Dagens Vibes — 22. oktober 2026

Fra X-feedet

Claude Duet: en kollega, der selv tager kontakt

Anthropics Opus 5.5-agent har morgenbrief, baggrundsarbejde og egen cloudbrowser; previewet er for Pro/Max-brugere i USA via venteliste. Efter to ugers brug kalder Dan Shipper den et reelt pro-værktøj — men den kan ikke dirigere hans lokale Claude-sessioner, og voice er kun transskription.

Dan Shipper@danshipper

BREAKING: @AnthropicAI just launched Duet, its always-on personal agent. It’s the first personal agent I’d call a pro tool. We’ve been testing it for two weeks. I’ve gone from mostly using ChatGPT to roughly 50/50 between ChatGPT and Claude—the most time I’ve spent with Anthropic in close to a year. • It lives in its own app, separate from Claude, on iOS, Android, and the web. This makes it feel clean and easy to understand. • It has channels for specific areas of your work and life rather than just a single chat • It has a todo list so you never lose what you’re working on and you can always see statuses • It allows you to easily write and edit documents, slides, and more all through the app. • It’s not a chief of staff yet. It can see the Claude sessions on my computers but can’t direct them. Its email drafts are worse than dot’s, and voice is transcription only—you can’t have a spoken conversation with it. My verdict: the vibes are good, and I’m opening Duet every day. It’s worth trying. Check out our agent comparison page: https://t.co/fZyTorNK4p Full Vibe Check coming soon on @every

♥ 604 · ↻ 29 · 💬 74 · 🔖 440

https://x.com/danshipper/status/2112997559463506059

Claire Vo giver design og onboarding topkarakter, men savner personalisering og fart. Proaktivitet er kun en fordel, hvis man husker, at assistenten findes.

claire vo 🖤@clairevo

my tl;dr on claude duet +++ design - no notes, v mindful, v demure ++ onboarding - clean and easy ++ models are good + connects w my other stuff + artifacts on mobile, what a time to be alive + channel concept - good design, but not good use of personalization (i created a sales channel and it asked what the channel was for) -- SLOW -- I forgot it

♥ 27 · ↻ 2 · 💬 1 · 🔖 3

https://x.com/clairevo/status/2113004851265519717

Vercel gør det eksisterende harness durable

Pi, Codex, Claude Code og OpenCode kan kobles til en vedvarende log og kø i stedet for at blive genopfundet. Afsluttede kald genbruges efter et nedbrud; afbrudte værktøjer behandles efter deres replay-politik — en uklar betaling bliver altså ikke bare sendt igen.

Pi@pidotdev

We have huge respect for the engineering at Vercel and are excited they are building with Pi-Durable

♥ 830 · ↻ 24 · 💬 11 · 🔖 253

https://x.com/pidotdev/status/2112910333744431146

Vercel@vercel

Introducing 𝚍𝚞𝚛𝚊𝚋𝚕𝚎-𝚑𝚊𝚛𝚗𝚎𝚜𝚜. Run Pi, OpenCode, Codex, Claude Code, eve, fx, or your own harness durably on Vercel. With 𝚍𝚞𝚛𝚊𝚋𝚕𝚎-𝚑𝚊𝚛𝚗𝚎𝚜𝚜, agents automatically resume work after failures and run for hours, days, or months. https://vercel.com/blog/making-every-harness-durable

♥ 1,357 · ↻ 101 · 💬 121 · 🔖 1,129

https://x.com/vercel/status/2112653833382027729

Flint: en åben kodemodel med Pi i træningen

Recursives første model har 39,8B parametre, 2,5B aktive og Apache-2.0-vægte. Den er trænet med Pi-harnesset og kan prøves gratis via Radius; Unsloths 18,5 GB-kvantisering passer i 24 GB RAM ved kort kontekst.

Pi@pidotdev

@Recursive_SI shipped its first model, Flint! Exciting to see they did extensive RL training on the Pi harness, so it works well out of the box. 40B MoE (2.5B active) for fast local coding. Apache 2.0. Available now for free in Pi via Radius.

♥ 466 · ↻ 30 · 💬 19 · 🔖 234

https://x.com/pidotdev/status/2112978311252979761

Richard Socher@RichardSocher

Introducing Flint, @Recursive_SI’s first model: a fast local coding model built with our automated research system, the Eureka Machine. • Best-in-class fast local coding • Up to 2x faster than comparable models: 125+ tokens/sec on a MacBook Pro, 150+ on RTX Spark • Built from scratch in under 4 months Private, unmetered intelligence on your own machine. Open weights available now on @huggingface, run it locally on your favorite framework or try it for free on @opencode or @pidotdev.

