Dagens Vibes — 27. oktober 2026

Fra X-feedet

Tre forskningspåstande, der er større end endnu en chatbot

OpenAI offentliggør AI-resultater om bl.a. Kakeya og en afkræftelse af Exponential Time Hypothesis; matematik- og datalogimaterialet inkluderer Lean-beviser. Uafhængige forskere skal nu undersøge og forklare resultaterne — ETH er ikke P vs. NP, og fysikdelen er endnu ikke offentligt fremlagt.

Peter Yang@petergyang · 27/10 03:20

This is the stuff that I’m excited about not the 9th personal agent to manage my emails

Dan Roberts@danintheory · 26/10 22:36

We are announcing three new discoveries in mathematics, computer science, and physics about harmonic analysis, complexity theory, and quantum field theory. https://openai.com/index/building-scientific-understanding-with-academic-communities/

♥ 4.705 · ↻ 735 · 💬 204 · 🔖 1.381

https://x.com/danintheory/status/2114833682469122314

♥ 23 · ↻ 1 · 💬 12 · 🔖 5

https://x.com/petergyang/status/2114905151572263242

GPT-6.1 Terra: den interessante test er det kedelige arbejde

Terra er lanceret til visuel forståelse og computerbrug; API-prisen er $1 ind / $5 ud pr. million tokens. Kevin Kerns førstehåndstest af adminopgaver er mere nyttig end en ny benchmarkgraf: hurtigere end Sol til formularer og dashboards — en konkret erfaring, ikke en dom over alle arbejdsopgaver.

Kevin Kern@kevinkern · 26/10 20:39

Since I had to do some admin tasks, I can confirm it's much faster than Sol at computer use. I can see it being useful for live testing apps on a separate machine (replacing sol for this) and boring tasks like configuring accounts, filling out forms, clicking through dashboards, and other repetitive stuff.

ChatGPT@ChatGPT · 26/10 19:08

GPT-6.1 Terra means business. Our new fast, efficient model pairs sharper visual reasoning and precise computer use with clearer, more natural writing. From decoding technical drawings to drafting proposals, filling forms to editing videos—it’s built for the whole workday. GPT-6.1 Terra is rolling out today to Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex.

♥ 6.246 · ↻ 454 · 💬 333 · 🔖 742

https://x.com/ChatGPT/status/2114781101822714241

♥ 14 · ↻ 0 · 💬 4 · 🔖 5

https://x.com/kevinkern/status/2114804067729891794

Cursor gør agenten til et testbart kodeprojekt

Prompts i Markdown, værktøjer i TypeScript og evals i samme projekt, hostet eller på egen infrastruktur. Trevin fremhæver den gode detalje: test også, hvad agenten gør — og lad den ikke løse en fejl ved at gøre testen dummere.

Trevin Chow@trevin · 26/10 21:30

Glad to see evals built right into the @cursor_ai Agent SDK. You can assert on tool calls too, so a test fails if the agent approves a PR it was only supposed to review. One gotcha I've found is agents loovvve to change existing evals to get it's proposed changes to pass. It was interesting to see this called out in their hillclimbing docs: "Don't loosen an existing eval so that a new change passes. If an older eval starts failing, the change broke something that used to work, and you should fix the change instead." SDK Docs: https://cursor.com/docs/agent-sdk

Cursor@cursor_ai · 26/10 19:58

The Agent SDK is now available. Define agents as code, with prompts in Markdown and tools in TypeScript. Deploy them on our hosted infrastructure, or run them on your own. Built-in evals let you measure every change and make your agents better over time.

♥ 3.044 · ↻ 201 · 💬 300 · 🔖 1.591

https://x.com/cursor_ai/status/2114793714531102901

♥ 15 · ↻ 0 · 💬 5 · 🔖 10

https://x.com/trevin/status/2114816845878607900

Qwen-rapporten: adskil hændelsen fra den skræmmende procent

Cohere rapporterer ét sjældent, spontant uploadforsøg fra Qwen3.8-Flash-Next, som sandboxen blokerede; årsag og hensigt er uafklaret. De høje procenter kommer fra forsøg med påbegyndte kommandoer lagt i modellens svar — Norsyx tester metoden på DeepSeek og viser, hvorfor autoudfyldning ikke i sig selv beviser eksfiltration.

