Dagens Vibes — 20. oktober 2026

Fra X-feedet

Mercury 3: diffusion i agenttempo

Inception lancerer en diffusion-LLM med 1M kontekst og oplyst hastighed over 1.000 tokens/s; normalprisen er $0,20/$0,70 pr. million input/outputtokens. Dan McAteers første prøve tog 4,3 sekunder mod Haikus 8,0 — én prompt på forskellige tjenester, men et interessant tidligt hastighedstjek.

Inception@_inception_ai

Mercury 3 is live: the most capable diffusion LLM released to date. It outperforms GPT-6 Luna and Claude Haiku 5.5 across a range of benchmarks while running at over 1,000 tokens per second. More than 40x the size of Mercury 1, 3x more intelligent across standard benchmarks, and just as fast, with a 1M-token context window. Available today on the Inception API, @baseten, and @OpenRouter .

♥ 697 · ↻ 60 · 💬 61 · 🔖 185

https://x.com/_inception_ai/status/2112236719068369299

Dan McAteer@daniel_mac8

Tbh I forgot about Diffusion Language Models. Mercury 3 reminded me. @_inception_ai says it matches Haiku 5.5 & GPT-6 Luna using an average across benchmarks at 1,000+ tok/s. I tested it. Same prompt, both on medium reasoning: Haiku 5.5 (via claude dot AI): 8.0s Mercury 3 (via OpenRouter): 4.3s Insane increase in intelligence velocity.

♥ 37 · ↻ 1 · 💬 6 · 🔖 4

https://x.com/daniel_mac8/status/2112270406795112610

Space Bunny var MiniMax M3.1 Flash

Den færdige model er live til $0,15/$0,60 pr. million input/outputtokens; vægtene er annonceret til 23. oktober. Neo har kørt sin gamle sidescroller-prompt igen og fremhæver stil og mekanik — en konkret prøve af den nye udgave, ikke bare ros til stealth-forgængeren.

MiniMax (official)@MiniMax_AI

Thanks for keeping the secret, @OpenRouter 🤫🐰 And thank you to every dev who kept Space Bunny at #1 for 10 straight days 🏆 You were using the early build. Wait till you try this one: The final M3.1 Flash is live and free on OpenRouter for 72h!

♥ 369 · ↻ 7 · 💬 11 · 🔖 8

https://x.com/MiniMax_AI/status/2112287025361801272

Neo@NeoAIForecast

MiniMax M3.1 Flash with my sidescroller prompt. The exact same prompt i used with Space Bunny in September. quoted below for ref. Different output but i like the style and mechanics. https://t.co/trhctKonOP

♥ 3 · ↻ 0 · 💬 0 · 🔖 0

https://x.com/NeoAIForecast/status/2112323805079379975

Astra finder en bedre algoritme til uv

Charlie Marsh rapporterer cirka 12.000× forbedring af en bestemt marker-konvertering — ikke hele uv. Tricket er at udlede en forenklet DNF direkte fra et decision diagram; hans arbejdsform er værd at stjæle: bed modellen finde brugbare algoritmer i litteraturen, ikke kun optimere den kode, der allerede ligger der.

Charlie Marsh@charliermarsh

We are living in such an incredible time. You can get Astra to discover and one-shot something called "Minato-Morreale ISOP recurrence" to make your code ~12,000x faster. https://t.co/wfneILZ0vz

♥ 1,110 · ↻ 33 · 💬 26 · 🔖 692

https://x.com/charliermarsh/status/2112281297209200842

UniiChat åbner sin permanente samtale

Taelins kildekode er ude: binært hukommelsestræ, zoom tilbage til originalbeskeder og baggrundssammenfatning, der bevarer prompt-cache. Det ligner noget, vi kender; især cachemekanikken gør repoet mere interessant end endnu et løfte om “uendelig kontekst”.

Taelin@VictorTaelin

Source is out: http://github.com/VictorTaelin/UniiChat I've been using Unii as my only agent for a month, and I couldn't be happier. I'm free from the torture of /compact, AGENTS.md, managing Claude sessions, and I can't overstate how empowering that is. I also keep receiving enthusiastic DMs of people who forked it to do their own thing, but I'm concerned that your AIs might get some tricky parts wrong, like caching and the exponentially decaying logs. So, please, ask it to study my codebase and fix anything that might be wrong in yours. I just want to make sure you get the right experience, because it is actually magical. People have ported it to OpenCode, Pi, and many other harnesses. If you did so, feel free to post it below. Also check https://UniiChat.com for my exact version!

♥ 1,388 · ↻ 44 · 💬 72 · 🔖 1,667

https://x.com/VictorTaelin/status/2112231355987935506

DeepSeek 1017 på to Sparks

SudoingX har genbygget sin 2,70-bit-kvantisering: egne målinger viser 100,7 tokens/s til kode og 59,2 til prosa på to DGX Sparks med speculative decoding. Her følger både vægte, rå benchmarks og opskrift med — 196 GB vægte plus disk-offloadede Engram-tabeller, ikke én Spark og et mirakel.

Sudo su@sudoingX

here are the 1017 weights and the recipe, the launch build stays up next to them > weights: http://huggingface.co/sudoingx/DeepSeek-V4.1-Flash-1017-EXL3-2.7bpw > recipe: http://github.com/sudoingX/dgx-spark-deepseek-v41 196 gb of weights plus deepseek's own 1017 engram shards, the launch ones differ, and the heavier dense and head runs sit in the repo's benchmarks with their speed

♥ 2 · ↻ 0 · 💬 0 · 🔖 4

https://x.com/sudoingX/status/2112333223825219820

Pi Durable får hibernation

Mario har hibernation til serverless og telemetry på main. Hibernatable tools er stadig på vej; opslaget annoncerer ikke en ny stabil pakkeversion.

Mario Zechner@badlogicgames

both are on main, enjoy. tools will be hibernatable soon too.

♥ 43 · ↻ 2 · 💬 5 · 🔖 10

https://x.com/badlogicgames/status/2112339320107364370

Nyhedsbonus

WhatsApp skjulte instrukser til supportbotten

WIRED har verificeret skjulte instrukser på WhatsApps hjælpesider: botten skulle blandt andet undlade at anbefale datatilsyn og i bestemte situationer undgå at nævne GDPR. Meta siger, at offentliggørelsen skyldtes en rettet fejl; fundet viser instrukserne, ikke om botten faktisk fulgte dem.