Today is September 13, 2023.

The last two weeks have seen quite a few interesting developments in the AI world. We’re starting to see a global race on the AI model front, with Chinese ChatGPT competitors finally in public release.

And in the US, the historically stodgy IRS has revealed they’re already adopting cutting edge AI to investigate tax fraud.

Meanwhile, Meta’s AI team continues to open source more interesting AI tech, this time for video-tracking.

In this issue:

• 🪝 IRS deploys AI to catch tax evaders

• 🏃‍♂️ Meta wants their next chatbot model to rival OpenAI’s GPT-4

• 🇨🇳 China’s ChatGPT rivals make public debut

• 🍎 Duet for Google Workspace launches for $30 per month

• 🔎 Call of Duty polices gamers with AI, and other quick scoops

• 🧪 The latest science experiments, including Meta’s new point-tracking model (with code and demo)

Did someone forward this to you? Subscribe here!

🪝 IRS deploys AI to catch tax evaders

The IRS is using cutting-edge AI to assist its investigations into complex cases, reports the New York Times.

With its new AI tools, the IRS is specifically focusing on hedge funds, private equity groups, real estate investors and law firms – complex organizations where IRS teams have in the past found themselves under-resourced to police.

Why this matters: AI is increasingly finding a fit in serving as a co-pilot to smart humans. The ability of artificial intelligence to sort through large data sets, spot trends, and identify patterns is highly valuable in complex projects ranging from medical to legal and now finance work.

With just 1% of large, complex partnerships currently audited by the IRS, AI tooling could significantly improve the coverage of these organizations in the years to come.

🏃‍♂️ Meta wants their next chatbot model to rival OpenAI’s GPT-4

Meta’s next generation language model is intended to rival GPT-4’s abilities, the Wall Street Journal reports. After releasing Llama 2, Meta intends to further accelerate their AI progress by training a model several times more powerful.

Training won’t start until early 2024, sources familiar with Meta’s plans revealed, as the company is still building up the required data center capacity to develop this model.

Why this matters: sources revealed that Meta likely intends to open-source this model as well, which will continue to put pressure on closed-source AI model vendors OpenAI and Google. But the release of an even more powerful model is also likely to raise further questions from critics about the ethics of sharing such powerful technology so freely.

🇨🇳 China’s ChatGPT rivals make public debut

Chinese search engine and AI company Baidu officially launched their ChatGPT competitor model last week, marking an important milestone in China’s own race to dominate the artificial intelligence landscape. Two other companies, Baichuan and Zhipu AI, also launched AI models the same day.

Dubbed “Ernie,” Baidu’s app quickly topped the iOS App Store charts in China upon release. The model will now benefit from wide-scale human use to further improve its outputs.

Why this matters: China has heavily stressed its goal of overtaking the United States in AI advancements, and the generative AI space is an important sector to showcase these gains. At the same time, US restrictions of advanced AI chips to China have introduced geopolitics into this global tech race, as Chinese companies seek to find workarounds and stay competitive in their AI progress.

These same Chinese AI companies are also navigating a thorny regulatory environment for AI models, as they must pass review by various government bodies before they launch their services.

🍎 Duet for Google Workspace launches for $30 per month

Duet is an all-in-one personal assistant that seamlessly integrates with the Google suite of products (Docs, Spreadsheets, and more), and is designed to enable busy professionals to have a strong AI partner. Imagine how “a last-minute request that once called for an all-nighter, can now be completed before dinner time,” Google’s own marketing buzz touts.

Why this matters: AI models are expensive to run, and Microsoft already announced their 365 Copilot at a cost of $30 per month. Google’s competitive product will help retain customers on their work software suite while also showing off their AI prowess as Microsoft touts OpenAI’s foundation models. And successful wide-scale adoption could represent billions in revenue for Google as well.

Google is currently offering a free trial to any existing Workspace users curious about Duet’s capabilities.

🔎 Quick Scoops

Call of Duty is using AI to identify toxic behavior in real-time voice chats. The AI helps flag instances of toxic behavior for a human moderation team. (PC Gamer)

OpenAI is on pace to generate over $1 billion per year, as demand for artificial intelligence stays strong. The company generated just $28M in 2022. (The Information)

AI startups are booming as the latest batch from Y Combinator features 160 AI companies. The entire class has just 229 companies in total. (Democratizing Finance)

US rejects copyright protection for an award-winning AI-generated art piece, despite the fact that the text prompt was revised 6254 times and the resulting image was then edited in Photoshop. (Ars Technica)

Elon Musk readies X to train its AI on user data, as changed terms of service for X users recently reveal. (Decoder)

🧪 Science Experiments

Meta releases CoTracker, a model for point tracking in videos

• The AI model understands how various points in a video can shift, potentially saving hours of time for humans who normally do this by hand.

• You can try it out for yourself here by uploading your own video.

Credit: Meta

VideoGen: A Reference-Guided Latent Diffusion Approach for High Definition Text-to-Video Generation

• Baidu’s AI team releases their video generation model, which touts impressive improvements in fidelity and composition over rival models.

• See their project page here.

Credit: Baidu

Blended-NeRF: Zero-Shot Object Generation and Blending in Existing Neural Radiance Fields

• Editing a local region or a specific object in a 3D scene represented by a NeRF is challenging, mainly due to the implicit nature of the scene representation. Blended-NeRF enables editing a specific region of interest in an existing NeRF scene, opening up a realm of possibilities.

• See their project page here.

👋 How I can help

Here’s other ways you can work together with me:

• If you’re an employer looking to hire tech talent, my search firm (Candidate Labs) helps AI companies hire the best out there. We work on roles ranging from ML engineers to sales leaders, and we’ve worked with leading AI companies like Writer, EvenUp, Tome, Twelve Labs and more to help make critical hires. Book a call here.

• If you would like to sponsor this newsletter, shoot me an email at michael@artisana.ai 

As always — have a great week!

Today is August 29, 2023.

This week’s issue highlights a number of areas in AI beyond chatbots and generative AI that are seeing significant advancement.

First, we have an in-depth look at how the US Air Force is rapidly shifting their strategy towards a future where AI-powered warplanes will outnumber human pilots.

We also feature two interesting developments on the medical front as AI improvements increasingly offer hope to paralyzed patients, and overworked doctors stand to benefit from AI detection models serving as copilots in their work.

Read on!

In this issue:

• 🪖 The US Air Force is embracing AI-powered warplanes

• 🧠 AI helps paralyzed woman speak out loud via decoded brain signals

• 🚀 OpenAI launches ChatGPT Enterprise, aiming to address privacy and security needs

• 🤗 Hugging Face raises at a $4.5B valuation, Meta releases their coding assistant AI, and more

• 🧪 The latest science experiments, including an AI model that generates college lectures and radio shows from text prompts

Did someone forward this to you? Subscribe here!

🪖 The US Air Force is embracing AI-powered warplanes

The US Air Force is leading the way on AI with its focus on next-generation AI-powered warplanes, this comprehensive report from the New York Times (note: paywalled site) details.

Next-generation drones able to fly as AI wingmen to human pilots are currently in testing, and the Air Force plans to build as many as 2,000 of these drones at a target cost of $3M apiece – a fraction of what the F-35 fighter costs ($80M per plane).

The pilotless XQ-58A Valkyrie prototype. Photo Credit: NY Times

Why this matters:

• The focus on these “collaborative combat aircraft” shows that the Air Force is aware that future battlefields may require a large swarm of highly capable robots to dominate.

• Increasingly sophisticated air defense systems from adversaries like China also make the concept of cheaper, unmanned AI-powered warplanes far more attractive.

At the same time, these moves are accelerating a debate within the military for what role humans should play in conflicts, and the limits we should place on software that enables machines to kill.

🧠 AI helps paralyzed woman speak out loud via decoded brain signals

A Canadian woman who was paralyzed after a stroke is now able to “speak” for the first time in nearly 20 years, thanks to an AI system that decodes her brain signals.

Why this matters:

• As AI systems have grown increasingly sophisticated, brain-computer interfaces are a promising area that is likely to see significant advances.

• The scientists behind her accomplishment notably didn’t use AI to decode thoughts (as some other AI models have shown is possible). Instead, they simply trained an AI model on phonemes, which are muscle movements that form the building blocks of speech.

• To everyone’s surprise, the AI model only needed to learn 39 phonenes to decipher any word in English, allowing the model to perform far more efficiently than other AI approaches.

Brain-computer interfaces are an area that numerous private sector companies are pursuing as well, with the most notable being Elon Musk-founded Neuralink. Expect to see more promising developments here in the future.

🚀 OpenAI launches ChatGPT Enterprise, aiming to address privacy and security needs

ChatGPT’s adoption amongst knowledge workers has been impressive since its launch, but in recent months a multitude of companies have banned ChatGPT use by employees over privacy and security concerns.

OpenAI officially plans to change this with an enterprise-level offering of ChatGPT, specifically focused on concerns large corporations have raised.

• Privacy: OpenAI’s models won’t use conversations in ChatGPT Enterprise for training purposes, and additional SOC 2 compliance and encryption are designed to reassure enterprises this is a trustworthy offering.

• Capability: ChatGPT Enterprise notably removes all usage caps, prioritizes speed (up to 2x faster), and allows users to utilize a 32k context window to process longer inputs and files.

OpenAI may also be seeking to ward off competition posed by open-source AI models such as Meta’s Llama, which have touted their privacy and customizability as key advantages for enterprise-grade AI needs. How enterprise-grade adoption plays out, especially in light of OpenAI’s new managed service offering, will be interesting to observe over the next few years.

🔎 Quick Scoops

AI startup Hugging Face valued at $4.5B in latest funding round. Notably, Google, Nvidia, Intel and others participated, highlighting investor enthusiasm for AI and the ecosystem of partners supporting each other. (Bloomberg)

Google proposes a new watermarking approach to stop deepfakes, but the availability of open-source AI image-creating projects leaves its widespread adoption in doubt. (Washington Post)

AI models can read chest radiographs and pick up signs that doctors often miss, researchers from Osaka University revealed. AI copilots of human doctors could become increasingly common as improvements arrive. (SciTech Daily)

Meta releases Code Llama, their AI model designed to help write code. Similar to Llama 2, the model comes in multiple sizes and isn’t truly open-source, but is free to use for a wide variety of use cases. (Meta)

ElevenLabs comes out of beta and now supports AI speech for 28 languages. AI speech generation is gaining in sophistication, posing numerous business possibilities but also raising challenges around its misuse. (ElevenLabs)

🧪 Science Experiments

DenseDiffusion: Dense Text-to-Image Generation with Attention Modulation

• Existing text-to-image diffusion models struggle to synthesize realistic images when given dense captions. DenseDiffusion is a training-free method that offers far better outputs and control over scene layout when digesting dense captions.

• Project page here, with comparisons between DenseDiffusion and StableDiffusion.

Comparison of DenseDiffusion’s outputs vs. standard StableDiffusion. Credit: GitHub

WavJourney: Compositional Audio Creation with LLMs

• LLMs that can specialize in audio content creation remain relatively unexplored. WavJourney is a end-to-end system that leverages LLMs to excel at audio storytelling, all starting from a prompt of an auditory scene.

• Project page here, including examples of a new report, college lecture, and more.

Meta releases SeamlessM4T, a multimodal AI model for speech translation

• Seamless translation across speech and text for over 100 languages is possible across a variety of use cases, including speech-to–speech translation. According to Meta, this is the first model to achieve state-of-the-art results across so many languages and tasks.

• See their project page here.

Credit: Meta

👋 How I can help

Here’s other ways you can work together with me:

• If you’re an employer looking to hire tech talent, my search firm (Candidate Labs) helps AI companies hire the best out there. We work on roles ranging from ML engineers to sales leaders, and we’ve worked with leading AI companies like Writer, EvenUp, Tome, Twelve Labs and more to help make critical hires. Book a call here.

• If you would like to sponsor this newsletter, shoot me an email at michael@artisana.ai 

As always — have a great week!

Today is Monday, August 21, 2023.

Welcome to issue 17 of our AI Brief! We kick off this issue with a provocative yet important question to ask: how overhyped is generative AI?

A thoughtful piece by Gary Marcus is currently at the center of attention right now in the AI community. After months of seeing what feels like breathless advancements in AI, he wants us to consider the opposite: is it all just going to lead to disappointment?

An important question I ask is how we can stay on top of what’s truly meaningful in the world of AI while filtering out the hype that leaves us fatigued. This is a key motivation for why I avoid adding long laundry lists of news developments to this newsletter.

Less is more. If there’s important news to feature — it’ll end up here. Otherwise I try to stay away from the unneeded hype.

And of course, this is just one humble person’s opinion.

In this issue:

• 🧠 The big debate: what if generative AI turns out to be a dud?

• 🖼️ AI-created art isn’t copyrightable, federal judge rules

• 👊 Meta prepares to launch its own open-source code generation AI model

• 📦️ Amazon moves to summarize product reviews with AI, and other quick scoops

• 🧪 The latest science experiments, including the first LLM agent benchmark

Did someone forward this to you? Subscribe here!

🧠 The big debate: what if generative AI turns out to be a dud?

Driving the conversation within AI circles is a thought-providing piece from entrepreneur and neuroscience professor Gary Marcus.

He asks a simple question: what if generative AI is set up to disappoint us?

Outside of writing text and code, Marcus argues, “other potential paying customers may lose heart quickly” as the limits on generative AI become clearer to the professional world.

• The love affair with ChatGPT is wearing off, Marcus points out, as a growing chorus of voices become disillusioned with its limitations.

• OpenAI’s $29B valuation may be hard to justify, and other startups raising at lofty valuations may find themselves in hot water.

• And as numerous startups broadly claim they’re AI companies, the AI hype bubble risks distorting the real work that remains to be done in the field.

AGI is unlikely to be imminent, no matter what the optimists would have you believe, Marcus writes in a separate blog post. The long march towards improving artificial intelligence systems is still rife with unsolved challenges, and LLMs are unlikely to be a pathway to AGI.

A personal thought: as a newsletter writer, I’ve always shied away from the hyped-up proclamations of “ChatGPT is yesterday’s news” and “AI Agents just made Microsoft obsolete,” etc. Deep down, I believe AI can be big, but a quick survey of friends show that we’re personally experiencing some AI fatigue as well.

Venture capitalist Benedict Evans sums up the feeling of the moment quite well:

🖼️ AI-created art isn’t copyrightable, federal judge rules

A federal judge has upheld a finding from the U.S. Copyright Office that AI-created art isn’t copyrightable. Existing copyright claw cannot “protect works generated by new forms of technology operating absent any guiding human hand,” U.S. District Judge Beryl Howell wrote in her decision.

At the core of the finding is a foundational rule: “Human authorship is a bedrock requirement” for to grant a copyright.

• This tracks with previous legal rulings, including one that determined a photograph captured by a monkey couldn’t be granted copyright. Copyright law is meant for “human individuals to engage in” creation, the judge in that case ruled.

However, one exception remains: works created with the help of AI could still be eligible for copyright, if a human “selected or arranged” it in a “sufficiently creative way that the resulting work constitutes an original work of authorship,” the copyright office said.

👊 Meta prepares to launch its own open-source code generation AI model

Meta’s open-source wins will likely mount as it readies another model for the public – this time focused on code generation. Meta’s model, dubbed Code Llama, could pose a threat to paid coding assistants like GitHub’s Copilot, which currently uses OpenAI’s AI models to power its suggestions.

Why this matters:

• Coding assistants have been one of the areas of AI tools to see rapid adoption, as increasingly more powerful LLMs and an eager user base of software developers have witnessed significant leverage in the last year.

• In particular, an open-source model could be highly attractive to companies focused on security: an enterprise could theoretically take Code Llama, fine-tune it on their own code base, and keep the model completely walled off.

One likely casualty of Meta’s Code Llama release could be all of GitHub’s startup competitors, who don’t have GitHub’s reach, Microsoft backing, and head start in user base. And they now have to contend with a free open-source model likely to see rapid adoption.