♥ 5,658 · ↻ 600 · 💬 460 · 🔖 5,378

https://x.com/RichardSocher/status/2112944385415426249

Første temperaturmåling fra 0xSero: tæt på Qwen-27B, men 3–4× hurtigere. Et lovende brugerindtryk, ikke en dokumenteret sammenligning med opgaver og målemetode.

0xSero@0xSero

New best model to run on 64gb Mac’s/sparks/slower GPUs It is very close to qwen-27B but 3-4x faster

♥ 157 · ↻ 7 · 💬 16 · 🔖 77

https://x.com/0xSero/status/2112968871930163314

Richard Socher@RichardSocher

Flint is the best-performing model in its class (MoE models under 50B) on agentic coding benchmarks like SWE-bench Pro v2, Terminal-Bench 2.1, and SWE-bench Verified. It also rivals similarly sized dense models that are much slower to run. https://t.co/8oX529mkoT

♥ 291 · ↻ 17 · 💬 6 · 🔖 73

https://x.com/RichardSocher/status/2112944391266386133

Opus 5.5 kan vinde opgaven og tabe nattevagten

Jamon Holmgren får bedre resultater end med Sol 6.1 i sit projekt, men Opus spørger konstant, hvad den skal gøre bagefter. Et nyttigt vibe check: kodekvalitet og selvstændig fremdrift er to forskellige egenskaber.

Jamon@jamonholmgren

I've had much better luck with Opus 5.5 recently (not sure why); in the past, Sol 6.1 has performed better. But it seems to have flipped in my project. However, one of the more aggravating things about Opus 5.5 is it's too eager to ask me what to do next. This is an overnight autonomous loop. I need it to pick a lane, we'll fix it in the morning if we need to. Anyone have suggestions?

♥ 22 · ↻ 0 · 💬 32 · 🔖 1

https://x.com/jamonholmgren/status/2112941614138319223

Voxtral møder Gustave Eiffel på en dårlig linje

En konkret opfølgning på Voxtral 3: Noctus kørte en støjende fonografoptagelse fra 1891 og rapporterer 100 af 107 ord rigtige mod INA’s offentliggjorte uddrag. Egennavne og gæsternes dialog drillede; de ekstra 75 transskriberede ord har ingen reference i testen.

Mistral AI@MistralAI

RT @noctus91: Can an STT model make sense of a 135 year old recording? I tested @MistralAI new Voxtral transcription model on a noisy Fre…

♥ 0 · ↻ 16 · 💬 0 · 🔖 0

https://x.com/MistralAI/status/2112807891094716723

Noctus@noctus91

Can an STT model make sense of a 135 year old recording? I tested @MistralAI new Voxtral transcription model on a noisy French phonograph recording of Gustave Eiffel from 1891. It matched 100 of 107 words in INA’s published excerpts (6.5% WER), while transcribing 75 additional words omitted from the reference. It handled Eiffel’s dictation well but struggled with proper names and some guest dialogue. A fascinating test of modern speech recognition on audio from 1891.

♥ 121 · ↻ 16 · 💬 6 · 🔖 21

https://x.com/noctus91/status/2112584933571227745

Se præcis, hvor harnessene er forskellige

Harness Lens sammenligner systemprompts, tools, komprimering og tilladelser side om side med kildehenvisninger til fastlåste commits. Godt modmiddel mod at diskutere “modellen”, når forskellen måske ligger i resten af maskineriet.

Victor M@victormustar

Harness Lens is a nice app to check and compare agent harnesses (claude code, pi, codex,etc...) > system prompts (with side by side diff) > list default tools for each > when do they compact context > what's theur default permissions > more like what edit format they use etc...

♥ 40 · ↻ 4 · 💬 15 · 🔖 25

https://x.com/victormustar/status/2112846503588819035

Nyhedsbonus

Efter PHP: Node har også en vigtig rettelse

Node.js offentliggjorde 21. oktober en sikkerhedsrelease for 22, 24 og 26. Brotli-fejlen CVE-2026-111497 kan give kodekørsel ved dekomprimering af angriberkontrollerede data; rettelserne er 22.23.4, 24.22.1 og 26.12.1. Det er inputbehandlingen, ikke bare en åben Node-port, der er angrebsfladen.

OpenAI flytter ind i Google Clouds modelhylde

Google Cloud får OpenAI-modeller og Codex-workflows via Model Garden; køb kan tælle mod eksisterende GCP-aftaler. Generel tilgængelighed er planlagt i de kommende uger. Modelkrig foran, fælles kasseapparat bagved.