Dagens Vibes — 15. oktober 2026

Fra X-feedet

Agenten er færdig. Men hvad venter på dig?

Claude Code Desktop får nested sessions, tydelig status og recap-kort, når en session kræver din opmærksomhed. Det interessante er ikke flere parallelle agenter, men mindre bogholderi i hovedet på mennesket.

ClaudeDevs@ClaudeDevs

The sidebar makes it easy to see which sessions are working and which ones need you. Inside a session, a live to-do list follows Claude's plan as it goes, a recap card sums up the turn when Claude needs you, and the side rail shows where it runs, its context and its changes.

♥ 471 · ↻ 7 · 💬 10 · 🔖 38

https://x.com/ClaudeDevs/status/2110445851466485786

Argon efter lanceringen: hurtig kernel, tvivlsom kodeforståelse

Elliot Arledge fik Gemini 4 Argon til at skrive en GPU-kernel på 27× PyTorch-baselinen, før sikkerhedsfilteret stoppede kørslen. Theo melder samtidig om over 50 % faktuelt forkerte påstande i næsten alle sine kodebase-tests: specialiseret optimering og forståelse af et eksisterende system er to forskellige eksamener.

Elliot Arledge@elliotarledge

Gemini 4 Argon [high] on KernelBench-Mega: It wrote a one-launch Kimi-Linear Decode megakernel for RTX PRO 6000 at 27.0x the optimized PyTorch baseline, 3rd of 35, behind Opus 5.5 (35.5x) and Sonnet 5.5 (32.1x). Gemini 3.8 Flash scored 2.7x here. For three hours it climbed the same curve as Sonnet 5.5. Then the session ended mid-profiling, when one of its own requests was blocked by Gemini's safety filter.

♥ 94 · ↻ 2 · 💬 3 · 🔖 18

https://x.com/elliotarledge/status/2110370046308679895

Theo - t3.gg@theo

This isn't the usual "oh it's 5% worse in this way, 3% better in others". In almost every test I ran, Gemini was factually wrong about over 50% of it's claims. This model is a straight up liability. Keep it away from your code.

♥ 412 · ↻ 9 · 💬 25 · 🔖 22

https://x.com/theo/status/2110461582237188207

Dine agenter, dine maskiner — én livscyklus

Factorys OpenCompute skiller agentproduktet fra maskinerne: fælles styring af opstart, pause og nedlukning på tværs af egne runtimes. Det er private preview med planlagt open source; den interessante arkitektur er durable intent og eksplicitte driver-kontrakter.

Factory@FactoryAI

Introducing OpenCompute: a self-hosted orchestrator for agent sandboxes on your choice of runtime. Adopting autonomous software development should not require handing over your valuable code and context to someone else’s cloud. Scale your fleet of remote agents while keeping them in your VPC.

♥ 647 · ↻ 46 · 💬 29 · 🔖 584

https://x.com/FactoryAI/status/2110403273903808988

Mindre “agenten fejlede”, mere “her gik det galt”

Claude Platform kan sende OpenTelemetry-spans til din eksisterende observability-stack. Metadata er standard; prompts og svar kræver et særskilt valg — en fornuftig forskel på fejlsøgning og endnu et chatarkiv.

ClaudeDevs@ClaudeDevs

Now in public beta, Claude Platform can send OpenTelemetry traces from the Messages API, Claude Managed Agents, and the Claude SDKs to your observability stack.

♥ 4,923 · ↻ 264 · 💬 223 · 🔖 2,200

https://x.com/ClaudeDevs/status/2110421732029063352

Den gode personlige agent efterlader noget brugbart

Banteg lod Codex lave en offline rejse-side med kalender, tidszoner og praktiske oplysninger, opdateret morgen og aften. Et lille, vedvarende værktøj slår endnu et langt svar, man skal lede efter i chatten.

banteg@banteg

i used it to make a companion site for a trip, it had the full itinerary and a calendar with a cursor, handled multiple timezones seamlessly, had info like phone numbers, sites and directions always handy, and it even worked offline. then i had codex update it in the morning in case some activities were disrupted or something better came up. also did that in the evening so the next day matches the energy level. overall i loved the experience with sites.

♥ 2 · ↻ 0 · 💬 1 · 🔖 0

https://x.com/banteg/status/2110279973445501318

Nyhedsbonus

GPT-4-generationen får en slutdato

OpenAI varslede i går API-lukning for de angivne GPT-4o-, 4o mini-, 4.1- og 4.1 mini-versioner den 14. januar 2027, inklusive deres fine-tunes. Den anbefalede afløser er GPT-6 Luna.

OpenAI · Den officielle liste

Stemmemodeller bliver bedre til at lade dig tale færdig

Grok Voice 3.0 topper Artificial Analysis’ speech-to-speech-indeks og scorer 97 % på testen af pauser og afbrydelser. Prisen er fortsat $0,08/minut; en mere interessant forbedring end endnu en chatbot med en ny frisure.