ChatGPT Work begynder at ligne et rigtigt arbejds-OS: cloud-computer, browser, persistent disk og subagenter — plus de første workflows, der viser hvorfor det betyder noget. Samtidig bliver de mindre modeller både billigere og mere uafhængige. Agenten er flyttet ind; nogen bør stadig tjekke elregningen.
Simon Willison har lavet den forklaring, OpenAI selv burde have skrevet. Work Cloud er ikke bare ChatGPT med arbejdstøj: det har internetåben kodekørsel, headless Chrome, fælles persistent filsystem, subagenter og Sites — samt den sædvanlige prompt-injection-elefant midt i kontoret.
Here's my attempt at explaining what ChatGPT Work can actually do - it's a deeply confusing but extremely powerful tool with a whole lot of useful features that aren't available in regular ChatGPT
Det bedste praktiske bevis: Emanuele bad Work køre en PR i sin sandkasse og sende screenshots af UI'et tilbage. Ingen lokal server, ingen varm Mac, bare en ret direkte vej fra kode til visuel kontrol.
ChatGPT Work is sensational for testing features I was building the ⌘+F feature for @trySynara, started by @bassa_dev I then asked Fable to tweak the UI a bit I did not want to sit there, run the server, heat my mac and click through how it looked So I told Work to run that PR in its sandbox and send screenshots... and it worked It's insane
Desktop-appen fik den kedelige forbedring, der i praksis føles som en ny model: lange tråde loader over 90 procent hurtigere og bruger over 90 procent mindre hukommelse. Sidechats bør blive næsten øjeblikkelige. Tre hundrede PR'er senere kan fanen endelig trække vejret.
Last week, I pushed a change to the ChatGPT desktop app that makes the loading of threads faster. Not just a little faster. A LOT faster. ✅ Long threads now load over 90% faster. ✅ Decreased the memory footprint of long threads by over 90%.
GLM 5.3 Flash ser ud til at lande torsdag og kører på én DGX Spark. Feedet gravede samtidig en tidlig Blender-vibe-check fra fredag frem: omtrent samme scene som 5.3, bedre promptfølgning og 6 cent mod 88. Det er én test — men 17× er svært at gemme under gulvtæppet.
GLM-5.3-Flash on 1x Spark Coming Thursday
Armin opdagede, at hans agent på egen hånd havde reverse-engineeret en specialbygget client-side crypto-challenge — uden overhovedet at køre JavaScript. Det interessante er ikke tricket, men at løsningen opstod som et ubestilt mellemtrin i rigtig research.
I currently have a little data scraping project going for the war in Ukraine. At no point in time did I tell the agent *how* to do something, but at one point I realized it reverse engineered a custom client side crypto JS challenge without invoking JavaScript.
Executor gør Codex-plugins som Computer Use og iMessage til almindelige MCP-værktøjer, så de kan bruges fra andre harnesses. Den rigtige platformskrig er tilsyneladende ved at ende med adaptere. Som al god infrastruktur.
Executor now supports some Codex plugins like iMessage and Computer Use - this lets you use them as a regular MCP in any harness I was tired of being mid session in a Claude thread and having to start a new Codex chat just for computer use - enjoy!
Og den bedste lokale model-anbefaling kom fra nogen, der faktisk mistede nettet: Qwen 3.8 27B på 24 GB VRAM holdt et coding-flow kørende offline på en laptop. Lidt mere eksplicit instruktion, meget mindre tilladelse fra skyen.
anon, let me share you about the day local ai stopped being a hobby for me. that day i was out with no internet but my ROG 5090 laptop, 24gb of vram, 64gb of ram, and i was mid flow on my own, head full of the problem and commit ideas. so i loaded qwen 3.8 27b dense from weights already sitting on my nvme, and i just worked. sure, it needed a little more explicit instruction than a frontier model needs, i had to say exactly what i meant. but it kept up with my flow, offline, on battery, on hardware that fits in a backpack. somewhere in that session it hit me. if ai development ends tomorrow, if every api shuts down tonight, i am already unstoppable. the weights are on my nvme. nobody can rate limit them, nobody can deprecate them, nobody can take them away. and you will never feel this until you are forced into it. kill your wifi for an afternoon and try to work. that is the day you find out what local actually means.
Caterpillar har den mindst powerpoint-agtige AI-strategi i dagens nyhedsstrøm: feltassistenter på 16 petabyte maskindata, digitale tvillinger og agenter til at modernisere og teste legacy-kode. Den dyre del er stadig arbejdsgangen — derfor bruger selskabet 100 millioner dollar på at træne 118.000 medarbejdere.