Two years ago, frontier models couldn't write a working Python script. Today they ship production software. Cursor writes features. Devin closes pull requests. Claude reasons through systems architecture better than most senior engineers.
The coding gap went from impossible to largely solved in 24 months.
The same kind of gap is about to close for quantitative finance. The argument is structural, not speculative.
What quant actually requires
Strip a quant role down to its hard prerequisites and there are two.
The ability to write code. Backtests, data pipelines, execution systems, risk overlays. Quant is a coding job that happens to have a P&L attached.
The ability to reason under uncertainty. Why is this signal real? What's the regime? What breaks this trade? When do you cut? You're making decisions with incomplete information against adversarial counterparties.
That's it. Everything else - the math, the market knowledge, the intuition for microstructure - sits on top of those two foundations. If a system can code and reason, it can in principle do quant.
Two years ago, LLMs failed at both
In 2024, GPT-4 could write a function but couldn't structure a project. It could answer a math problem but couldn't trace why its answer was wrong. It generated code that looked right and failed in production. Reasoning was shallow. Coding was brittle.
The gap to “quant researcher” looked enormous because both prerequisites were missing.
Today, LLMs pass both
Coding. Frontier models now solve SWE-Bench tasks at the level of senior engineers. They write production features end-to-end. Every major hosted IDE is built around them.
Reasoning. The same models score at the top of mathematical olympiad benchmarks. Multi-step tool use, planning, self-correction, working memory across long horizons - all of it works now.
The two prerequisites that take human quants five to ten years to develop, frontier models acquired in 24 months. The capability ceiling is no longer the bottleneck.
So why can't they trade yet?
We ran the first rigorous evaluation of frontier AI models on quantitative trading. Six models. Thirty independent trials. A multi-stage pipeline with hidden out-of-sample data and an 80% net-exposure constraint that prevents the directional bets models default to.
Best Sharpe across all six models: −2.098 average. Pass threshold: +2.0. None passed. Only one trial out of thirty produced a positive out-of-sample Sharpe.
Models that ship pull requests for Anthropic cannot construct a hedged portfolio. Models that solve IMO problems cannot read an order book. The capability is there. The performance is not.
The answer is training data, not capability
There are roughly 1,000 software engineers for every quantitative researcher. Engineers publish on GitHub, Stack Overflow, and in open-source libraries. Quants never share theirs. Trading strategies are trade secrets. Execution logs sit locked inside prime brokerage systems. The knowledge that separates a profitable strategy from a losing one has never been digitised, never been published, never been uploaded.
LLMs can code because the internet is full of code. They can reason about math because the internet is full of math. They can't trade because the internet has no alpha.
This is a data problem, not an intelligence problem. And data problems are tractable.
What changes
The capability ceiling is high enough. The two hard prerequisites are met. What's missing is the training signal - and the training signal can be built.
Build realistic market environments that LLMs can learn from. Build evaluations that can't be gamed. Capture expert decisions from people who have actually traded. Train on the data the internet doesn't have.
This is what RAETH exists to do. We're building Enigma - a foundational AI for markets, from first principles. AlphaZero for markets.
Two years ago, AI couldn't code. Today, it codes. Today, AI can't trade. By the same logic that closed the coding gap, the trading gap closes next.
Quant is next. Machines making machines making money.
