Dagens Vibes — 16. oktober 2026

Fra X-feedet

Måske er neurale netværk mindre rodede, end vi troede

Nick Cammarata udfordrer en grundantagelse i interpretability: mange tilsyneladende sammenfiltrede neuroner kan forstås direkte med bedre værktøjer. Hans foreløbige Qwen3-32B-audit finder ét hovedjob i omkring 90 % af neuronpolerne — en stor forskningspåstand med konkrete eksempler, ikke en færdig løsning på alignment.

Nick@nickcammarata

interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works

♥ 13,765 · ↻ 1,750 · 💬 620 · 🔖 13,393

https://x.com/nickcammarata/status/2110757819427574056

Meta åbner vægtene. Skrivebordet får brug for venner

Muse Spark Open er Metas nye model til kode og agentarbejde: 816B parametre, 42B aktive og 1M kontekst. Meta demonstrerer lokal kørsel på to M5 Ultra Mac Studios; “på din egen computer” har fået en ret rummelig definition. Vægtene udgives under Muse Model License med begrænsede kommercielle vilkår.

AI at Meta@AIatMeta

Today we’re releasing Muse Spark Open, the most capable US open-weight model to date. Built for coding and agentic work, it excels at long-horizon tasks, complex software engineering, and multi-step reasoning. Key Model Specs: → 816B total / 42B active parameters → 1M token context window → Natively multimodal across text, image, video, audio

♥ 9,684 · ↻ 1,010 · 💬 361 · 🔖 2,016

https://x.com/AIatMeta/status/2110779754274292212

Agenten får en tidsmaskine til de svære bugs

Antithesis lader agenter genafspille en fejl deterministisk, spole tilbage og undersøge alternative forløb. Demoen med en Raft-fejl er mere interessant end endnu en agent, der erklærer sin egen diff flot: den kan undersøge den konkrete race og teste rettelsen.

geoff@GeoffreyHuntley

Do you remember when @bcherny said Claude couldn't build distributed systems? We have a fix for that. 🫡

♥ 175 · ↻ 7 · 💬 10 · 🔖 198

https://x.com/GeoffreyHuntley/status/2110731428573180351

Argon-vibes: bedre end gårsdagens dommedag

Efter gårsdagens hårde kodekritik kommer to konkrete modbilleder: Chong-U fik multiplayer med lobbyer og disconnect-håndtering på to prompts. Soumitra fremhæver usædvanligt god prosa, men ujævn faglig ræsonnering — samme model, meget forskellige opgaver.

Chong-U@chongdashu

Gemini 4 Argon shines the most in complex 'coding' tasks. It made my entire game multiplayer - with lobbies, matchmaking, bots, optimisation, disconnection handling All in just 2 prompts.

♥ 24 · ↻ 1 · 💬 3 · 🔖 9

https://x.com/chongdashu/status/2110792827097813502

Soumitra Shukla@soumitrashukla9

First impressions of Gemini 4 Argon (aka "big boi"): - incredible writer. MUCH better than either Opus 5.5 or GPT 6 Astra. I found myself reading (and re-reading) its replies. Just smooth, human-like writing. First time I have seen something like this with the model. - paper reading/ideation: very similar to Opus 5.5 and GPT 6 Astra. But also weirdly jagged, as it will sometimes miss something basic in OLS, but then also surprise you with a really deep insight or intuition. To be clear, the "bad" behavior here is 1/10 otucome. Overall, awesome awesome model. My workflow is now 80% Astra, 10% for Opus 5.5 and Argon each (previous share: 80% Astra, 20% for Opus 5.5).

♥ 196 · ↻ 15 · 💬 7 · 🔖 70

https://x.com/soumitrashukla9/status/2110792951899046343

“Kør det på den anden Mac, og hent resultaterne hjem”

Mitchell Hashimotos Rex-beta rammer remote-worker-idéen helt rent: genkør gårsdagens benchmark på en anden maskine og sammenlign lokalt. Det interessante er én arbejdsgang på tværs af maskiner, ikke endnu et sted at kopiere prompts hen.

Mitchell Hashimoto@mitchellh

Rex beta testers are figuring it out. 😍 You can also do stuff like "rerun the benchmarks I did last night on my Mac Studio, but run them on Jack's Mac instead, then bring both results to my local Mac to compare." And it does what you'd expect. We actually aren't running evals on this stuff yet, but plan to. We want to make sure that all harnesses/models out there can be productive with Rex in as little turns as possible and do what you want them to do.

♥ 572 · ↻ 6 · 💬 70 · 🔖 148

https://x.com/mitchellh/status/2110767244506767795

Flere agenter er let. Din opmærksomhed er flaskehalsen

Ryan Browns bb-workflow skiller agenter, der arbejder, fra dem, der venter på svar, og lader mennesket svare direkte eller parkere dem. Det er præcis forskellen mellem en liste over kørsler og et overblik over det, du mangler at tage stilling til.

Ryan Brown@NotRyanB

for the last ~2 months i’ve been using an orchestrator within @get_bb_app as a single entrypoint to manage my agents with the launch of cursor projects and @nicbstme’s viral post about AI feeds, figured it was time to have opus vibe up a hype video for my workflow at this point there are a lot of ways to have one agent orchestrate a bunch of other agents to do work, but there are a few things an ideal experience needs to solve for: observability: you should easily be able to talk directly to any of the spawned agents; especially for me as an fde there’s still a decent amount of in-the-loop work where going through the manager would be an unnecessary layer of translation attention: there’s only so much i can focus on at once. the split in bb feed makes it easy: agents still working are in a separate section, so the ones that actually need responses are clearly emphasized. if you have too many that need you, you can “pause” some to set them aside (bottom section) for later interaction: support various levels of interaction with the agents - let the manager handle it, talk to the manager directly and have it talk to the agents, reply to the agent directly without leaving the feed (either based on the summary line or expanding the card to see the full most recent message), or fully click into the agent session there might be other things out there that have all three of these but i’m a tinkerer at heart and constantly customizing so bb’s plugin system makes this seamless, shoutout @sawyerhood for building a great product

♥ 12 · ↻ 2 · 💬 3 · 🔖 1

https://x.com/NotRyanB/status/2110811717437440501

Nyhedsbonus

GitHub udvider sikkerhedsscanningen uden for CodeQL

AI Scan er generelt tilgængelig til blandt andet Shell, Dockerfiles og Terraform. Fundene er rådgivende og blokerer ikke merges; en Copilot-licens er ikke påkrævet, men funktionen bruger Code Security AI Credits.

FLUX 3 Klein: billedredigering på et mindre grafikkort

BFL udgiver 5,6B-modellen med åbne vægte, op til ti referencebilleder og afgrænsede redigeringer. Den kvantiserede version kører fra 8 GB VRAM; BFL opgiver cirka ni sekunder for et 1MP-billede på RTX 4070. Gratis kommerciel brug gælder under $5 mio. i årlig omsætning.