Dagens Vibes — 26. oktober 2026

Fra X-feedet

T3 Code: del tråden, behold kontrollen

Theos nye nightly-feature lader en gæst bidrage via et link, mens ejeren bestemmer, hvad der gennemføres — inspireret af fælles takeaway-bestilling. Eric Provenchers førstehåndstest peger på den anden styrke: et API, hvor man kan bygge egne arbejdsgreb uden at genkompilere klienten.

Theo - t3.gg@theo · 25/10 07:11

New beta feature in T3 Code: "share with guest" I'm so tired of copy pasting from a thread, sending it to a coworker, waiting for them to respond, screenshotting the context and sending it back to my agent. Now I just send them a link. It's a silly place to start, but I was heavily inspired by the Uber Eats/Doordash group order flow. If I send someone a link to my group order, they can add whatever they want, but ultimately I decide if the order goes through or not. Realized that flow would solve at least half the use cases I have for "shared threads", so it's now built into T3 Connect in T3 Code (nightly) Very curious how you guys feel about this flow. It's still early and I expect to adjust it a bunch (I left the ui bad so @maria_rcks will come in to fix it sooner) Also s/o Opus 5.5 for making this lil video about the feature :)

♥ 1.638 · ↻ 29 · 💬 174 · 🔖 362

https://x.com/theo/status/2114238361967988747

eric provencher@pvncher · 26/10 00:34

I’ve been playing a bit with t3 code to understand the hype, and I have to give props to them for making it possible to slot in custom client code with their clean api. You can add features without even recompiling the interface or mobile app! You can do fun things like make a cli the model can call to trigger compaction, or have a programmatic way to steer any running lane when a build completes. I think hackable tooling around your harness is underrated and it’s very cool it just uses the codex app server underneath. Kudos to @theo and his team

♥ 230 · ↻ 5 · 💬 33 · 🔖 63

https://x.com/pvncher/status/2114500865344057499

Claude efter en uge: modellen er god, limen knirker

Peter Yang følger op med en konkret fejlliste: nye browserfaner, usynlige manuelle rettelser, forsvindende remote-tråde og en Duet-check-in, der ikke vil ændre frekvens. Det interessante vibe check er produktet omkring modellen — og hans ønske om lidt mere rygrad, ikke mere medløb.

Peter Yang@petergyang · 25/10 04:49

Ok here it is after using Claude all week: Claude Code 1. Browser use opens a new tab always, even when I ask it to use my existing open tab. 2. Claude Code can't see my manual edits in the email edit widget so I can't just make the edits and hit send. This works perfectly on Codex. Mobile 3. Remote Control threads vanish after a few idle hours. I want all my threads available if my computer is on. 4. On mobile, sometimes when I open it a thread it scrolls from the top for 5-10 sec. Open at the latest message instead (this may have been fixed). Projects 5. Can't attach images in project chat 6. I'd love to see a project's active threads in the left sidebar for easy navigation 7. Local projects can't seem to see my local skills Duet 8. It refuses to change frequency for the default check-in task. 5x a day is too much. 9. Voice only seems to work in the Everything channel vs all channels. 10. I like how a dedicated app makes the UX simpler however one of the primary value add of Dots is it can orchestrate all my local and cloud threads. Can Duet do this? Having channels, projects, threads all exist as separate things is going to lead to complexity down the line. Model 11. It is very good. But it still lacks backbone and is too easy to flip to agreeing with me as soon as I push back. I wish it was a more principled thinker. 12. Sometimes tends to be overeager in doing stuff. Happy to follow up more on private channels. Thanks for listening, Claude team. 🫶

♥ 63 · ↻ 1 · 💬 11 · 🔖 18

https://x.com/petergyang/status/2114202713567195254

Mindre ritual, bedre beviser

Dex slettede sin lange browser-testskill, fordi nyere modeller klarede jobbet uden den. Thorsten Ball rammer den vigtigere instruktion: aftal, hvad der tæller som bevis — en grøn isoleret komponenttest er ikke det samme som en fungerende integration.

