Dagens Vibes — 13. oktober 2026

Fra X-feedet

Muse Wonder: prøv den rigtige model

Meta lancerer Wonder og Wonder Full til agentarbejde; Full findes kun i Muse Code, ikke API’et. Theos praktiske erfaring er positiv — han bruger Wonder til review og arkitektur — men hans topplacering gælder specifikt Full.

Theo - t3.gg@theo

I should specify that this is Muse Wonder Full. Standard Wonder (the one over API) is significantly worse. These are very different models. The naming sucks. The restriction of Full being subscriber-only sucks even more.

♥ 706 · ↻ 7 · 💬 51 · 🔖 41

https://x.com/theo/status/2109784568672985520

Riley Brown gik tilbage til Codex og Claude Code efter fem timer. Første vibe check er splittet; det cirkulerende 52 %-benchmark er desuden trukket tilbage af Dan Shipper på grund af en fejl i kørslen.

Riley Brown@rileybrown

Used Muse Wonder for 5 hours. Was gonna make a video. No need, back to codex and claude code.

♥ 294 · ↻ 2 · 💬 63 · 🔖 21

https://x.com/rileybrown/status/2109832818536890732

Agenten bliver hængende efter push

Amp sender PR-kommentarer, reviews og CI-resultater tilbage til tråden, der skrev koden. Arbejdet fortsætter med den oprindelige kontekst, frem for at starte forfra med “her er en fejl”.

Tim Culverhouse@rockorager

Amp now gets GitHub reviews, comments, and CI results in the thread that wrote the code. Create a PR from an Amp thread and it will automatically subscribe. Never leave the Amp app again. https://ampcode.com/news/github-listeners https://t.co/tVT00PZTaE

♥ 119 · ↻ 7 · 💬 11 · 🔖 24

https://x.com/rockorager/status/2109735737356939510

DeepSeek lokalt med en ordentlig målelog

Sudo rapporterer 103 tokens/s på kode med DeepSeek V4.1 Flash på to DGX Sparks. Repoet viser målemetode, kvalitet mod originalmodellen og de forsøg, der gik galt — mere interessant end endnu en bar uden akser.

Sudo su@sudoingX

i took deepseek v4.1 flash from 39 tok/s to 103.0 tok/s on 2x dgx sparks in under 4 days this is every step that moved it 🐋 > oct 8, the build i started on, 39 tok/s > oct 9, my own weights serving for the first time, 60.1 tok/s > oct 10, my two cuda kernels + 5 draft tokens a round, 77.1 tok/s > oct 10, draft depth picked every round + a faster all-reduce, 85.6 tok/s > oct 11, overlap + fewer bytes to stream, 99.9 tok/s > oct 11, a 4-bit output head + a re-measured cost table, 101.0 tok/s > oct 12, the drafter's experts on my kernel, 103.0 tok/s every bar on the chart is the code prompt, one stream, greedy, dspark spec decode on, measured on my 2x dgx sparks with the same client, except the 39 tok/s, which is the start build's own number. near 100 tok/s it turns into a fight for single digits, the last two steps add 3.1 tok/s together. the chart doesn't show what broke in between, my first bit plan died at layer 1, my first build failed ifeval and i re-encoded all 15,360 experts, one spark hard hung five times. the full grind, every dead end with its number, goes to subscribers post soon.

♥ 89 · ↻ 14 · 💬 8 · 🔖 63

https://x.com/sudoingX/status/2109689984940957784

Tests skal fastholde adfærd

Kyle Mistele bruger git-historik til at finde tests, der hele tiden skal omskrives. Høj churn er et spor til dårlige fixtures eller tests af intern struktur — en undersøgelsesmetode, ikke en automatisk sletteknap.

Kyle Mistele 🏴‍☠️@0xblacklight

The other thing that’s interesting here (I can post the loop but you can also use our design-control-loop skill to do this) Is that our loop that’s pruning tests is primarily looking for high-churn test files by checking the git history, biased by recency The more often a test file changes the less useful it tends to be Good tests record invariants / ground truth & establish correct behavior If a test file changes a lot, EITHER it needs better fixtures so the test file doesn’t have to worry about signatures changing that are orthogonal to the test cases OR more often, high churn means that it’s making your code harder to modify and probably isn’t establishing a ground truth about how the code works And should be rewritten or deleted

♥ 25 · ↻ 2 · 💬 3 · 🔖 32

https://x.com/0xblacklight/status/2109715875464871983

Agenten på dine egne maskiner

Dax prioriterer koordinering mellem brugerens devices før cloud i sit kontoløse produkt. Et ret direkte ekko af spørgsmålet om én assistent på flere laptops.

dax@thdxr

pretty much every other agent product is well positioned to pursue running agents in the cloud we're going to keep exploring the extent of coordinating them across all your devices since we're primarily an account-less product will slowly work up to cloud but will be last

♥ 542 · ↻ 5 · 💬 85 · 🔖 39

https://x.com/thdxr/status/2109528502802538641

Vite: konkret Windows-fix

Vite offentliggør flere dev-server-advisories, herunder command execution via visse editors på Windows. launch-editor-advisoryen beskriver et filsti-tjek, der kan omgås og kombineres med en escaping-fejl.

Vite ⚡@vite_js

We've published multiple advisories affecting the dev server. The main one is an arbitrary command execution when using some editors or through launch-editor on Windows. Update now to the patched versions: - Vite 8.3.3, 8.2.4, 8.1.6, 7.3.7, 6.4.4 - Vite+ 1.1.0 Report links 👇🏻

♥ 262 · ↻ 35 · 💬 3 · 🔖 30

https://x.com/vite_js/status/2109573811465023661

Nyhedsbonus

Python: mindre opstart, bedre profiler

Simon Willisons aktuelle gennemgang af Python 3.15 fremhæver eksplicitte lazy imports og en sampling-profiler, der kan kobles på en kørende proces. Især nyttigt til CLI-værktøjer, hvor ét tungt plugin gør alle kommandoer langsomme.