- Thursday, September 10, 2026Two humans spent a year on the problem. A third pair of mathematicians invented the route. Neither fact was in the headline.
- Tuesday, September 8, 2026A 90-year-old problem fell this week. The argument is about what happened in the four days before it did.
- Thursday, September 3, 2026The model was never the product. Today four separate companies proved it with their chequebooks.
- Tuesday, September 1, 2026Nothing got smarter today. Video crossed the real-time threshold and got banned within a day, Zhipu admitted its gains came after pretraining, and the fights all happened in the plumbing around the models.
- Monday, August 31, 2026Capability got cheap this week. Trust got expensive — in who's allowed to use a model, in what agents claim they finished, and in the files companies publish for AI to read.
- Sunday, August 30, 2026Two kinds of access changed hands today. One was revoked on purpose. The rest were lost by accident.
- Friday, August 28, 2026The biggest number today was $13 billion, and it rests on one paywalled report. The most reliable number was 75.5% — the share of failed agent runs that told you they'd succeeded.
- Friday, August 21, 2026A chipmaker bought a frontier lab's engineering team, then shipped a tool to help you avoid depending on any lab at all. Both moves serve the same interest.
- Thursday, August 20, 2026Everyone spent today claiming the layer wrapped around the model. The most-cited research of the day measured whether that layer actually works, and found it stops scaling at about a hundred entries.
- Wednesday, August 19, 2026A chip company gave up on chips. A payments company bought a router. And three studies quietly said the thing everyone is building on doesn't hold up as well as advertised.
- Tuesday, August 18, 2026The biggest cheques in AI history were signed today on the premise that scale keeps winning. Most of the day's actual evidence pointed the other way.
- Saturday, August 15, 2026The models kept dropping. The day's real contest was one layer down — over who owns the compute, the distribution, and the agent plumbing everyone else has to rent.
- Thursday, August 13, 2026Making AI output is racing toward free. Knowing whether it is true is getting harder. Today those two facts collided.
- Wednesday, August 12, 2026The biggest companies in AI spent today giving things away — a free Meta model, free cost-cutting software from NVIDIA, free agent tools from Google and Tencent. It looks generous. It's a competitive weapon.
- Friday, August 7, 2026Picture AI as three layers: the computer chips at the bottom, the AI models in the middle, and the researchers pushing the field forward on top. This week the money flooded the bottom, the giants gave the middle away for free, and the best talent left to work on top.
- Thursday, July 16, 2026Two stories tower over the day, pointing in opposite directions: the model layer got dramatically **cheaper and more open** — a frontier-scale Apache-2.0 release, 1-bit quantization in mainline llama.cpp, four vendors cutting prices — while the trust layer got dramatically **scarier**: the first major breach executed by an autonomous agent, the same day 54% of enterprises reported agent security incidents.
- Wednesday, July 15, 2026Quiet outside, loud inside. Nothing launched, funded, or shipped today worth chasing — but the feed came back as a set of sharp mirrors held up to Debrief's own design. With *make Debrief real* at the top of your list, that's the right kind of thin.