4:55 PM
Today in brief.
Robots learn from one video, and Pi opens an API
Gemini 4 Argon taught Apptronik's Apollo humanoid cornhole from a single human video by writing its own physics sim, and Physical Intelligence opened an API so anyone can fine-tune its robot foundation models.
🔧 Tech
Anthropic outgrew git at 30,000 merges a day
Anthropic went from about 2,500 merged code changes a day in May to over 30,000 in September, so it built its own version control system, Scion, and moved its monorepo off git in five months.
💼 Business
Xbox makes AI game ports official
Microsoft formed Xbox Labs to embrace the wave of AI-assisted decompiles and recompiles that are porting old games to Macs, browsers and new hardware, even when it challenges its own business model.
🔧 Tech
Try today: private Android apps, Runway in Replit, $8 Grok bundle
A cluster of build-it-yourself doors opened yesterday: private vibe-coded Android apps on Google Play, Runway assets inside Replit, and a cheap Grok plus Cursor plus X Premium bundle.
🔧 Tech
Dead startups are selling their data to AI labs
Shutdown brokers now sell dying startups' code, Slack, Notion and email archives to AI labs, and 47% of SimpleClosure's deals include a code or data sale; Character.AI's own wind-down sold assets to Disney for $20M.
🔧 Tech
Run models on hardware you already own
Three signs that inference is escaping the datacenter: Qwen 27B on a jailbroken PS5, Apple's paper on pooling idle Macs into a GPU cloud, and four 9-year-old V100s beating new desktop AI boxes.
🤖 AI
Gemini 4 Argon ships to everyone
Google finally released Gemini 4 Argon globally, claiming it beats GPT-6 Astra, Fable 5.1 and Opus 5.5 across coding, cyber, science, finance and law, and early testers are already building whole apps and decks from one prompt.
🏛️ Policy
Skeptic corner: hidden capabilities and OpenAI's FFT claim
A new paper argues AI models can provably hide arbitrary capabilities from evaluators, and a close read of OpenAI's faster-Fourier-transform claim checks what the math actually shows.
💼 Business
Claude's oboe concerto passed for human 60% of the time
A listener test reportedly had 60% of people judge a Claude Opus 5.5 oboe concerto as human-composed, above real Handel (58%) and Vivaldi (43%), the same day a Harvard physicist reported 36 manuscripts in three months with Claude.
No stories in this category for this edition.