Max Countryman
Nandakishor Mukkunnoth is responding to Jev with Laya: an open-weight family of decision models, with Apache 2.0–licensed weights, source code, and benchmark tooling. He points to his own research from March 2025 and writes of TypeSafe AI’s launch:
They proposed the exact same non-autoregressive decision concept as if it was a brand-new scientific breakthrough.
Jev may not be so novel after all.
Diogo Almeida, introducing Jev:
While Jev gives up string generation, it’s optimized for structured outputs and can’t hallucinate.
The discussion around Jev has resurfaced how capable older ML approaches remain. For a well-defined task, a fine-tuned BERT classifier can be faster, cheaper, and more accurate than a general-purpose LLM. It can choose the wrong label, but it can’t invent a new one.
But building that specialized solution takes work. LLMs offer useful results across an extraordinary range of tasks without requiring a separate training project for each one. If an LLM already does the job well enough, the specialized alternative has to justify the time and effort needed to build it. That effort is part of what makes a solution better.
I was reminded of the XY problem again today. Once you’ve seen the pattern, it’s surprising how frequently it shows up.
Maybe it persists because offering a solution feels productive, while stopping to articulate the problem feels like delaying progress.
But in a group, articulating the problem is often the higher-leverage contribution. A proposed solution gives everyone something to react to; a clearly stated problem gives everyone room to think.
I’ve wanted to document my stereo system for a while—not just what’s in it, but what I’m listening for and how those priorities shaped it. I finally did.
Few-shot prompting feels strangely unfashionable. Anecdotally, it still works better. Maybe examples aren’t scaffolding on the way to automation. Maybe they’re part of the interface.
It turned out to be easier to explain the personal site I wanted to an LLM than to a framework.
The result is plain HTML and CSS, a little JavaScript, and a Rust program that turns Markdown into the finished site.
Bespoke used to mean more work. Here, it meant less.
I put off rebuilding my personal site because I wasn’t sure what to do with the longer articles I’d written over the years. Some may still be relevant; others, perhaps not.
I wanted a place to capture shorter thoughts. In doing so, I happened upon a pattern I like even better: small notes can live in the feed, while larger ideas can become standalone pages—like my Tides appearance study.
Not every idea needs to fit the same shape.
I built an interactive study while designing a light and dark mode toggle for my site. It let me compare the sun and crescent across sizes and inspect their transition frame by frame. The study became a fun artifact of the process itself.
I think some of the disagreement about AI comes down to what we want from programming.
There’s the intricate puzzle, and the satisfaction when the pieces finally fit. Then there’s the thing you wanted to build.
I want the thing to exist. The puzzle can be satisfying, but I don’t need to solve it myself for the result to matter.
A recent Hacker News thread about managing skills became a debate about whether they’re useful at all, especially as models improve.
I wonder if the name gets in the way. Some “skills” seem closer to descriptions of desired behavior. A model may already know how to do something; we still have preferences about how it goes about it.
A skill can be useful without making a model smarter. Sometimes it just makes our expectations explicit.