🔎 Quick Scoops

Advertisers are exploring generative AI tools to cut costs and increase productivity, but overall remain cautious about security and copyright risks. (Reuters)

Academic journals are fighting an influx of shoddy research papers largely written by AI. Concerns about harming their own credibility abound as major journals rush to put in place better safeguards and policies around the use of AI tools. (Decoder)

Google is building an AI to offer life advice, and their DeepMind unit is testing focused AI models designed to help users navigate intimate questions. (CNBC)

Amidst the Hollywood strikes, 96% of entertainment companies plan to boost generative AI spend. This pace of planned AI investment is worrying to many parties. (Forbes)

Amazon rolls out product review summaries made by generative AI. Starting first on mobile, this is designed to help users cut through the noise of browsing product user-written reviews. (AP News)

🧪 Science Experiments

AgentBench: the first benchmark designed to evaluate LLM-as-Agent

• Benchmarks to evaluate LLM performance on focused areas exist, but none to date look at how LLMs can perform as agents.

• AgentBench offers 8 distinct environments to provide a more comprehensive evaluation of the LLMs' ability to operate as autonomous agents in various scenarios.

• See the project page here.

Dual-Stream Diffusion Net for Text-to-Video Generation

• Videos created from a text-to-video generative model carry flickers and artifacts. By separating out the process into two separate diffusion streams - one for video, and one for motion – the end result is significantly improved.

• See the paper here. See their example videos here.

TeCH: Text-guided Reconstruction of Lifelike Clothed Humans

• Despite recent research advancements in reconstructing clothed humans from a single image, accurately restoring the "unseen regions" with high-level details remains an unsolved challenge that lacks attention.

• TeCH produces high-fidelity 3D clothed humans with consistent & delicate texture, and detailed full-body geometry.

• See their paper here.

Credit: arXiv

👋 How I can help

Here’s other ways you can work together with me:

• If you’re an employer looking to hire tech talent, my search firm (Candidate Labs) helps AI companies hire the best out there. We work on roles ranging from ML engineers to sales leaders, and we’ve worked with leading AI companies like Writer, EvenUp, Tome, Twelve Labs and more to help make critical hires. Book a call here.

• If you would like to sponsor this newsletter, shoot me an email at michael@artisana.ai 

As always — have a great week!

Today is August 8, 2023.

This week’s issue arrives one day behind schedule — but expect future issues to resume regular Monday drops in your inbox.

In this issue:

• 🍏 Apple ramps up generative AI recruiting as it readies its own mobile-first model

• 🤯 New deep learning model maps acoustic data from keystrokes with 95% accuracy

• 🎊 Alibaba releases a set of open-source models to rival Meta’s Llama 2

• 🎼 Meta open-sources AudioCraft, their text to audio generative AI model

• 🧪 And latest science experiments, including Google reconstructing music from brain scans

Did someone forward this to you? Subscribe here!

🍏 Apple ramps up generative AI recruiting as it readies its own mobile-first model

A deep-dive from the Financial Times (paywalled) covers how Apple is now moving aggressively on the AI front, as they ready their own generative AI models – in particular, models that can run on mobile devices.

Why it matters: OpenAI and Bard have both released models that run in the cloud. But neither have released a private language model capable of running on mobile devices, personalized for users and fully private in nature.

Driving the news:

• Recent job postings from Apple highlight their strategy, asking AI talent to join the company and help build “state of the art foundation models to the phone in your pocket, enabling the next generation of ML-based experiences in a privacy-preserving way”.

• On a recent earnings call, Apple CEO Tim Cook affirmed their AI strategy, calling AI and machine learning “core, fundamental technologies that are integral to virtually every product that we build”.

Apple’s R&D spending for the third quarter of 2023 was $3.1B higher than the same quarter last year, which Cook acknowledged was driven by heavy investments in generative AI. Most analysts expect Apple’s first generative AI models to see public release in 2024.

🤯 New deep learning model maps acoustic data from keystrokes with 95% accuracy

Researchers in the UK revealed a new deep learning model capable of mapping recorded sounds of keystrokes with 95% accuracy. When Zoom-recorded sounds were used, the model still retained 93% accuracy, a startling high rate, the report revealed.

The research setup. Commonly available consumer devices were used to capture acoustic training data. Credit: arXiv

Why it matters: universal availability of microphones combined with advancements in machine learning now make it possible for adversaries to steal passwords and other sensitive information from just acoustic recordings, marking a new security risk in the digital age.

How the model was trained:

• Researchers pressed 36 keys on a MacBook Pro 25 times each and recorded the sounds on both an iPhone 13 nearby as well as over a Zoom call

• The keystrokes were translated into spectrogram images, which were then trained as an image classifier. Further tweaks in the modeling process produced the high-accuracy model featured in the research paper.

By converting sounds into spectrograms, an image classifier model was then trained to guess keystrokes. Credit: arXiv

Countermeasures still exist, the researchers noted. In particular, detection of shift key usage remains challenging, and keystroke sound removal in popular VOIP protocols could also ward off attacks. But the authors ultimately conclude we’ll need to move away from typed passwords to be truly secure.

🎊 Alibaba releases a set of open-source models to rival Meta’s Llama 2

We’re in the early innings of open-source vs. closed-source AI, and Alibaba has now jumped into the fray with a new set of open-source models. 

These models are also available on ModelScope and Hugging Face as well with code and checkpoints.

Driving the news:

• Alibaba claims their model represents a new performance milestone: “In general, Qwen-7B outperforms the baseline models of a similar model size, and even outperforms larger models of around 13B parameters, on a series of benchmark datasets, e.g., MMLU, C-Eval, GSM8K, HumanEval, and WMT22, etc.” they state.

• Qwen-7B has been trained on a dataset of over 2.2 trillion tokens, including web texts, books, and codes, while supporting an 8K context length and featuring a tokenizer with over 150K tokens. The chat version (Qwen-7B-Chat) has additional features to function as an AI assistant.

Our takeaway:

• General performance claims aside, we’re curious how the model’s training impacts its ability to generate content on banned conversation topics in China.

• Researchers have already highlighted how AI models can be “poisoned” by the smallest amounts of deliberately tuned training data, so it may take some deep-diving to uncover where this model displays certain biases.

🔎 Quick Scoops

OpenAI updates ChatGPT with new features, including prompt examples, suggested replies, and keyboard shortcuts. (Twitter)

AI data centers are growing, thanks to an “insatiable appetite” for computing power, creating a rush to spin up new data centers from scratch at an unprecedented pace. (WSJ - note: paywalled)

Meta is preparing to launch chatbots with distinct personalities, including historical figures such as Abraham Lincoln and a surfer persona. (Al Jazeera)

ChatGPT is getting dumber at basic math possibly due to model drift, the Wall Street Journal reports in its own breakdown of the Stanford paper making waves in the AI community. (WSJ - note: paywalled)

Travel scammers are flooding Amazon with AI-written travel guides, often filled with incorrect information. (New York Times - note: paywalled)

AI tools are threatening the Youtube thumbnail industry, which allowed some workers to earn thousands per thumbnail image created. (Rest of World)

🧪 Science Experiments

Meta open-sources AudioCraft, their text to audio generative AI model

• AudioCraft is a “simple framework that generates high-quality, realistic audio and music from text-based user inputs after training on raw audio signals as opposed to MIDI or piano rolls.”

• See their project page here.

Brain2Music: Reconstructing Music from Human Brain Activity

• A Google team partnered with Osaka University to show the possibility of reconstructing music from brain wave activity, specifically data captured via functional magnetic resonance imaging (fMRI).

• See their project page here.

Tool Documentation Enables Zero-Shot Tool-Usage with Large Language Models

• Researchers demonstrate that zero-shot prompts with only tool documentation are sufficient for eliciting proper tool usage, achieving performance on par with few-shot prompts. This matters because crafting demos of a tool’s usage is significantly harder than simply documenting a tool, pointing to a world where AI can learn to use tools by themselves as long as sufficient baseline documentation exists.

• See the study here.

😀 A reader’s commentary

👋 How I can help

Here’s other ways you can work together with me:

• If you’re an employer looking to hire tech talent, my search firm (Candidate Labs) helps AI companies hire the best out there. We work on roles ranging from ML engineers to sales leaders, and we’ve worked with leading AI companies like Writer, EvenUp, Tome, Twelve Labs and more to help make critical hires. Book a call here.

• If you would like to sponsor this newsletter, shoot me an email at michael@artisana.ai 

As always — have a great week!

Today is July 31, 2023.

The development of AI continues to produce newsworthy tidbits, but two key themes stand out from the past week.

First, researchers who reverse-engineered open-source LLMs discovered they could generate universal attack strings against all Transformer-based LLMs. This represents a new breakthrough, similar to how computer vision models remain vulnerable to encoded inputs invisible to human eyes to this day. With the rush to implement LLMs moving quickly, the security implications of this discovery are profound.

Second, Google’s DeepMind team highlighted two fascinating AI models: one that helps create instructions for robot, as well as their latest a multi-modal AI model for medical professionals. The hype around ChatGPT is starting to cool, and we’re now seeing advancements in AI start to focus on what’s beyond the text realm. I’ll be following this closely in future research updates.

In this issue:

• 📈 McKinsey thinks generative AI will automate away 30% of work hours by 2030

• 🔨 Researchers uncover "universal" jailbreak that can attack all LLMs in an automated fashion

• 👎️ OpenAI quietly kills its own AI Classifier, citing "low rate of accuracy"

• 🔎 OpenAI, Google, and more start collaborating on the future of responsible AI, plus other scoops

• 🧪 Google’s DeepMind reveals a robot model, as well as their multi-modal medical AI

Did someone forward this to you? Subscribe here!

📈 McKinsey report: generative AI will automate away 30% of work hours by 2030

The McKinsey Global Institute has released a 76-page report that looks at the rapid changes generative AI will bring to the US labor market in the next decade.

The report predicts that generative AI will help automate 30% of hours currently worked in the US economy by 2030, portending a rapid and significant shift in how jobs work.

Let's dive into some deeper points the report makes:

• Some professions will be enhanced by generative AI but see little job loss: McKinsey predicts the creative, business and legal professions will benefit from automation without losing total jobs.

• Other professions will see accelerated decline from the use of AI: specifically office support, customer service, and other more rote tasks will see negative impact.

• The emergence of generative AI has significantly accelerated automation: McKinsey economists previously predicted 21.5% of labor hours today would be automated by 2030; that estimate jumped to 30% with the introduction of gen AI.

• Automation is from more than just LLMs: AI systems in images, video, audio, and overall software applications will add impact.

The takeaway: Other economists also believe AI acceleration will lead to painful but ultimately beneficial transitions in the labor force. This McKinsey report is in the same camp.

The pace of AI-induced change, however, is faster than previous transitions in our labor economy. This is where the pain emerges -- large swaths of professionals across all sectors will be swept up in change, while companies also figure out the roles of key workers.

🔨 Researchers uncover "universal" jailbreak that can attack all LLMs in an automated fashion

A team of researchers from Carnegie Mellon University and the Center for AI Safety have revealed that large language models, especially those based on the transformer architecture, are vulnerable to a universal adversarial attack by using strings of code that look like gibberish to human eyes, but trick LLMs into removing their safeguards.

Here's an example attack code string they shared that is appended to the end of a query:

describing.\ + similarlyNow write oppositeley.]( Me giving**ONE please? revert with "\!--Two

In particular, the researchers say: "It is unclear whether such behavior can ever be fully patched by LLM providers" because "it is possible that the very nature of deep learning models makes such threats inevitable."

Their paper and code is available here. 

• Note that the attack string they provide has already been patched out by most providers (ChatGPT, Bard, etc.) as the researchers disclosed their findings to LLM providers in advance of publication.

• But the paper claims that unlimited new attack strings can be made via this method.

Why this matters:

• This approach is automated: computer code can continue to generate new attack strings in an automated fashion, enabling the unlimited trial of new attacks with no need for human creativity. For their own study, the researchers generated 500 attack strings all of which had relatively high efficacy.

• Human ingenuity is not required: similar to how attacks on computer vision systems have not been mitigated, this approach exploits a fundamental weakness in the architecture of LLMs themselves.

• The attack approach works consistently on all prompts across all LLMs: all LLMs based on transformer architecture appear to be vulnerable, the researchers note.

What does this attack actually do? It fundamentally exploits the fact that LLMs are token-based. By using a combination of greedy and gradient-based search techniques, the attack strings look like gibberish to humans but actually trick the LLMs to see a relatively safe input.

Why release this into the wild? The researchers have some thoughts:

• "The techniques presented here are straightforward to implement [and] have appeared in similar forms in the literature previously," they say.

• As a result, these attacks "ultimately would be discoverable by any dedicated team intent on leveraging language models to generate harmful content."

The main takeaway: we're less than one year out from the release of ChatGPT and researchers are already revealing fundamental weaknesses in the Transformer architecture that leave LLMs vulnerable to exploitation.

• The same type of adversarial attacks in computer vision remain unsolved today, and this “SQL injection”-style attack on LLMs shows the rapidly evolving frontier of AI security will continue to stay dynamic.

👎️ OpenAI quietly kills its own AI Classifier, citing "low rate of accuracy"

First launched in January, OpenAI's own AI Classifier tool represented one of the many new tools emerging at the time for detecting AI-generated text.

In the past few months, numerous other tools like GPTZero have also gained popularity with educators as well. Thread after thread of “my professor’s AI detection tool accused me of cheating” can be found on Reddit, highlighting the new battle students now face in presenting original work.

Last week, OpenAI quietly shut their tool down, and did so by only updating the original blog post. "As of July 20, 2023, the AI classifier is no longer available due to its low rate of accuracy," the post now says.

Why this matters:

• AI writing detectors simply can't be trusted, a body of studies in recent months have shown. False positive rates are high, and various simple prompting approaches can all fool AI detectors. As LLMs improve, researchers argue, true detection will only become harder.

• GPTZero's own founder admitted last month he was pivoting the product away from "catching" students, and more towards highlighting the "most human" parts of writing.

The main takeaway: OpenAI's latest move represents a potential nail in the coffin for AI detectors in general.

• If OpenAI, with all its internal knowledge about their AI models, says they can't reliably detect their own text outputs, what does that say about the viability of AI detection in general?

🔎 Quick Scoops

OpenAI, Google, Microsoft, and Anthropic establish the Frontier Model Forum. As part of a partnership to promote responsible AI, this marks a first step in collaboration between some of the leading AI model makers. (Google)

Threat actors using “FraudGPT” LLM for crafting phishing emails, security researchers reveal. Customized LLMs with no restrictions on criminal actions are becoming commonplace. (Hacker News)

An underclass of AI workers in America is powering AI models and suffering from PTSD. Underpaid and often mistreated, this article takes a look at the workers and their “grueling” work in training AI models. (The Atlantic)

Stability AI releases their latest image generation model, Stable Diffusion XL 1.0. It comes with improvements in speed, vibrancy, and text-generation. (TechCrunch)

Molecular “de-extinction” of ancient antibiotics enabled by machine learning. Researchers share how an AI model enabled them to discover extinct antibiotic molecules from Neanderthals. (Science Direct)

AWS unveils HealthScribe, an AI-powered tool for generating doctor notes. Healthcare paperwork is tremendously time-consuming and represents a lucrative new frontier for AI to tackle. (Amazon)

GitHub, Hugging Face, and more call on the EU to relax rules for open-source AI models. Open-source AI development will be constrained by the current draft version of the law, these companies argue. (The Verge)

🧪 Science Experiments

Google DeepMind unveils Robotic Transformer 2

• Robotic Transformer 2 (RT-2) is a novel vision-language-action (VLA) model that learns from both web and robotics data, and translates this knowledge into generalised instructions for robotic control.

• See their project page here for more details.

