Aug 2, 2026
Today in brief.
OpenAI's unreleased Astra model cracked ten open math problems
OpenAI says its next model solved ten problems that had each been open for a decade or more, and shipped machine-checkable Lean proofs so you don't have to take their word for it.
DeepSeek V4-Flash is cheap enough that $2 lasts a full day
V4-Flash went GA yesterday; today the receipts landed - a full agent task for seven cents, and local runners now match a March frontier model.
The EU AI Act's labelling rules take effect today
As of August 2, anything you publish in the EU that looks authentic but isn't has to say so - and this is the first rule that touches ordinary creators, not just labs.
A 26,000-student study found AI's learning cost takes two years to show up
Thirty months of panel data on 26,000 students found that heavy AI use cost up to 30 percent of learning gains - and the damage didn't become visible until about two years later.
Nvidia paid $20B for Groq's people and IP without buying the company
Nvidia is paying $20 billion in cash for Groq's inference talent and an IP licence - and deliberately not acquiring the company, which is the part worth noticing.
Gallup says workplace AI adoption jumped 6 points in one quarter
Adoption went from 41 to 47 percent in a single quarter, Gallup's sharpest jump ever - while 56 percent of CEOs report no measurable return at all.
Someone scanned 7.6 petabytes of HuggingFace training data for secrets
Truffle Security scanned every public HuggingFace dataset for live credentials, and the answer to 'are people leaking keys into training data' is yes.
Figure's F.03 climbed stairs for two hours straight with zero real-world training
Figure's humanoid ran stairs continuously for two hours on a policy trained entirely in simulation, transferred zero-shot with no real-world fine-tuning.
Reddit stock fell 23% and the CEO is publicly questioning Google's AI Overviews
Reddit lost nearly a quarter of its value as AI answers eat the traffic that used to arrive as clicks, and its CEO has stopped pretending the deal works.
Two things worth opening: Flint charts and an 8GB post-training kit
Microsoft shipped a visualization language built for LLMs to write, and someone published post-training experiments that fit on an 8GB GPU.
No stories in this category for this edition.