AI Roundup 066: AlphaFold 3
May 10, 2024.
AlphaFold 3
Google DeepMind has unveiled AlphaFold 3, a model that aims to reveal biological secrets and boost drug search efforts.
Between the lines:
Previous versions of AlphaFold only predicted the shape of proteins, but the new model can work with DNA, RNA, and "all life's molecules."
It uses a diffusion technique (similar to Stable Diffusion) to take input molecules and predict how they’ll combine, with seemingly a big jump in accuracy.
The model was built with Isomorphic Labs, DeepMind's drug discovery spin-off. CEO Demis Hassabis believes AI-powered drug discovery is a $100B opportunity.
Elsewhere in the FAANG free-for-all:
Microsoft plans to add auto-complete features to Copilot in the coming months to help with prompt engineering.
Meta's AI tools for advertisers have new upgrades, and users can now make full images (not just backgrounds).
Apple's new M4 chip is focused on AI, though the company is reportedly also working on an AI-specific chip for data centers.
And TikTok plans to start auto labeling AI-generated user content, including images from DALL-E and Adobe Firefly.
Elsewhere in model mayhem:
ElevenLabs previews a music-generating AI model, complete with human-sounding lyrics.
A study by Meta researchers suggests LLMs are more efficient when trained to predict multiple tokens at once.
And Microsoft is working on a top-secret air-gapped LLM (to share with intelligence agencies) and a massive 500B parameter LLM (to compete with OpenAI and Anthropic).
Model Spec
OpenAI published its Model Spec, a document describing the desired behavior for its API models and ChatGPT.
Why it matters:
The full Model Spec document outlines the objectives, rules, and default behaviors that models should have. Yet, clearly, this is a difficult problem - many of the rules have the potential to be contradictory.
It's worth reading to get a deeper sense of how exactly ChatGPT’s RLHF logic works - and why it might reject your seemingly reasonable requests.
There's some prior art here. Anthropic has long touted its Constitutional AI approach, which bakes values into the model from an early stage.
Elsewhere in OpenAI:
After much speculation, more outlets are reporting the imminent launch of an AI-powered search tool to challenge Google and Perplexity.
Stack Overflow and Dotdash Meredith have signed deals to supply OpenAI with content in exchange for new integrations using the company's models.
The mysterious gpt2-chatbot returns to the LMSYS Arena, and is seemingly confirmed to be an OpenAI model.
The company joined C2PA (the Coalition for Content Provenance and Authenticity) and released new tools to detect AI-generated images.
And a future tool called Media Manager will let creators identify their works and opt out of training.
Uncanny companions
While chatbots are nothing new, they're now at a point where many people could believe they're talking to another human. While the first applications are mostly productivity-focused, many consumer/companion applications are in the works.
What's the latest:
Teenagers are making friends with AI chatbots, though some feel their companions are "addictive."
Some (adults) are using chatbots to prepare for difficult conversations with friends and family in a form of pseudo-therapy.
And some companies are working to create "digital twins" of individuals to work as fashion models, focus group members, and clinical trial participants.
Elsewhere in AI anxiety:
Reddit has released a new content policy that bans those AI data licensees from using deleted posts or comments.
Some Singaporean writers are resisting the government's request to use their works to train AI.
And the US Commerce Department is considering a new push to restrict China's access to foundation models.
Last week's roundup
Things happen
The last stock photographers await their fate under generative AI. Meet AdVon, the AI-powered content monster infecting the media industry. Sundar Pichai lays out his AI roadmap. Mistral is reportedly raising ~$600M at a $6B valuation. Microsoft is investing $3.3B to build AI in Wisconsin. AI comes for YouTube’s thumbnail industry. Softbank is in talks to buy troubled AI chip firm Graphcore. Red Hat unveils RHEL AI to democratize enterprise AI. CRISPR-GPT: an agent for automated gene-editing experiment design. Defense think tank MITRE to build AI supercomputer with Nvidia. Trump's former campaign manager is building a new AI company. Amazon commits $9B to expand its cloud infrastructure in Singapore. The Techbro turf war over AI's more hardcore hacker house. Hugging Face launches LeRobot, an open-source robotics code library. Audible now has over 40,000 audiobooks read with AI voices. Researchers are training AI models on headcam footage from toddlers. AI Copilots are changing how coding is taught. US Air Force plans a fleet of 1,000+ AI-controlled jets. Runway's second International AI Film Festival. The man who made the world fear AI asks, "what if it fixes everything?" The news industry is divided over whether to license content for AI, and for how much. AI climbing coach – visualize how to climb any route based on your body. AI doesn't threaten humanity - it's owners do. Slop is the new name for unwanted AI-generated content. Recruiters are going analog to fight the AI job application overload.