Large language models are remarkable at synthesis, reasoning, and explanation — and completely wrong for some parts of a trading system. Using them well means knowing exactly where they belong and, more importantly, where they don't.
Where an LLM helps
LLMs excel at synthesis: taking many disparate inputs — structure, volatility, positioning, the character of the tape — and weaving them into a coherent read, in plain language. They're strong at contextual judgment and at explaining a situation clearly. That maps well to the "perception and reasoning" layer of a system — forming and articulating a view of the market.
Where it absolutely doesn't
An LLM has no place on the millisecond trigger. It's comparatively slow, non-deterministic (the same input can yield different output), and can occasionally produce confident nonsense. You never want a model improvising the actual order — the size, the price, the moment. That belongs to fast, deterministic rules with hard limits.
Let the LLM write the market's story. Never let it press the buy button.
The correct division of labor
The robust pattern: the LLM (and other AI) informs a read — a contextual view of the market — on a sensible cadence, while a deterministic engine handles the fast, repeatable execution within pre-set risk. The intelligence shapes understanding; the mechanics act, always inside guardrails. That's the design behind NoVo — AI for the read, a disciplined machine for the trade.