Models
The pick-from-options model is ChatGPT Luna tuned for speed. Independent decision indexes rank it 8th and 14th, behind TypeSafe's Jev and Liquid's D1 Ultra.
Why it matters: If an agent step is really a classifier (route, approve, label), a decision model can cut a one to three second LLM call to about 150ms, so benchmark it against Jev on your own workload before you swap.
Deep Dive · FridayModels
Stories that change something.
Models
Why it matters: If an agent step is really a classifier (route, approve, label), a decision model can cut a one to three second LLM call to about 150ms, so benchmark it against Jev on your own workload before you swap.
Tools
Why it matters: Skills you write once now sit under neutral governance, so a skill folder is a safer bet as the portable format for team workflows across Claude, Cursor and Copilot.
Tools
Why it matters: Model-found bugs now arrive faster than humans can triage them, so maintainers of widely used packages should decide now whether to enroll and how they will vet machine reports.
Tools
Why it matters: Attacks driven by agents compress the detection window to minutes, so stolen service credentials and missing MFA on vendor integrations are now emergencies, not backlog items.
Research
Why it matters: Review is now the bottleneck, so teams adopting agents should invest in review capacity, smaller PRs and automated checks or the extra code just piles up in the queue.
Models
Why it matters: A design-native model with an API and MCP hook is a cheaper route to text-heavy graphics than general image models, and its layered output fits real editing workflows.
Tools
Why it matters: Pasting passwords into agent prompts leaves copies you cannot control, and this gives a per-login approval pattern worth copying for any agent you build that touches user accounts.
Tools
Why it matters: Custom post-training of open-weight models no longer requires running your own GPU cluster, which lowers the bar for tuning a small model on your own agent tasks.
Hype we checked so you don't have to. How we judge
Microsoft's AI 'beats doctors'? Only generalists, and on cases meant for specialists
What we checked: The Verge reports the Science paper's 83% versus 36% and 22% figures on Microsoft's own SDBench; TechCrunch notes specialists were excluded and that GPT-5 alone scored 78.9%, with DxO adding about four points and little on emergency cases.AI agents 'started playing' on an island, after a prompt told them they were curious
What we checked: TechCrunch's report on the arXiv paper notes the agents' master prompt called them 'naturally curious' and quotes Berkeley cognitive scientist Eunice Yiu saying the paper cannot show whether they are genuinely playing.The Switch 2 was 'hacked with AI'? One modder credits Codex for a tethered jailbreak
What we checked: The Verge's report attributes Codex's role to modder pepsipu alone, describes the jailbreak as software-only and tethered, and notes its full impact is still being debated by the Switch 2 hacking community.OlmoEarth v1.3 'state of the art' for planet embeddings? On Ai2's own 27 tests
What we checked: Ai2's Hugging Face post reports the 20 of 27 wins and introduces the PLANTEUR benchmark alongside the model; we found no independent leaderboard confirming the state of the art claim, though weights, data and code are open for anyone to check.Signal only. One email, weekday mornings.