The training methodology is more interesting than the result. They didn't teach the model what to read. They taught it what to skip. The expert judgment they captured wasn't "this paragraph matters." It was "these 200 pages don't." That's the knowledge that can't be prompted. You can tell a frontier model to act like a macro investor. You can't tell it which 80% of a 10-K is boilerplate without already knowing yourself.
The cost number is the one that matters at scale. 13.8x cheaper per task turns AI from an experiment into infrastructure. At Bridgewater's query volume, the difference between frontier pricing and fine-tuned pricing is probably eight figures annually. That's the gap between "we're exploring AI" and "AI is in the production stack." Most firms are stuck on the wrong side of that gap because they're paying frontier prices for tasks that don't need frontier capability.
Next market event predictions from AI will have about the same accuracy as humans because...AI is simply regurgitating data and biases fed by humans. Am I missing something? Or has Bridgewater discovered AGI before the labs?
Hey sorry for the delay. I think of what Bridgewater has done more as a way to boost efficiency and the quality of AI outputs. It doesn’t make sense to have AI, or humans for that matter, read a bunch of legalese that doesn’t’ really say anything. It’ll just end up confusing the agent. And when you’re using AI scale costs add up quickly..
The training methodology is more interesting than the result. They didn't teach the model what to read. They taught it what to skip. The expert judgment they captured wasn't "this paragraph matters." It was "these 200 pages don't." That's the knowledge that can't be prompted. You can tell a frontier model to act like a macro investor. You can't tell it which 80% of a 10-K is boilerplate without already knowing yourself.
The cost number is the one that matters at scale. 13.8x cheaper per task turns AI from an experiment into infrastructure. At Bridgewater's query volume, the difference between frontier pricing and fine-tuned pricing is probably eight figures annually. That's the gap between "we're exploring AI" and "AI is in the production stack." Most firms are stuck on the wrong side of that gap because they're paying frontier prices for tasks that don't need frontier capability.
Next market event predictions from AI will have about the same accuracy as humans because...AI is simply regurgitating data and biases fed by humans. Am I missing something? Or has Bridgewater discovered AGI before the labs?
Hey sorry for the delay. I think of what Bridgewater has done more as a way to boost efficiency and the quality of AI outputs. It doesn’t make sense to have AI, or humans for that matter, read a bunch of legalese that doesn’t’ really say anything. It’ll just end up confusing the agent. And when you’re using AI scale costs add up quickly..