> The timescale of the project needs to be shorter than the characteristic timescale of the exponential

So the advice is to pick a really fast project. But I feel like there’s a big caveat here: this is the timescale of improvement specifically for general purpose LLMs.

So you could also just… pick a project in an area that has a longer characteristic timescale for improvement. One where you’re not competing with industry researchers backed by literally hundreds of billions of dollars in funding.

AI for the physical world is not advancing nearly as fast. Pair with a robotics group and come up with a better way to generate robust datasets and use reinforcement learning to do real-time control.

I did a lot of my PhD work in image vectorization for complex material microstructures. Very niche. Still unsolved by AI over a decade later and not showing any signs of being solved soon.

Text processing is probably going to be a solved problem and hard to keep up with, I don’t disagree with that. It’s far from the only field of AI, and much less so the overall R&D.