OpenAI poached back a rival's co-founder, while DeepMind quietly picked apart the team behind a Nobel laureate's work. Plus record OpenAI revenue, a price crash on its models, and rumors of a Tesla-China split.
🤖 AI
OpenAI got Lilian Weng back — she couldn't keep up with Thinking Machines' pace
Weng, co-founder of Mira Murati's startup (Murati is OpenAI's former CTO), left Thinking Machines citing stress and almost immediately returned to OpenAI to lead the team working on recursive self-improvement of models.
DeepMind quietly took apart the team behind AlphaFold
Almost a quarter of the authors of the AlphaFold paper that won DeepMind its Nobel Prize have left the company — including laureate John Jumper, who left for Anthropic. The rest got reassigned to Gemini.
Thinking Machines shrank its flagship model fourfold without losing quality
Inkling-Small, at 276B parameters with 12B active, trails the 975B Inkling by less than a point — and even beats its predecessor at coding. The weights are up on Hugging Face.
💻 Software
GitHub opened up stacked pull requests to all developers
The feature splits large changes into a chain of small pull requests that preserve dependencies — each gets reviewed separately, and they all merge with one click. GitHub is on pace for 14 billion commits in 2026 — double last year's number.
OpenAI found a bug in its own benchmark harness — and tripled its score
GPT-5.6 Sol was only scoring 13.3% on ARC-AGI-3: the harness was wiping its reasoning after every move. By turning on memory and context compression, engineers pushed the score up to 38.3% while using six times fewer tokens.
⚙️ Robots and hardware
Google taught Gemini Robotics to run entire teams of robots
Gemini Robotics ER 2 breaks a voice command down into steps and taps into Google Search, while the new Gemini Robotics 2 controls a humanoid's whole-body movement for the first time, not just its arms.
💰 Business
OpenAI's annualized July revenue topped its entire Q2
CFO Sarah Friar told employees that annualized revenue in July topped all of Q2 — the growth came from GPT-5.6, ChatGPT Work, and Codex.
OpenAI slashed prices on GPT-5.6 Luna by 80%
Luna dropped from $1/$6 to $0.20/$1.20 per million tokens, Terra fell 20% to $2/$12, and Sol stayed put. OpenAI cited model-development efficiency as the reason, though competitors see pressure from cheap Chinese models at play too.
Tesla is weighing selling off its China business to merge with SpaceX
According to the WSJ, Tesla's leadership has been told to prepare a split of its China operations — a spin-off, sale, or shutdown are all on the table, since Beijing could block a deal over SpaceX's ties to the US defense establishment. Musk called the report fake.
#digest