Hacker News Reader

Comment by nopinsight | Hacker News Reader

What the cases so far tell us: Once an AI beats humans who have spent 10,000 hours practicing a skill, it is a matter of time before it beats the best professionals in the field.

Cases it already happened: board games such as Chess and Go, Poker, diagnostics of certain diseases using medical images

Cases where AI is still clearly inferior: video understanding, natural language understanding, motor control esp of hands and legs, general medicine, driving

Hard-to-classified cases (AI is better for some instances, worse for others): image tagging and classification, speech recognition (speech-to-text), diagnostics of certain other diseases using medical images (which might need to take into account other information outside of images)

More examples esp counter examples are welcome.

View on HN

> Poker

This is a misnomer.

The only variant of Poker where AI beats humans is heads-up (two-player) variant, which is simplest form of poker and also rarely played. The AI was (marginally) beating humans there by playing a game-theory-optimal strategy. For poker games with 3+ players, the GTO strategy (Nash equilibrium) no longer exists, so AIs need to use more standard techniques (search-based, reinforcement learning etc.), which are, at the current state of the art, laughably weak at poker.

Not to mention, that in poker the actual hierarchy of players' skill is not 100% obvious. You could distinguish at least two areas of skills:

- play vs other experts

- play vs amateurs/weaker players. Here' the goal is not to come out ahead (which, in long term, is a given), but to _maximize_ the dollar amount taken from these players, which is a skill in itself.

Thanks for the clarification. I do not know much about poker.

The observation applies to a given variant of poker (or any other domain). So if an AI beats humans with 10000-hour experience in that variant, the best experts in that specific variant are not far-off targets.

Superficially similar problems might in fact require very different techniques to solve as your example illustrates.

Modern AI may "beat" humans at certain tasks, but I doubt they are as efficient as humans are at those tasks: it makes sense that throwing more compute at problem will yield better results. But is it possible to get better results while constraining training time?

Edit: I looked up how much time it takes to train: "OpenAI Five plays 180 years worth of games against itself every day, learning via self-play." [1]

[1] https://blog.openai.com/openai-five/

If you can get hundred years experience in 24 hours by wall clock, why not use it? Maybe it's better to teach AIs to create simplified, but usable models, they can use to simulate and train on.

I view the problem in terms of "computational complexity": if the AI of today requires O(2^n) time to perform at a human level, it is possible to get that down to O(n^2)?

I believe there's a connotation with problems, that if the best algorithm to solve it is exponential (brute-force search), then we truly don't understand the (underlying structure of the) problem.

I don't think it is a fair comparison. We have the ability to build somewhat working strategy from known parts, and then we still need thousands of hours (or man-years in the case of Go) to improve it and to teach the brain to do some parts without conscious attention. Alpha Zero, OpenAI Five and others build strategy (mostly) from scratch.

I have a feeling that it is not possible to reduce algorithmic complexity of finding optimal solutions in most of the intellectual tasks (those that are in NP complexity class and above).

Most likely it is a trade-off. Quickly cobble up suboptimal strategy / build better strategy from scratch avoiding all time-saving benefits of using known parts, and avoiding all the pitfalls of not reevaluating utility of those parts in the current situation.

AIs surely will need to use all the spectrum to compete with humans.

OpenAI is far from beating "the best professionals" of dota in a full dota match, though.

Yes. My prediction from the observation above, which agrees with predictions from many others, is that it is a matter of time before that happens (likely within 1-2 years if they continue to work on it).

I actually have an intuition around using modular neural networks with dynamic topology to tackle more complex disease cases, general medicine, and other complex, hierarchical problems. I'm working towards trying to use them in my thesis for school

If anyone has any reading suggestions or research I should look into, I'm all ears

Reply on news.ycombinator.com