Models
About 42B parameters fire per token, with a 1M context and text, image, audio and video input. Meta, not a leaderboard, calls it the top US open model.
Why it matters: A frontier-scale coding model you can download, fine-tune and run on your own hardware changes the build versus rent math for any team wary of API lock-in.
Deep Dive · FridayChips
Microsoft rebuilt its inference chip around moving fewer bytes.
Stories that change something.
Models
Why it matters: A frontier-scale coding model you can download, fine-tune and run on your own hardware changes the build versus rent math for any team wary of API lock-in.
Policy
Why it matters: Outside evaluators with real access and real publishing rights is the template regulators will point to when they ask every other lab why it has none.
Tools
Why it matters: An assistant that sees every logged-in tab is a new data exposure path, so security teams need extension policies before employees install it.
Research
Why it matters: Proofs that a program ran correctly, written in languages engineers already use, could let agents and vendors show their work instead of asking for trust.
Models
Why it matters: Running guardrails and transcription off one hosted model instead of three services cuts GPU spend and latency for teams building enterprise agents.
Research
Why it matters: It shows agentic coding tools crossing into lab science, though the candidate still has to survive wet lab validation before it means anything.
Models
Why it matters: Bounding boxes in the prompt give product teams repeatable layouts from an open model they can host and fine-tune, without a closed image API.
Hype we checked so you don't have to. How we judge
Maia 300's '40% more tokens per megawatt' never says which GPUs or what tokens
What we checked: InfoQ's report quotes Microsoft's posts: the rival GPU systems go unnamed, tokens are undefined, and Guthrie says 'expects' where Nadella says 'delivers'. We covered the Maia 300 launch itself on Oct 13.Nvidia's 'frontier accuracy at one fifth the size' is measured against its own model
What we checked: NVIDIA's launch post says Super VL matches 'NVIDIA's largest reasoning model' at a fifth the parameters; the 5x speed edge is limited to its own size class.Third Moment's '97% faster backtests' came from a caching fix, not smarter agents
What we checked: OpenAI's Third Moment post: the 97% figure is a backtest going from about 3 hours to 3 to 5 minutes after a cache fix; gains to the trading system go unquantified.The White House 'restored' column photo carries an AI watermark
What we checked: The SynthID result is The Verge's test, not ours: it ran the Truth Social image through Google's SynthID Detector and got an AI flag, while the other photos came back clean.Signal only. One email, weekday mornings.