LLMs still cannot design a real institutional trading stack without extensive firm context

ivan_bezdomny · x · 2026-07-21

A quoted discussion argues that no LLM has seen the inside of an institutional trading firm, so it cannot naturally infer how to design a real trading system.
The claim is that the only viable way to use LLMs here is to feed them copious context so the output starts to resemble an actual institutional trading stack.
The post frames trading infrastructure as something that encodes a firm's “soul,” and says current models still cannot produce a coherent architecture on their own, even when prompted to build a world-class system without mistakes.
The quoted thread uses a Jane Street-style interview problem as background to show how domain-specific and operationally subtle these systems are.

More from Companies & People