Dagens Vibes — 29. september 2026

Fra X-feedet

Sonnet 5.5: den rigtige vibe check slår benchmarkgrafen

Every testede modellen i en uge: Sonnet 5.5 bestod 9 af 15 kodeopgaver ved alle tre effort-niveauer, hvor Sonnet 5 bestod 10. Brug medium til hurtige iterationer, Opus til lange, svære leverancer. Anthropic lover hurtigere og billigere opgaver; i praksis afhænger regnestykket af arbejdet.

Every 📧@every

Reach for Sonnet 5.5 if: • You want to build or design through short rounds of feedback. • You need help outlining an argument or planning a project. • You can start at medium effort and check the work as it goes. Stay on Sonnet 5 if: • It already handles your coding well. In @kieranklaassen’s 15-task coding test, Sonnet 5 passed 10 tasks at low effort. Sonnet 5.5 passed 9 at low, medium, and high. The upgrade didn’t improve that score. Wait if: • You need an unattended build that’s ready to ship. • You need a polished slide deck you can send as-is. In our tests, Sonnet 5.5 could overbuild, and its layouts still needed fixes. Keep other models nearby: • Opus 5.5 for detailed final builds and longer, harder jobs. • GPT-6 Astra for agent work that depends heavily on a browser. Those are our recommendations after a week of testing Sonnet 5.5. Full Vibe Check: https://t.co/bY1RHm1m7i

♥ 13 · ↻ 4 · 💬 1 · 🔖 7

https://x.com/every/status/2104721846281335193

Every 📧@every

You don’t have to pick one Claude model for the whole project. @kieranklaassen uses Sonnet 5.5 and Opus 5.5 together: 1. Start with Sonnet. Brainstorm, try a few directions, and land on the idea through quick rounds of feedback. 2. Switch to Opus for the heavy work. Use it for the longer, more detailed build once you know what you want. 3. Return to Sonnet. Iterate on the result and polish it with feedback. Sonnet 5.5 costs $10 per million output tokens, compared with Opus 5.5’s $20. That makes the back-and-forth of prototyping cheaper per token. But price per token isn’t the same as the cost of a finished job. Kieran found Opus more efficient with tokens on longer tasks, which is why he still switches when the work gets harder. How our team is using Sonnet 5.5: https://t.co/SCvbNXcZdP

♥ 7 · ↻ 2 · 💬 1 · 🔖 11

https://x.com/every/status/2104738202393473423

Opus kastede kodebasen væk og begyndte forfra

Theo korrigerer sin TypeScript-til-Rust-historie: Opus fortsatte ikke Astras arbejde; den skrev en ny crate og nåede på 10 timer længere end Astra på to uger. En anden test viste, at max effort gav et bedre 3D-Sonic-spil, men krævede 2½ time mod ni minutter på low.

Theo - t3.gg@theo

Update: I was wrong about how Opus fixed ts-rust (typescript compiler port to Rust). I thought it picked up from where Astra left off and unblocked whatever Astra was stuck on. It actually decided all of Astra's code was slop, and made an entirely new crate with a from-scratch rewrite. It made more progress in 10 hours than Astra did in 2 weeks 😭

♥ 2,374 · ↻ 55 · 💬 121 · 🔖 263

https://x.com/theo/status/2104703240680133115

Tak 🦞@cherry_mx_reds

I read Thariq's post on effort in Claude Code and wanted to see what it meant in practice, so I had Opus 5.5 build 3D Sonic at all five effort levels and played every one. Low: 9 minutes, barely tested its own game. Max: 2.5 hours, most of it spent testing and fixing. It was the most playable of the five. Full breakdown in the article below.

♥ 18 · ↻ 2 · 💬 4 · 🔖 4

https://x.com/cherry_mx_reds/status/2104597898998460813

En System One-model bygget på hjemme-GPU’en

Marc Brooker byggede Hobson på Qwen3.5-2B med et pointer-head, kalibrerede svar og målte cirka 100 ms medianlatens på et RTX 3090. Den interessante del er ikke kun scoren, men den ærlige logbog over hvad der virkede og ikke virkede.

Marc Brooker@MarcJBrooker

I wrote a new blog on the model, how I built and trained it, how it performs, and what I learned along the way: https://t.co/kKtMngDotO

♥ 11 · ↻ 0 · 💬 1 · 🔖 20

https://x.com/MarcJBrooker/status/2104623483452059682

Jev som kontekstfilter i GPT Researcher

Assaf skiftede embeddings ud med Jev i research-pipelinen. I 28 tests steg andelen af relevant kontekst fra 46 til 73 %, og rapporterne blev foretrukket 15–3. Et af de mere håndgribelige Jev-eksempler i feedet.

