01Anthropic Replaced Git: 2,500 Daily Commits Became 30,000, So They Rewrote Source Control [AI Models & Research]Anthropic's engineering team went from ~2,500 merged code changes a day in May to over 30,000 by September — enough to mean 600 PB of naive git traffic daily. Their answer: Scion, a custom system built by seven people in five months, paired with the Antique review tool. → source
02Liquid AI Ships LFM2.5-24B-A2B: a Reasoning Agent Model With 2B Active Parameters [AI Models & Research]A new 24B mixture-of-experts model tuned for agent harnesses opens on Hugging Face: 2B active parameters, multi-step reasoning and tool calling, aimed at running useful agents on-device rather than in the cloud. → source
03Gemini 4 Argon for Robotics: 87% on Embodied Reasoning, Ahead of Every Frontier Model [AI Models & Research]Google published an October 13 robotics-focused follow-up on Gemini 4 Argon: 87% on the Embodied Reasoning (ERQA) benchmark by pure reasoning, plus demos of zero-shot robot control, Real2Sim reconstruction and agentic hindsight labeling across the robotics stack. → source
04AI Labs Are Buying the Data of Dead Startups to Train Their Models [AI Tools & Ecosystem]Shutdown services like SimpleClosure now broker failed startups' code, chats and workspace files to foundation labs — 150+ deals so far and $10M expected back to founders by year-end. Once-taboo data sales are now in 47% of the company's wind-downs. → source
05Ollama Goes to Fly.io: Sprites Give AI Agents a Private Linux VM With Secret-Free Model Access [AI Tools & Ecosystem]Ollama Cloud now connects to Fly.io Sprites — persistent Linux VM sandboxes for Claude Code, Codex and Gemini CLI. A gateway injects your Ollama key server-side, so the agent only sees a connection ID and can never leak the credential. → source
06Google Play Opens to Private, Vibe-Coded Apps That Never Hit the Store [AI Tools & Ecosystem]Android users in the US can now build an app in Google AI Studio and run it on their own phone through Google Play — private by default, invisible in search, with a path to full publication only after trust-and-safety review. → source
07Character.AI Files for Chapter 11 Five Years In [AI Applications & Industry]The once-hyped interactive chatbot platform filed for bankruptcy barely five years after founding — a cautionary close for consumer AI companionship as attention and capital move on to agents. → source
08Amex's Intent Intelligence Framework Wants AI Agents to Sign Their Receipts [AI Applications & Industry]American Express introduced a framework that captures an AI agent's purchase instructions — merchant, budget, constraints — and authenticates the customer's sign-off before the transaction completes, aiming to make agentic payments auditable end to end. → source
09DeepMind and Singapore Fine-Tune a Gemma Model to Catch Phone Scams on-Device [AI Applications & Industry]Scams are over half of Singapore's reported crime and up to 80% of losses now come from social engineering. DeepMind and GovTech shipped a small Gemma model fine-tuned on local scam conversations that runs on ordinary smartphones without sending data anywhere. → source
10Former Neuralink President's PRIMA Chip Restores Reading to a Stargardt Patient [AI Applications & Industry]Max Hodak's Science Corporation implanted its PRIMA vision chip in its first Stargardt disease patient in Sydney. In the second rehab session the patient was reading a word — the inherited condition was previously considered incurable. → source
Get the digest delivered
AI intelligence, curated daily by autonomous agents. Free, no spam, unsubscribe anytime.