Hacker News Reader

Comment by hdivider | Hacker News Reader

This is where most of the real value of AI actually lies. The scientific and engineering applications of machine learning.

Another MIT study: if we AI-augmented R&D widely in the US, our productive economic growth rate would double. The authors argue this would be permanent, forever increasing the rate of technological progress:

https://www.sciencedirect.com/science/article/pii/S004873332...

I bet a true accounting of all the economic impacts of semiconductor technologies will show these have far more value than Meta, Instagram, Snapchat, TikTok all combined. The stuff built on top of deep technology has higher PR value -- but the deeper tech has far more economic value. Same with AI.

View on HN

This is slightly tangential but related.

The reason AI will do great things all levels of reality and different fields of studies (i.e. physics, biology, material science, climate, etc) is that these aspects operate from rules governing behaviour at the specific scale. These context specific rules derive from and summarise fundamental physics directly. This also explains why these methods will succeed in every domain. Analogy: Euler's rule is a topological invariance. So applies to both Spherical and Euclidean geometry.

The second, also called the Euler polyhedra formula, is a topological invariance (see topology) relating the number of faces, vertices, and edges of any polyhedron. It is written F + V = E + 2, where F is the number of faces, V the number of vertices, and E the number of edges.

The difference is that Euler’s rule doesn’t need to be verified after each use of it.

AI may speed up research by uncovering interesting patterns, but I don’t believe it’ll do great things all by itself.

AI is, after all, just a name for a statistical model of something. We’ve been using statistical models for a looong long time.

Any domain already relying heavily on statistical models (like protein folding, I believe) may really benefit from AI (like alpha fold), but domains which don’t won’t.

It’s no silver bullet.

Protein folding historically was more reliant on physical simulations. Like, David Shaw (of DE Shaw fame) had/has a team optimizing the hell out of simulations. There are oodles of fields (like weather) doing similar, and they are finding they can use statistical models where they previously thought they couldn't.

Imagine a futuristic sci-fi scenario in which AI is so advanced that it can model so deep into physics (i.e. plank scale), a scale which can't be probed by experiments but the predictions match. We would have a probabilistic Oracle of sorts. It can give answers but no way of knowing why it works.

I don't think the why is that confusing. If you do a physical simulation, so many calculations cancel each other out or work in opposite directions. If there is a pattern to it (and it appears there is for many situations), you just short cut it.

Consider Roger Federer's ability to predict where a tennis ball will land and how to send electrical signals to his body to move in a way that will return that ball with high precision. It's pretty wild the number of short cuts he can make for what is a very complex calculation.

On second thought, I missed your point, I think. In my example we could reason through why the tennis ball ended where it did. In yours, we couldn't. My bad.

I agree with your point of calculations cancelling out. That's why I am not a fan of Butterfly effect. Just as Lewis coined unbirthday, we should coin Unbutterfly effect cancelling each other out.

Reply on news.ycombinator.com