← The Brief
69 ·

Twelve models in twenty days

Releases are now arriving faster than anyone can evaluate them. That changes what a model decision even is.

What caught my eye

Roughly a dozen models shipped in the first twenty days of this month, from seven or eight different providers. I want to be precise about why that matters, because "the pace is accelerating" is a boring observation on its own. The problem is that a serious evaluation of one model against your own task set takes a week or two of real work. The release cadence is now faster than that. So for the first time you are structurally unable to have an informed opinion about the current frontier, and anyone who tells you they do is describing vibes. The practical consequence: stop trying to pick the best model. Build the harness that lets you swap models in an afternoon, and treat the choice as reversible.

One model to watch

Z.ai shipped GLM-5.3 on the 14th and GLM-5.2 Turbo on the 17th. Three days apart, which tells you how they weigh capability against cost. Google put out Gemini 3.7 Flash on the 13th, ByteDance released Seed 2.1 Turbo on the 10th. What every one of them has in common: multimodal input is the default now, not a premium tier. If your product still assumes text in and text out, your competitors are about to ship something you cannot match on inputs alone.

The regulation I'm tracking

Three weeks into EU AI Act enforcement and the interesting question has shifted. It is no longer whether the Commission can act. It is which provider gets used to set the example. My read is that the first action will be about documentation rather than model behaviour, because documentation gaps are trivially provable and behaviour is not. If you serve EU users, go and look at your training-data summary this week and ask whether it would survive being read out loud by a regulator.

One thing I'm building

Spent the week on the apply worker in Tessera, which submits job applications through real browser automation. Two independent gates have to be open before anything gets sent in a user's name, and I wrote the test that proves the open case genuinely submits. That last test is the one that matters. A safety stop you have never seen fire is indistinguishable from dead code.

Get the next one

One email a week. No noise, easy unsubscribe.

← Nº 68
The open-weight frontier stopped being a compromise