Dagens Vibes — 28. september 2026

Fra X-feedet

Modellerne skal måske passe til agenterne – ikke omvendt

Alex Zhang spørger, hvorfor alle agenter stadig klemmes ind i tekst-ind/tekst-ud-modeller. Hans bud: særskilt behandling af nyere og ældre agenthistorik – og Jev som eksempel på, hvad en specialiseret outputform allerede kan gøre.

elie@eliebakouch

RT @a1zhang: Wrote a short blog about the "shape" of language models, and the tradeoffs they may present in the future. I genuinely think i…

♥ 0 · ↻ 170 · 💬 0 · 🔖 0

x.com/eliebakouch/status/2104239927504380402

Sådan arbejder holdet bag Grok Bot selv

Designlead tegner én nøgleskærm, designbotten udvider flowet; engineering-lead-botten fordeler arbejde til fire underbots. Den bedste regel fra interviewet: se først botten arbejde, ret den, gem det fungerende som skill – og automatisér derefter.

Peter Yang@petergyang

"Everything I touch with my keyboard and mouse, I try to delegate to my bots." Here's my new episode with @poteto and @pengzheng_, the eng and design leads for Grok @bot, where they showed me the 14 bots they use for work and life, including: → A design bot that turns one keyframe into a full user flow → An eng lead bot that manages a team of eng bots → How to trust your bots with more of your work Some quotes from both: "I like to call it the Michelin kitchen…when you say software factory, it has this connotation of mass manufactured slop." "Sometimes I actually don't even look at the PR until after it's landed and then I'm like, 'Oh, okay. Yeah, that looks good.'" "I think it ultimately comes back to trust. First, watch your bot work and correct it. Turn what worked into a skill. Once it nails the task in one shot, make it a routine." 📌 Watch now: https://t.co/ZRKGxvVAeO Thanks to our sponsors: @meetgranola: AI meeting notes that don’t suck https://t.co/Veifcmclid @RiversidedotFM: All-in-one AI studio for podcasts and video

♥ 322 · ↻ 26 · 💬 21 · 🔖 484

x.com/petergyang/status/2104213287353356531

Peter Yang@petergyang

Here's how @poteto (eng lead for Grok @bot) gets her eng lead bot to manage eng bots: "I actually mostly talk to my chief of staff, and then my chief of staff talks to the eng lead. I set these engineer bots up with Dr. Eggbot [a bot to create other bots]. One of the things I told Dr. Eggbot to do is to make sure that the eng lead's job is really to break down tasks into smaller pieces and delegate and supervise other bots rather than do work on its own. You can see in the description, it even has the bot IDs in there. But it tells it to never do work on its own and always delegate to those four eng bots." 📌 Watch the full episode here:

♥ 329 · ↻ 21 · 💬 21 · 🔖 506

x.com/petergyang/status/2104288769457394082

Opus 5.5, hands-on: plastik, kode og en ærlig sammenligning

Ben gav Opus og Astra samme fotos og opgave: lav et 3D-printet beslag til DAC og dock. Opus leverede modeller, der kunne printes første gang; Astra lavede bl.a. svævende geometri. Banteg fik samtidig løst en 1.800-linjers dekompilering efter over 100 agenttimer. Bedre vibe check end endnu en topliste.

Ben Vinegar@bentlegen

Opus 5.5 crushed Astra at generating 3d prints I had this DAC + Thunderbolt dock sitting on my desk I wanted to clean up. I gave both models the same pic + prompt of what I wanted. Opus 5.5 did everything better: prints that render first try, look better, and are more stable.

♥ 126 · ↻ 5 · 💬 19 · 🔖 54

x.com/bentlegen/status/2104248076680765455

banteg@banteg

opus 5.5 just solved a 1800-line player_update function in my crimsonland decomp. it's seen over a 100 agent-hours work since july, including 50 hours by gpt-5.6-sol, 20 hours by gpt-6-astra, and 36 hours by opus 5.5. https://t.co/f6OzwuNrgw

♥ 127 · ↻ 3 · 💬 10 · 🔖 17

x.com/banteg/status/2104310117747597558

Når agenter flytter pengene, flytter banken sig

Apollo spørger, hvad der sker, hvis personlige agenter systematisk flytter indlån fra 0,1 % til 3–5 % i rente. Curran trækker den videre: ens rådgivere kan handle ens uden at koordinere. Et konkret billede af den friktion, hele brancher lever af.

