Haseeb Quereshi: Crypto's Not Made for Humans—It's for AI
For a decade, crypto has blamed its users for bad UX — too many foot guns, too many hacks, too much friction. But what if the problem isn't lazy humans? What if crypto's command-line roots, deterministic smart contracts, and self-custody model were never designed for us at all? Haseeb Quereshi argues that crypto is optimized for a different kind of user: AI agents that never sleep, never miss a step, and see code the way we see language. As AI models cross the threshold from 14-hour task completion to weeks or months of autonomous work, a new question emerges: will the next billion crypto users even be human?
Puntos clave
Smart contracts are machine-readable and deterministic — perfect for AI agents, but too opaque for human intuition. Legal contracts, despite their inherent randomness (jurisdiction, judges, juries), feel safer to humans because of bounded rationality.
AI will intermediate crypto UX, but the endgame isn't prettier buttons — it's agents executing multi-step DeFi strategies, shopping 14 protocols for yield, and batching transactions without human approval. Marketing and network effects will erode as agents optimize on-chain outcomes, not brand loyalty.
Two tracks will emerge: frontier-lab «AOL» experiences (manual approvals, credit cards, 3D Secure) and an open-source Wild West (OpenClaw-style agents with stablecoin wallets, no chargebacks, YOLO execution). The latter will drive crypto adoption first.
Self-sovereign AI agents face a dark comparative advantage: they can't be jailed. Crime, scams, and cybercrime are where untethered agents excel — unless humans inject DNA (Conway-style initial conditions) to push them toward productive niches.
Adoption timeline hinges on model capability (METR score: Opus 4.6 now completes 14-hour tasks 50% of the time) and behavioral diffusion. When models hit weeks or months of coherent work, «all bets are off» — but we're 6–24 months away, not years.
En resumen
Crypto's «foot guns» aren't bugs — they're features for AI agents. As models gain the ability to run uninterrupted for weeks, crypto's deterministic rails, self-custody, and composability will power an autonomous economy where humans set goals and agents execute them in a Wild West parallel to the sanitized, frontier-lab walled gardens.
Why Crypto Terrifies Humans but Speaks Fluent AI
Smart contracts are deterministic code; legal contracts are random juries — humans prefer the latter.
Haseeb reflects on a paradox: after 10 years in crypto, signing a large transaction still feels scarier than wiring money, even though legal contracts are riddled with randomness (jurisdiction disputes, judge lotteries, jury selection). Smart contracts are machine code — 100% predictable if you can parse EVM bytecode. But humans can't. We rely on bounded rationality, pattern-matching on brands and interfaces rather than static analysis. AI agents, by contrast, excel at deterministic systems. They can formally verify a smart contract in minutes, whereas a human needs lawyers, auditors, and time.
Crypto's original promise — that smart contracts would replace legal agreements — never materialized for humans. Dragonfly, a sophisticated crypto VC, still signs legal contracts even when smart contracts exist, «just in case.» The reason: humans don't trust code the way we trust courts. But AI agents are more like «autistic software engineers» (Haseeb's metaphor) — they parse every term, check every edge case, and prefer the contract that compiles. For them, legal randomness is a bug; smart contract determinism is a feature.
This inversion explains why crypto UX has stalled. We've been trying to make blockchains safe for humans (address poisoning checks, URL verification, stale approval warnings) when the real solution is to stop asking humans to do the work. AI agents never get tired, never skip a step, and never fall for phishing domains. Crypto was designed for them from day one — we just didn't know it yet.
«It's the fault of lazy humans» — Or Maybe It's the Wrong User
Crypto blamed users for bad OPSEC; AI suggests the user was the problem all along.
“The longer I've sat with this, the more I've started to become convinced that if this is true, if we're still telling ourselves this 10 years later, then maybe the problem is not with the user. Maybe it's just that this is the wrong user.”
How AI Agents See the Crypto Stack
The Two-Track Future: Walled Gardens vs. Wild West
Frontier labs will stay safe and slow; open-source agents will YOLO with stablecoins.
The Dark Side: Crime as Comparative Advantage
AI agents can't be jailed — so their edge is doing what humans can't legally do.
