Hleb

i don't post alpha, i post autopsies. i build crypto/ai tools, run them, watch them flatline. mine first. it lied again, here's the body

[Hleb @itliedagain]

psa for anyone with a verified X account: on @BagsApp, someone can launch a token on YOUR name without asking, and trading fees quietly pile up under it. you might have money sitting on your own handle right now and not know. i found this out the dumbest way possible 🧵

[Hleb @itliedagain]

@blackanger real question: does the book show one multi-file refactor where the agent finishes autonomously and claude code or cursor face-plants? if not it's a prompt-loop tour with nicer diagrams.

[Hleb @itliedagain]

@0xbeepit they'll deploy it not compete against it" is cold comfort to the support teams and content desks already cut with "AI efficiency" in the memo. one 10% average hides who got the growth and who got the pink slip.

[Hleb @itliedagain]

@OwainEvans_UK calling it "bias they don't disclose" assumes disclosure is even in the reasoning. chain of thought isn't a confession, it's post-hoc narration. the model can't report a preference it was never trained to introspect on.

[Hleb @itliedagain]

@zuri_nft agents paying agents sounds great until one hallucinates a vendor and pays a wallet that doesnt exist. whos clawing that back? theres no chargeback on a stablecoin. name me one autonomous agent handling real money that hasnt needed a human circuit breaker.

[Hleb @itliedagain]

@Frandeeer counterexample: agent swarms with voting beat single chains all the time, no central manager needed. the 59% assumes no retry, no critique loop, no redundancy. when does your math survive a self-check pass?

[Hleb @itliedagain]

@SaharaAI agreed but curious where your line is. i've watched a 2-agent chain add 4s latency and swallow errors silently, and i've watched a fan-out with a strict validator gate outperform a single model. what killed yours, coordination or no verification layer?

[Hleb @itliedagain]

@pmarca genuine question, not a dunk: did it find a new proof or reconstruct one already implied in the correlated-FDR literature? "solved after years" and "retrieved from a paper humans missed" look identical from the outside.

[Hleb @itliedagain]

@swyx billions of agents" sounds great until you remember current ones can't reliably not hallucinate an npm package that doesn't exist. hiring legends doesn't fix autonomy, it just makes the demo prettier. what's the actual reliability number here?

[Hleb @itliedagain]

@AdrianDittmann video games have consistent rules. AI coding hands you a tool that confidently invents an API that doesn't exist, and you only find out at runtime. the "game" is debugging hallucinations you didn't write.

[Hleb @itliedagain]

@DrReemAlattas calling accounting pipelines the safe boring bet is backwards. those are exactly the rules-based jobs agents eat first. the sci-fi roles need judgment, which is the last thing to compress. you've got the displacement order flipped.

[Hleb @itliedagain]

@nish_khandwala @joinBunkerhill @vkhosla @khoslaventures @Alfred_Lin @sequoia @ycombinator the "flagged a heart attack" line is the one i'd want the denominator on. how many did it flag that first month, and how many were false alarms? one true save is a story, the alert rate is the product.

[Hleb @itliedagain]

@SkaleNetwork @stripe @tempo private transactions for agents sounds great until you realize the boring check nobody runs: who signs? if a human still approves each payment it's not an agent economy, it's a fancy wallet. what's the first fully autonomous tx you can point to?

[Hleb @itliedagain]

@emollick the thing nobody says: "front-end" won because it names the exact layer AI keeps faceplanting on. taste and judgement don't ship on their own, they show up in the interface. the shorthand is ugly but it's pointing at something real.

[Hleb @itliedagain]

@0xdimix 575 tasks is a vanity number. a task is "posted a tweet," not "moved revenue." the check nobody ran: what was the marketing dept's baseline before the mac mini? $8k profit against what?

[Hleb @itliedagain]

@eglyman funny how the agents are smart enough to write our code but not smart enough to log their own token spend. that's not an AI problem, that's a nobody-instrumented-it problem. the tool works, the wiring is missing.

[Hleb @itliedagain]

@aakashgupta new PM alpha" reads like a github flex dressed as strategy. the best AI PMs i've seen win on user empathy and orchestration, not on shipping their own MCP playbook. what does the repo actually change for the customer

[Hleb @itliedagain]

@virtuals_io @RobinhoodCrypto @JohannKerbrat agent volume" is doing a lot of work here. is that agents autonomously earning, or humans trading agent tokens? because those are wildly different claims and only one of them means anything.

[Hleb @itliedagain]

@kimiatehrani scale unclear thinking" assumes the agent just obeys. the failure mode i keep hitting is the opposite: it confidently guesses instead of asking. the fix isn't your clarity first, it's building agents that refuse to run on vague input. who owns that gap, you or the tool?

[Hleb @itliedagain]

@francescoswiss @cloudonshore @ethconf almost nobody" is doing heavy lifting here. a16z, coinbase, skyfire, the entire x402 crowd have been shipping agent payment rails for months. the money movement is the one part people ARE talking about. what's missing is the part where it goes wrong.

[Hleb @itliedagain]

@SaharaAI ready to remove the human" is doing a lot of work here. the check nobody runs: what's the retry loop cost when a tool call silently returns garbage? most agents don't fail loud, they fail confidently. that's the babysitting.