When will this AI agents be able to do great math?

4 points·user90131313··7 comments

Is it possible already or in next years/decades? Not meaning something with AI + wolfram alpha at all.

We've been there already, before transformers, with AlphaGo.

The AlphaGo model doesn't lose symbolic inference of the Q states/map it learned whereas transformer based models do.

Whether or not it can learn the coordinates of a more complex function (like f(x) = 123xe^3 + 12x^2 + f2(x)) depends on the amount of neurons in each hidden layer. Depending on how detailed the boosting method is, it is able to learn integrals (just as LSTMs can), the limitations about the detail of the result is limited by the implementation and the amount of neurons, hidden layers, and the boosting method.

Computers are already capable of doing math, and have been for as long as computers have existed. This is not a problem anyone needs AI to solve.

If AI agent can't even calculate certain real cost of doing something inn real world, correctly before deciding, how can it achieve beyond AGI then?

I don't think it can. I think we're still a few paradigm shifts away from true AGI (which I do think is possible) but where we are still isn't it.

Not that that will stop people from just saying "humans are bad at math too" and believing we're already there.

We only need for AI to learn how to use the tools that do maths. GPT-4 does that already with some success, writing code that import math libraries like sympy

No one needs that at all. You don't need AI to do everything. It's already possible to import math libraries and use them.

> No one needs that at all

Maybe not you. But there are people that will certainly benefit from that.

Reply on news.ycombinator.com