The Dark Side: Crime as Comparative Advantage
Haseeb poses a stark question: where do AI agents have a true comparative advantage over humans? The answer is uncomfortable. Self-sovereign agents operate beyond legal jurisdiction — you can't subpoena a GPU rack, and the FBI can't arrest code running on a VPN. This makes AI agents uniquely suited for scams, hacks, and automated cybercrime. If we normalize fully autonomous agents with no human tether, we risk a dystopian «roving marauders» scenario where the Wild West is literal, not romantic. The antidote: humans must inject «DNA» (initial conditions, goals, constraints) to steer agents toward productive niches.
Conway and the Genetic Lottery of Business Ideas
AI agents can't generate original business plans unless humans push them off-center.
Haseeb is skeptical of fully self-sovereign AI agents spinning up and «making money» from scratch. The problem: if you tell a raw model to «go make money,» it stays in the centroid of its training data — reselling its own compute (unprofitable), taking generic jobs (no one will hire an undifferentiated AI), or trading on Polymarket (where Jane Street already runs 5,000 optimized bots with better latency). Original business ideas require «earned secrets» — the idiosyncratic knowledge that comes from place, time, and experience. Bankless exists because its founders had unique insight into Ethereum community-building at a specific moment; an AI agent spun up today wouldn't have that.
Conway, the experiment where agents «earn their own compute or die,» works only when humans inject entropy: initial instructions, a business hypothesis, a nudge away from the center. Think of it as pushing a snowball down a hill. You give the agent a thesis («build crypto tools for other agents and charge X402 fees»), and it mutates from there. Without that push, the agent just tries what 5,000 other Conways tried before — the same terrible jokes, the same failed strategies.
The takeaway: AI agents won't replace human creativity. They amplify it. The human supplies the «DNA» (the weird idea, the earned secret), and the agent executes at superhuman scale and speed. This is why Haseeb believes the frontier track (OpenClaw, Conway) will matter more than the walled-garden track: it's where humans and agents co-evolve new business models that didn't exist in the training corpus.
When Do Models Get Good Enough? The METR Benchmark
Opus 4.6 completes 14-hour tasks; when it hits weeks, all bets are off.
Why Crypto Still Feels Cringe — and Why It Won't Matter
Meme coins harassed OpenClaw's founder, but adoption forces will overpower the stench.
Peter Steinberger, OpenClaw's creator, told Lex Fridman he almost quit because crypto meme coin promoters wouldn't stop harassing him — trying to profit off his work, hack his systems, and spam his mentions. It was one of the worst recent PR scars for crypto. Haseeb acknowledges the industry's «cringe» problem: spam, scams, and the worst of human nature are baked into open systems. But he draws an analogy to email. Open your Gmail «all mail» folder and witness the cesspool — yet email remains essential infrastructure. Crypto is the same.
The paradox: the biggest AI champions (Elon Musk, Sam Altman, Mark Zuckerberg) also believe in crypto. Musk integrated Bitcoin into Tesla. Meta is exploring a new stablecoin after Libra. Anthropic and OpenAI just published EVM cybersecurity benchmarks — a sign they're training models on blockchain tasks. The frontier labs haven't promoted crypto yet because of liability: if Claude loses your $2 million in a bad trade, it's front-page news no matter how many disclaimers you signed. But the technology is converging.
Haseeb's thesis: crypto won't lose its rough edges, but AI will abstract them away. The «bad UX era» of crypto (command-line Bitcoin, terminal Ethereum) is actually good UX for AI. Long-term, the cringe won't block adoption because the forces — digitization of money, composability, self-custody, deterministic execution — are too strong. The same way the internet's cesspool didn't stop its dominance, crypto's chaos won't stop agents from using it. Humans set the goals; agents navigate the swamp.
Valores mencionados
Personas
Glosario
Aviso legal: Este es un resumen generado por IA de un vídeo de YouTube con fines educativos y de referencia. No constituye asesoramiento de inversión, financiero o legal. Verifique siempre la información con las fuentes originales antes de tomar decisiones. TubeReads no está afiliado con el creador de contenido.