Hacker News Reader

Comment by rappatic | Hacker News Reader

I don’t have any kind of ML background but I have always thought of sample efficiency as the great unsolved problem of AI. We humans have unbelievably good sample efficiency; often we can durably learn something on just a single example or two. This is the main area in which LLMs are vastly, vastly behind us.

View on HN

the caveat is that we are not learning those small number of samples from scratch, since we're coming in with a large amount of training already, much of it from before we were even born

I've seen a few videos about people having their girlfriend/dad/etc play very simple video games for the very first time. It's a really great example of this. The most basic, simple, mechanics are difficult for them, and take surprising repetition to remember, because they have zero framework around any of it. It's really similar to watching a genetic algorithm learn to play Mario, but on fast forward.

No, the unsolved problem of AI is continuous learning. We never stop learning, we don't have a "training phase". You are always updating your world model even when you sleep. Also more quality training data does lead to greater learning efficiency as you have more priors to work with.

Why not both?

>We never stop learning, we don't have a "training phase".

We kind of do, if you count critical periods in childhood.

Probable mechanism: "What we found is a rapid increase in GABA in children, associated with learning, ..." [1]

[1] https://www.brown.edu/news/2022-11-15/children-learning

Reply on news.ycombinator.com