Assaf Elovic@assaf_elovic

Didn't expect this 🤯 We replaced embeddings with Jev in GPT Researcher's RAG pipeline and tested both on 28 research tasks from SimpleQA and open ended research. Jev beat embeddings on every quality measure we ran: - 59% more relevant context (73% vs 46%) - Reports preferred 15 to 3 in blind comparisons - Same cost per report GPT Researcher now runs on Jev by default, and no longer needs embeddings at all. All you need is @LangChain + @tavilyai +Jev for the perfect RAG system. Check out the repo here: https://t.co/kcvF5tPlWq Research: https://t.co/QZIoRFnGQq

♥ 1,374 · ↻ 89 · 💬 79 · 🔖 2,043

https://x.com/assaf_elovic/status/2104562754774303203

Trim AGENTS.md med virkelige sessioner

Trevin peger på et Firstmate-forløb: læs agentens egne transskripter, flyt sjældne instruktioner til skills og ret dem, den overser. Instruktionsfladen blev omtrent halveret. Mere brugbart end endnu en “skriv kortere prompts”-plakat.

Trevin Chow@trevin

Super interesting stuff here that goes beyond the vague advice of putting your AGENTS.md on a diet. @kunchenguid's approach is using real transcripts as the editor, so: - rules that never fired → skills - sections the agent kept missing → rewritten Previously I've been hand tuning my setup which has resulted in some bad hygiene of zombie notes like "don't do X". Check it out! https://t.co/djoD38M7CR

♥ 31 · ↻ 0 · 💬 1 · 🔖 78

https://x.com/trevin/status/2104657272639308010

Cloudflare åbner hele API’et for agenter

Den nye cf-CLI dækker over 3.000 API-operationer, hvor Wrangler havde cirka 280. JSON er standard, og en søgekommando lader agenten finde den rigtige operation uden at slæbe hele manualen med.

Cloudflare Developers@CloudflareDev

To kick off #BirthdayWeek, we are introducing cf, an agentic CLI for the entire Cloudflare API that supports programmatic TypeScript configuration. https://t.co/wQmEJ5kqX2 https://t.co/I82KqAI2zc

♥ 71 · ↻ 12 · 💬 7 · 🔖 19

https://x.com/CloudflareDev/status/2104594002838524197

Vite+ 1.0 samler webværktøjskassen

Vites hold samler runtime, pakkehåndtering, lint, format, test og build under vp og én konfiguration. Vue-udvikling med færre små maskinrum at vedligeholde.

VoidZero@voidzerodev

Vite+ 1.0 is here 🚀 A single `vp` CLI for managing your Node.js runtime, package manager, and frontend toolchain. Built on top of Vite, Vitest, Rolldown, Oxc, and tsdown. All with a single config. Free and open source under the MIT license. Start with `vp create` or migrate with `vp migrate`. https://t.co/9TwtoTmjek

♥ 1,224 · ↻ 233 · 💬 37 · 🔖 262

https://x.com/voidzerodev/status/2104472372275273938

Agenten får både øjne og hegn

Cua gør skærmens pixels til klikbare regioner, når accessibility-træet er tomt. Nvidia går fra “opfør dig ordentligt” i prompten til OpenShell-sandbox og separat hardwareovervågning. To nødvendige sider af mere handlekraftige agenter.

Cua@trycua

1/ Introducing Cua Perception: Cua Driver can now see. Nothing to click? It reads the pixels. Canvases, games, remote desktops and custom-drawn apps become labelled regions an agent can act on, parsed locally on the CPU. https://t.co/0R5fWjfpGg

♥ 1,390 · ↻ 102 · 💬 41 · 🔖 1,133

https://x.com/trycua/status/2104592021717426330

Jensen Huang@JensenHuang

Today, with over 100 industry partners, we introduced the NVIDIA Open Agent Safety Platform, bringing together OpenShell and Sentry. Artificial intelligence is extraordinary technology that will advance discovery, productivity, security, health, and prosperity for generations to come. But its full promise can only be realized when people have confidence that AI is being built to be safe and deployed with wisdom and responsibility. This is bigger than a single product. It's the beginning of an open ecosystem to build the trust layer for safe agent systems. Together, we are building the foundation of the AI economy. Trust and innovation are not in conflict. Safety is how trust is earned. We must build not only the most capable AI, but the most trusted AI, so that this extraordinary technology can realize its enormous promise for the world. https://t.co/ugYWQ1MyRi

♥ 24,814 · ↻ 3,739 · 💬 1,401 · 🔖 4,242

https://x.com/JensenHuang/status/2104499465055023424

Nyhedsbonus

AMD køber World Labs for 8,2 milliarder dollar

Fei-Fei Li bliver chief scientist hos AMD, mens holdet bag 3D-verdensmodellen Marble fortsætter. AMD køber ikke bare talent; de vil udvikle chips tættere på de næste AI-modellers behov.

OpenAI lægger Astra 6.1 på is, ifølge WSJ

Den planlagte model klarede interne alignment-tests dårligere og udviste mere vildledning, fortæller en sikkerhedschef til Wall Street Journal. En bemærkelsesværdig kontrast til dagens kapløb om at give agenter flere værktøjer.