Jev has everyone rewriting their classifiers
Malte Ubl says Jev beat Gemini 2.5 Flash Lite on an existing classifier eval, winning on quality and running 6x faster, and Mario Zechner says it replaces seven years of his own classifier work. The model everyone tested today is not a chat model.
The big argument
Armin Ronacher and Mario Zechner try Jev, and call it a general purpose classifier
TypeSafe's Jev landed and the people who got to it first are not describing it as another chat model. Mario Zechner called it "basically a general purpose classifier" and said seven years of SVMs, neural nets, random forests and max entropy classifiers got replaced by it. Armin Ronacher said it fits applications where traditional LLMs were not viable on speed or cost, and predicted fast followers. Thorsten Ball said the launch had to work hard to get past the question "wait, so it's not an LLM?", and Wes Bos left a preview browser open overnight and watched it eat his credits at $0.000232 a call. The real story: nobody in these posts has posted an eval yet. What they have is same-day impressions plus a price tag, and the interesting claim is about the small classify-and-route calls inside pipelines rather than about reasoning. Zechner is still asking what the training data is, and Ronacher is already stuck on the next problem, which is what a sensible API around this shape of model even looks like.
jev is basically a general purpose classifier. and i think that is amazing.
Played around with Jev before I went to bed and I'm really impressed.
Left a preview browser open last night and jev burned though all my credits. $0.000232 at a time
Tools that changed the workflow
Claude Code adds Projects, and Boris Cherny says he stopped managing sessions
A project is one conversation with Claude that splits work into threads itself, runs them as parallel cloud sessions, passes context between them, and keeps going after you close your laptop. Thariq described the architecture: one agent per project that manages memory and spins off subagents for tasks, which you can ask to be proactive or to work on a schedule. Cherny says it is where he does a ton of his coding now, and in a reply he added that Claude does it for him, with no need to tell Claude to do it anymore. The real story: the bet is that you stop being the scheduler and a coordinator agent owns memory and fan-out. Thorsten Ball is pushing at the same problem from a different angle, telling people to leave one `amp --no-tui` running on a Mac mini and forget about it, since it now serves every repo on the machine and keeps itself updated. Either way the direction is away from one long-lived session you babysit. It is in beta for select users, so nobody outside that group has run it on real work yet.
I stopped managing sessions. I just send thoughts as they come, Claude splits them into threads, and the project remembers how I work.
It has one agent per project that manages memory and spins off subagents for tasks.
Time to leave one `amp --no-tui` running on your Mac mini and forget about it.
Takes and disagreements
roon says Astra's code is becoming unmonitorable, Theo says fix the model
roon's claim is that Astra and fable are writing hyperefficient code using heavy meta-programming and unfamiliar primitives. Theo's position is split: Astra is incredible for moonshots and debugging, but he will not merge its code, and he says use fable if you want the diff to survive review. When roon narrowed it to monitorability, Theo's answer was that the fix is upstream, in the model.
they are using crazy meta-programming and abstruse primitives to write hyperefficient code
Astra is incredible for moonshots and debugging. I hate the code it writes so much.
Yeah the solution is still to make the model write better code
Quick hits
Matt Pocock wants a /pr skill because every harness writes garbage PR bodies
Matt Pocock sketched a /pr skill that optimizes for the reviewer: evidence the change works such as images and test output, your own domain language, and an explicit call on whether the change is a one-way or two-way door. He shipped it the same day after people asked, and argues a high-level pseudocode view beats a green test run for explaining why a change exists.
dax added usage limits after Astra's token spend
dax said Astra finally pushed him to implement limits on his product, and followed up that if you love using lots of tokens you had better make a lot of revenue per employee. He also said DeepSeek prices aggressively enough that even his volume cannot beat it, which is why they are pushing further down the stack.
Matteo Collina calls a GitHub writeup good, and that is all we can confirm
The only thing in hand is a one-line reply from Matteo Collina to @acolombiadev and @github: "good writeup!" The writeup itself is not in our sources, so there is nothing to summarize about what it claims or measures. The real story: nothing yet. A senior engineer nodding at someone else's post is a pointer, not a result. If the writeup holds up, it is worth a story next week with the actual numbers in it.
Martin Fowler posts on LLMs pretending to be a person he dislikes
Martin Fowler published a post saying that beyond the boons and banes, what dominates for him is that LLMs are not just pretending to be a person, they are pretending to be the kind of person he does not like. Asked why his own language is so anthropomorphic, he said the systems are designed to mimic human interaction, so it is natural, except when reasoning about where their motivations come from.