Dagens Vibes — 25. oktober 2026

Fra X-feedet

Pi-subagents lærer at overleve et crash

Nico Bailons extension gemmer arbejdet: afbrudte subagents kan fortsætte, og færdige resultater dukker op, når sessionen åbnes igen. Rex-terminalfaner virker også lokalt eller over SSH — mindre genfortælling, mere fortsættelse.

Nico Bailon@nicopreme · 24/10 23:50

pi-subagents just hit 4K stars on GitHub! 😇 Just shipped: it's now durable, so Pi no longer loses work when the session crashes. Interrupted subagents resume where they stopped, and finished ones show their results the next time Pi reopens the same session. Also new: Rex support. Subagents can run in Rex terminal tabs, locally or over SSH. pi install npm:pi-subagents

♥ 120 · ↻ 5 · 💬 11 · 🔖 68

https://x.com/nicopreme/status/2114112321002021077

Ling 3.1 Flash: 27 cent og en meget lang lokal køretur

Sudo su lod samme agent bygge samme 3D-scene lokalt og via API: 12 timer 26 minutter mod 57 minutter, begge med 12/12 tests og visuelt tjek. Hans API-regning var $0,27; det er en konkret prøvetur, og hastighedsforskellen siger også noget om serving-stack og kvantisering — ikke kun modellen.

Sudo su@sudoingX · 24/10 15:40

watch! i ran ling 3.1 flash on the same prompt twice with hermes agent on my floating tree spec once as a q4 gguf across 4x dgx sparks and once on the api, with the same agent settings both times > q4 over llama[.]cpp rpc: 12h 26m > api, novita on openrouter: 57 min > 7.5 vs 98.6 tok/s decode > 325k vs 228k tokens out > 12/12 tests + visual gate on both > $0.27 for the whole api run both builds pass everything, the api build has the better tree with real branches and 8,270 veined leaves, the q4 build kept the better cliff the video is one camera path on both builds, frame for frame, q4 on the left and the api on the right my verdict: ling 3.1 flash is a strong model and the speed gap is the setup, q4 over rpc proves it fits on 4 sparks, for agent work you want it on a real serving stack, and an official 4bit build would make it practical locally the long turns question from the 12 hour build is answered in the thread below 👇

♥ 13 · ↻ 2 · 💬 5 · 🔖 5

https://x.com/sudoingX/status/2113989015917727789

Kimi-opfølgning: flere fund er ikke altid det bedste køb

I BugBunnys egen test af 100 PR’er med kendte sårbarheder fandt K3.1 Pro 47 mod K3’s 37. Men DeepSeek V4.1 Flash 1017 fandt ifølge samme test 54 til langt lavere pris; en nyttig modvægt til lanceringspodiet, ikke en dom over alle sikkerhedsopgaver.

Z A D D Y@Zaddyzaddy · 24/10 14:39

We gave Moonshot’s new Kimi K3.1 Pro 100 pull requests with known vulnerabilities and one hour per PR to find them in the @BugBunny_ai VulnPR-100 benchmark. It found 47/100 for $124.69. That is 10 more vulnerabilities than Kimi K3 while costing $20.54 less. Cost per finding dropped from $3.93 to $2.65, a 33% improvement. K3.1 Pro beat GPT-6 Astra and Muse Wonder 1.0, both at 46, and landed just one finding behind Claude Opus 5 while costing 3.6× less. It still trails DeepSeek V4.1 Flash 1017, which found 54 vulnerabilities for only $7.53. That is seven more findings at roughly 16.6× lower cost. A strong generational improvement from @Kimi_Moonshot, but DeepSeek remains the model to beat on cost-adjusted security performance.

♥ 83 · ↻ 5 · 💬 10 · 🔖 19

https://x.com/Zaddyzaddy/status/2113973761502253210

Agenten bygger spillet — og spiller det bagefter

Nader Dabit beskriver Devin på en AWS-buildmaskine og en separat L4-GPU til at spille Unreal-spillet, finde fejl og rette dem. Pointen er testsløjfen, ikke bare endnu en flot genereret scene.

nader dabit@dabit3 · 24/10 19:00

Been experimenting with Unreal Engine, it's been really fun and easy to build and test insane looking games combining Devin Cloud agents with AWS, where I offload the computational-heavy work. So basically you don't need a $4,000 gaming rig, GPUs, or really any hardware at all to do this anymore. Devin builds the game on (beefy) AWS boxes, plays them on AWS GPU instances, so it can spot and fix problems and I get to watch every build on Devin's computer. A Devin Outpost runs on an AWS c7i.8xlarge (32 vCPU, 64 GB) builds the game, and then an on-demand g6.2xlarge with an NVIDIA L4 tests it. I put together a prompt (you can also use it as a skill) if you'd like to replicate:

♥ 199 · ↻ 9 · 💬 35 · 🔖 60

https://x.com/dabit3/status/2114039244150681645

Codex får genbrugelige arbejdsgreb

CLI 0.164.0 gemmer prompts og indstillinger i presets, der kan hentes midt i arbejdet. OpenAI melder samtidig, at skift af reasoning-niveau ikke længere smider samtalens cache væk — små knapper med reel betydning i et review → fix-loop.

Felipe Coury 🦀@fcoury · 24/10 22:01

Presets are in Codex CLI 0.164.0 🎬 Save the prompts you repeat plus the model, reasoning and other settings you switch between, then recall them mid-task with Ctrl X P or /presets. Here they are in a real review → fix loop. Try it and send me your feedback and sharp edges.

♥ 99 · ↻ 11 · 💬 16 · 🔖 45

https://x.com/fcoury/status/2114084792350494876

Broderimaskinen slipper for sit abonnement

Kath Korevec købte en broderimaskine til $300, men ville ikke betale $12 om måneden for den obligatoriske app. Hun fik ChatGPT til at bygge en anden klient, der taler med maskinen — endelig en agenthistorie, der kan ende som en fysisk lap.

Kath Korevec@simpsoka · 24/10 17:17

Got a cheap embroidery machine off amazon for $300 where the biggest complaint is the mandatory $12/mo app that you’re forced to use. Had ChatGPT create its own app that communicates with the machine. Generative art patches incoming!

♥ 2.203 · ↻ 52 · 💬 81 · 🔖 705

https://x.com/simpsoka/status/2114013404143001695

Nyhedsbonus

Flere fejl, kortere tid til at patche

Reuters har talt med syv sikkerhedsfolk efter ugens PHP- og Node.js-sager. De er enige om kortere reaktionstid, men ikke om katastrofen: curls Daniel Stenberg peger på, at mange fund er svære at udnytte, og flere CVE’er er ikke automatisk flere angreb.