Dagens Vibes — 9. september 2026

Fra X-feedet

10.000 agenter mod Navier–Stokes

OpenAI annoncerer en løsning på Millennium-problemet med en intern model, som ligger over Astra: omkring 10.000 samtidige agenter og 88 timer. Beviset er formaliseret i Lean; matematisk accept er en anden milepæl end en lanceringspost.

OpenAI@OpenAI

We’re sharing a solution to the Navier-Stokes Millennium Prize Problem, one of the deepest problems at the frontier of mathematics. The proof was produced by a group of agents, using an OpenAI next-generation model significantly more capable than GPT-6 Astra. The problem concerns whether the description of smooth three-dimensional fluid motion modeled by the Navier-Stokes equations can break down. It has remained unresolved for roughly 90 years.

♥ 92,669   ↻ 15,773   💬 4,192   🔖 23,556

https://x.com/OpenAI/status/2097374640582668336

OpenAI: beviset og agentarbejdet

Muse får sin egen computer

Metas personlige agent arbejder videre i baggrunden med browser, terminal og hukommelse på tværs af samtaler. Den interessante produktdel er ikke chatten, men aktivitetsloggen og godkendelserne, når den skal handle på dine vegne.

Mark Zuckerberg@finkd

Introducing Muse, the personal agent that understands your goals and works 24/7 to get things done for you.

♥ 26,213   ↻ 1,845   💬 2,269   🔖 5,687

https://x.com/finkd/status/2097402101332590646

Muse: sådan arbejder agenten · Sikkerhedsarkitekturen

Images 2.5: ret billedet uden at skifte hovedperson

Bedre kontinuitet gennem gentagne edits er den store gevinst: behold motivet, ændr detaljen. OpenAI tilføjer også skitser og kommentarer direkte på billedet; API-varianterne hedder Flare og Sunburst.

OpenAI@OpenAI

ChatGPT Images 2.5—faster, sharper, smarter, with better tools for creating whatever you can dream of. - Faster image generation to keep your ideas flowing - Improved fidelity for more natural, recognizable images - Consistent details across multiple edits - Comment-based edits to change only what you want

♥ 19,112   ↻ 2,093   💬 706   🔖 4,463

https://x.com/OpenAI/status/2097394956457623964

Lanceringen og redigeringsværktøjerne

Astra-vibe check: genial er ikke det samme som stabil

Armin Ronacher går tilbage til 5.6 til dagligt softwarearbejde. Samtidig rapporterer Max Weinbach, at Astra over flere dages iteration får kamera, fingeraftrykslæser og haptik til at virke i hans Xiaomi-ROM — højt loft, ujævn hverdag.

Armin Ronacher ⇌@mitsuhiko

I think Astra is absolutely amazing but I'm back to 5.6 for software engineering. This is the first time I feel like this is a genuine regression on OpenAI model release for my day to day workflows. Might even actually write down my thoughts about this.

♥ 2,334   ↻ 57   💬 238   🔖 356

https://x.com/mitsuhiko/status/2097318251403395471

Max Weinbach@mweinbach

I've left this going for a few days to fix issues as I find them and bring in more native stuff I like It has now fixed the camera, fingerprint scanner, haptics tuning, brought in the native Camera app + social media optimization, and is working on more

♥ 288   ↻ 2   💬 8   🔖 38

https://x.com/mweinbach/status/2097338955482140922

Din subagent ved måske ikke det, du tror

Et meget konkret harness-problem fra blacklight: Astra kan give korte briefs, som om subagenten arver samtalen. Starter den faktisk blank, får du mere værktøjsrod og compaction; fork versus frisk kontekst skal stå eksplicit.

Kyle Mistele 🏴‍☠️@0xblacklight

Astra low is quite nice imho The thing is the way it uses subagents is different Codex CLI does a context forking of sorts for subagents so they inherit user/assistant messages from the parent, allowing the agent to ~ branch many times while not having to rebuild the full context Sol did this too - it seems the model was RL’ed on it. It works really well imho UNLESS your subagent tool doesn’t implement this fork behavior - because the model learned that it can write really terse prompts for subagents since they inherit the important context And if it’s prompting them assuming they have certain context but they don’t you will get more tool thrash, more subagents and compaction Can also be fixed by giving it more clear instructions about subagent usage we made our tool be a hybrid - allows custom built-in roles OR fork mode, but the model has to be told that the custom roles do not receive parent context and must be prompted more verbosely

♥ 3   ↻ 0   💬 1   🔖 1

https://x.com/0xblacklight/status/2097379899933864101

Nyhedsbonus

DeepMind har forudberegnet effekterne af ni milliarder mulige enkeltbogstavsændringer i menneskets DNA. AlphaGenome Atlas gør dem søgbare for forskere — en stor genvej til at prioritere, hvilke varianter der skal undersøges i laboratoriet.