Are there any alternatives to supervised (mainstream) AI? | Hacker News Reader

Are there any alternatives to supervised (mainstream) AI?

1 pointsirolimus3 comments

Training, training and more training. Is there any new research on unsupervised learning/or something else which isnt moving in the same direction as 99% of the rest of the AI industry? Any research on something else than neural nets and training statistical predictive and classificstion models? Actual «real» AI?

View on HN

No, not really. When you say "alternatives" you really mean "cool tech demos you can show investors to make them money-horny" like ChatGPT. Investors have no interest in sponsoring a sequel to ELIZA, a "racist Tay" from Microsoft revival, or really even computer vision unless you make it look sexy or put it in a particularly expensive weapon.

There are other things, but they're mostly toyish novelties like LLMs are. Unsupervised learning is almost entirely undesirable in most business applications. If your notion of "real" AI precludes neural nets and statistics classifiers then I am sorry to report that such «real» AI remains a science-fiction myth.

Nothing like live unsupervised learning? That’s the only «real»-like AI that I can «see» being AI-y. We dont have enough and will never have enough data to make the supervised trained models generalizable. We will need something different, because the mixture of experts and alike wont suffice.

FWIW, there's a lot of chatter in the AI world right now about the need for another "big" discovery like transformers were in order to take things to the next level. There could well be a paper out right now that points us to the future, the dots just haven't been connected yet..

For those of us non-academic types working with what we have, though, we'll just have to wait and continue to experiment with the tools at hand. There's no shortage of things to explore, though, such as LLM as judge, eval techniques, and novel prompting approaches.

Reply on news.ycombinator.com