Superhuman Free

🤖 1. Anthropic's Claude Opus 4.5 Sets New Coding Record (And It's 3x Cheaper Than Before)

Why This Matters

On November 24, Anthropic launched Claude Opus 4.5 with a specific claim: "best model in the world for coding, agents, and computer use." The benchmarks actually back it up.

• 80.9% on SWE-bench Verified (beats Google's Gemini 3 Pro at 76.2% and OpenAI's GPT-5.1 at 77.9%)
• 66.3% on OSWorld (computer use benchmark, best in class)
• $5/$25 per million tokens (down from $15/$75 for the previous Opus 4)

This is the first model to break 80% on SWE-bench Verified, which tests whether AI can autonomously solve real GitHub issues—fixing bugs, adding features, refactoring code. That's not trivial code completion, that's actual software engineering.

What Changed

Opus 4.5 doesn't just autocomplete or suggest fixes—it plans multi-step solutions, coordinates changes across multiple files, and handles ambiguity without constant handholding. Early testers report it can complete tasks that Sonnet 4.5 couldn't even attempt.

The pricing drop is massive: $5 input / $25 output versus the old $15 / $75. That makes Opus-level intelligence actually affordable for production use, not just occasional experiments.

Anthropic gave Opus 4.5 the same coding test they give prospective performance engineering candidates. The AI outperformed most human applicants. That's not a benchmark gaming trick— that's a real-world signal that AI coding has crossed a new threshold.

Real-World Applications

• GitHub Copilot integrated it immediately as the default agent model for users on Business and Enterprise plans
• Cursor, Warp, and other AI coding tools upgraded within days
• Lovable, a no-code platform, reported that Opus 4.5's reasoning depth "transforms planning" for code generation

The practical impact: developers are using it for multi-hour autonomous coding sessions, complex refactors that span dozens of files, and architectural decisions that require understanding trade-offs.

Bottom line:

The gap between "AI can help with code" and "AI can write production code autonomously" just got a lot smaller. This isn't replacing developers yet, but it's replacing a lot of the grunt work— and doing it faster and cleaner than most humans can.

🧨 2. OpenAI Declares "Code Red" and Fast-Tracks GPT‑5.2 to Fight Back

The Situation

On December 2, OpenAI CEO Sam Altman issued an internal "code red" directive. The reason: Google's Gemini 3 topped the leaderboards last month, and Anthropic's Claude Opus 4.5 followed immediately after. OpenAI is now scrambling to respond.

What's Happening

• GPT‑5.2 release moved up to December 9 (originally scheduled for later this month)
• All “side projects” paused including advertising tests, shopping agents, and experimental features
• Internal focus shifted to core model quality over new product features
• Developer memo warned of “various errors during December” as infrastructure gets rebuilt at speed

Sources report that GPT‑5.2’s internal benchmarks edge out Gemini 3 on reasoning tasks, but that's unconfirmed. What’s clear: OpenAI is treating this as an existential challenge, not a routine release.

The Competitive Context

ChatGPT’s market share is slipping:

• Over the last four months, ChatGPT’s share of AI chat app users dropped 3 percentage points
• Gemini gained 3 points and users’ daily time doubled to ~11 minutes
• ChatGPT’s daily time per user increased only 6% and actually declined in November
• Perplexity and Claude posted triple‑digit year‑over‑year growth

This isn't about benchmarks anymore—it's about whether ChatGPT can maintain its position as the default AI assistant. Google has distribution through Search and Android. Anthropic has enterprise contracts. OpenAI has brand recognition, but that’s not enough if competitors catch up on quality.

What GPT‑5.2 Will Focus On

Reports say the update prioritizes “speed, reliability, and customizability” rather than flashy new features. OpenAI wants to close the gap Google created with Gemini 3 while preserving ChatGPT’s ease of use.

The “code red” framing matters: it signals a shift from long‑term strategic planning to rapid‑response execution. This is the AI industry moving from multi‑month release cycles to continuous competitive sprints.

Bottom line:

The AI race is accelerating. OpenAI, Google, and Anthropic are all releasing major updates within weeks of each other, not months. For users, this means better models faster— but it also means more fragmentation and constant learning curves as tools rapidly evolve.

1. 📚 Few‑Shot Prompting: Teach AI By Example (Not By Explanation)

The technique: Instead of describing what you want in words, show AI 2–3 examples of the exact output format you need. AI learns the pattern and replicates it.

Why it works: AI is better at pattern matching than following abstract instructions. Examples eliminate ambiguity—AI sees exactly what “good” looks like.