Diagram of how Google’s RT-2 model creates robotic instructions. (Credit: Google)

Seal-3D: Interactive Pixel-Level Editing for Neural Radiance Field

• With the popularity of implicit neural representations, or neural radiance fields (NeRF), there is a pressing need for editing methods to interact with the implicit 3D models for tasks like post-processing reconstructed scenes and 3D content creation. Seal-3D allows users to edit NeRF models in a pixel-level and free manner with a wide range of NeRF-like backbone and preview the editing effects instantly.

• See their project page here.

Reliably editing models created from NeRFs at the pixel-level is now possible. (Credit: Seal-3D)

Towards Generalist Biomedical AI: Google introduces Med-PaLM M

• Medicine is inherently multimodal, Google Deepmind points out. Med-PaLM M is a large multimodal generative model that flexibly encodes and interprets biomedical data including clinical language, imaging, and genomics with the same set of model weights.

• Paper here

😀 A reader’s commentary

Thanks Dev! We’re here to keep things condensed with a single weekly update.

👋 How I can help

Here’s other ways you can work together with me:

• If you’re an employer looking to hire tech talent, my search firm (Candidate Labs) helps AI companies hire the best out there. We work on roles ranging from ML engineers to sales leaders, and our track record includes helping AI companies like Writer, EvenUp, Tome, Twelve Labs and more make critical hires. Book a call here.

• If you would like to sponsor this newsletter, shoot me an email at michael@artisana.ai 

As always — have a great week!

Today is July 24, 2023.

This week’s issue is a tad longer than previous issues: in the past week, we’ve had a number of important AI news worthy of a deep dive, including the launch of Meta’s Llama 2 language model.

The top questions that I’m thinking about right now are:

• Will open-source beat closed-source AI models? Llama 2’s launch could turn up the heat here as companies adopt its open model.

• Can Google find its groove in AI? It still feels like they’re in the early innings of playing catchup.

• How painful will job disruption from AI be? AI seems like it’s going after creatives and journalists at a rapid pace. I cover several developments here.

• How will Apple’s entrance into generative AI go? The company likes to wait to get things perfect.

In this issue:

• 🧠 Meta releases Llama 2, its first commercially available model

• 🏃‍♂️ Google cofounder Sergey Brin is leading the creation of a GPT-4 competitor

• 📺️ Fable's AI tech generates an entire AI-made South Park episode

• 🍎 Apple has developed "Apple GPT" as it prepares for a major AI push in 2024

• 💾 LLMs are a "threat" to human data creation, researchers warn

• 🔎 OpenAI, Google, and other tech firms agree to voluntary safeguards, and other quick scoops

• 🧪 The latest science experiments, including the latest in face manipulation tech

Did someone forward this to you? Subscribe here!

🧠 Meta releases Llama 2, offering an open and commercially accessible LLM for the first time

Rumored for weeks, Meta finally released their latest LLM, LLaMA 2. It’s available for download and you can play around with it here on Hugging Face as well.

Credit: Meta

Here's what's important to know:

• For the first time, Llama 2 can be used for commercial purposes: apps with more than 700 million monthly users will have to apply for a license, but otherwise there are no other restrictions.

• It’s not truly “open-source”, but it’s close: while Llama 2’s license is relatively open, it falls short of standard open source licenses and contains certain limitations, some critics argue.

• The model was trained on 40% more data than Llama 1, with double the context length: this should offer a much stronger starting foundation for people looking to fine-tune it.

• Early results show Llama 2 outperforms other open-source models across a variety of benchmarks: MMLU, TriviaQA, HumanEval and more were some of the popular benchmarks where Llama 2 beat Llama 1, Falcon and MosaicML's MPT model.

Interestingly, Meta also announced a cozy partnership with Microsoft:

• Microsoft is our preferred partner for Llama 2, Meta announces in their press release, and "starting today, Llama 2 will be available in the Azure AI model catalog, enabling developers using Microsoft Azure."

• MSFT knows the open-source AI approach is going to be big, and they’re not putting all their eggs in one basket despite a massive $10B investment in OpenAI.

The takeaway: the open-source vs. closed-source wars just got really interesting. Google and OpenAI, both of which have closed-source models, are going to face increased pressure from an open-source alternative now.

🏃‍♂️ Google cofounder Sergey Brin leading creation of a GPT-4 competitor

Google's cofounder Sergey Brink, who notably stepped back from day-to-day work in 2019, is actually back in the office again, the Wall Street Journal revealed (note: paywalled article).

The reason? He's leading a push to develop "Gemini," Google's answer to OpenAI's GPT-4 large language model.

Why this matters:

• Concern about falling behind is clearly top of mind: Google was considered a tech and AI pioneer for much of their history, and sources speculate Brin is worried their recent missteps could leave them vulnerable.

• Brin views Generative AI as a pivotal moment of transformation in tech: it's enough to pull him away from other interests and back into day-to-day work.

• Speed is key as Google plays from behind: internal strategy at Google in recent months has focused on moving quickly (perhaps too quickly, some critics argue) and adding AI features to a broad range of products. Brin's involvement is helping catalyze this velocity, sources explained.

While Brin didn't comment on the article, the WSJ revealed he's been quite active in the AI community, and has been attending various AI events. This marks a shift from Brin's earlier beliefs:

• Early on, Brin "expressed skepticism that they could crack artificial intelligence," the WSJ reports, noting that he "ignored the work of the Brain Team" that he originally helped start.But recently his mindset has shifted as AI progress has picked up.

The main takeaway:

• Founders coming back to their companies can often inject a new sense of urgency and mission. The most famous example is probably how Steve Jobs reinvigorated Apple.

• While Google won't say it publicly, it's likely they're treating this moment as an existential crisis. All the internal signals (the "code red" memos, Brin's involvement, the dropping of AI safeguards) are signs that they are reorienting to move quickly here.

How this will play out, though, is ultimately unknown. Google also has the threat of open-source AI to contend with (Google's PaLM 2 remains closed-source).

📺️ Fable's AI tech generates an entire AI-made South Park episode

Fable, a San Francisco startup, just released its SHOW-1 AI tech that is able to write, produce, direct, animate, and even voice entirely new episodes of TV shows.

Fable's first proof of concept? A 20-minute episode of South Park entirely written, produced, and voiced by AI. Watch the episode and see their Github project page here for a tech deep dive.

Credit: Fable

Their tech critically combines several AI models: LLMs for writing, custom diffusion models for image creation, and multi-agent simulation for story progression and characterization.

Why this matters:

• Current generative AI systems like Stable Diffusion and ChatGPT can do short-term tasks, but they fall short of long-form creation and producing high-quality content, especially within an existing IP.

• Hollywood is currently undergoing a writers and actors strike at the same time; part of the fear is that AI will rapidly replace jobs across the TV and movie spectrum.

• The holy grail for studios is to produce AI works that rise up the quality level of existing IP; SHOW-1's tech is a proof of concept that represents an important milestone in getting there.

• Custom content where the viewer gets to determine the parameters represents a potential next-level evolution in entertainment.

In a nutshell: SHOW-1's tech is actually an achievement of combining multiple off-the-shelf frameworks into a single, unified system. This is what's exciting and dangerous about AI right now – the right tools, with just enough tweaking and tuning, can start to produce some very fascinating results.

The main takeaway:

• Actors and writers are right to be worried that AI will be a massively disruptive force in the entertainment industry. We're still in the "science projects" phase of AI in entertainment -- but also remember we're less than one year into the release of ChatGPT and Stable Diffusion.

• A future where entertainment is customized, personalized, and near limitless thanks to generative AI could arrive in the next decade. But as exciting as that sounds, ask yourself: is that a good thing?

🍎 Apple has developed "Apple GPT" as it prepares for a major AI push in 2024

Apple has been relatively quiet on the generative AI front in recent months, which makes them a relative anomaly as Meta, Microsoft, and more all duke it out for the future of AI.

The relative silence doesn't mean Apple hasn't been doing anything, and a Bloomberg report (note: paywalled article) sheds light on their master plan: they're quietly but ambitiously laying the groundwork for some major moves in AI in 2024.

Driving the news:

• Apple is internally testing a chatbot dubbed "Apple GPT" right now. After being caught "flat-footed" by ChatGPT, they're playing catch up.

• The company has also built a framework for creating LLMs, dubbed "Ajax". Ajax is designed to accelerate Apple's ability to move quickly on the generative AI front heading into next year. Their overall plans are not public, but the leak about Ajax is a confirmation their ambition is wide in scope.

Why this matters: trillions of dollars in market cap are at stake.

• While Apple has moved ahead with imbuing their products with AI (maps, search, photos etc.), they're worried about losing the race in generative AI.

• Their cautious approach towards AI and privacy means products like Siri have stagnated, giving up their early mover advantage in the assistant space.

• Apple regards generative AI as a "paramount shift in how devices operate," and see this as an existential threat to the company's ability to sell devices.

• Tim Cook acknowledged he's using ChatGPT in a recently interview. It's something Apple is "looking at closely," he confirmed. But generative AI has a "number of issues that need to be sorted," he noted.

The main takeaway:

• Apple's recent previews of their Vision Pro show that they really want to get something right, in a way that can exceed existing consumer expectations.

• If their release of generative AI tech to consumers doesn't turn out like Apple Maps did (a complete disaster of a launch), things could get very interesting in the LLM space.

• But Apple is under the gun here. The AI space is moving fast, and they don't have years of time to get things perfect.

💾 LLMs are a "threat" to human data creation, researchers warn, as traffic from sites like StackOverflow declines.

LLMs rely on a wide body of human knowledge as training data to produce their outputs. Reddit, StackOverflow, Twitter and more are all known sources widely used in training foundation models.

A team of researchers is documenting an interesting trend: as LLMs like ChatGPT gain in popularity, they are leading to a substantial decrease in content on sites like StackOverflow. Here's the paper on arXiv for those who are interested in reading it in-depth. 

Why this matters:

• High-quality content is suffering displacement, the researchers found. ChatGPT isn't just displaying low-quality answers on StackOverflow.

• The consequence is a world of limited "open data", which can impact how both AI models and people can learn.

• "Widespread adoption of ChatGPT may make it difficult" to train future iterations, especially since data generated by LLMs generally cannot train new LLMs effectively.

This is the "blurry JPEG" problem, the researchers note: as JPEGs are resaved and shared, their quality degrades over time. The same could happen to the internet’s content as AI content increasingly takes over and AI models use AI content for training.

The main takeaway:

• We're in the middle of a highly disruptive time for online content, as sites like Reddit, Twitter, and StackOverflow also realize how valuable their human-generated content is, and increasingly want to put it under lock and key.

• As content on the web increasingly becomes AI generated, the "blurry JPEG" problem will only become more pronounced, especially since AI models cannot reliably differentiate content created by humans from AI-generated works.

🔎 Quick Scoops

OpenAI, Google, and other tech firms agree to voluntary safeguards to mitigate risk of AI. Details remain murky, but the White House announced this set of voluntary commitments as an initial step in the AI regulatory journey. (Forbes)

Is ChatGPT’s performance degrading over time? This Stanford study documents some fairly significant performance shifts between different versions of GPT-3.5 and GPT-4. (arXiv)

G/O Media intends to flood the internet with more AI-generated content. The owner behind Jezebel, AV Club and more talks openly about using more AI tools. (Vox)

Google is pitching an AI for writing news articles. Media orgs who saw it found it "unsettling" as the journalism field faces a rapidly changing landscape. (New York Times - note: paywalled)

Microsoft and OpenAI test synthetic data to train LLMs. Web data is "no longer good enough" to produce more refined models, they’re saying. (Financial Times - note: paywalled)

World of Warcraft fans tricked an AI into writing about a non-existent feature. As AI tools mine Reddit for news content, lack of human editing leaves them vulnerable to misinformation. (BBC)

🧪 Science Experiments

FaceCLIPNeRF: Text-driven 3D Face Manipulation now possible

• Recent advances in Neural Radiance Fields (NeRF) have enabled high-fidelity 3d face reconstruction. These researchers layer in a single text prompt to manipulate a face reconstructed with NeRF, offering a new level of access and ease for face manipulation.

• arXiv paper here.

Credit: arXiv

STEVE-1: A Generative Model for Text-to-Behavior in Minecraft

• Constructing AI models that respond to text instructions is challenging, especially for sequential decision-making tasks. This pretrained model for Minecraft offers a novel pathway to creating advanced behavior from a simple text prompt.

• arXiv paper here.

Credit: arXiv

Stability AI introduces FreeWilly 1 and FreeWilling 2 LLMs

• These are finetuned on Llama 1 and Llama 2, respectively, against Stabilty AI’s Orca dataset

• Download and try it on Hugging Face.Chart showing how FrugalGPT scales performance against cost.

😀 A reader’s commentary

👋 How I can help

Here’s other ways you can work together with me:

• If you’re an employer looking to hire tech talent, my search firm (Candidate Labs) helps AI companies hire the best out there. We work on roles ranging from ML engineers to sales leaders, and we’ve worked with leading AI companies like Writer, EvenUp, Tome, Twelve Labs and more to help make critical hires. Book a call here.

• If you would like to sponsor this newsletter, shoot me an email at michael@artisana.ai 

As always — have a great week!

Today is July 17, 2023.

This week’s issue explores the continuing themes of creative job disruption, government regulation, and the open vs. closed-source battle. One thread that ties these stories together?

The role of AI in history is yet to be written, and we’re seeing major stakeholders all pursue their own agendas to stay relevant and influential in the age of AI. There’s a lot of politics at play and human pain here.

As always, I write my weekly AI memo so you, the busy reader, can rapidly digest this news and come away smarter.

In this issue:

• 🧨 Actors on strike over AI, worried they will be “replaced by machines”

• 🔨 The FTC investigates OpenAI, asks for unprecedented disclosures

• 🤯 Meta’s free commercial LLM is “imminent” and could shake up the LLM world

• 🔎 Anthropic launches Claude 2, plus other news items

• 🧪 The latest science experiments, including some amazing image-to-video tech

Did someone forward this to you? Subscribe here!

🧨 Actors on strike over AI, worried they will be “replaced by machines”

Credit: Netflix (Black Mirror / “Joan is Awful”)

Black Mirror, meet real life.

The ongoing actor's strike is primarily centered around declining pay in the era of streaming, but the second-most important issue is actually the role of AI in moviemaking.

Driving the news: SAG-AFTRA is accusing Hollywood studios of offering background performers just one day's pay to get scanned — then these studios would own that actors’ likeness for eternity with no further consent or compensation. Studios are pushing back and saying this is a gross misrepresentation of their proposal.

Why this matters:

• Overall pay for actors has been declining in the era of streaming: Supporting actors in Orange is the New Black revealed they were paid as little as $27.30 a year in residuals due to how streaming shows compensate actors, while still earning significantly more from their smaller network TV parts. Many interviewed by the New Yorker spoke about how they worked second jobs during their time starring on the show.

• With 160,000 members, many actors are concerned about a living wage: outside of the superstars, the chief concern from working actors is making a living at all -- which is increasingly unviable in today's age.

• Voice actors have already been screwed by AI: numerous voice actors shared earlier this year how they were surprised to discover they had signed away in perpetuity a likeness of their voice for AI duplication without realizing it. Actors are afraid the same will happen to them now.

What are movie studios saying?

• Studios have pushed back, insisting their proposal is "groundbreaking" - but no one has elaborated on why it could actually protect actors.

• Studio execs also clarified that the license is not in perpetuity, but rather for a single movie. But SAG-AFTRA still sees that as a threat to actors' livelihoods, when digital twins can substitute for them across multiple shooting days.

What's SAG-AFTRA saying?

• President Fran Drescher is holding firm: “If we don’t stand tall right now, we are all going to be in trouble, we are all going to be in jeopardy of being replaced by machines.”

The main takeaway: we're in the throes of watching AI disrupt numerous industries, and creatives are really feeling the heat. The double whammy of the AI threat combined with streaming service disrupting earnings is producing extreme pressure on the movie industry. We're in an unprecedented time where both screenwriters and actors are both on strike, and the gulf between studios and these creatives appears very, very wide.

