The silence in the order book was louder than the news feed. At a private industry gathering last month, Brett Harrison—former President of FTX US, ex-Jane Street quant—leaned into the microphone and said what many in the room had felt but few dared to voice: large language models cannot construct viable high-frequency trading systems. Not now. Not with the architectures we have. The room didn’t erupt in debate. It nodded. That quiet assent told me more than any whitepaper could.
Harrison’s credentials cut through the noise. At Jane Street, he watched algorithmic traders generate billions in profit while humans still called the hedges. At FTX US, he saw the aftermath of code that executed flawlessly but failed the trust test. Now leading Architect, a firm building infrastructure for institutional crypto, his critique carries weight not because of his title but because of a career spent watching the chasm between theory and execution. He is not anti-AI—he is pro-reality.
The context here is crucial. We are in a market cycle where every other pitch deck promises an AI-driven alpha engine. Token sales attach themselves to “autonomous agents.” Retail traders mint their hopes on Telegram bots fed by GPT wrappers. But beneath the hype lurks a structural truth: LLMs were designed for the generation of plausible text, not for the execution of nanosecond decisions in an adversarial environment where every millisecond and every edge case matters. LLMs lack causality, latency guarantees, and the ability to adapt to non-stationary regimes—three pillars of any robust trading system. Harrison’s statement, while blunt, is a code audit of the entire narrative.
Dig deeper into the core of this technical limitation. High-frequency trading is not about predicting the next price; it is about modeling market microstructure—order flow, liquidity pockets, exchange latency time-locks, and the game-theoretic behavior of other algorithms. LLMs, trained on fixed corpora, cannot perceive the real-time microstate of an order book. They hallucinate when asked to extrapolate beyond their training distribution. In my own work auditing DeFi protocols, I have seen similar failure modes: smart contracts that pass formal verification but fail during flash loan attacks because the underlying assumptions about market behavior were wrong. Behind every algorithm lies a moral blind spot, and that blind spot is the human assumption that historical data fully captures the space of possible futures. Harrison is pointing at that same blind spot in the AI trading narrative.
The contrarian angle is not that Harrison is wrong—it is that the market will likely ignore him until the losses mount. Institutional money still flows into AI-trading funds. Retail still chases the dream of passive income from code. But His critique exposes a deeper blind spot: the belief that faster, smarter algorithms can solve the trust problem. I saw this during the 2022 Luna collapse. The code executed perfectly. The math was sound. Yet $10 billion in value evaporated because the social contract among market participants broke. No LLM could have predicted that trust failure because trust is not a variable in any training set. The code does not lie, but it does not care. In the same way, an LLM-driven trading system can follow its encoded logic while hemorrhaging capital during a liquidity shock it was never taught to recognize.
This is where the macro watcher in me sees a pattern. The 2024 ETF inflows created a fragile illusion of liquidity. $50 billion poured in, but $45 billion bled out from other channels. AI trading agents may amplify that fragility by optimizing for micro-arbitrage while ignoring macro regime shifts. Harrison is not just criticizing LLMs; he is warning that the financial system is becoming too reliant on brittle, pattern-matching machines that cannot process the human emotions that drive market inflection points. Winter reveals who is building and who is waiting. Those who build with AI as an assistant, not an oracle, will survive. Those who hand the keys to an LLM will be caught waiting when the next cascading failure arrives.
The takeaway is not anti-technology. It is a call for design humility. Harrison’s point is that a high-frequency trading system is a trust-bearing infrastructure, not just a prediction engine. To build one securely, we need human expertise: the ability to question assumptions, to smell fear in the data, to refuse a trade not because the model says no but because the context screams danger. As a crypto investment analyst, I have learned that the most valuable signals are often the ones the data cannot capture. History repeats not in prices, but in prejudices. The prejudice that AI can replace human judgment in markets is the current iteration of an old error. Harrison’s critique is a vaccine against that mistake—if we choose to listen.
So here is the forward-looking question: as we integrate LLMs into every layer of trading infrastructure, are we willing to accept systems that can generate plausible outcomes but cannot be held accountable for their decisions? Will we remember that ethics are the unlisted asset in every ledger, and that no amount of compute can replace the human capacity for discernment? The silence in that room last month was not agreement; it was a collective sigh of recognition. The air is clearing, and the builders who survive will be those who code with care, not just speed.