People slowly realizing,

“You can't just ask Claude to build you a profitable trading strategy from scratch. This surprises people, because you can ask it to build pretty much anything else from scratch and given enough tokens it'll do a decent job.”

“The difference is that unlike most technical domains, trading is intrinsically competitive. And since anything your agent can build on its own can be built equally well by everyone else talking to the same model, strategies produced this way have no real edge.”

100%

If it’s easy and obvious for your lazy ass do with Claude, then it’s easy and obvious for everyone else to do with Claude, so then zero money zero edge

Just as these guys point out, you need either,

(1) External mining and search organization system of some kind, beyond the vanilla model — e.g. like OpenAI is doing, putting literally $10 million dollars or more into each big math problem they do, along with huge amounts of compute on the Lean search system, along with using the prove2me search orchestration system

Or

(2) Original data that no one else has access to or has thought to use — e.g. …like OpenAi has for math problems because they can see all the math pros’ private chats

Or both

— And yes this is exactly what we did with ClankerQuant, where (1) and (2) are substantial systems now approaching 1M lines of code after 9,214 prompts and 431,905 prompt words over the past 9 months, and now we’re up ~11x on real forward trading

No, you were never going get EV+ out of “trade bitcoin for me claude, make no mistakes”

Rather you need a compute-scalable general mining system that you dump large amounts of compute into, or highly original data with alpha, or both

Bitter lesson

Btw access to ClankerQuant private chat: here

🄳🄾🄾🄼🄿🄾🅂🅃🄸🄽🄶