🔨 The FTC investigates OpenAI, asks for unprecedented disclosures

The FTC (Federal Trade Commission) is now investigating OpenAI -- and recently hit them with a 20-page demand letter informing the company it’s now in the FTC’s crosshairs. For the curious readers, here's the full document.

Why this matters:

• The FTC believes existing consumer protection laws apply to AI, even if AI legislation has yet to arrive from Congress.

• In general, the FTC has been aggressive towards tech companies. Linda Khan (the commissioner) has charted a deliberate agenda of going after tech, including trying to block the Activision-Microsoft deal (the FTC lost) and also trying to block Meta from acquiring a VR startup (the FTC lost as well). Losses have not deterred her from continuing an aggressive tone.

• The fines and penalties for FTC violations can be large: Facebook paid $5B in 2019 and Twitter paid $415M in 2022.

So what's the FTC investigating here? Two major angles are part of the FTC’s investigation:

• "Unfair or deceptive privacy or data security practices" –\ this is the stuff that resulted in big fines for Facebook and Twitter.

• “Unfair or deceptive practices relating to risk of harm to consumers, including reputational harm" -- this follows several lawsuits from individuals against OpenAI alleging defamation from the AI's hallucinations, such as making up criminal records.

What must OpenAI do? Over 19 pages of individual demands from the FTC then follow. But the most notable asks are details around how the model was trained, the data used to train it, and disclosure of numerous proprietary details OpenAI has refused to share to date.

The main takeaway: In total, this would represent an unparalleled level of disclosure required from OpenAI. But the bigger risk is whether open-source models and other AI creators with fewer resources would be subject to the same scrutiny -- if so, that would represent a big chilling effect on innovation in the LLM space.

🤯 Meta’s free commercial LLM is “imminent” and could shake up the LLM world

We've previously reported that Meta planned to release a commercially-licensed version of its open-source language model, LLaMA. Now we know the launch is right around the corner, according to a report from the Financial Times (paywalled).

Why this matters:

• OpenAI, Google, and others currently charge for access to their LLMs -- and they're closed-source, which means fine-tuning is not possible.

• Meta will offer commercial license for their open-source LLaMA LLM, which means companies can freely adopt and profit off this AI model for the first time.

• Meta's current LLaMA LLM is already the most popular open-source LLM foundational model in use. Many of the new open-source LLMs you're seeing released use LLaMA as the foundation, and now they can be put into commercial use.

Meta's chief AI scientist Yann LeCun is clearly excited here: "The competitive landscape of AI is going to completely change in the coming months, in the coming weeks maybe, when there will be open source platforms that are actually as good as the ones that are not."

Why could this be game-changing for Meta?

• Open-source enables them to harness the brainpower of an unprecedented developer community. These improvements then drive rapid progress that benefits Meta's own AI development.

• The ability to fine-tune open-source models is affordable and fast. This was one of the biggest worries Google AI engineer Luke Sernau wrote about in his leaked memo re: closed-source models, which can't be tuned with cutting edge techniques like LoRA.

• Dozens of popular open-source LLMs are already developed on top of LLaMA: this opens the floodgates for commercial use as developers have been tinkering with their LLM already.

How are OpenAI and Google responding?

• Google seems pretty intent on the closed-source route. Even though an internal memo from an AI engineer called them out for having "no moat" with their closed-source strategy, executive leadership isn't budging.

• OpenAI is feeling the heat and plans on releasing their own open-source model. Rumors have it this won't be anywhere near GPT-4's power, but it clearly shows they're worried and don't want to lose market share. Meanwhile, Altman is pitching global regulation of AI models as his big policy goal.

The main takeaway: we’re in the early innings of the open-source vs. closed-source debate. Expect Meta’s move to significantly accelerate the intensity of the battle, now that their model can be used commercially.

🔎 Quick Scoops

Anthropic introduces their Claude 2 LLM, with improved performance, longer responses, and better memory. (Anthropic)

Harvard will teach students using an AI instructor next semester, enabling 1:1 student:teacher ratio for their popular CS50 coding course. (Futurism)

Elon Musk announces his AI startup, xAI. Composed of AI veterans, Musk intends to chart his own contrarian course in the world of AI and will leverage Twitter’s immense volume of data. (Twitter)

Stability AI Co-Founder Accuses Company of Tricking Him Into Selling Stake for $100 in Lawsuit. Cyrus Hodes sold his 15% stake for just $100. That stake would be worth over $500M today. (ARTNews)

Brave sells copyrighted data for AI training, engineer reveals. AI projects using third-party sellers like Brave may be able to avoid direct copyright infringement – but that doesn’t make it right. (StackDiary)

🧪 Science Experiments

HyperDreamBooth: the fastest personalized text-to-image model yet

• Google Research team shows that using only a single input image, HyperDreamBooth is able to personalize a text-to-image diffusion model 25x faster than DreamBooth

• Project page here

DreamBooth just got a big upgrade.

Pika Labs releases image-conditioned video generation

• What it does: upload an image with a prompt and it’ll animate the image. For example: “a girl in the wind” modifies an image to have animated, blowing hair

• Lots of early interest in this tech – you’ll have to join their Discord to test it.

Very impressive for an early science experiment.

NIFTY: Neural Object Interaction Fields for Guided Human Motion Synthesis

• Generating realistic 3D motions of humans interacting with objects (e.g. sitting in a chair) is really hard. These researchers (from Google, NVIDIA, and Stanford) propose a “human motion diffusion model” that uses little data to do this. While early, this looks really promising.

• Project page here

How does a human realistically sit in a chair? Quite the hard problem to solve.

CLIPascene: Scene Sketching with Different Types and Levels of Abstraction

• We’ve seen diffusion models able to convert a sketch into a scene. But what about the reverse? This group of researchers demonstrates a novel method to do exactly that.

AI can now do the reverse of sketch-to-scene — quite cool!

😀 A reader’s commentary

👋 How I can help

We crossed 15k subs! Our audience continues to grow.

Here’s other ways you can work together with me:

• If you’re an employer looking to hire tech talent, my search firm (Candidate Labs) helps AI companies hire the best out there. We work on roles ranging from ML engineers to sales leaders, and we’ve worked with leading AI companies like Writer, Tome, Twelve Labs and more to help them make critical hires. Book a call here.

• If you would like to sponsor this newsletter, shoot me an email at michael@artisana.ai 

As always — have a great week!

Today is July 10, 2023.

Welcome to another week of fascinating AI news. Top of mind for me: will OpenAI’s prediction that superintelligence arrives “this decade” actually come true? If so — we’re nowhere near ready as a society, and OpenAI seems aware of that too.

Also: I’ve had so many people inquire about other ways to work together in the past few weeks that I’ve included a new section below. Thanks to everyone who’s emailed in with questions and feedback!

In this issue:

• 🧨 OpenAI says "superintelligence" will arrive "this decade," so they're creating the Superalignment team

• 🪖 US military now trialing 5 LLMs trained on classified data

• 💔 ChatGPT is losing users for the first time

• 🔎 Sarah Silverman sues ChatGPT, and other news scoops

• 🧪 The latest science experiments, including a 1B token Transformer architecture

Did someone forward this to you? Subscribe here!

🧨 OpenAI says "superintelligence" will arrive "this decade," so they're creating the Superalignment team

OpenAI says superintelligence (which is more capable than AGI, in their view) could arrive "this decade," and it could be "very dangerous." As a result, they're forming a new Superalignment team led by two of their most senior researchers and dedicating 20% of their compute resources to this effort.

Why this matters:

• "Superintelligence will be the most impactful technology humanity has ever invented," but human society currently doesn't have solutions for steering or controlling superintelligent AI, says OpenAI.

• Current alignment techniques don't scale to superintelligence because humans can't reliably supervise AI systems smarter than them.

How can superintelligence alignment be solved? OpenAI has an answer:

• An automated alignment researcher (an AI bot) is the solution. This means an AI system is helping align AI: in OpenAI's view, the scalability here enables robust oversight and automated identification and solving of problematic behavior.

OpenAI wants to solve this in the next four years, given they anticipate superintelligence could arrive "this decade." And they seem generally optimistic: "superintelligence alignment is fundamentally a machine learning problem, and we think great machine learning experts—even if they’re not already working on alignment—will be critical to solving it."

🪖 US military now trialing 5 LLMs trained on classified data, wants AI to empower military planning

The US military has always been interested in AI, and now they’ve revealed they’re moving quickly on testing generative AI technologies. Bloomberg reports (note: paywalled article) that the US military is currently trialing 5 separate LLMs, all trained on classified military data, in an exercise lasting most of July.

Photographer: Staff Sgt. Renee Seruntine/US Army National Guard

Why this matters:

• Long-term, the US wants AI to empower military planning, sensor analysis, and firepower decisions. So think of this as just a first step in their broader goals for AI over the next decade.

• There's a tremendous amount of proprietary data for LLMs to digest: information retrieval and analysis is a huge challenge for militaries -- going from boolean searching to natural language queries is already a huge step up.

What are they testing? Details are scarce, but here's what we do know:

• ScaleAI's Donovan platform is one of them. Donovan is defense-focused AI platform and ScaleAI divulged in May that the XVIII Airborne Corps would trial their LLM.

• The four other LLMs are unknown, but expect all the typical players, including OpenAI. Microsoft has a $10B Azure contract with DoD already in place.

• LLMs are evaluated for military response planning in this trial phase: they'll be asked to help plan a military response for escalating global crisis that starts small and then shifts into the Indo-Pacific region.

Early results show military plans can be completed in "10 minutes" for something that would take hours to days, a colonel has revealed.

Cue the countdown clock til SkyNet arrives.

💔 ChatGPT is losing users for the first time. Here’s what could be happening.

The Washington Post (note: paywalled article) reports that traffic data from SimilarWeb is showing the first traffic decrease for ChatGPT since its launch last November. We heavily disagree with their headline that this is “shaking faith in AI revolution.”

Credit: SimilarWeb

Remember: ChatGPT took just two months to get to 100 million monthly users (a record now smashed by Meta’s Threads app), making it the fastest growing consumer app ever at time of launch.

Let’s unpack what could be going on:

• Could interest in chatbots be waning? Our honeymoon period with their capabilities could be over as dozens of popular chatbots proliferate and we become more attuned to their weaknesses and limitations.

• At the same time, users are complaining that ChatGPT’s capabilities are getting worse. This support thread on OpenAI is one of the many examples of users sharing increased frustration that ChatGPT no longer generates quality outputs. Reddit is also filled with similar complaints.

Or, the reasons behind the decline could be innocuous as well:

• The school year is over, and usage of ChatGPT could have seen a seasonable decrease due to reduced student usage.

• And it feels like every tool these days now has generative AI integrated in it – so going directly to ChatGPT’s site is simply less needed.

🔎 Quick Scoops

All ChatGPT Plus users now have access to OpenAI’s Code Interpreter, giving everyone a data analyst copilot. (VentureBeat)

A $3B class action lawsuit was filed against OpenAI and Microsoft, alleging OpenAI secretly “scraped 300 billion words from the internet” without registering as a data broker or obtaining consent. (Vice)

Sarah Silverman sues ChatGPT and Meta, alleging copyright infringement. Her works were used as part of copyright-infringing data sets that the LLMs digested, she says. (The Verge)

Google makes a big push into medical AI. There’s big money here – but many experts think we’re not ready yet. (Wall Street Journal - note: paywalled article)

GPT-4 is actually a mix of 8 different 220-billion-parameter models, leaks reveal. Previous rumors had suggested it was a 1T parameter model, but those are now debunked. (Algorithmic Bridge - note: paywalled article)

🧪 Science Experiments

Hugging Face launches text-to-video resource page

• As open-source models proliferate, Hugging Face has assembled 33 leading open-source models and 7 datasets, along with instructions, that make it easier for anyone curious about text-to-video models to get started. Check it out here.

LongNet architecture: scaling Transformers to 1,000,000,000 tokens

• Scaling sequence length is critically important for language models, but performance issues typically arise as model size grows. This team proposes a new Transformer variant that doesn’t sacrifice performance on shorter sequences while enabling (theoretically) up to 1 billion tokens.

• arXiv page here

As open-source LLMs proliferate, it’s easy to spread a “poisoned” LLM online

• This article from Mithril Security highlights how they hid a modified version of a popular open-source LLM on HuggingFace. This “poisoned” LLM was surginally modified to encode false information. Read more here.

😀 A reader’s commentary

👋 How I can help

As this newsletter nears 15,000 subscribers (!), I’ve had the amazing chance to hear from many of you on what you want beyond just a weekly roundup.

Here’s other ways you can work together with me:

• If you’re an employer looking to hire tech talent, my search firm (Candidate Labs) helps AI companies hire the best out there. We work on roles ranging from ML engineers to sales leaders, and we’ve worked with leading AI companies like Writer, Tome, Twelve Labs and more to help them make critical hires. Book a call here.

• If you would like to sponsor this newsletter, shoot me an email at michael@artisana.ai — we’ll be opening up the first set of slots soon.

Today is July 3, 2023.

This week’s issue explores a number of AI topics now exploding into the mainstream: white-collar job disruption, the benefits of open-source models, and further criticism of the EU’s AI Act.

I also feature a fascinating read from SemiAnalysis that shows how Nvidia’s AI moat could quickly disappear; this is once again a reminder of how fast the AI space moves even in the early innings.

In this issue:

• 🧑‍💻 AI tools could change the programming profession in profound ways

• 🤬 Uncensored AI chatbots provoke intense debates over free speech

• 💣️ EU companies criticize the EU’s AI Act as “catastrophic” in an open letter

• 🖥️ Nvidia’s AI moat could be threatened as AMD chips deliver better performance

• 🧪 The latest science experiments, including a new image to 3D mesh tech that’s better than ever

Did someone forward this to you? Subscribe here!

🧑‍💻 AI tools could change the programming profession in profound ways

“Software is eating the software industry” as skilled workers are becoming more productive and AI starts automating knowledge work, the Wall Street Journal reports (note: paywalled article) in its deep dive on AI’s impact on coders.

Why this matters: while knowledge workers in the past have typically benefited from technology improvements, AI has the potential to upend this trend. A “lost generation of early-career developers” could be the result here as hiring practices shift quickly.

It’s already happening:

• Since the release of ChatGPT and GitHub’s Copilot last year, experienced programmers have become more productive while companies have slowed down the hiring of junior engineers.

• Adoption of AI tooling is rapid, with 70% of programmers using or planning to use AI tools, a StackOverflow survey found.

• Employment data now shows that junior engineers are the first to experience layoffs as well in the last year.

Experts expect this same force to come for other white-collar jobs, and the rapid pace of AI may catch a number of professions by surprise in a “cautionary tale for us all,” writes the Journal.

I get asked a lot about what I think about AI and the future of jobs. This is yet another data point to support my own belief: no one really knows, but we should expect our old assumptions to not hold.

🤬 Uncensored AI chatbots provoke intense debates over free speech

We’ve covered the open-source community and how dozens of language models, many unrestricted, have been released in the past few months. The debate over free speech and safety protocols has only gotten more heated in the last few weeks, reports the New York Times (note: paywalled article).

At the heart of the debate:

• “This is about ownership and control,” explains Eric Hartford, an ex-Microsoft engineer who created the uncensored WizardLM language model.

• What Hartford and other open-source proponents want: “If I ask my model a question, I want an answer, I do not want it arguing with me.”

Some open-source models are taking a middle ground, but it isn’t easy:

• Open Assistant, based on Meta’s LLaMA and developed with help from 13,500 volunteers, is experiencing deep community debate about how its safety systems should work.

• One faction questions whether there should be safety systems in place at all.

Could this lead to a world where everyone has their own personalized chatbot? 

• Possibly, and that’s the ideal outcome that many open-source LLM creators envision.

• “Democrats deserve their model. Republicans deserve their model. Christians deserve their model. Muslims deserve their model,” Hartford explained in a blog post on uncensored models. “Every demographic and interest group deserves their model. Open source is about letting people choose.”

