Ask HN: What are the hottest areas of *non*-LLM AI work currently?
10+ years ago, "AI" would likely refer to work in RL, evolutionary/genetic algorithms, etc.
Nowadays most of the spotlight seems centered on LLMs, CNNs, and other methods that are either human-labeled or at least reliant on human-created data, and have a static separation between "learning" and "inference".
I know that there are still non-LLM, non-CNN, non-anthropocentric topics of AI development currently, in RL and in other areas. Which would you say are the most prominent or promising today, or likeliest to come to fruition?
In order of nicheness
- Singular learning theory
- Vector-symbolic architectures
- Homomorphic learning
BTW, on the topic of the fading of genetic algorithms, here is an interesting recent take: https://statmodeling.stat.columbia.edu/2025/04/17/what-happe...
The quoted paper is from 2018. Evolutionary programming is IMO similar to a search algorithm. The biosphere is not. It's not searching; it's removing unfit matches. Which is far less efficient.
I'm a big fan of evolutionary programming; it's just inefficient in the past. I think LLM agents might just be the little advantage they need, like guided missiles with GPS.
Evolutionary programming is hard though. I think it might answer OP's question - it's something that's difficult enough for most people to avoid, but there would be greatly increased interest in it.