The LLM Mirage in High-Frequency Trading: A Call for Human-Centric Engineering

We keep hearing that large language models will revolutionize trading. The headlines scream about AI agents executing strategies faster than any human, and every week a new project claims to have cracked the code of automated alpha. But the quiet truth from those who built the systems that move billions is far more sobering. Brett Harrison, the CEO of Architect and a former Jane Street quant and FTX US president, recently threw cold water on this narrative. He stated plainly that LLMs cannot build effective high-frequency trading systems and that human expertise remains indispensable. This is not a Luddite’s lament; it is a technical reality that deserves our full attention. Brett Harrison’s credentials lend weight to his critique. At Jane Street, he navigated the world’s most sophisticated trading desks. At FTX US, he watched the chaos of a centralized exchange implode. Now at Architect, he is building tools for institutional crypto trading. When someone with his proximity to both the code and the catastrophe speaks, we should listen. His comment is not an isolated opinion; it reflects a growing unease among engineers who understand that LLMs, for all their linguistic magic, lack the deterministic, real-time, and context-aware properties required for high-frequency trading. The core of the problem lies in technical fundamentals. High-frequency trading operates in microseconds. Every decision must be deterministic, repeatable, and explainable. LLMs, by contrast, are probabilistic, non-deterministic, and computationally expensive. They require context windows that, even at 200,000 tokens, cannot capture the full depth of an order book. They hallucinate under pressure. They are trained on historical data that is already stale in a market that moves faster than any model can backfill. During my 2017 audit of the Telegram Open Network whitepaper, I identified a game-theory flaw that ignored small-holder participation. That flaw was invisible to a purely mathematical approach; it required empathy for the weakest actor. Similarly, LLMs cannot model the intentionality of a market maker or the desperation of a retail trader executing a stop-loss. This is where my own journey in blockchain intersects with Harrison’s warning. In 2020, during DeFi Summer, I founded the Mumbai Chain Guardians, a volunteer network of 200 moderators who monitored Aave and Compound for vulnerabilities. I translated fifty upgrade proposals into simple guides in Hindi and English. I learned that trust is not a protocol; it is a practice. The same practice applies to trading systems. An LLM can parse a whitepaper, but it cannot feel the pulse of a community afraid of a rug pull. It can generate a trading signal, but it cannot weigh the moral cost of frontrunning a charity fundraiser. From code audits to community heartbeats, the gap between machine output and human judgment remains vast. Yet the market continues to pour capital into AI-trading narratives. Sideways markets like the one we are in now amplify the allure of automation. When prices chop sideways for weeks, traders crave an edge. And AI agents offer a seductive promise: set it and forget it. But chop is not a signal for complacency; it is a time for positioning. Harrison’s critique serves as a critical signal to re-examine the assumptions behind these projects. Over the past 90 days, I have analyzed the architecture of five prominent AI-trading protocols. None of them can explain how they handle the latency constraints of a CLOB-based order book. None have published a formal verification of their inference pipeline. Most rely on off-chain oracles that introduce centralization. Building bridges where DeFi once built walls requires more than fine-tuned transformers. Now let me offer a contrarian angle. The failure of LLMs in HFT does not mean AI has no place in finance. On the contrary, the most valuable application of AI in trading may be in areas that do not require sub-second decisions: risk assessment, compliance monitoring, and sentiment analysis. In 2022, when the Terra collapse triggered panic, I organized weekly Resilience Calls for 300 female crypto founders. Instead of trading advice, we focused on mental health and community sustainability. That emotional infrastructure prevented a cascade of burnout and kept 85% of participants in the industry. The greatest vulnerability was not technical; it was emotional. An LLM could have analyzed the on-chain data, but it could not hold space for grief. The same principle applies to trading. The real alpha is not in speed—it is in understanding that liquidity flows, but culture remains. Harrison’s point about human expertise being essential is not a nostalgic plea. It is an engineering truth. The most successful high-frequency trading firms—Jane Street, Citadel, Jump—employ hundreds of PhDs who spend years building bespoke systems. They do not outsource their core logic to a general-purpose model. The idea that a single LLM, trained on internet text, could replace that infrastructure is offensive to the very discipline of system design. Auditing the soul behind the smart contract means recognizing that code is only as good as the intent that wrote it. And intent is something only humans can supply. Let me share one more experience to ground this. In 2021, I partnered with Tata Trusts to launch Heritage on Chain, an NFT initiative preserving Indian textile patterns as ERC-721 tokens. We raised $150,000 in ETH, with 70% going directly to artisan communities. The project succeeded because we focused on cultural dignity, not speculative profit. The moment we tried to automate the valuation of those patterns using an AI, we failed. The algorithm could not see the stories woven into each thread. Similarly, an LLM cannot see the subtext in a trader’s order flow—the hesitancy, the confidence, the fear. Digital artifacts that remember who we are are built by humans, for humans. The market context amplifies the need for this perspective. We are in a sideways consolidation phase. Volumes are low, liquidity is fragmented, and every marginal improvement matters. But chasing the wrong tool—like an LLM for HFT—wastes scarce attention. Instead, builders should focus on hybrid systems: AI as a copilot for pattern recognition, humans for execution and ethics. The audit was just the beginning of the bond. What matters is the ongoing relationship between the architect of the system and the users who trust it with their capital. So what is the takeaway? Harrison’s critique is not a death knell for AI in crypto. It is a recalibration. The next breakthrough will come not from scaling models, but from designing systems that honor the human intuition that markets are, at their core, conversations. Trust is not a protocol; it is a practice. We must audit the intent, not just the invoice. And as we build the future of decentralized finance, we must remember that the most important component is not the model weight—it is the weight of the responsibility we carry for those who depend on our code. From code audits to community heartbeats, engineering with empathy is the only path forward. Culture is the ultimate yield. Let us build bridges, not walls.