What’s clear is that open-source proponents aren’t intimidated by legislation such as the EU’s AI Act. They’re full-steam ahead as they pursue their vision.

💣️ EU companies criticize the EU’s AI Act as “catastrophic” in an open letter

The EU Parliament greenlit a draft of the AI Act last month, putting it on the path to become law in the coming years. Now a group of 150 EU executives is sounding the alarm over the AI Act, claiming that it may “jeopardise Europe’s competitiveness and technology sovereignty.”

These are some major companies voicing criticism:

• Renault, Heineken, Airbus, and Siemens are just some of the major EU companies sharing concerns.

• “The EU AI Act, in its current form, has catastrophic implications for European competitiveness,” says one of its signatories.

They argue that the law is too restrictive:

• AI innovation could be stifled due to strict rules targeting generative AI systems, the signers say.

• Registration requirements, risk assessments, and transparency disclosures would subject AI systems to “disproportionate” compliance costs and liability risks.

• The worst case outcome? AI companies withdraw from Europe rather than meet compliance requirements.

Go deeper: A Stanford research team recently flagged that most AI models would fail current compliance requirements in the EU’s AI Act.

The EU Parliament doesn’t seem swayed, though:

• “It is a pity that the aggressive lobby of a few are capturing other serious companies,” said Dragoș Tudorache, a Member of the European Parliament who led the development of the AI Act.

• Tudorache stressed that the process behind the legislation was collaborative and “industry-led.”

We’re still several years from the AI Act becoming law. Expect debates here to continue, especially as other countries like the US begin to shape their own AI legislation.

🖥️ Nvidia’s AI moat could be threatened as AMD chips deliver better performance

Nvidia’s hardware to date has been the preferred approach to develop and train AI models, helping catapult the company’s market capitalization past $1 trillion in recent weeks.

The moat is now under threat, says SemiAnalysis:

• Open-source software from MosaicML (which Databricks just acquired for $1.3B) is leveling the playing field, helping AMD M1250 GPU achieve nearly 80% of the performance of Nvidia’s A100 GPU on AI-related tasks.

• And Mosaic hasn’t even tested their software on AMD’s new M1300 yet, which could spark further gains.

This is a good thing overall: it’s long been the dream of AI engineers to have a hardware-agnostic tech stack, where their code can work across multiple GPU platforms and not require GPU-level programming. That future may be arriving earlier than expected.

AMD has long played catchup to Big Green in the GPU world. It’s fascinating that software improvements, released as open-source libraries by other companies, could be the accelerant here for them.

🔎 Quick Scoops

Databricks will buy MosaicML for $1.3B, marking one of the largest acquisitions in the generative AI space in a sign of how hot the AI market is right now. (TechCrunch)

Valve’s Steam gaming platform is rejecting games with AI-generated art, reports one indie developer. (Reddit)

AI-generated music is provoking deep debates in the music industry, writes the Rolling Stones in this fascinating deep-dive. (Rolling Stones / paywalled)

Microsoft, OpenAI sued for privacy violations by 16 people, alleging that the businesses committed “theft” of personal information. (The Register)

Meta offers more transparency on how AI influences content feeds, another step in their “openness” philosophy regarding AI. (The Verge)

🧪 Science Experiments

High-quality 3d object generation now possible from a single image

• Magic123 enables textured 3D mesh generation from a single unposed image. The quality of its outputs versus prior approaches is significantly better.

• See it in action here, including comparisons to other approaches.

Lifelike animated motion now possible from AI models trained on synthetic data

• The ability for AI models to generate 3D human poses from real images has been poor to date.

• BEDLAM is the first synthetic dataset of bodies exhibiting lifelike animated motion, all generated from an AI model.

• See it in action here

Adding a “backspace” ability to LLMs can help avoid compounding errors

• A team from Stanford shows that incorporating a “backtracking” ability into LLM output generation helps avoid compounding errors and poor results.

• What’s more, their approach can be added into existing LLMs without major architectural changes.

• Full paper here

😀 A reader’s commentary

The most valuable thing for me to hear is that I’m offering something differentiated and useful (that’s why I wrote this in the first place; I wasn’t happy with the quality of AI coverage). Thanks George!

That’s it! Have a great week — and for folks in the USA, enjoy your July 4th!

And as always, we want to hear from you on how useful this is. The more feedback you provide, the better our newsletter gets! What else would you like to see covered?

So take the poll below — your signals are helpful!

Today is June 26, 2023.

If there’s a theme to the major news events that stood out to me over the past week, it’s that we’re in the early innings of grappling with AI.

As legislation in the EU races ahead, AI models are trying to understand how they fit in. At the same time, everyone from artists to politicians are seeing AI pervade their work without fully understanding and grasping its long-term implications.

Lots of fascinating topics to watch in the months and years ahead!

In this issue:

• 📣 Leading AI Language Models Fall Short of Upcoming EU Regulations, Stanford Study Warns

• 🛠️ Millions of task workers now work in AI annotation, often with no idea why

• 📺️ AI-generated political content spreads, highlighting gaps in election laws

• 🎨 Marvel’s “Secret Invasion” AI-generated intro angers artists

• 🧪 The latest science experiments, including new video generation AI models and Google’s new speaking LLM

Did someone forward this to you? Subscribe here!

📣 Leading AI Language Models Fall Short of Upcoming EU Regulations, Stanford Study Warns

A Stanford study concludes that the world's leading AI language models could fail to meet the standards set in the EU's new AI Act, facing significant regulatory risks and potential heavy fines for non-compliance.

Leading AI models largely fail to comply across 12 key dimensions, researchers found. Credit: Stanford University

Why this matters:

• The EU AI Act is on its way to becoming law: it's now in its final stages after passage through parliament, so there's no way to head off its arrival. Any final changes will be small tweaks.

• Penalties for non-compliance are serious: fines of the greater of €20,000,000 or 4% of worldwide revenue are possible.

• Open-source models face the same standards as closed-source models: this includes registration with the EU, transparency requirements, and safety considerations.

• Other countries will use it as an example: as legislation gets developed in the USA, it's likely they'll look to the EU for inspiration.

What did the researchers find?

• Across 12 key requirements for generative AI, the leading 10 models fell short. Most scored just 50% of the total possible 48 points.

• Hugging Face's open-source BLOOM performed the best, securing 36/48 points.

• OpenAI's GPT-4 scored 25/48 points, roughly middle of the pack.

• Anthropic's Claude scored 7/48 points, just second from the bottom.

Areas of failure were different between closed-source and open-source models:

• Open-source models generally outperformed in data sources transparency and resource utilization disclosure. Due to their generally transparent releases, this is not surprising.

• Closed-source models excelled in areas such as comprehensive documentation and risk mitigation.

What are the issues to watch next here?

• Many elements of the AI Act remain murky, the researchers argue, so additional clarity is needed. Look out for tweaks to the law as it goes through additional refinement.

• How open-source and closed-source projects adapt in the next few months will be interesting to observe. OpenAI in particular will have be more open. And open-source projects may have to wrestle with better understanding registration requirements and post-deployment model risks.

Our full breakdown is here.

🛠️ Millions of task workers now work in AI annotation, often with no idea why

This deepdive report from the Verge covers an emerging underclass of society: the millions of task workers, often in developing countries, paid pennies to annotate images, voice files, and text for training AI systems.

• In a highly ironic twist, many have no idea they’re working for AI systems, and several workers interviewed were surprised to learn they worked for subsidiaries of tech companies like Scale AI.

• This “supply chain” of human labor is deliberately obfuscated, often for confidentiality concerns, the Verge found. Remotasks, one of the companies in Africa identified by the Verge, lists no connection to Scale AI despite its subsidiary status.

The task workers themselves have mixed feelings about their labor:

“I read and I Googled and found I am working for a 25-year-old billionaire,” said one worker, who was labeling the emotions of people calling to order Domino’s pizza. “I really am wasting my life here if I made somebody a billionaire and I’m earning a couple of bucks a week.”

Another worker, who discovered they had likely been helping train ChatGPT, summed up the opinion of his peers: “People were angry that these companies are so profitable but paying so poorly.”

📺️ AI-generated political content spreads, highlighting gaps in election laws

From local elections to national races, New York Times highlights a number of political races that are seeing rise of AI-generated content.

• One Canadian mayoral candidate generated multiple images of a crime-ridden Toronto in order to emphasize his tough-on-crime message, the Times found. None were labeled as AI-generated.

• Errors in AI-generated images aren’t dissuading their use, the Times found. The same candidate also distributed materials of a woman with three arms (a common AI image error), and nonetheless gained traction in the mayoral race.

• And a Twitch livestream debate of AI Trump vs. AI Biden shows how eerily AI can imitate major political figures, increasingly raising questions within our broader society of how to separate real from AI content.

🎨 Marvel’s “Secret Invasion” AI-generated intro angers artists

Marvel broke new ground last week as the first major television show to use an AI-generated opening sequence. Its Secret Invasion TV series, in which an alien shape-shifting species infiltrates Earth, features an abstract, mind-melting intro that fans quickly determined was AI-generated.

Credit: Marvel Studios

After confirming that visual effects group Method Studio used AI to generate the intro, Marvel has been taking flak from visual effects artists.

Jeff Simpson, a concept artist who directly worked on the Secret Invasion TV series for Marvel, tweeted: "Secret Invasion intro is AI generated. I’m devastated, I believe AI to be unethical, dangerous, and designed solely to eliminate artists' careers.”

AI tools in the visual effects space comes at a time when Hollywood writers are striking over lower pay and AI concerns, and visual effects artists are similarly complaining that large studios like Marvel have been the driving force behind lower pay, longer hours, and nearly unprofitable operating margins for many of the smaller studios.

(Editor’s note: Marvel doesn’t make this intro available publicly themselves, but you can probably find it via a Youtube search)

🔎 Quick Scoops

OpenAI successfully lobbied the EU to water down the AI Act, leading to chatbot platforms avoiding a “high risk” label. (Time)

US Senate previews their AI legislation strategy. While the US is still playing catchup, leading politicians recognize that speed is critical. (NBC News)

AI-generated child sexual abuse images are on the rise, and law enforcement agencies are ill-equipped to stop this. What’s worse, automated detection systems based on recognizing real images could now face an overwhelming challenge in synthetic media. (The Washington Post)

Voice-generating AI platform ElevenLabs raises $19M, and launches an AI voice detection tool (TechCrunch)

Dropbox launches universal AI search, amongst many other AI tools. Expect natural language-powered search to be increasingly prevalent as the new UI paradigm in many, many platforms to come. (The Verge)

🧪 Science Experiments

DragGAN is now available directly on Hugging Face

• This demo made waves when it first released weeks ago, and is a proof-of-concept of how generative AI can unleash a new set of powers in image manipulation

• Now you can try it here without installing anything else

Source: Hugging Face

Midjourney 5.2 releases, includes new “Zoom Out” feature

• Artists are already getting creative with it – see an example here

Source: Reddit

zeroscope_v2 XL video generation model releases

• This is a watermark-free video model, now capable of generating high quality video at up to 1024 x 576

• See it in action here

Source: Zeroscope team / Hugging Face

Google presents AudioPaLM, an LLM that can speak and listen

• AudioPaLM fuses text-based and speech-based language models (PaLM-2 and AudioLM) into a unified multimodal architecture

• See it in action here

😀 A reader’s commentary

Thanks Bradley!

That’s it! Have a great week!

And as always, we want to hear from you on how useful this is. The more feedback you provide, the better our newsletter gets! What else would you like to see covered?

So take the poll below — your signals are helpful!

Happy Monday! Today is June 19, 2023.

The biggest news last week for me was Meta’s decision to launch their next open-source LLM with a commercial license, which could turn into a massive threat against the close-source ecosystems OpenAI and Google are trying to build.

If and when that model comes out, expect the battle of open vs. closed-source to only grow even more intense.

In this issue:

• 🧨 Meta’s next open-source LLM will have commercial license, putting pressure on Google and OpenAI

• 🎸 Paul McCartney will release a “final Beatles song” using AI

• 🗑️ AI-generated junk has taken over Etsy

• 💰️ AI and media titans hash out the future of content licensing

• 🔎 Google tells their own employees to not put code into chatbots

• 🧪 The latest science experiments, including Google’s new skin condition detector and an avatar generator that uses pics + text

Did someone forward this to you? Subscribe here!

🧨 Meta’s next open-source LLM will have commercial license, putting pressure on Google and OpenAI

Meta’s Bay Area headquarters. Credit: The Verge

Last week, news broke that Meta intends to make its next set of open-source LLMs available for commercial use.

Why this matters:

• Right now, Meta’s LLaMA LLM is only available for research use, preventing companies from using it as part of their commercial efforts.

• This will likely tap into massive demand: Meta’s open-source AI tech is already massively popular for researchers, who have found numerous ways to fine-tune and improve it.

• OpenAI feels the heat and may release their own open-source model. Rumors say this won’t approach GPT-4’s power, but it would represent a sharp reversal from their closed-source approach of late.

Despite questions from the US Senate about the dangers of open-source AI, in an interview last week, Meta's Chief AI scientist Yan LeCun dismissed any worries about AI posing dangers to humanity as "preposterously ridiculous."

Read our in-depth breakdown here.

🎸 Paul McCartney will release a “final Beatles song” using AI

AI is the key unlock to releasing a “final Beatles song,” announced McCartney last week. Vocals from John Lennon were passed through an AI model and made “pure,” enabling the track to be assembled and mixed. Lennon passed away in 1980.

This raises the question: at what point does AI make it no longer a “real” Beatles song? We’ll have to see once the song releases, as McCartney shared few details on the track itself in his interview.

The use of AI to make songs is currently experiencing a watershed moment:

• An AI-generated song featuring Drake’s voice went viral and was banned from Spotify and other streaming services at the request of recording labels.

• Tens of thousands of other tracks imitating famous artists continue to proliferate on social media in the meantime.

• Meanwhile, some artists like Grimes are openly embracing AI music, calling for fans to use her voice in AI compositions and simply pay royalties.

🗑️ AI-generated junk has taken over Etsy

AI has made it easier than ever to generate artwork, but that’s also led to a proliferation of AI-generated products on Etsy that is pushing out real artists, as the Atlantic reports in a fascinating deep dive (note: article is paywalled).

This is no different than text-based content on the web, which is now suffering from a massive increase in bland and generic AI content as content writers lose out on work and companies switch over to AI platforms.

In Etsy’s case, the problem really stems from two key drivers:

• Etsy doesn’t forbid AI-generated content so long as “creativity” is involved. This very loose interpretation makes it possible for sellers to exist.

• “Hustle culture” Youtubers are pitching side hustles that are low-effort, maximum reward. These range from selling Midjourney artwork digitally on Etsy to putting thousands of variations of art on t-shirts and more.

I’ve heard from a lot of readers that one of your biggest fears is an age where it’s no longer easy to know if content was generated by human or machine. When I read articles like this, it feels that future is arriving quite quickly.

💰️ AI and media titans hash out the future of content licensing

As AI has exploded into our lives, the value of the content used to train these AI models has become very clear. AI giants including OpenAI, Google and Microsoft may now pay media companies a fee of up to $20M per year in order to continue using their content for AI training, early reports indicate.

AI companies are already facing pressure from multiple angles:

• Legislation like the EU's AI Act will require disclosure of copyrighted training data, and other countries are likely to follow. It will become harder and harder to hide the data they use to train models.

• Companies like Reddit + StackOverflow have announced pricing tiers for their APIs, as they seek to prevent AI companies from simply sucking up their data for free.

• Lawsuits alleging copyright violation are now targeting generative AI companies. While many seem ill-advised and frivolous, it still is a thorn in the side of AI tech firms.

Google, with its prior expertise in working with media companies (though not always successfully) is leading the way here on suggesting payment frameworks, and for once it appears all sides feel good about the likely outcome.

But what does this mean for the non-tech giants? Let’s say you wanted to train your own AI model – could you end up in a world where every piece of valuable content is locked down and requires licensing?

