Brett Harrison, former president of FTX US and current CEO of Architect, just dropped a truth bomb that the AI-hype machine desperately needs. In a recent interview, he stated bluntly: Large Language Models (LLMs) cannot build effective high-frequency trading systems. The market, of course, continues to price AI-trading tokens as if the opposite were true. Read the code, not the pitch deck. The pitch says LLMs will revolutionize quant finance. The code — or rather, the fundamental architecture of LLMs — says otherwise.
This is not a contrarian hot take from a Twitter influencer. Harrison spent years at Jane Street, one of the most sophisticated quantitative trading firms on the planet, before running the US arm of FTX. He lived through the collapse of his previous employer, witnessing firsthand how over-reliance on opaque algorithmic models can lead to catastrophic failure. His current venture, Architect, is building institutional-grade trading infrastructure. He has skin in the game — and he is publicly warning that the emperor has no clothes.
Let me be clear: I am not a trader. I am a crypto security audit partner with a background in applied mathematics. But I have spent the last seven years dissecting financial protocols, from DeFi lending markets to NFT wash-trading rings. I have seen the same pattern repeat: a narrative inflates, capital floods in, and then the underlying technical reality asserts itself — usually at the expense of latecomers. The AI-trading narrative is currently in that euphoric phase. Harrison’s critique is the first credible signal that the turning point may be approaching.
The Context: AI Trading Mania Meets Harsh Reality
Over the past twelve months, the crypto ecosystem has been awash in projects promising autonomous trading agents powered by LLMs. Tokens like $OLAS, $TAO, and various AI-themed meme coins have soared on the premise that GPT-4 or its successors can decode market microstructure better than human quants. Venture capital firms have poured hundreds of millions into startups claiming to replace human traders with chatbots. The thesis is seductive: if an LLM can write poetry, surely it can predict the next price move.
But trading — especially high-frequency trading (HFT) — is not poetry. It is a brutal, low-latency war of microseconds where deterministic execution matters more than linguistic creativity. Harrison’s critique zeroes in on this fundamental mismatch. He argues that LLMs lack the determinism, latency characteristics, and causal reasoning required for real-time market making or arbitrage. I have seen this same disconnect in DeFi protocols that promised algorithmic stability but collapsed under stress — Terra/Luna being the most egregious example. The lesson from that $60 billion black hole: when the code does not match the narrative, the code always wins.
The Core: A Systematic Teardown of LLMs for HFT
Let me deconstruct the technical reasons why Harrison is right. This is not an opinion; it is a mathematical and engineering reality that anyone who has built low-latency systems will recognize.
Latency — The Unforgiving Constraint
HFT operates at timescales measured in microseconds. An LLM inference request to a cloud API, even the fastest one, takes hundreds of milliseconds. That is six orders of magnitude too slow. Local inference with a quantized model might reduce that to tens of milliseconds, but still orders of magnitude slower than a custom C++ strategy running on an FPGA. In my audit experience, I have seen projects claim they can run ML models on-chain. The gas costs alone make it laughable. Complexity hides the body — the body here being the latency budget that no LLM can satisfy.
Determinism — The Hidden Betrayal
LLMs are fundamentally probabilistic. Given the same input, they produce different outputs each time (temperature > 0 is standard). In HFT, non-determinism is a death sentence. A strategy that behaves differently on every tick cannot be backtested, stress-tested, or risk-managed. Regulators require explainability; an LLM that cannot articulate why it executed a trade (beyond generating a plausible sentence) is a regulatory time bomb. I recall auditing a DeFi protocol that used a neural network to adjust interest rates. The output varied by 2% between calls with the same input. That protocol suffered a 40% loss in liquidity within a week because market makers could not hedge the stochastic behavior.
Context Window — The False Memory
HFT relies on market microstructure — order book depth, trade sequencing, latency arbitrage signals — over very short horizons. An LLM’s context window, even at 128k tokens, is designed for text documents, not for the continuous, high-frequency data stream of a limit order book. By the time the LLM processes the last 10,000 trades, the market has moved on. The model is always looking at history, not the present. This is a structural flaw that no prompt engineering can fix.
Causal Reasoning — The Fundamental Gap
LLMs learn correlations from text, not causal structures from market physics. They do not understand order flow imbalance, the impact of a large iceberg order, or the network topology of colocation servers. They can mimic the language of trading but not the mechanics. In my 2020 deep dive into Curve Finance, I discovered that the bonding curve math was sound, but the price oracle response was vulnerable to high-frequency manipulation. A human quant would see the attack vector; an LLM trained on whitepapers would miss it because the attack pattern was not in the training data.
The conclusion is clear: LLMs are not just bad at HFT — they are fundamentally incapable of it at the current architectural level. Harrison’s statement is not exaggeration; it is an engineering fact.
The Contrarian Angle: What the Bulls Got Right
To maintain intellectual honesty, I must address the counterarguments. LLMs do have value in trading — just not in the way the hype suggests.
Natural Language Processing for Sentiment
LLMs excel at parsing news, social media, and regulatory filings to gauge market sentiment. A well-tuned LLM can identify subtle signals in Fed speeches or earnings calls faster than a human analyst. This is a legitimate edge for swing trading or event-driven strategies, but it operates on timescales of minutes to hours, not microseconds. The mistake is conflating this with HFT.
Backtesting and Strategy Generation
LLMs can assist in generating and documenting candidate strategies, but they cannot execute them. A human quant can use an LLM as a brainstorming tool, then write deterministic code to implement the strategy. The AI is a co-pilot, not the pilot. This distinction is lost in most marketing materials.
Risk Management Reporting
LLMs can generate human-readable risk reports from complex data, aiding compliance and oversight. Again, this is an auxiliary function, not core execution.
What the bulls got right is that AI will transform the trading industry — but through augmentation, not replacement. The most successful trading firms will be those that combine human expertise with AI-assisted analysis, not those that hand the keys to a chatbot. Harrison’s own company, Architect, seems to be building exactly that: infrastructure that lets humans leverage AI tools without ceding control. That is the pragmatic path.
The Takeaway: Stop Believing the Pitch, Start Reading the Performance
This article is not a prediction of a crash. It is a call for accountability. If you are considering investing in an AI-trading token or protocol, ask for a technical whitepaper that addresses the latency, determinism, and causal reasoning issues. Demand backtest results on real market data, not simulated environments. Look for evidence that the team has actual HFT experience — not just AI research backgrounds. Most importantly, read the code, not the pitch deck.
I have seen too many projects fail because their economics were built on fragile technical assumptions. The Terra/Luna collapse taught me that no amount of user adoption can save a protocol with a broken technical foundation. The AI-trading narrative is currently sitting on a foundation of sand. Harrison’s critique is the first gust of wind. The question is whether investors will pay attention before the tide goes out.
As for me, I will continue to focus on what matters: empirical data, structural integrity, and the cold, hard reality of how systems actually behave under stress. The LLM delusion will pass — but not before extracting wealth from those who mistake a chatbot for a quant.