Copy‑paste template:

Here are 3 examples of the format I need:

Example 1:
Input: [sample input 1]
Output: [desired output 1]

Example 2:
Input: [sample input 2]
Output: [desired output 2]

Example 3:
Input: [sample input 3]
Output: [desired output 3]

Now apply this same format to:
Input: [your actual input]
  

Example – Email Tone Matching:

Example 1:
Input: Reject a meeting request
Output: "Hey Sarah, thanks for reaching out! I'm slammed this week, but I'd love to connect. How's next Tuesday at 2pm?"

Example 2:
Input: Request a deadline extension
Output: "Hi Mike, quick heads up—I'm running a day behind on the Q4 report. Could we push to Friday? Happy to share a draft Thursday if helpful."

Now write in the same style:
Input: Thank a colleague for covering your shift
  

AI will match the casual‑but‑professional tone automatically.

When to use: Specific tone requirements, data formatting, structured outputs, maintaining consistent style.

Tools:ChatGPT | Claude | Gemini

🔗 2. Chain-of-Thought + Tool Use: Build AI Workflow Chains

The technique: Break complex tasks into a sequence of AI-powered steps where each output feeds into the next input. Like an assembly line for AI operations.

Why it works: AI excels at individual tasks but struggles with complex multi-step processes. Chaining steps together maintains quality at each stage while automating the entire workflow.

Copy-paste template:

Complete this workflow in steps:

Step 1: [First task]
Output labeled as "STEP 1 OUTPUT:"

Step 2: Use STEP 1 OUTPUT to [second task]
Output as "STEP 2 OUTPUT:"

Step 3: Use STEP 2 OUTPUT to [final task]

[Provide initial input]
  

Example - Content Repurposing Chain:

Transform this blog post into social media content:

Step 1: Extract the 3 most important insights
Output as "KEY INSIGHTS:"

Step 2: Using KEY INSIGHTS, write 3 LinkedIn posts (150 words each)
Output as "LINKEDIN POSTS:"

Step 3: Using KEY INSIGHTS, write 3 Twitter threads (5 tweets each)
Output as "TWITTER THREADS:"

Blog post: [paste content]
  

Each step builds on the previous output, ensuring consistency across formats.

Example - Research to Report:

Step 1: Find 5 recent articles about [topic] and list key findings
Output as "RESEARCH FINDINGS:"

Step 2: Using RESEARCH FINDINGS, identify 3 main trends
Output as "KEY TRENDS:"

Step 3: For each trend, write 2-3 sentences on business implications
Output as "EXECUTIVE SUMMARY:"

Topic: [your topic]
  

Power move: Save successful chains as templates you can reuse with different inputs.

When to use: Content repurposing, research synthesis, multi-step analysis, any workflow where each stage matters.

Tools:ChatGPT | Claude (best for long contexts) | Gemini

3. 🧠 Context Window Hacking: Make AI Remember 10x More

The technique: Use strategic summarization and context compression to make AI work with much more information than its context window technically allows.

Why it works: Modern AI has huge context windows (200K+ tokens), but quality degrades past a certain point. Intelligent compression gets better results than dumping everything in.

Copy-paste template:

I'm providing information in chunks. For each chunk:
1. Read and understand
2. Summarize key points in 2-3 bullets
3. Wait for next chunk

After all chunks, I'll ask you to use all summarized information.

Chunk 1: [First section]
  

Example - Processing Multiple Documents:

I have 5 research papers. For each, I'll paste the abstract and conclusion.

Extract for each:
- Main finding (1 sentence)
- Methodology (1 sentence)
- Key limitation (1 sentence)

Store as PAPER 1, PAPER 2, etc.

After all 5, create a comparison table.

Paper 1: [paste]
  

Example - Long Conversation Memory:

When a conversation gets too long, compress it:

Before we continue, create a "memory summary":
- 5 main points we've agreed on
- 3 open questions still being explored
- Any specific examples we referenced

Then continue using this as foundation.
  

Power move - Progressive detail:

Start broad, then zoom in:

Round 1: Give 10,000-foot overview in 3 sentences
Round 2: Expand to 10 bullet points
Round 3: For bullets 2, 5, and 7, give detailed examples

This controls depth without overwhelming context.
  

When to use: Long documents, multiple sources, ongoing projects, research synthesis.

When NOT to use: When you need verbatim quotes, legal/compliance work, when nuance matters more than summaries.

Tools:ChatGPT | Claude (200K tokens) | Gemini

🔗 Useful Links:

• Claude Opus 4.5 Announcement | Model Card
• Superhuman | AI Features
• SWE-bench Verified Leaderboard

📚 More Resources:

• ChatGPT | API Docs
• Claude | Developer Console
• Gemini | AI Studio

A quick note before you go

Thanks for reading this week’s Brain Bytes — I hope something here helped you move faster or think better.

How’d this one land?

• A ✨ Super Helpful!
• B 🤔 Some Good Stuff!
• C 😕 Could be better.

See you next week, — Oliver