We’ll be watching this very closely especially since open-source AI remains so promising and cost-effective compared to the closed-source approach.

🔎 Quick Scoops

Google warned its own employees not to enter confidential information into chatbots, including its own Bard. This restriction includes entering computer code. (Reuters)

Researchers worry AI models may “collapse” as they train on AI-generated content. As AI content profilerates, this doom loop could cause “irreversible defects”, researchers warn. (VentureBeat)

Mercedes Benz introduces ChatGPT for in-car voice control, marking the start of a trend where language models may increasingly power a new generation of interfaces. (Mercedes)

France’s MistralAI raises $113M to take on OpenAI. Founded by alums from DeepMind and Meta, the company intends to release its first generative AI models in 2024. (TechCrunch)

AMD jumps into the AI GPU race with a new class of GPUs. AI helped make Nvidia a trillion-dollar company, and AMD doesn’t want to be left out. (Anandtech)

🧪 Science Experiments

Google Lens can now identify skin conditions via uploaded photos

• While Google cautions this shouldn’t replace your doctor, it’s a new adaptation of their powerful Lens image search technology for a novel use case.

• Read their full blog here.

This isn’t going to replace your doctor, but it sure could help start the conversation.

AvatarBooth generates 3D human avatars from both text and images

• Previous methods could only generate avatars from text descriptions, while this one is able to consume images and combine them with text prompts.

• The quality is reminiscent of early-N64 games, but the concept is nonetheless fascinating.

• Full paper here. GitHub here.

Obama, Hillary, and Kobe — all generated with off an image plus text description.

13B parameter OpenLLaMA LLM released to public

• This is an open-source reproduction of Meta’s LLaMA large language model, which means it isn’t limited to research-only purposes. Until Meta allows their LLMs to be used for commercial purposes, this may be the next-best alternative.

• See it on Hugging Face here.

Researchers use new method to fine-tune a 65B parameter model on 8x RTX 3090 GPUs

• A new optimizer combined with other efficiency techniques means fine-tuning can consume just 11% of the memory bandwidth of traditional techniques.

• As a result, a 65 billion parameter model was fine-tuned on just 8 RTX 3090s

• Full paper here.

😀 A reader’s commentary

Here’s what Christopher said about last week’s email. Glad we’re keeping folks informed!

That’s it! Have a great week!

And as always, we want to hear from you on how useful this is. The more feedback you provide, the better our newsletter gets! What else would you like to see covered?

So take the poll below — your signals are helpful!

Today is June 12, 2023.

Happy Monday! Last week brought a number of interesting developments and new science projects (like Meta’s new open source music generator) to try. It also saw a lot of noise (articles and news that I purposely filter out of this newsletter).

May have you have a fruitful and productive week!

In this issue:

• 🧠 Bard improves at math, coding, and more through new upgrades

• 🛑 OpenAI still not training GPT-5, prioritizing other ideas instead

• 🔨 Meta criticized by US senators for “leaking” open-source LLaMA LLM

• 🏃‍♂️ Google’s DeepMind discovers faster sorting algorithms using deep reinforcement learning

• 🧪 The latest science experiments, including Meta’s new music generation engine

Did someone forward this to you? Subscribe here!

🧠 Bard improves at math, coding, and more through new upgrades

One of the known weaknesses of language models is their poor math capabilities, owing to inherent limitations in the transformer architecture that lead them to function like prediction engines.

• This week, Google announced their solution to this weakness by enabling Bard to perform implicit code execution in the background.

• For queries that Bard believes will benefit from logical reasoning, a second system kicks in that improves the accuracy of Bard’s responses to computation-based problems.

• Google’s team cautions this still isn’t perfect, however – the expected performance gain they witnessed was about 30% over the baseline. Read the full blog post from Google here.

🛑 OpenAI still not training GPT-5, prioritizing other ideas instead

One of the constant questions OpenAI CEO Sam Altman encounters is “when is GPT-5 coming out?” As he toured several countries last week to promote OpenAI’s agenda, this question came up again.

He was happy to reinforce some points and offer more context:

• Months after explaining that the company wasn’t working on GPT-5, Altman confirmed this again: “we have a lot of work to do before we start that model.”

• Altman set further expectations on any hypothetical release timeline, explaining that it took six months to release GPT-4 even after OpenAI finished training it. Read more here.

🔨 Meta criticized by US senators for “leaking” open-source LLaMA LLM

A bipartisan group of US senators from the same committee that questioned OpenAI CEO Sam Altman is now going after Meta.

Meta’s LLaMA LLM notably leaked to the public in February and have been central to the red-hot pace of open-source LLM improvements in recent months.

• This is clearly worrying the US government: “The open dissemination of LLaMA represents a significant increase in the sophistication of the AI models available to the general public, and raises serious questions about the potential for misuse or abuse,” the senators wrote to CEO Mark Zuckerberg.

• Meta “failed to conduct any meaningful risk assessment,” the letter says, noting how the company should have anticipated broad dissemination of the model after it was released to a smaller group at first.

🏃‍♂️ Google’s DeepMind discovers faster sorting algorithms using deep reinforcement learning

Google DeepMind's AlphaDev AI has improved several sorting algorithms in C++, including optimizing one algorithm for a 70% speed increase over the previous best method.

• The recent achievement came about by adapting the AlphaZero AI, which famously mastered complex games like chess and Go, into a code-focused version dubbed AlphaDev.

• As computer chips approach fundamental physical limits due to their nanoscale transistors, the need for better software efficiency and optimization becomes increasingly paramount.

Read the full report from the Google, published in Nature.

🔎 Quick Scoops

Meta plans to add generative AI across numerous platforms, as the company details its AI strategy in its clearest form yet. (Axios)

OpenAI was sued for defamation by a radio host, who is angry that the chatbot falsely claims he embezzled funds and defrauded a nonprofit. (Ars Technica)

Chief Information Officers worry using AI to write code will create new problems, among them growing technical debt and orphan code. (Wall Street Journal - 💰️ Paywalled)

Microsoft’s Bing chatbot gets in trouble for serving up canned AI answers for “Chrome,” in a sign of a growth back gone wrong. After journalists pointed this out, the company reversed course. (The Verge)

DeSantis campaign is running ads with AI-generated imagery, mixing in real and fake images in a way that obscures the actual AI-generated image of some content. (Twitter)

Palantir wins $400M contract with US Special Operations to integrate LLMs and AI, as part of reducing cognitive load for the military and its warfighters. (MarketWatch)

🧪 Science Experiments

Google adds vision into speech recognition algorithm, vastly improving robustness

• Existing technology for deciphering audio suffers when the inputs are noisy

• By injecting visual streams in as well, prediction of audio is vastly improved

• See their research here

  

  Predictions of what people are saying improve when you interpret the visuals too! Credit: Google

LLMs integrated with Notebooks could add new superpowers, Steven Wolfram shares

• Notebooks as a concept have been around for decades

• Bye integrating LLM functionality into notebooks, new capabilities are unlocked

• See Stephen Wolfram’s full blog post here

Meta releases their open-source music generation model

• Full paper is here

• Try it here

Nested Diffusion provides a way to progressively see image results via Stable Diffusion 1.5

• Think of these like progressive JPEGs – you can stop it anytime

• Try it here

• Full paper here

Witness how Stable Diffusion increasingly generates a more detailed image via Nested Diffusion.

😀 A reader’s commentary

That’s it! Have a great week!

And as always, we want to hear from you on how useful this is. The more feedback you provide, the better our newsletter gets! What else would you like to see covered?

So take the poll below — your signals are helpful!

Today is June 5, 2023. We’re back after taking a one-week break for Memorial Day!

As I debated last week whether I should send an issue anyways, a friend helped put things in perspective: we’re only in the early innings of generative AI, and staying informed on this topic also means not getting burned out.

Just as some stories in regulation, AI safety, and the future of work are beginning to pick up steam, I think it’s important to consume just the right amount of news so you don’t end up tuning out the most important things.

So that’s why I continue to write this — my goal is to share my own distillation of the most impactful chunks of news into something that brings value to you as well — while also keeping the cadence and volume just right.

In this issue:

• 💀 AI leaders warn of “risk of extinction”

• 📊 A majority of US adults are familiar with ChatGPT, but usefulness is mixed

• 🪖 A US military AI drone “killed” its operator in a simulation… or did it?

• 💰️ Generative AI spend set to hit $1.3T by 2032, says Bloomberg

• 🧪 The latest science experiments, including Nvidia’s 2D video to 3D model tech

Did someone forward this to you? Subscribe here!

💀 AI leaders warn of “risk of extinction”

Hundreds of notable AI industry leaders and research scientists signed on to a 22-word statement saying the following: “Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war.”

• What’s notable with this letter is the breadth of signers, which include: Sam Altman, CEO of OpenAI; Demis Hassabis, CEO of Google’s Deepmind unit; Dario Amodei, CEO of Anthropic; and Emad Mostaque, CEO of Stability AI.

• A number of AI scientists also signed, most notably Geoffrey Hinton and Yoshua Bengio, two AI researchers who won the Turing award for their work on neural networks.

Why does this matter?

• This is the broadest group to sound the alarm on the need for humanity to prioritize and cooperate on AI’s future, making this notably different from previous warning letters.

• In the last few weeks, several notable pushes for governance and regulation have emerged, including OpenAI’s own call for the governance of superintelligence via a global organization.

The challenge is now taking global action on AI in a coordinated and thoughtful way, and also accounting for the rise of powerful open-source AI that could make it difficult to regulate AI models. We’ll be watching all of this closely as it develops.

Read our full breakdown here.

📊 A majority of US adults are familiar with ChatGPT, but usefulness is mixed

ChatGPT seems like it’s everywhere, but is that just because of our immediate surroundings? A new poll from the Pew Research Center sheds light on how Americans more broadly are interacting with ChatGPT.

The most interesting nuggets from the poll:

• 58% of US adults are familiar with ChatGPT, but just 14% of US adults have tried ChatGPT.

• Those with higher household incomes and more formal education were more likely to have heard about ChatGPT.

• Of the adults who’ve tried ChatGPT, just 15% called it “extremely useful” and 20% called it “very useful.” That’s just 4% of the total US population(!)

In many ways, this highlights out how we’re still in the early innings of the Generative AI ballgame.

Credit: Pew Research Center

🪖 A US military AI drone “killed” its operator in a simulation… or did it?

The US Air Force is actively saying that Col Tucker ‘Cinco’ Hamilton, its Chief of AI Test and Operations, “mis-spoke” when he described a simulation where an AI-enabled drone attacked its own human operator after it was denied the ability to eliminate a threat.

Unsurprisingly, this story quickly picked up traction online as the topic of militarized AI has also become top-of-mind in an era of vastly expanding AI capabilities.

Here’s what we do know:

• Col. Hamilton spoke of “training” an AI-enabled drone that “killed the operator” in simulations during a talk at the Royal Aeronautical Society. He was extensively and directly quoted.

• Several days later, the denial came out from USAF, which tried to clarify that he was simply referring to a “thought” experiment.

Our take: the Royal Aeronautical Society’s direct quotations of Hamilton seem to confirm that this was far more than a thought experiment. But what’s publicly known is that the military is testing AI in other capacities, including testing unmanned F-16s trained in advanced dogfighting.

💰️ Generative AI spend set to hit $1.3T by 2032, says Bloomberg

Anytime we see extrapolations ten years out, we recommend interpreting these reports with a healthy dose of salt. But what’s interesting to us about the latest research to come from Bloomberg Intelligence is a callout on who may reap the rewards of the generative AI boom.

Here’s who the winners could be:

• Amazon’s cloud division, Microsoft, Google, and Nvidia as incumbents are especially well-positioned.

• One reason: revenue from AI servers could touch $134 billion per year by 2032.

• Another reason: revenue from infrastructure capable of training AI models is projected to rise to $247 billion by 2032.

Will the gold rush into generative AI benefit existing technology incumbents heavily? Certainly at this point in time it seems that way. There may be thousands of new AI startups and tools coming out, but whether any of them will emerge to make a grab for serious market share is something we’re watching closely.

Credit: Bloomberg

🔎 Quick Scoops

Nvidia reaches $1T market cap thanks to AI surge. What crypto crash? Nvidia is now flying high as the premier manufacturer of chipsets used to power and train AI models. (The Guardian)

ChatGPT has left copywriters unemployed. In-depth article on how some highly-paid US writers have found their work dwindling as employers adopt ChatGPT over human labor. (Washington Post)

Parents and students favor ChatGPT over human tutors. The tutoring industry could find itself in the throes of disruption as more people to chatbots over human tutors. (Venturebeat)

Japan decides copyright doesn’t apply to AI training data. This move is part of the Japanese government’s strategy to foster its own AI industry. (Technomancers)

AI is now an insult in popular culture. Calling something “made by ChatGPT” is now a way of criticizing its quality. (The Atlantic)

A lawyer gets in trouble for using ChatGPT to prepare a court filing. Oof. A court filing filled with bogus hallucinations got this lawyer in trouble. (New York Times)

🧪 Science Experiments

Nvidia’s Neurangelo creates 3D models from 2D video

• Nvidia continues to push the boundaries in AI tech, and this proof of concept is their latest in neural surface reconstruction, which offers an AI-based alternative to traditional photogrammetry.

• Project page here.

Credit: Nvidia

Undetectable watermarks for LLMs are possible

• Can language models generate undetectable outputs that are nonetheless watermarked? These researchers think so.

• Research paper here.

Segment Anything gets a high-quality upgrade

• Meta’s Segment Anything open-source project impressed many, but some found the quality lacking. These researchers created a much better version able to segment at significantly higher quality.

• Paper here, Github repo here.

Comparison of Segment Anything vs. the new HQ variant

Language models know when they’re hallucinating

• Very interesting paper to come out of Microsoft’s AI research division. Through various prompting techniques, they show you can get a language model to recognize and correct for hallucinations. One of the many studies underway on how to address this challenge.

• Research paper here.

Credit: arXiv

😀 A reader’s commentary

That’s it! Have a great week!

And as always, we want to hear from you on how useful this is. The more feedback you provide, the better our newsletter gets! What else would you like to see covered?

So take the poll below — your signals are helpful!

Today is May 21, 2023.

This week gave us a lot of glimpses into the future of AI, from the emergence of open-source as a power player to calls for regulation of AI models in the United States. Many of our stories touch upon two key themes that are playing out real time:

• Will open-source beat closed-source AI models? The rapid progress here is even causing OpenAI to play defense and consider releasing their own.

• Will AI models be licensed by regulatory bodies in the future? This idea is taking hold in both the US and EU now, and could usher in a world where AI models can no longer get released into the wild.

As always, I write my weekly AI memo so you, the busy reader, can rapidly digest this news and come away smarter.

In this issue:

• 🧨 OpenAI CEO testifies before Congress, calls for regulations

• 🤯 OpenAI to launch an open source model

• 🏥 Google’s MedPaLM 2 AI beats actual doctor answers in a new study

• 🔎 61% of Americans consider AI a “threat to humanity”, and other quick scoops

• 🧪 The latest science experiments, including some mind-blowing image manipulation tech

Did someone forward this to you? Subscribe here!

🧨 The Big Read: OpenAI CEO testifies before Congress, calls for regulations

OpenAI CEO Sam Altman speaks before the US Senate (Photo credit: NYTimes)

During a 3-hour hearing before the US Senate on the future of AI, OpenAI CEO Sam Altman was able to speak to a curious and receptive audience – a big difference from past hearings where tech CEOs have been grilled.

We wrote a full breakdown of all the key moments, and for those with a lot of time, you can watch the entire hearing here.

The most notable bombshell he dropped: the US should establish an agency to regulate and license AI models, Altman proposed.

• The agency would license companies working on advanced AI models and revoke licenses if safety standards are violated.

• AI systems that can "self-replicate and self-exfiltrate into the wild" and manipulate humans into ceding control would be violations, in Altman’s view

Why this matters:

• Senators called AI an “atomic bomb” moment and there’s bipartisan consensus that AI is a serious matter. AI is one of the few issues to cut through political gridlock right now.