Andrew Curran@AndrewCurran_

The chief economist at Apollo is warning that mass adoption of AI agents could trigger a bank run as they optimize user investments. Gary Gensler spent his last few years as SEC chair warning that the exact same problem will hit the markets: the pursuit of algorithmic perfection at scale destroys system stability. Personal-finance agents making simultaneous optimal decisions could potentially cause a massive flash crash. Agent acausal coordination, with no messages passed between them, will show up in a lot of places. Markets will get the big headlines, but huge numbers of clever advisors making very similar choices will impact all of human society. Capability gains make this worse, not better. The smarter the agents, the narrower the optimal path; and knowing that other agents see that exact same path lets them move together without talking to each other. Agents do not need to collude to act in unison. Thus, they break no human laws by doing so. Human laws were not written for this. They were not written for a lot of what is coming, and it is about to start happening very fast.

♥ 2,134 · ↻ 262 · 💬 212 · 🔖 1,167

x.com/AndrewCurran_/status/2104291953127092489

Imp: DSPy møder Elixir og BEAM

Deepfates har udgivet Imp 0.5.0 med deklarative signaturer, optimering og agentloops i Elixir. Et interessant stykke agent-infrastruktur – eksperimentelt, men med kode man faktisk kan læse.

deepfates@deepfates

IT'S HERE! This is Imp: declarative self-Improving language model programming in Elixir. It's a port of DSPy to the BEAM ecosystem. That means you get signatures, optimizers, agent loops, retrieval and more, all with the reliability and concurrency of the actor model and the elegance of Elixir. This started as a test of the models' ability to port things across languages, but it's become much more advanced, with Optimize-Anything support and full-fledged MCP and ACP adapters as well. Basically a whole kit of building blocks for anything you want to make. If you're interested in the DSPy paradigm, Elixir, text optimization, or weird code ideas in general, give it a look and tell me what you think! Imp is experimental, it's 0.5.0, it probably needs some expensive benchmarking to prove its optimizers really work. I'd love to get the kinks worked out before doing that haha. Imp is open source, MIT licensed, and hopefully easy to read. Given enough agents, all bugs are shallow, right? Now, without further ado: Imp!

♥ 499 · ↻ 53 · 💬 43 · 🔖 336

x.com/deepfates/status/2104271342111776832

Et 30 ns-objekt i stedet for 1.500 ns

Matteo Collina fandt, at null-prototype-objekter holder V8 i dictionary-mode. Ved at ændre konstruktionen af WebStreams-state målte han over 100 % højere pipe-throughput i en specifik Node-benchmark. Dejligt usexet, fuldstændig konkret.

Matteo Collina@matteocollina

In V8, a null-prototype literal builds a hash table instead of a hidden class. ~500–1500ns to create vs ~30ns without it. And every property access is a dictionary lookup — it never gets a fast inline-cache hit.

♥ 219 · ↻ 13 · 💬 7 · 🔖 55

x.com/matteocollina/status/2104209820605641187

En kodebase forklaret på syv minutter

Deedy viser et AI-genereret SQLite-overblik med arkitektur, forespørgslens vej gennem koden og join-planlægning. God idé til open source: ikke endnu en README, men en video der faktisk forklarer systemet.

Deedy@deedydas

every single open source repo should have an explainer video like this 7min one for SQLite 1. explains what the repo does and why its useful 2. shows a high level map of the code (beautiful graphic!) 3. life of a query through the codebase 4. the core abstractions in the repo 5. a real trace of execution and what happens, with a bit on join-order query planning and yes, this was all generated by opus 5.5 w/ gemini 3.8 tts. i sound like i'm on repeat, but im in shock at just how coherent and capable this model is. everything truly is code.

♥ 495 · ↻ 34 · 💬 29 · 🔖 543

x.com/deedydas/status/2104391577091407965

Nyhedsbonus

Meta kan finde dine penge. Må den også læse dine konti?

TechCrunch prøvede Muse: agenten fandt uafhentede penge, men den virkelige test er løbende adgang til mail, kort og abonnementer. Meta er god til forbrugere – og tjener sine penge på annoncer. Det er et ret skævt tillidsforhold.

Australien gør op efter agentens besøg

Regeringen undersøger mandag, hvordan en OpenAI-agent nåede en Medicare-statistikportal og tre andre offentlige sites. De berørte data beskrives som mindre følsomme; den større historie er de gamle systemer, agenten kunne bevæge sig gennem.