We didn’t spend years auditing the blind spots of algorithmic trading just to hand the keys to a language model. That’s the gut reaction I had when I read Brett Harrison’s recent critique of using Large Language Models (LLMs) to build trading systems. The former FTX US president and current CEO of Architect didn’t mince words: LLMs, he argued, are fundamentally incapable of constructing effective high-frequency trading (HFT) systems. Human expertise, he insisted, remains non-negotiable. For anyone who’s watched the crypto market’s love affair with AI narrative explode over the past year, this is a much-needed splash of cold water.
Let’s step back. Harrison isn’t just another talking head. His resume includes years at Jane Street, the quantitative trading behemoth, followed by leading FTX US through its most aggressive growth phase. He’s seen the sausage-making of market-making and the fatal fragility of automated systems—firsthand. So when he says LLMs fall short, it’s not a tweet from a Bitcoin maximalist; it’s a signal from someone who has built and tested million-dollar trading engines. LLMs like GPT-4 and Claude are brilliant at parsing text, summarizing reports, and even generating code snippets. But they are not designed for the microseconds of a latency-sensitive trade or the deterministic math required to predict order flow.
The core issue is architectural. LLMs operate on probabilistic completions—they guess the next token based on patterns in their training data. High-frequency trading, by contrast, demands precise, repeatable responses to market microstructure events: the size of a bid-ask spread, the rhythm of a particular liquidity provider, the subtle drift in an order book. An LLM cannot model these phenomena with the necessary computational efficiency or consistency. I recall auditing a DeFi protocol in 2021 that attempted to use a neural network to optimize arbitrage paths. The model would hallucinate trades that didn’t exist. The team spent months trying to fine-tune it, only to revert to a deterministic solver. The lesson: probabilistic reasoning is dangerous when your latency budget is a handful of milliseconds.
Open source isn’t just a license; it’s a philosophy of transparency. In trading, that philosophy applies to the logic behind every execution. LLMs introduce a black-box layer that regulators and risk managers despise. If you can’t explain why an order was placed at exactly 10:32:01.457, you can’t defend it during an audit. Harrison’s critique aligns with what I’ve seen in the compliance departments of several crypto prime brokers: the demand for explainable AI in trading is not a preference—it’s a regulatory requirement. The SEC and CFTC are already circling; they will not accept a system that says “the model felt it was a good time to sell.”
But here’s where the contrarian angle gets interesting. What if the real problem isn’t LLMs themselves, but how we’ve tried to shoehorn them into the wrong role? Harrison is right: LLMs cannot build an HFT system. But they might be excellent as auxiliary tools—for example, scanning legal documents for regulatory changes, generating human-readable trade summaries for clients, or even flagging anomalies in log data that a Python script might miss. The mistake is confusing “language understanding” with “market intuition.” A metaphor I use in my newsletters: asking an LLM to trade is like asking a librarian to run a 100-meter sprint. It has the knowledge but not the physiology. Decentralization is not a tech stack; it’s a philosophy of distributing trust. Similarly, AI should be a layer in a stack, not the entire stack.
There’s also the unspoken subtext: Harrison is building his own trading platform, Architect. By publicly setting boundaries on what AI can and cannot do, he positions his product as the “human-centered” alternative—a kind of pragmatic institution that respects the limits of machine intelligence. This is classic Evangelist framing: use a crisis (AI overhype) to define a new standard (hybrid expertise). I’ve done similar things myself. When I pivoted my educational platform to focus on on-chain risk after the FTX collapse, I made the conscious choice to emphasize human judgment alongside code audits. The market rewarded that balanced narrative because it felt honest.
Art isn’t just about creation; it’s about who owns it. For AI-generated trading strategies, the question of ownership is messy. If an LLM suggests a trade that later causes a loss, who is liable? The model provider? The developer who wrote the prompt? The trader who executed it? HFT firms have insurance policies for such scenarios, but those policies are built on clear lines of human accountability. LLMs muddy those lines. This is a risk that most AI-crypto cheerleaders conveniently ignore.
Let me give you a concrete example from my own audit days. In 2022, I reviewed a project that claimed to use a GPT-based model to predict liquidations on Uniswap. The model had a 72% accuracy in backtesting. In production, it was a disaster. Why? Because the market microstructure shifted when a new L2 incentive program changed the liquidity landscape. The model had never seen that pattern in its training data. No real-time adaptation. A human trader would have felt the change after the first 10 minutes and adjusted. The LLM kept firing bad signals for an hour before someone pulled the plug. That’s the difference between pattern recognition and market intuition.
So where does this leave us in the bull market of 2024? We’re already seeing a wave of AI-agent tokens that promise autonomous trading. Most of them are pure narrative plays. Harrison’s critique gives us a technical lens to separate signal from noise. The signal is that LLMs will never replace the wizardry of a Jane Street trader who can read the order book like a cardiogram. The noise is the hype saying “AI will do everything.” My recommendation: treat any project that claims “fully autonomous LLM-powered trading” with extreme skepticism. Instead, look for teams that use LLMs as copilots, not autopilots—where humans review and sign off on every material trade.
A philosophy of transparency requires systems that can be inspected, tested, and audited. LLMs, by their nature, resist that. Harrison’s warning is not just about trading—it’s about the fundamental mismatch between probabilistic consumer AI and deterministic financial infrastructure. As someone who builds educational bridges between these worlds, I’d add: don’t let the excitement of a new tool blind you to the reality that some tools are not built for all tasks.
The takeaway? Harrison is right to push back. But the battle isn’t over. The future likely belongs to systems that combine the speed of machines with the wisdom of human oversight. We didn’t engineer our way out of market risk; we barely learned to manage it. Let’s not delegate that responsibility to a black box. Stay curious, stay skeptical, and always keep a human in the loop.