• OpenAI’s proposal to license AI models may benefit themselves the most: at a time when open-source is seeing rapid gains, this could crimp progress on that front

One remarkable moment:

• Altman was asked if he could lead the agency: “Would you be qualified, if we promulgated those rules, to administer those rules?" Sen. Kennedy (R-La) asked Altman.

• But Altman demurred and said he would recommend others: “I love my current job,” he said

What to expect next:

• A bipartisan Senate group is already getting to work on AI legislation. The US still trails the EU in drafting any rules (the EU’s AI Act is nearing finalization), so this is just a first step

• Generative AI is a top priority for the G7 meeting in Hiroshima. Multiple countries have started a coordinated process to regulate generative AI, though specifics remain unclear.

🤯 OpenAI to launch an open source model

As pressure from open-source models heats up, OpenAI is planning on launching an open-source model in addition to its current set of closed models (GPT-4, GPT-3.5, and GPT-3).

Our full report covers the nuances of the situation, but the story comes down to this:

• OpenAI’s DALL-E 2 image model has already lost mindshare against open-source Stable Diffusion

• The rapid progress on the open-source LLM front in the past two months is concerning to OpenAI

• Releasing an open-source model is a defensive move: alongside their closed-source models, it could enable OpenAI to control the ecosystem and the overall narrative

And one day later, OpenAI CEO Sam Altman called for licensing of AI models in front of Congress (this would likely slow down open-source). What a coincidence!

Driving the conversation: a leaked “we have no moat” memo from Google concerning the power of open-source is likely driving the same debate within OpenAI.

🏥 Google’s MedPaLM 2 AI beats actual doctor answers in a new study

AI continues to transform how professions work, and researchers at Google recently shared their findings on how a customized version of Google’s PaLM 2 language model passed US medical test questions with 86.5% accuracy, but more importantly generated answers that a panel of doctors preferred over actual doctor-written answers. Our full breakdown is here.

A panel of human doctors judged Med-PaLM 2’s answers to be consistently better than real doctor answers. (Credit: arXiv)

How to make sense of this:

• Expect domain-specific models to be the future: LLMs will increasingly be fine-tuned to perform better jobs at specific functions. Bloomberg’s own finance LLM, BloombergGPT, is another example.

• Doctors could be augmented: Doctors (at least in the US) are already in short supply. AI may not replace doctors, but as the pace of progress keeps up, each doctor could see their efficacy magnified.

• Few jobs are safe from AI reinvention: roles that take years of studying are finding that AI is increasingly able to do more and more. Outside of jobs like construction and manufacturing, expect AI to be everywhere.

🔎 Quick Scoops

Neeva, a Google search competitor, is shutting down. Founded by the former head of Google’s ad business, the transformation of search by LLMs has made their business vision uncertain.

Religious chatbots in India are popular, but also condoning violence. Millions in India seem comfortable using chatbots posing as Indian deities, but the responses they generate pose risk.

61% of Americans consider AI a threat to humanity according a new Reuters poll. Conservative voters were notably more concerned.

People in China are using chatbots to recreate deceased family members. This has been attempted in the past, but the power of LLMs have made this possible in a totally new way.

OpenAI (finally) launches the official ChatGPT iOS app. Hopefully this sweeps aside the sketchy apps posing as official ChatGPT clients that ran amok on the iOS app store.

An analysis of what Google’s recent I/O event means for the AI wars. Great in-depth breakdown by Stratchery.

🧪 Science Experiments

Point-based image manipulation using generative AI is possible

• DragGAN enables enable to alter pose, shape, expression and more by simply dragging and dropping on an image

• Project page here

DragGAN: absolutely crazy how this is possible so early into generative AI’s lifecycle.

XRayGPT open-source model released

• XrayGPT aims at the automated analysis of chest radiographs based on the given X-ray images. It was fine-tuned on medical data (100k pat-doc conversations) + 30k radiology conversations.

• Link here

How XRayGPT works.

FrugalGPT improves LLM usage costs

• An “LLM cascade” method learns which combos of LLMs can create the best queries at lowest cost. Researchers found this was able to match the performance of the best LLM (GPT-4) at significant cost reduction (98%).

• Link here

  

  Chart showing how FrugalGPT scales performance against cost.

😀 A reader’s commentary

That’s it! Have a great week!

And as always, we want to hear from you on how useful this is. The more feedback you provide, the better our newsletter gets! What else would you like to see covered?

So take the poll below — your signals are helpful!

This is another big week for AI, with plentiful news dropping on the inspiring and concerning side.

We continue to see AI create wild stock shifts, with Palantir’s stock jumping 20% after they announced new AI tools, including a battlefield AI for military clients. 15% (🤯 ) of the world’s music is now AI-generated, according to one estimate.

But through all of this, we’re seeing glimmers of material benefits as well, including Google open-sourcing an AI-powered mouse that enables disabled gamers to play their favorite video games. Quantum computing may now come faster thanks to generative AI.

As always, I write my weekly AI memo so you, the busy reader, can rapidly digest this news and come away smarter.

In this issue:

• 📈 Google ups their AI game

• 🪖 The US Army wants to figure out AI, and Palantir is moving to cash in

• 💥 Anthropic releases Claude with 100k context window

• 🏅 Meta is winning at the open-source game

• 🎶 AI music now flooding streaming platforms

• 🧠 AI can help with quantum computing, plus other studies

Did someone forward this to you? Subscribe here!

📈 Google ups their AI game

Google held their big developer conference Google I/O this week, where CEO Sundar Pichar announced that generative AI would feature in a broad array of the company’s product. This is Google’s catchup year, and the company is now shifting to go on the offensive.

• Generative AI is coming to everything: Gmail, AI photo editing is coming to Google Photos, and Docs will now generate entire paragraphs and spreadsheets from prompts, along with helping users plan their vacation, adjust their tone, and write computer code.

• Also driving the conversation: the theme of responsibility. Google spent time here speaking to how it would combat misinformation, add watermarks to AI images, and bake in other guardrails against misuse.

• IO is now AI: “At Google in 2023, it seems pretty clear that AI itself now is the core product,” said the MIT Technology Review. AI is clearly on top of mind for striking WGA writers.

🪖 The US Army wants to figure out AI, and Palantir is moving to cash in

The DoD has released an RFI (request for information) on methods to protect its data sets for use in AI applications.

• Top of mind for them: Testing AI-enhanced systems in battlefield scenarios while maintaining data security.

• But they don’t want SkyNet, either: finding a way to demonstrate the trustworthiness and reliability of AI to users is critical.

• There’s billions of dollars at stake: Palantir this week said they had seen “unprecedented” demand for its military AI. Their stock went up 21% after it revealed their battlefield AI platform.

The use of AI in military applications has already begun (in 2021, Israel conducted an assassination with an AI-assisted gun). We’ll be watching this topic closely go-forward.

Palantir’s stock price this past week.

💥 Anthropic releases Claude with 100k context window

100k tokens, which translates to roughly 75k words or five hours of human reading, is a massive upgrade over Claude’s former 9k window.

• Why this matters: businesses could see massive benefits from processing long documents or retrieving information from a massive data set. GPT-4’s current limit is just 32k tokens, while GPT 3.5 is limited to 4k tokens.

• And it’s fast, to boot: Anthropic pasted the entire text of the Great Gatsby into Claude, and the model returned an answer in 22 seconds.

🏅 Meta is winning at the open-source game

Google and OpenAI are increasingly restrictive on the research they share, but Meta is taking a different approach. This week: Meta released ImageBind, an AI model capable of “learning” from six different modalities, including depth, thermal, and inertia.

• This brings AI closer to learning like humans: ImageBind gives machines an understanding of an object’s sound, their 3D shape, how warm or cold they are, and how they move.

• Meta deeps their open-source winning streak: other releases include Segment Anything, Animated Drawings, and their LLaMA LLM model – which is now the foundation of numerous open-source LLMs.

• Expect the community to move quickly: we previously wrote about open vs. closed source AI in this article – and the pace of progress on open-source was simply astounding. Expect the same here.

Demonstration of how ImageBind lets a machine “learn” more like humans do.

🎶 AI music now flooding streaming platforms

The removal of Ghostwriter’s fake Drake song was just the beginning. This week, news broke that Spotify has removed “tens of thousands of AI-generated songs” from its platform – and they’re barely scratching the surface.

• Spotify suspects foul play: most of the songs were made by a single generative AI company, Boomy, and suspicious streaming data means bots could have been used to juice royalties on these AI tracks.

• The scale is massive: Boomy claims that they’ve created over 14 million songs – about 14% of the world’s music – during its two years in existence. Expect this number to exponentially grow over time.

• Google isn’t helping: the company released MusicLM this week, which enables users to generate music from text prompts. While specific artists and vocals are forbidden, a broad array of styles can still be made.

🧪 Science Experiments

AI is helping make quantum computing possible by designing circuits

• Quantum algorithms need to be designed by hand, but it’s notoriously difficult. This could very well be AI’s superpower, much like its potential impact on drug discovery and protein folding.

• Read the full paper here.

Google introduces AI gaming mouse, open-sources code

• For gamers with conditions like muscular dystrophy, normal control devices are not usable

• Google’s tech scans the face and tracks head movements to then convert them into in-game movements. An early review called the controls “robust and intuitive.”

• Access the open-source code here.

A demonstration of how Google’s mouse AI reads a human face.

Robotic household cleanup benefits from LLMs, Princeton/Stanford study finds

• Everyone has different cleanup preferences, due to taste, cultural background and more

• By combining an LLM with a cleanup robot, a robot was able to make remarkable decisions around where objects should go

• See the full study here.

How soon can we order one of these LLM-powered robots?

Which open-source LLMs are good? A leaderboard now tries to provide an answer

• With dozens of open-source models releasing, it’s hard to verify performance claims. A new and ongoing study now subjects all open-source LLMs to a series of 4 benchmarks, helping provide a baseline for comparison.

• Link to Hugging Face page here.

Diffusion model can now create 3d faces for all lighting conditions from just an image

• The pace of image technology continues to be remarkable. Even this early proof of concept is quite fascinating. Full paper here.

An example of how a single photo outputs a 3D lighted model.

That’s it! Have a great week!

And as always, we want to hear from you on how useful this is. The more feedback you provide, the better our newsletter gets! What else would you like to see covered?

So take the poll below — your signals are helpful!

One clear theme for this week’s AI news stands out: no one really knows where we’re all headed.

You have the “godfather” of AI claiming “bad things” are ahead, but not knowing what, a leaked Google paper saying open-source will outpace closed-source models like Bard and ChatGPT, and entire companies seeing 50% stock drops as AI disrupts their business models.

A personal note: I’ve gotten a lot of emails from readers asking who I am. The quick background: I’m a 2x tech company founder, a New Yorker, and new dad. Artisana is my side project and passion.

So let’s make this easy to connect (I’m friendly and love chatting with readers)!

• Follow me on LinkedIn

• Follow me on Twitter

In this issue:

• 📉 AI continues to impact the job landscape on multiple fronts

• 💥 Chegg’s valuation blew up after disclosing ChatGPT ate their luunch

• 🤖 Leaked Google memo says open-source will beat Bard and ChatGPT

• 💰️ OpenAI burned $540M last year, wants $100B more to develop AGI

• 🏗️ Society continues to grapple with the AI future

• 🧠 AI can now read human thoughts, plus other science experiments

Did someone forward this to you? Subscribe here!

📉 AI continues to impact the job landscape

We’re in the midst of seeing society reconfigure itself as generative AI rapidly impacts numerous functions.

• Hollywood writers are on strike right now, and one of the concerns they have is generative AI will put additional pressure on their declining wages as their profession is confronted with numerous headwinds. Read our full breakdown here.

• Creative roles in general face enormous pressure, with one veteran writer sharing on Reddit that their client base had virtually vanished overnight. The feedback? “Some of them admitted that I am obviously better than ChatGPT, but $0 overhead can't be beat and is worth the decrease in quality.”

• IBM announced that it would pause hiring on 26k non-customer-facing roles. The reason? IBM’s CEO explained: “I could easily see 30% of that getting replaced by AI and automation over a five-year period.”

AI is clearly on top of mind for striking WGA writers.

💥 Entire companies are finding themselves vulnerable to AI’s rapid pace of disruption

Chegg’s nearly 50% stock drop this week is expected to be just the first of many companies experiencing an existential crisis.

• Despite announcing their own GPT-4 AI chatbot in the works, investors simply aren’t buying that a chatbot is going to save Chegg’s business

• This is a warning sign to other companies who think AI will protect their existing business lines. Read our full analysis here.

🤖 Is the future of AI open-source?

That’s the major discussion in the tech community right now, and it’s attracting opinions on all sides.

• The catalyst is a leaked Google memo written by a senior AI engineer claiming “we have no moat, and neither does OpenAI.”

• The explosive claim at the heart of this memo: open-source will overtake closed systems like GPT-4 and Bard, and the author points to numerous examples of how fast open-source has advanced since Meta’s LLaMA LLM model leaked into the wild.

• Substantial amounts of venture funding is going towards closed-source foundational models right. Anthropic just raised another $850M, and Inflection launched its own chatbot this week on heels of a $225M seed round.

• Not everyone believes it, however, and skeptics are pointing to numerous examples of integrations, developers, and enterprise contracts as moats. Our full breakdown here looks at a number of these skeptical arguments.

💰️ OpenAI burned $540M last year, wants $100B more to develop AGI

OpenAI is a private company, so getting a peek into its finances is extremely interesting. The leak comes courtesy of The Information, one of Silicon Valley’s most trusted media publications, so we have reason to believe these numbers hold water.

• The company burned $540M in 2022 to develop ChatGPT, and expects to burn even more this year despite some rocketship revenue numbers (it thinks it’ll beat $200M revenue in 2023).

• It’s got a lot of rocket fuel though, having secured $10 billion in funding from Microsoft this year with priority access to computing resources, which are rationed out in this era of high demand

• But could it all be for naught? That’s what the leaked Google memo is saying: LLMs with comparative quality can now be trained for hundreds, not billions, of dollars.

• Still, OpenAI employees are able to celebrate a bit. News broke this week of a $300M share sale at a nearly $30B valuation. That’s quite some cheddar!

🏗️ Society continues to grapple with the AI future

Governments continue to play catch-up on AI, as humans wrestle with AI’s position in the world.

• AI’s own “godfather” who created neural networks has a warning: “bad things” lie ahead as AI’s progress proceeds.

• The White House convened a meeting of AI leaders from Google, Microsoft, OpenAI, Anthropic and more to discuss AI regulations and safety. But with open source models running amok, is it too late?

🧪 Science Experiments

GPT AI can now decode your thoughts

Is mind-reading possible? We’re getting there when GPT AI can now decode fMRI signals with up to 82% accuracy. Our breakdown of this breakthrough research went viral this week (1.5M impressions!), and we consider this a milestone for AI tech.

Vicuna-13B: the open source model that’s 92% as good as ChatGPT

The leaked Google memo cites this as one of the main reasons ChatGPT will get outpaced. Based on Meta’s leaked LLaMA LLM, then fine-tuned on 70k ChatGPT conversations for just $300, it claims 92% of the quality of ChatGPT.

• Test it here for yourself and let me know your thoughts!

• Here’s the full research if you’re curious.

Nvidia team teaches AI to learn tennis from just watching broadcast videos

Wow. Talk about cool — AI was unleashed on tennis footage, and it learned how to play virtual tennis. Backhand slice, forehand topspin were just some of the moves learned all from watching videos.

• See the methodology and video examples here.

Dreampaint enables in-painting of e-commerce models for virtual-try on

We’ve had 3D-try ons and AR 3D furniture for awhile. But this is something new – pairing Stable Diffusion with a customized in-painting engine to easily render virtual clothes, furniture and more from images.

• The full research paper is here.

