I think this is a very valid hypothesis but it's hard to control for in experiments, since if these traits are necessary to become (or stay) famous, we don't really have a control group.
LLMs are largely used by developers, who (in some sense or the other) supervise what the LLM does constantly (even if that means for sum committing to main and running in production). We do already have a lot of tools: tests, compilation, a programming language with its harsh restrictions compared to natural language, and of course the eye test, this is not the case for a lot of jobs where GenAI is used for hyperautomation, so I am really curious in which way it will or won't get adopted in other areas.
I just open-sourced: https://github.com/cloudexplain/xaiflow, a mlflow plugin to get interactive xai (particularly shap values) as mlflow artifacts.
Furthermore looking into causal AI, especially dowhy.
Thanks for the question, there are a couple of existing solutions:
- There is already a mlflow builtin tool to log shap plots. This is quite helpful but becomes tedious if you want to dive deep into explainability, e.g. if you want to understand the influence factors for 100s of observations. Furthermore they lack interactivity. Here's the link to the builtin tool: https://mlflow.org/docs/latest/ml/evaluation/shap
- There are tools like shapash or what-if tool, but those require a running python environment. This plugin let's you log shap values in any productive run and explore them in pure html, with some of the features that the other tools provide (more might be coming if we see interest in this).
- Project manager mathematicians: Tao draws a future where mathematical insights are "produced" like anything else in our society. He attributes the lack of this productionalization of mathematics to a lack of tools in this area but AI and proof assistants might be revolutionary in this regard (proof assistants already are). Human interaction and guidance still needed
- Implicit Knowledge: he points out that so much knowledge is not in papers, e.g. intuition and knowledge of failures. This is crucial and makes it necessary even for top mathematicians to talk to one another to not make mistakes again.
- Formalization of mathematics: one would think that mathematics is pretty formalized already (and it is) but in the papers a lot of common knowledge is taken for granted so having proofs formalized in such a way that proof assistants can understand will help more people actually understanding what is going on.
I think this just shows how Tao always searches for new ways to do mathematical research, which I first came across in his talk about one of the polymath projects.
It's a nice read but I doubt that the idea delivers what it promises.
I think one could downsize the idea to: if you learn one topic deeply you (probably) touch surrounding topics. That does not sound as elegant and holistic as in the article but is IMO closer to the truth. I would call it a huge exaggeration to say I've learned things about the world when in reality I just became a really good software engineer that knows how to interact with OSS communities.
These ideas are repeated often and I lean more to the specificity side of things: you only get good at what you learn/train. You won't become better at decision maker by learning chess/poker, you won't (or just become a slightly better) endurance swimmer by becoming a good runner, you won't understand human psychology by getting good at coding.
I remember a talk of top-notch mathematics where they were asked about related mathematical topics and most of them would just answer something like: "I just try to understand my field of mathematics well, I can't say much about something else". This was the discussion from the Breakthrough Prize in Mathematics 2015: https://www.youtube.com/watch?v=eNgUQlpc1m0&list=PLyF3OMOiy3...