Dagens Vibes — 23. oktober 2026

Fra X-feedet

Decisions API: et svar behøver ikke være en roman

OpenAI gør Decisions API generelt tilgængelig: vælg en handling, klassificér et input eller giv en score uden et fuldt genereret svar. Nyhederne omfatter prompt-cache, native JSON og 400 spørgsmål pr. kald; input koster $0,10 pr. million tokens, cachede input $0,01.

OpenAI Developers@OpenAIDevs · 22/10 21:07

Add near real-time decisions to your workflow with the Decisions API, now generally available. Build faster, more cost-efficient apps and agents that can classify text and images, evaluate inputs, and determine what happens next without generating a full model response.

♥ 1.030 · ↻ 69 · 💬 124 · 🔖 349

https://x.com/OpenAIDevs/status/2113346642787086538

MemT: genstart modellen, behold erfaringen

MemT gemmer hukommelse som modeltilstand frem for gammel tekst i prompten. Sudo testede Nano lokalt, gemte 42 MB, genstartede med nul beskeder og fik sine Rust-regler tilbage i koden — et konkret første forsøg; modelkortet siger samtidig, at stærke tekstbaserede memory-systemer stadig fører samlet.

Sudo su@sudoingX · 22/10 14:02

i saw this very interesting memory native model and thought i'd test it and share here MemT 1.0 just came out, and it writes what you tell it into a memory state inside the model for later answers, while the current message still goes in the context window so i ran it in bf16 on my 128gb strix halo i saved its memory to a 42 mb file, killed the process, started a new one from that file, and it still wrote code my way in the first session i gave it my rust house rules and one lesson: > one error enum called BeastError > no unwrap, ? everywhere > a missing config must never crash > fall back to defaults + a warning the new process had the same weights and that memory with 0 messages in context, and i asked it for a config loader what it did: > BeastError on every error path > ? everywhere, no unwrap > missing file: defaults + a warning watch anon! this is a new kind of technology i found from the @MemOS_dev team, and it was so interesting i had to share it with you, at least i had never seen anything like it, and it opens a new way of building better models, where memory is part of the model itself the bigger 288b version is built on deepseek v4 flash, and the weights are open on both links below: 👇

♥ 34 · ↻ 3 · 💬 4 · 🔖 31

https://x.com/sudoingX/status/2113239498419224670

MemOS@MemOS_dev · 20/10 15:30

Can large language models retain information from past interactions and reuse it in the next task? Introducing MemT-1.0: Towards Memory-Native Evolution. This time, we've built memory directly into the model architecture. Our MemT-1.0-Nano and MemT-1.0-Flash make persistent memory part of model computation, carrying information across interactions.

♥ 34 · ↻ 3 · 💬 4 · 🔖 39

https://x.com/MemOS_dev/status/2112536909969035761

Skills får et neutralt hjem — og kan læres ved at kigge

Anthropic donerer Agent Skills til AAIF; over 70 platforme har allerede taget formatet til sig. Den praktiske side findes hos Hermes: optag et browserflow, og lad /learn lave en genbrugelig skill — udgivet som eksperimentelt community-plugin.

ClaudeDevs@ClaudeDevs · 22/10 19:55

We're donating Agent Skills to the Agentic AI Foundation (@AgenticAIFdn), the same foundation we donated MCP to. Since December 2025, 70+ agent platforms have integrated the standard, including ChatGPT, Cursor, GitHub Copilot, and VS Code. Here's what that means for you:

♥ 6.571 · ↻ 295 · 💬 215 · 🔖 1.825

https://x.com/ClaudeDevs/status/2113328526413996100

Opus 5.5 + Pi får de gode hverdagskarakterer

Mario har været GPT-fan i næsten et år, men kalder Opus 5.5 virkelig rar at arbejde med. Sawyer har skiftet til Pi + Opus og oplever Claude Code-kvalitet med cirka halvt tokenforbrug på første prompt — en brugererfaring, ikke en kontrolleret måling.

Mario Zechner@badlogicgames · 22/10 18:12

been a gpt shill for almost a year, but opus 5.5 is just so nice to work with.

♥ 1.077 · ↻ 13 · 💬 72 · 🔖 29

https://x.com/badlogicgames/status/2113302439164281308

Sawyer Hood@sawyerhood · 22/10 16:37

i have switched to using pi w/ opus 5.5 for everything and have felt no perceivable difference from using claude code but except for burning 50% less tokens on first prompt

♥ 531 · ↻ 7 · 💬 56 · 🔖 139

https://x.com/sawyerhood/status/2113278469534789779

Amp lader agentens arbejdscomputer bo hos dig

Self-hosted orbs er gratis for individer og teams: din infrastruktur starter og pauser miljøerne, mens Amp står for oplevelsen. Provideren forbinder udad, så du ikke skal åbne inbound-porte — egen compute, ikke en selvhostet Amp-tjeneste.

Thorsten Ball@thorstenball · 22/10 07:29

You can now self-host orbs. It's free for individuals and teams. https://ampcode.com/news/self-hosted-orbs https://t.co/DSXGjOGC85

♥ 1.167 · ↻ 80 · 💬 71 · 🔖 943

https://x.com/thorstenball/status/2113140757410521159

Gem JavaScript midt i en løkke

dquickts er en QuickJS-fork med TypeScript og snapshots af hele heapen, inklusive en suspenderet call stack. Programmet kan fortsætte i en anden proces med samme engine-build; filer og anden host-state kræver stadig eksplicit håndtering.

Can Bölük@can1357 · 23/10 00:03

An interesting primitive for those working on durable harnesses: https://t.co/PbtkM77Rzf QuickJS with typescript support out of the box as well as a snapshottable & resumable heap. You can persist the state and resume on a different machine, or fork, even in the middle of a loop.

♥ 193 · ↻ 10 · 💬 15 · 🔖 123

https://x.com/can1357/status/2113390699441701305

Komprimér historikken, ikke promptcachens livslyst

Nico Bailons Pi-extension bevarer prompt, tools og beskedstruktur, når den beder om en opsummering. Hans 667k-token Opus-eksempel kostede $0,19 mod estimeret $1,40 med Pi-komprimering, regnet ved API-priser; små sessioner kan være dyrere med extensionen.

Nico Bailon@nicopreme · 22/10 19:58

New Pi extension: pi-cached-compaction. Pi's built-in compaction swaps the system prompt, drops the tools, and flattens the messages, which breaks the prompt cache. This one keeps all three as they are and just appends the summary request, so ~99% is read from cache. And if you step away, it compacts just before the cache expires, so your next message doesn't pay to cache the whole session again. pi install npm:pi-cached-compaction https://t.co/iQmjPV8i6e

♥ 372 · ↻ 24 · 💬 23 · 🔖 430

https://x.com/nicopreme/status/2113329221196218564

Nyhedsbonus

Nej, Claude har ikke opfundet en modgift

The Verge får forskere til at skille Citrinis scorpion-historie ad: 17 designede proteiner bandt til toksiner, men neutralisering og behandlingseffekt er ikke testet. Det interessante er den autonome opdagelsesproces; en binder er stadig ikke medicin.