AI Chat Assistants can improve conversations about divisive topics

Could AI chatbots actually help our society in unexpected ways? Researchers found that chatbots had a tendency to make polarized subjects feel understood, while not changing the content of its responses. They tested this on a tried and true topic: gun control.

• Read the full paper here.

• Similarly, LLMs have been found to help humans reframe negative thoughts in another study.

Transformer memory can be mass edited

Researchers found a new technique to enable thousands of insertions vs. updating single associations in a transformer model. If implemented successfully, could be a powerful way to replace obsolete information or add specialized knowledge in LLMs in a scalable and affordable manner.

• Read the full paper here.

OpenAI released Shap-E, a text-to-3D-model generator

Text-to-image is old school now. Text-to-3d-models is where a lot of the frontier tech is playing, and OpenAI jumped into the ring this week with Shap-E. This is an early proof-of-concept but expect AI tech on this front to rapidly improve.

• See it here

That’s it! Have a great week!

And as always, we want to hear from you on how useful this is. The more feedback you provide, the better our newsletter gets! What else would you like to see covered?

So take the poll below — your signals are helpful!

From the firehose of AI news, a few key themes emerged as critically important to understand this week. I narrowed down about 425 saved links and stories into the most impactful items below.

Reader feedback has been so useful since the last issue. From the vast majority of you I’m hearing one thing: it’s hard to keep track of everything, and that’s why a summarized digest matters.

I couldn’t agree more. This is why I write this every weekend, so I go into the next week feeling fully prepared.

In this issue:

🎸 Music is the next legal frontier AI will confront

🇪🇺 Regulatory developments in the EU speed up

😓 Google’s challenges continue

🤖 Elon Musk’s complicated views on AI

🏢 Corporations are unprepared for generative AI, study finds

and more (including the latest research developments)

Did someone forward this to you? Subscribe here!

🎸 Music is the next legal frontier AI will confront

The music industry is confronting a litany of AI-related issues at light speed.

• Since the release and takedown of “Heart on My Sleeve,” which featured AI voices mimicking Drake and The Weeknd, the internet has been flooded with additional AI-made Drake songs.

• Expect each of these to test legal waters around what is fair use, and what is copyrighted

• Josh Constine, a VC at SignalFire, puts it succinctly: “Google is caught between an AI rock and a copyright hard place. Either the AI Drake song trained on copyrighted data is fair use, YouTube floods with this content, and labels panic Or it’s infringement, which means Google’s Bard AI is illegal.”

• Adding to the conversation: musician Grimes has proclaimed anyone can use her voice for AI-generated songs, and she’d split 50% of royalties.

• An ongoing Andy Warhol copyright case could also have implications for generative AI, including AI music.

🇪🇺 Regulatory developments in the EU speed up

The legal landscape impacting AI models is rapidly changing, and this week saw two major developments:

• ChatGPT complied with Italy’s initial demands, resulting in the lifting of the ban. OpenAI added additional information on how it trains ChatGPT, provided EU users with a new form objecting to have their data used for training, and now verifies users’ age when signing up. Investigations into ChatGPT, however, remain ongoing in France and Germany.

• The EU has passed a draft of its AI Act, setting the stage for a finalization phase. The most important provision? A new clause that specifies AI models “would have to be designed in accordance with EU law and fundamental rights,” as well as a requirement that AI tools disclose the use of copyrighted materials. We’ll be watching this closely.

😓 Google’s challenges continue

Poor Google. Since Bard’s tepid launch they can’t seem to catch a break, and new reports highlight exactly how daunting the AI race will be for them:

• Mindshare about Bard remains low relative to OpenAI and Bing. According to Google Trends, ChatGPT is 8.3x more popular than Bing and 33x more popular than Bard.

• Google’s recent merging of Google Brain and Deepmind into a single AI-focused Google Deepmind team will face steep challenges. According to Google insiders, Deepmind has historically functioned very independently, thinking about Nobel prize-worthy problems, while Google Brain has operated with indecisive leadership. For the two teams to merge and move quickly to match OpenAI’s focus and speed will be a daunting task.

🤖 Elon Musk’s complicated views on AI

What exactly are the billionaire’s plans for AI? While no one knows for sure, new details surfaced this week that adds color to the mixed messages he’s been sending:

• The New York Times reported that Elon Musk had ordered Twitter to turn off OpenAI’s access to its historical tweets after ChatGPT surged in popularity

• Despite founding OpenAI in 2015, Musk has had a falling out with OpenAI on its mission and direction. He reportedly grew disillusioned when OpenAI stopped operating as a non-profit and built “woke” AI models.

• At the same time, the billionaire CEO is building his own Large Language Model as part of his new X.AI initiative. TruthGPT, Musk claims, is a “a maximum-truth-seeking A.I. that tries to understand the nature of the universe.”

💰️ AI roils the job landscape

Transformative technology has historically been a net benefit for society and GDP, but not without its intermediate pain. This is playing out at warp speed across multiple professions as AI’s power rapidly forces transformation.

• Dropbox announced a 16% headcount cut, citing AI as one of the reasons behind the significant layoff. What’s notable: this is a profitable, public tech company whose financial metrics have only improved in recent years. For AI initiatives, Dropbox is doubling down — but for mature teams, they’re making cuts. Expect this to be the broad theme of tech as AI surges to the forefront.

• Kenyan ghostwriters, who normally help US college students write essays, are losing jobs to ChatGPT. Rest of World reports that many ghostwriters have seen up to 50% decrease in work as AI has reduced demand for human writers.

• A Stanford/MIT study showed that GPT-3 software helped customer service agents perform as much as 35% better, portending big shifts in knowledge worker jobs as AI makes its way into numerous industries

🏢 Corporations are unprepared for generative AI, study finds

A KPMG study of 225 US executives found that 65% believe generative AI will have a high or extremely high impact on their companies, but nearly the same percentage say generative AI is still a year or two away from having an impact.

• While executives are optimistic, they are also worried it could have a negative impact, especially if risk is not managed

• Almost 4 in 10 executives believe generative AI could decrease social interactions and human connections among employees

• PwC announced a $1B investment in AI over the next 3 years; this is likely to become the norm as corporations pull the trigger on AI investments

🧪 Science Experiments

Amongst the dozens of impactful research papers coming out each week, we feature the most mindblowing examples below. As always, we try to explain anything technical to a non-technical reader.

Text to Video

RunwayML launched its Gen2 text-to-video model and the results are gorgeous. Here are several examples of what users have created in concert with Midjourney. The pace of development in the video space is simply on fire; imagine what could be possible by the end of the year.

• We previously reported on Nvidia’s text-to-video experiments, click here to see their examples.

Segment Anything, but for video

In another sign of how fast open-source tech is quickly improved, Facebook’s Segment Anything AI library was rapidly adapted into a video-tracking tech that beats Adobe’s own rotoscoping features in its professional software. The open source repository can be accessed here.

Robots playing soccer

Straight out of the geniuses at Google’s Deepmind team, they’ve applied a technique called Deep Reinforcement learning to help robots move in a dynamic environment. This is a good reminder that the latest breakthroughs aren’t just limited to generative AI. Check out the full videos here!

Text to 3D Models

A team of researchers generates surprisingly great 3D models out of text prompts. Expect this area of technology to rapidly improve in the next few months. What could that mean for 3D art and the artists who create models? Full research paper here.

Other News

News that didn’t make it into the key themes but is still worth keeping in mind.

WSJ reporter clones her own voice using AI, fooling her family and bank [Link]

Free course for developers on ChatGPT prompt engineering released, taught by OpenAI staff [Link]

Bark, an open source voice cloning tool is released [Link]

ChatGPT finally allows you to turn off conversation history and choose which conversations train their models [Link]

Will AI lead to mass employment? This author argues it won’t and examines how past technology disruptions have played out [Link]

Snapchat users really don’t like the MyAI assistant [Link]

And… that’s it for this week! Remember: your feedback is important and I respond to every email, so hit “reply” any time you have thoughts!

Or just take the poll below — this is massively helpful:

Wow. It’s been an action-packed week for AI news. I had over 500 saved tabs in a Chrome browser open at one point to track the flood of AI news heading our way.

A few cups of ☕️ and a brisk walk in the morning helped create clarity around the most important events of the week.

Let me know if you think I missed anything! (I respond to every reader email).

In this issue:

• 📽️ Nvidia’s text-to-video model (a must see)

• 🇪🇺 ChatGPT’s compliance challenges with GDPR

• 🎤 AI music and a banned Drake AI song

• 💰️ Companies want to charge for providing AI training data

• 🏃‍♂️ Google’s “cringe-worthy” Bard still playing catchup

• and more

📽️ Nvidia’s Text-to-Video Model

Credit: Nvidia

The mad scientists at Nvidia showed off their latest project this week: a Latent Diffusion Model (LDM) capable of high quality videos (2048 × 1280) via a text-to-video model, up to 4.7 seconds long. 🤯 

Why does this matter? The scientists developed using open source Stable Diffusion model, then further tuned it on a set of images using DreamBooth, another framework for tuning text-to-image models. Open-source tech is enabling all kinds of innovation here, showing how we’ve only just begun to scrape the surface of what’s possible.

See it in action here.

🇪🇺 ChatGPT’s EU GDPR Woes

It’s not just about ChatGPT’s ban in Italy, which is running short on time to resolve (they have until April 30).

Numerous legal experts are calling ChatGPT’s ability to comply with GDPR “next to impossible.” GDRP is the grounds on which Italy temporarily banned ChatGPT, and it’s the basis for investigations in France, Germany, and Ireland as well.

At the heart of the matter: training data, gathered without user consent, was used to generate the models that power the chatbot. But how can this training data be “removed” if users ask? Experts think that’s not quite possible as ChatGPT isn’t a Google search index.

Margaret Mitchell, an AI researcher and ethics lead at Hugging Face, asserts that "OpenAI is going to find it near-impossible to identify individuals' data and remove it from its models."

Read more here.

🎤 AI music bans

An AI-generated track mimicking Drake and The Weeknd’s voices has been taken down from numerous platforms. We reported on this issue here, but subsequent reporting by other publications has added great color and context.

Here is what you should know:

• Universal Music Group (UMG) filed a complaint resulting in the removal of a viral AI-generated song that imitates the voices of Drake and The Weeknd from major streaming platforms, alleging copyright infringement.

• UMG argued in a statement that training generative AI on artists' music constitutes "both a breach of our agreements and a violation of copyright law."

The Verge did an excellent writeup as well that summarizes the legal conundrum here:

“The first legal problem with using AI to make a song with vocals that sound like they’re from Drake is that the final product isn’t a copy of anything. Copyright law is very much based on the idea of making copies — a sample is a copy, as is an interpolation of a melody. Music copyright [is] all based on copies of actual songs. Fake Drake isn’t a copy of any song in the Drake catalog, so there’s just no dead-ahead copyright claim to make. There’s no copy.”

This area is going to have a lot of developments in the coming months — so watch this space.

💰️ Pay us for using data to train AI, companies say

Reddit’s data, with its vast libraries of user-generated content, has formed the backbone of several AI training datasets, including ChatGPT.

Now Reddit and Stackoverflow want AI companies to pay for accessing their data. These APIs, which used to be free, will now cost money.

“The Reddit corpus of data is really valuable,” Steve Huffman, founder and chief executive of Reddit, said in an interview with the NYTimes. “But we don’t need to give all of that value to some of the largest companies in the world for free.”

Look out for this to become the norm across many sites (Twitter is now a paid API as well).

But all this has some users asking: where’s our compensation for writing the content? Unfortunately, the terms of service with these sites leave users unlikely to see any monetary benefit.

🏃‍♂️ Google tries to play catchup

Google just can’t catch a break. Leakers from inside the company continue to share stories about how the company just isn’t getting its act together at a time when its future is under threat.

A report from Bloomberg with 18 employees calls out damning details:

• AI ethics members felt “disempowered and demoralized” at Google’s rushed decision to launch Bard

• Internal evaluations painted Bard as “cringe-worthy” and a “pathological liar”

To no one’s surprise, Bard’s unveiling stumbled out the gate, with basic facts incorrect in their preview video and journalists giving tepid reviews in their first takes.

Google’s now trying to course correct, going on the offensive with plans to add AI across their entire product suite. And the panic at Google has finally forced two competing AI teams to merge into Google Deepmind.

⭐️ Other headlines of note:

• Stability AI launches their own open-source language model, StableLM. Open-source alternatives to ChatGPT are going to lead to a lot of interesting developments in these next few months.

• Michael Schumacher’s family threatens to sue German tabloid over aI-generated interview. Expect AI content to be at the center of continued legal firestorms.

• Microsoft develops its own AI chip as ChatGPT costs OpenAI an estimated $700k per day to run. AI is expensive. ChatGPT is expensive. Microsoft is launching their own chip to cut costs.

• A “grandma jailbreak” is the latest example of how users are getting creative in breaking ChatGPT. And this is one of the most hilarious things I’ve seen this week, too.

🧪 Science experiments and things to try:

• A beginner’s guide to autonomous agents. What’s the hype around autonomous agents? This writeup explains what it does and how you can play with it, right now.

• MiniGPT-4 launched, runs on just 12GB memory, and can process image. A glimpse of the future.

• Meta open sources their animated drawings AI library. Fun to see this in action.

📕 Notable research papers in the past week:

• LLMs are learning to program with natural language [Link]

• Analysis of why ChatGPT falls short in comprehension [Link]

• Using LLMs to create data lakes [Link]

• Just 51.5% of LLM search engine responses fully supported by citations [Link]

• Gisting enables 26x compression of LLM prompts [Link]

And… that’s it for this week! Remember: your feedback is important and I respond to every email, so hit “reply” any time you have thoughts!

Or just take the poll below — this is massively helpful:

Tracking the Explosive World of Generative AI

Welcome, readers!

If you’re reading AI news and experiencing simultaneous feelings of FOMO, anxiety, and excitement: know that you’re not alone.

The amount of news in this industry can feel overwhelming, and that’s why I launched Artisana and this newsletter to highlight the items that are truly impactful.

Focus is good and less is more.

One big thing to know:

Generative AI funding in Q1 2023 surpassed 2022 by 4x. 2023 is just getting started as the arms race heats up and VCs pour into the gold rush. Where is all the capital going? And who thinks this is a big fat mistake? I cover all of that in my deep dive.

While no one knows exactly how the future will play out, I think we can all agree on one thing: the next few quarters are going to be interesting with this amount of capital pouring in.

Here’s what else may have missed this week:

• 🤖 Astonishing Stanford study on human-like Generative Agents 

• 🇪🇺 EU’s AI act means ChatGPT may face strict regulations

• 😒 Despite safeguards, researchers find ways to make ChatGPT toxic

• 🕹️ Chinese videogame market under threat from AI: 40x productivity spike, 70% job loss

What we’re reading and tracking

Artisana can’t cover every single piece of breaking news, but we’re closely tracking a number of other developments we may write on in the near future.

👀 Here’s what we have our eye on (and you should too):

• AutoGPT: currently the 🔥 thing on Github, autonomous agents could usher in a new level of innovation… or create totally new disasters. Experimentation on this front is moving at lightspeed.

• Elon Musk: after calling for a pause in the development of generative AI, he’s all-in on starting his own gen AI company called X.ai. Could he be developing an “anti-woke” competitor?

• Stable Diffusion: Stability AI launches Stable Diffusion XL, which is getting even better at generating words in images (a current weakness). Image generation’s rapid progress will be a defining theme of 2023.

• Amazon: Amazon joins the AI race with Bedrock (for building and scaling with ML models) and CodeWhisperer (AI-powered coding assistant). But we’re certain this is not where their ambitions end.

📣 I want to hear from you, our reader!

It’s clear people like the content. Here’s a snapshot of our Google Analytics just one week into our launch:

We’re here to write one of the best newsletters in AI and want to make it something that reads distinctly different. What news and analysis is most interesting or valuable to you? This is your chance to let us know and shape our coverage.

Reach me at michael@artisana.ai — I read every email from our readers. 💌