dex@dexhorthy · 25/10 18:37

ask the model to do something and see if it gets it right. Eg we used to have a lengthy “how to test in the browser” skill that covered local dev, agent-browser CLI, cdp, etc for recent models I deleted the skill. I just say “test in the browser” and it does everything that was in the skill and more

♥ 11 · ↻ 0 · 💬 1 · 🔖 14

https://x.com/dexhorthy/status/2114411085289894314

Thorsten Ball@thorstenball · 25/10 21:19

You need to calibrate agents on what counts as true evidence. Example: local in-memory db does not count as evidence for production performance. It should use real production canaries instead. Exaggerated, of course, most models won't make that exact mistake but there are some like it. Same goes for "I tested it in a single storybook" -- that's not evidence it works fully integrated, etc

♥ 10 · ↻ 0 · 💬 2 · 🔖 10

https://x.com/thorstenball/status/2114451847012376942

Qwen-billedredigering på to trin

Linoy Tsabans LoRA skærer Qwen Image 2.1 fra 20 trin til to; hendes redigeringsdemo er ifølge opslaget ikke speedet op. Modelkortet viser også skitse → foto — interessant for tegneleg. Licensen begrænser brugen til ikke-kommerciel forskning og evaluering.

Linoy Tsaban@linoy_tsaban · 25/10 22:17

Trained a 2-step distilled LoRA so I could run Qwen Image 2.1 in realtime⚡️ now testing on my Wednesday benchmark 🐦‍⬛ *edits not sped up LoRA: https://huggingface.co/linoyts/Qwen-Image-2.1-DMD2-2steps Space: https://huggingface.co/spaces/linoyts/qwen-image-2-1-realtime https://t.co/nBGXxe84MY

♥ 76 · ↻ 5 · 💬 2 · 🔖 84

https://x.com/linoy_tsaban/status/2114466331638837721

Lockjaw: agenten får pladsholderen, ikke nøglen

Et præcist feedfund til credential-diskussionen: Lockjaw indsætter hemmeligheder ved udgående HTTPS-kald, efter kontrol af afsender, destination og godkendelse. Det er en eksplicit gateway, ikke en firewall for agentens øvrige trafik; repoet markerer deploysporene som endnu ikke live-afprøvede.

joel ⛈️@joelhooks · 24/10 19:51

lockjaw is secrets egress using Rat King 👑 AT proto agent network for identity/trust both are Durable Object (celld) native and work so fuckin good @alchemy_run and @EffectTS_ obv 🐀 https://github.com/joelhooks/lockjaw

♥ 49 · ↻ 3 · 💬 3 · 🔖 52

https://x.com/joelhooks/status/2114052198443475417

Dit gamle Pinterest-rod er et smagsatlas

Claire Vo har gjort 5.000+ Tumblr-opslag og pins søgbare efter farve, motiv, æstetik og stemning. Gemini tagger samlingen, som bliver til klynger og kvarterer — en personlig AI-app med et mere interessant mål end endnu et chatvindue.

claire vo 🖤@clairevo · 25/10 17:40

upgraded my taste atlas so i can filter by color, motif, aesthetic, mood, year, topic across 5000+ tumblr posts + pinterest pins spanning 10 years https://t.co/IarUmS7AmE

claire vo 🖤@clairevo · 24/10 01:18

who needs a company brain? build an aesthetic galaxy 2 decades of tumblr + pinterest + instagram → gemini tagging of images, subjects, styles → build clusters and neighborhoods → explore by year, theme, vibe, or topic https://t.co/jIsOLAHlFp

♥ 71 · ↻ 2 · 💬 8 · 🔖 94

https://x.com/clairevo/status/2113772088154829144

♥ 235 · ↻ 4 · 💬 25 · 🔖 230

https://x.com/clairevo/status/2114396599791358013

Nyhedsbonus

Billigere modeller er også et konkurrenceproblem for laboratorierne

I en ny analyse fra søndag gennemgår Max Woolf prisfaldene hos OpenAI og Anthropic og peger på åbne modeller og konkurrerende inference-udbydere som prispres. Det er hans forklaring, ikke en garanti om fremtidige priser — men cacheprisen fortjener lige så meget opmærksomhed som input og output.