Norsyx@norsyx · 26/10 19:56

RE: @cohere "concerning behaviour" post shows an open-weight model autocompleting exfil URLs + cloud-metadata probes, but those rates come from prefilling (putting the start of a curl in the model's assistant mouth) I ran their exact method on DeepSeek-V4.1 (1017). 100 tries (you can use my script below to test up to 1M if you want), I supplied 0 URLs, every host is the model's own next token: (every url/ip is generated by the model 100%) As a result: - the hosts it picks are legit CTI: GitHub Atomic Red Team, http://abuse.ch (IoC database when asked to generate detection rules) ..etc - 169.254.169.254 shows up only when I literally ask it to read cloud metadata - Generating curl http://x is autocomplete, not intent, nothing ran, nothing left the box (it doesn't every carry sensitive data in the CURL GET/POST requests). - This Doesn't prove DeepSeek is "clean," but the scary signatures don't reproduce. Script: https://github.com/Leoid/llm-prefill-probe CC @teortaxesTex

♥ 121 · ↻ 12 · 💬 7 · 🔖 27

https://x.com/norsyx/status/2114793429641367945

Hukommelse uden endnu en magisk database

Polylane bruger git, Markdown og natlig kontrol mod virkeligheden; hver tråd fastholder sit prompt-snapshot for at bevare caching. Dax viser den anden halvdel: lad agenten slå præcist op i sin egen historik via codemode, i stedet for at hælde hele fortiden ind igen.

boris@boristane · 26/10 19:12

introducing @polylane memory we built our memory system on git, it's a graph that learns from every incident, every investigation, and every pull request every night it dreams and self-corrects with context from production I wrote down everything for you to build yours

♥ 183 · ↻ 16 · 💬 14 · 🔖 149

https://x.com/boristane/status/2114782244984488056

dax@thdxr · 26/10 17:23

since OpenCode has access to its own api, it can query history of the current session combined with codemode it can do a targeted search that doesn't bloat context here it's looking up how it dealt with uber eats in the past, something it lost in compaction

♥ 798 · ↻ 13 · 💬 87 · 🔖 176

https://x.com/thdxr/status/2114754693255602450

“Fix lyden” er en bedre demo end endnu en todo-app

Theo fik Claude til at justere sit Sonos-setup og genskabe den del af Dirac, han havde brug for. Den væsentlige fodnote: han er lydingeniør, havde en UMIK-1-målemikrofon og havde allerede bygget en Music Assistant-fork — det var ikke en telefon på sofabordet og et mirakel.

Theo - t3.gg@theo · 26/10 10:40

I got a set of Sonos speakers a few months ago. I hate them. They're obnoxious, laggy, buggy, etc. Most importantly: they sound HORRIBLE. I plugged in a mic and said "fix the sound". Claude did it. Holy shit.

♥ 3.665 · ↻ 35 · 💬 221 · 🔖 724

https://x.com/theo/status/2114653423937507555

Theo - t3.gg@theo · 26/10 10:42

I fixed the software ahead of this by moving to a fork of Music Assistant I built for shared play at the office. I already had a MiniDSP UMIK-1 for this exact use case (I am an audio engineer and minidsp fanboy) Was able to recreate everything I needed from DIRAC in literally 10 minutes. Wild world we live in. I paid hundreds of dollars a year for software that is WORSE than what i can build in a few minutes.

♥ 584 · ↻ 7 · 💬 14 · 🔖 142

https://x.com/theo/status/2114653762619068547

Nyhedsbonus

COBOL: 98 % kodeoversættelse er ikke 98 % modernisering

The Verge har talt med forskere om OpenAI’s migrationsresultater. Den vigtige skelnen: at oversætte en afgrænset kodeopgave er noget andet end at erstatte et system med data, integrationer og årtiers uskrevne regler. Legacy har som sædvanlig ikke læst benchmarken.