First off, I know this is a controversial topic but it should be discussed as it is an inevitable fate for anyone in quant.
Who would you say is the value add of a quant for an asset manager?
It’s a cost center so this is a serious question.
Is it raw knowledge (PhD+)?
Maybe, but most of it is encoded in LLMs now.
Is it understanding structure and how to solve problems?
Yea, that seems more aligned. But even then, LLMs and context-aware agents like Cursor for programming is bridging this gap rapidly.
Many quants scoff at the idea of using Cursor to build their models and trading strategies but I think they are using it incorrectly or their ego refuses to accept that years of academic training is becoming commoditized.
Here are some common guidelines that has helped me cured myself of LLM disease syndrome (LDS).
1. LLMs don’t understand context very well. That’s why you need something like Cursor to inject codebase context. It does a good job.
2. Yes, you absolutely have to refactor your codebase for any real project. Adding docstrings or asking the LLM to do so is also critical. This allows you want to work in chunks and not have to iteratively go back and forth.
3. No you shouldn’t write large prompts. You need to iteratively use the LLM. It’s human in the feedback loop that makes the coding powerful. You avoid taken unintended code jumps that are suboptimal.
The way I think about this is that you have a global reward function – your ideal model or program – which you are asking the LLM (along with a context engine) to find in an insanely high dimensional vector space. Yes this is not 100% accurate technically, but a good mental picture.
To guide it properly down the valley you need to do it gradually and iteratively to avoid getting stuck in suboptimal valleys (this is what happens when you inject large prompts usually). It’s like hiking down a large mountain, you don’t jump down, you do it gradually.
If you approach LLM agents with this perspective, maybe you can keep your job for another 10 years.
