. . AI: Big Model Releases from Google and (Potentially) OpenAI (5.8.24)

Today's edition covers news over the last four days, and they were certainly a busy four days! Google has released a new version of what was arguably the biggest development in drug research this century (AlphaFold V3). A mysterious model, claiming to be based on GPT-4.5 architecture, is making the rounds by creating video games from just a few images and prompts. We played around with the new model and can confirm it is quite impressive, especially its reasoning ability. It is also rumored that OpenAI is preparing to launch a new search product any day now. We've also distilled the most interesting headlines in case you missed them. Enjoy!

— Sasha

Google Releases AlphaFold 3 and Enhances Biomolecular Predictions

First impacted: biomedical researchers, pharmaceutical scientists

Google DeepMind has launched AlphaFold 3, an AI model that predicts the structures and interactions of a broad spectrum of biomolecules. This new model extends to proteins, DNA, RNA, and ligands, and is available for non-commercial use through the newly launched AlphaFold Server, enhancing global scientific research and drug discovery. [AlphaFold 3 predicts the structure and interactions of all of life’s molecules] Share this story by email

GPT-2 (Potentially GPT4.5) Chatbot Crafts Flappy Bird Clone on First Try

First impacted: game developers, AI researchers

A user on X has created a Flappy Bird game clone using a straightforward prompt on the mysterious GPT2 model in the Lymsys chatbot arena. Despite the lack of detailed public technical specifications, the model claims that it is based on GPT-4.5. [via @minchoi] Share this story by email

OpenAI Launches Model Spec for AI Behavior

[openai.com] Share this story by email

Tesla Shares New Optimus Robot Video

[via @Tesla_Optimus] Share this story by email

DrEureka Automates Robot Training, Cuts Expert Involvement and Allows Robot Dogs to Balance on Yoga Balls

[GitHub - eureka-research/DrEureka] Share this story by email

In-Context Learning Outperforms Finetuning in AI

[arxiv.org] Share this story by email

Robot Training Method Achieves High Success with Minimal Data Trials

[Finetuning Offline World Models in the Real World] Share this story by email

Andrej's Ilm.c Project Accelerates GPT-2 Model Training by 46%

[State of the Union [May 3, 2024] · karpathy llm.c · Discussion #344] Share this story by email

Altera Raises $9M for AI to Accelerate AI Agent Development

[Altera’s game-playing AI agents get backing from Eric Schmidt] Share this story by email