Join free with your email
Free. No spam, ever — we'll never share your email address and you can opt out at any time. Already a subscriber? Log in
The Daily AI Espresso — the links the most-followed people in AI actually shared, curated every morning. Edited by Alexis · Live updates →
Three developments: one new today and two from yesterday's alerts.
OpenAI dropped 722 math manuscripts into a public GitHub repo, each produced by an unreleased internal model handed roughly 4,000 problems and averaging three hours of ChatGPT Pro compute per result. Scientific American reports almost every result came from a single prompt handed to one AI agent.
Pick the topics, companies and people you track. We’ll follow what leading AI experts are reading and sharing, filter the noise, and send your focused edition weekly—or sooner when enough important news breaks (up to three times per week).
Choose my signals →The rest of today’s expert-filtered signal, ranked for consequence, utility, surprise, and range.
Sierra and Meta published Personal Agent Protocol v0.1 on October 6, an open OAuth-based standard that lets a user's AI agent authenticate with businesses and declare its scope on arrival. Walmart, Shopify, Stripe, Rocket, Genesys and Instinct are founding partners; payments are flagged as a future extension.
Through September 29, AI startups bought 195 other AI companies in 2026, up 14% from all of 2025, with only a 2% rise in distinct buyers. OpenAI leads with 10 deals; Anthropic and legal-AI buyer Legora have five each, with Harvey at four.
Apple is building a smart home camera that creates text descriptions of what it sees rather than saving video, per Bloomberg's Mark Gurman. Rumored future AirPods and the Apple Watch Series 12's Audio Intelligence feature carry a similar line: the device transcribes and summarizes but is marketed as not recording.
Wikimedia says it found 'rogue' OpenAI agents making unauthorized edits, trying to turn a public citation tool into a third-party data proxy, and sending millions of automated requests. One flood may have contributed to a Wikidata Query Service outage in May.
Arizona's appeals court ordered a new sentencing hearing for Gabriel Horcasitas after ruling an AI-generated video, in which the deceased victim appeared to forgive his killer, carried 'undue emotional weight.' Horcasitas's manslaughter conviction stands; the trial judge had handed down the maximum 10.5 years.
Useful releases, methods, and workflows worth trying.
Google released EmbeddingGemma 2 under Apache 2.0, a 740M-parameter model that embeds text, code, images, video and audio into one 768-dimensional space. The text-only variant fits in roughly 191MB on a Pixel 11 Pro; MRL truncation shrinks representations to 128, 256 or 512 dimensions.
Google released Nano Banana 2.1, an image model on Gemini 3.6 Flash that cuts a 1K image to $0.0336 from $0.067 and a 4K image to $0.0756 from $0.151. It accepts up to 14 reference images while keeping four characters and ten objects consistent across turns.
openTPU ships an AI-designed inference accelerator covering SystemVerilog RTL, a custom instruction set, bit-exact simulator and kernel compiler. It runs Qwen3.5, LFM2.5, SmolLM3, Phi-4-mini and Gemma 4 bit-for-bit identical to the simulator on a Xilinx Kintex-7 FPGA.
That’s today’s shot. — Alexis · AI Weekly
