AISep 2, 2026Hugging Face
Fine-tuning a 350M Model for Better Structured Outputs in 100 GRPO Steps
A Hugging Face how-to shows a 350M model jumping from 22.6 to 29.7 percent on IFStruct after 100 GRPO steps — form, not brains, on a free-tier GPU.
The file
Latest
Sourced AI briefs in publish order — newest first. Citeable originals live on the homepage under From Signal Desk.
AISep 2, 2026Hugging Face
Training a coding model to paint watercolours with TRL and OpenEnv
A Hugging Face engineer open-sources the viral p5.js watercolour recipe: RL over taste, a 178-image hand-rated pool, and three reward mixes that actually learn.
AISep 2, 2026OpenAI
Safety overview: GPT-6 Astra
OpenAI says GPT-6 Astra is its most capable broadly deployed model and the first to hit Critical cybersecurity capability—then spends equal space on stronger guards and weaker CoT monitorability.
AISep 2, 2026TechCrunch
TechCrunch Disrupt 2026’s new Real World AI Stage features Nvidia, robots, and extinct animals
TechCrunch is using its own conference blog to sell a second AI stage: robots, defense, edge, and a de-extinction CEO, October 13–15 in San Francisco.
AISep 2, 2026TechCrunch
OpenAI’s new reasoning technique alarms AI safety experts
TechCrunch's Astra piece is about a method, opaque recurrence, that safety people think can eat chain-of-thought monitoring if anyone scales it.
AISep 2, 2026The Verge
Google says its new Gemini 3.8 Flash model ‘works harder’ but might cost more
The Verge's Gemini 3.8 story is the bill: same $0.75 / $3.75 sticker as 3.7, about 40 percent more expensive per task because it talks more.
AISep 2, 2026WIRED
Trump Administration Sides With OpenAI in New York Times Copyright Lawsuit
WIRED's version of the DOJ letter is the one with Joan Didion at the typewriter, a $1.5 billion Anthropic contrast, and a judge who does not have to listen.
AISep 2, 2026WIRED
These Russian Mathematicians Taught AI Models How to Talk to Each Other Without Using Words
WIRED's Mostik profile is a lab demo: weights talking to weights, a 753B–4B hybrid at one-twentieth the cost, and a Fields Medalist as chief scientist.
AISep 2, 2026Ars Technica
Google releases Gemini 3.8 Flash, its third Flash model in six weeks
Ars reads Gemini 3.8 as cadence covering for a missing Pro: three Flashes in six weeks, OSWorld still trailing Claude Opus, Cyber as a niche.
AISep 2, 2026Ars Technica
Trump may be forced to reveal secret rules feds use for AI safety testing
Protect Democracy is suing four agencies for the unclassified AI safety-review framework Trump's White House will describe but will not publish.
AISep 2, 2026NVIDIA Developer
The Modern CUDA Toolbox in Practice: A Step-by-Step Optimization Walkthrough
Nvidia's developer blog turns a broken 6.8-second image pipeline into 23 milliseconds, and the moral is stop writing the kernel CUB already shipped.
AISep 2, 2026TechCrunch
US government sides with OpenAI on issue of training LLMs on copyrighted material
TechCrunch's shorter take is the procedural one: a 20-page brief invoking last year's AI executive order, filed in a court the White House does not run.