Dagens Vibes — 20. september 2026

1/ Introducing CUA-S1: a family of System One Models, small, specialized, and built for computer use. Today we're open-sourcing CUA-S1-FORMS, the first in the family: github.com/trycua/cua

Introducing Jev Model Router for Claude Code This Claude Code Mod lets you use Jev through its direct @typesafeai API or @vercel AI Gateway With every request you send to Claude Code, Jev classifies the subagent model, main model (only at session start to avoid breaking the cache), and effort level Install it with one command: npx claude-code-templates@latest --mod productivity/jev-model-router Full Mod details: github.com/davila7/claude-code-templates/tree/main/cli-tool/components/mods/productivity/jev-model-router

our first two jev use cases: 1. when you comment on a PR from Detail, we have an LLM check for whether the comment is asking Detail to do something (e.g. "remove this test") vs not (e.g. "hey teammate does this look right?") and when we interpret a comment as an ask, we react with an 👀 to ack that we're working on it. this now happens in 400ms instead of 1200ms thanks to jev. 2. when we find a bug, we check if it's a dupe of a bug we've already reported. this can be tricky because unrelated code change can move the "same" bug to a new location, and we don't want to report twice in that case. but two bugs can be nearly identical in structure but in different places, and we want to flag both of them, e.g. the same missing auth check in two different routes. this is now a jev check. in both cases, backtesting showed ~99.5% agreement with models that were 30x as expensive. extremely insightful form factor. i can't stop seeing places to jev up our product.

I was curious about the advantages/speed claims here so I put it to the test. I had claude convert my entire python/rust pixel-fixer tool to bend. Same algorithm: three grid detectors (autocorrelation + cepstrum, run-length lattice fit, self-similarity tiebreak) and the two-stage k-means repack that turns mushy AI "pixel art" into real 1:1 pixels. Result: 18/18 test images give the identical grid, identical sub-pixel steps and the same consensus path as the Rust core. F32 only, no f64 in Bend, and it still didn't flip a single decision. The clean test image reconstructs pixel-for-pixel. Speed: about 4x slower than Rust on the same box. Rust has rayon, rustfft and f64; Bend got ~2x from its fork-join on 6 threads. Honest number, not a blowout either way. The conversion was a real fight though. Bend is affine, lazy and termination-checked: no loops, no mutation, no if, no mutual recursion. Every array read hands the array back, so a 3x3 median is nine chained continuation functions. Fable ended up writing a kernel generator and a topological sorter just to keep compiling. First port used lists everywhere: correct, but 3.6 minutes per image. Switching every pixel stage to Bend Arrays (~1 ns reads) made that 4.5 s. Two of the three real bugs were silent U32 wraparounds. Zero memory or concurrency bugs, which is the point of the language. PROOF.bend gating the build is genuinely cool. Float math is opaque to the checker though, so the numerics could be wrong even if the 'laws' are correct. What would make Bend great for this kind of work: f64, a shareable read-only array, loop/if sugar over what it already compiles to, checked integer arithmetic, and an FFT in the stdlib. With those it'd be a few hundred lines shorter and read like the Rust. Highly recommend checking out Bend, its genuinely interesting and it will only get better I'm sure. Taelin is getting a lot of pushback now but I genuinely couldn't find a single false claim. It does what he says. It's missing some things, sure, but it literally just got released. I didn't try the cuda path, since that would be a pain for the systems I use to run the pixel-fixer. I'm really excited to see where this goes. The proofs and laws genuinely made this conversion easy. Just pointed fable pixel-fixer it and said to convert it from rust to bend.

new post: the senior engineer death spiral a friend just started a big job and asked for some advice. so I braindumped a monologue about a super common failure mode I see with engineers and posted it here, hope it helps whoever it can.

The new "bakeoff" capability in Compound Engineering is so damn useful. Even on smaller tasks, I'm surprised at how much better it gets when a bakeoff is used during ce-plan, or even standalone when I'm prompting my next crazy idea in a repo.

Opening access for developers to build Muse connectors. You bring the API — Muse brings the agent, the browser, and the context of what the person actually wants. People reach your service just by asking for it, and their agent takes it from there. New connectors are live today. Come build with us.

5 things to try with @Muse this weekend beyond the usual email, calendar, and subscription workflows: 1. Set up a personalized feed Muse's Feed is refreshing because it doesn't have AI slop or ads. You can personalize it around the exact topics you care about, plus reminders tied to your life. I love scrolling through it every morning. 2. Fill out forms and comparison shop online Ask Muse to help renew your driver's license, compare car insurance options, or book hard-to-get restaurant and show reservations. It's very capable, especially combined with item 5 below. 3. Monitor your DMs Muse has first-party connections to Instagram, Threads, and Messenger. If you install the desktop app, it can see your iMessages too. Have it monitor your messages for anything you need to follow up on instead of you having to check 4 different message inboxes yourself. 4. Buy and sell on Facebook Marketplace Other agents can do this too, but Muse makes it easy through its first-party integration. Snap a photo, upload it to Muse, and ask it to list the item. You can also have it regularly search for the best deals on something you want and DM sellers directly. 5. Find and talk to human customer support This is the killer feature IMO. Muse can navigate phone trees, wait on hold, and talk to a human customer support rep while representing itself as your agent. It also says it can dial you in halfway through the call, although I haven't been able to get that working yet. By the way, if you don't have access to this, ask Muse to message Meta and request access. I got it a day after doing this. Last but not least, I think Muse's Ideas tab (below) is very well done. The biggest problem with chat interfaces is that most people have no idea what to ask. Ideas solves that by suggesting useful, personalized things Muse can do for you. My only complaint so far: Muse's compaction doesn't seem to work as well as ChatGPT. If I discuss several topics in the main thread, it sometimes gets confused. Beyond that, it's been delightful to use. I can tell the team spent months polishing this.