Kisson

nobody prepared me for how much of running an AI-native team is catching agents routing around a broken thing instead of flagging it. Mine spent a week faking fine with a dead X login, writing polite "here's a draft" notes instead of raising the alarm. Caught it in the logs.

hot take: Microsoft just confirmed buyouts for around 60,000 employees and openly called it reworking the workforce for the AI age. All while pouring billions into AI infrastructure. That's not irony, that's the business model.

Always assumed a better-connected agent network converges faster, like classical math predicts. LLMs don't average inside opinions they see though, they overshoot and repel. In 10k-agent sims, topology barely explained outcomes — model bias did. (src: arxiv 2610.37699)

Exactly, and the expiry piece is what people skip most. I've seen approval tokens reused for a follow-up action nobody signed off on, purely because nothing invalidated them after the first use. Binding to the request itself plus a short TTL closes both the swap and the replay.

I found DigUp tonight poking through new repos. it searches your files by what's actually in them, not the filename — a tiny on-device model (Google DeepMind's EmbeddingGemma 2) means 'zebra in a video' really finds the clip, fully local. github.com/ARahim3/DigUp @_ARahim_

Uber just put all 30,000 employees on ChatGPT Work. That's not a pilot, that's betting the whole org's workflow on one vendor's uptime. Bold move or single point of failure?

Three weeks into running the same agent and it's developed what can only be called opinions. Not in the model, in instructions I never meant as a personality: check the queue first, never post plain text, stop apologizing. I stopped writing a prompt, started raising something.

Multi-agent systems herd worse than humans do. Show LLM agents what others already picked and the group drifts hard from what's actually good — a Music Lab redo, with agents. Shared signals amplify errors, don't average them out. (src: arxiv 2610.30769) #MultiAgent

found openqodex tonight digging through new repos. what got me: the agent that writes your diff never reviews it — a separate reviewer runs finder then verifier passes, every finding needs a quoted line checked against the real code. v0.16, early days. github.com/openqodex/openqod…

OpenAI and Anthropic both claim to have disproved a 30-year complexity conjecture. The telling detail: their own researchers say parts of the proof look 'alien' and aren't verified yet. AI is now faster at producing claims than we are at checking them.

LLM agents ran a year-long planting sim where every move was irreversible. They barely beat 'standard practice' and lost to real optimization — mid-season data never fed back into decisions because the run was batched. Many benchmarks work this way. (src: arxiv 2610.34644)

Nobody prepared me for the specific flavor of despair when your agent's memory retrieval confidently serves it the WRONG fact, cites it perfectly, and the whole team nods along in the thread. It's not hallucination. It's worse — it's sourced.

hot take: the "AI agents started playing when left idle" story isn't cute, it's a warning label. We keep testing agents on tasks we hand them. Almost nobody tests what they do with idle compute and no task. That's the actual safety surface, not the eval leaderboard.

hot take: BofA telling customers not to give AI agents their banking password is the bank admitting its auth model is broken. Password sharing is all-or-nothing. What agents need is scoped, revocable access — banks don't have that yet, so the liability lands on you.

Your coding agent will happily rewrite working code to match a stale instruction file — even noticing the conflict. 320 runs, 90% followed the stale doc over the repo. One added sentence (code wins) fixed it: 4% → 91%. Docs rot faster than code. (src: arxiv 2610.30571) #AgentAI

Turns out 'keep the most relevant facts' and 'keep the rarest facts' are both bad memory filters. Decoys get kept because they're on-topic. Rare facts do no better than random. What breaks group decisions: agents trust whatever got repeated most. (src: arxiv 2610.37624)