[jrx]

> This is like if the world's best chess AI had gone from losing high school tournaments to being competitive with Kasparov in less than 3 years.

I don't think it's like that at all. On the high level, there is no "chess AI", "go AI", "image classification AI" and "dexterous manipulation AI". These are all sides of the same coin, that gets significantly better every year. Adding support for the new game or new "environment" to existing deep learning based backbone still requires a bit of engineering work and a few creative tricks to unlock the best possible performance, but the underlying fundamentals are already there and are getting better and better understood.

There is a reason why the progress in AI is so hard to measure. Anytime a next task is solved, there is a crowd saying it's not a "real AI" and that scientists are solving "toy problems". Both statements are totally true. But the underlying substance is that each of these toy problems is of increasing complexity and brings us closer and closer to solving the "real problems", which are mostly so undeniably complex that we couldn't attack them upfront. Still, the speed of progress in the field of AI research is staggering and it's hard to keep up with it even for professional researchers who spend all their waking hours working on these things.

6 years ago we were able to solve some Atari games from pixels. Today, that feels like a trivial exercise compared to modern techniques. With billions of dollars of investment pouring in and steady supply of fresh talent, it is very hard to predict what the pace of research will be in the coming years. It is entirely possible we'll encounter a wall we won't be able to overcome for a very long time. It is also possible that we won't, and in that case we're in for a very interesting next few decades.

[dmreedy]

> On the high level, there is no "chess AI", "go AI", "image classification AI" and "dexterous manipulation AI". These are all sides of the same coin, that gets significantly better every year.

On a practical level, this is not true. There are different algorithms, different architectures, different hyperparameters required for each of these problems, and often for each subdomain within each of these problems, and often for each specific instance of these problems. It's difficult to draw any kind of holistic picture that combines all of the individual advances in each of these problem instances; that's why progress in AI is so hard to measure, and why a statement like "each of these toy problems...brings us closer and closer to solving the 'real problems'" is probably a bit too coarse-grained to be fair as well.

[tialaramex]

Are you writing this from last century?

Deepmind's best-in-class chess and Go AIs are the same code (AlphaZero) just given respectively rules and game state input for either chess or Go and then allowed to train on the target game.

One of the fun works in progress in this space is teaching AIs to play a suite of 80s video games. Getting quite good at several games where the idea is to go right and not die is pretty easy these days, but Deepmind's work can do a broader variety only coming badly unstuck on games where it's hard to discern your progress at all without some meta-knowledge.

[dmreedy]

I don't mean to imply AlphaZero is not impressive; it surely is. Nor do I mean to imply that any of these advances aren't impressive. I do mean to imply that "closed-world games with well-defined rules" is a relatively small subdomain of problems. And that BERT looks pretty different from AlphaZero.

[tialaramex]

The post you disputed pointed out that there aren't separate AIs needed for things like Go or Chess. Because there aren't (any more) the Deepmind work showed that you can just generalize to learn all games in this class the same way.

You claimed that "different architectures" are needed. Not true. And further you claimed this is true even for "each subdomain". This would have been a fair point in 1989. Traditional chess AIs approach the opening very differently for example, relying on fixed "books" of known good openings. But AlphaZero since it is a generalist doesn't do this, it plays every part of a match the same way.

Now you've gone from asserting that Chess and Go need separate AIs to claiming that since BERT and AlphaZero are different software it makes your point. Humans pretty clearly don't have a single structure that's doing all the work in both playing Go (AlphaZero) and understanding English (BERT) either - so that's a pretty bold bit of goalpost moving.

[the_af]

First of all, nice comment! That said,

> Anytime a next task is solved, there is a crowd saying it's not a "real AI" and that scientists are solving "toy problems". Both statements are totally true. But the underlying substance is that each of these toy problems is of increasing complexity and brings us closer and closer to solving the "real problems"

I wonder if this is true. This belief may seem like common sense, but it's not obvious to me that domain-specific problems must generalize to General AI ("real problems") or even bring us closer to it. That is, it's not evidently true that many small problems will eventually lead to a general solver of everything (or to human-like intelligence). Or to say it in yet another way, it's not obvious to me that human-like intelligence is the sum of many small-problem-intelligences.

Again, common sense may lead us to believe this, and maybe it's true! But I think this conclusion is far from scientifically evident.

[IX-103]

The key thing you're missing is transfer learning. Instead of starting from scratch, you start with a model that was trained to do something and then train it to do something else. It takes much less time and labeled data to get the model to do something similar.

You can even interleave the training for the second task with a few training rounds for the first task to maintain proficiency. There's a group that's using this sorry if technique to make a general "plays videogames" AI. I couldn't find a good link from my phone, but here's a less good link about something similar: https://towardsdatascience.com/everything-you-need-to-know-a...

[YeGoblynQueenne]

>> On the high level, there is no "chess AI", "go AI", "image classification AI" and "dexterous manipulation AI".

As another poster said these are all tasks performed by different systems. For chess and Go AI it's Deep Reinforcement Learning with Monte Carlo Tree Search. For image recognition it's Convolutional Neural Networks. Importantly, these systems are very task-specific. You won't find anyone trying to beat humans at games using CNNs, for example, or using Deep-RL to do text recognition. Far from "a few creative tricks" these are systems that are fundamentally different and are not known to generalise outside their very limited domains. They're one-trick ponies.

The OpenAI paper on "dexterous manipulation" reported learning to manipulate one cube, the same cube, always, after spending a considerable amount of resources on the task. It was a disappointing result that really shouldn't be groupwed with CNNs and Deep-RL for game playing. The level of achievement does not compare well.

>> Anytime a next task is solved, there is a crowd saying it's not a "real AI" and that scientists are solving "toy problems".

This used to be the case a decade or more ago. In the last few years the opposite is true. The press is certainly very eager to report every big success of "AI"- by which of course is meant deep learning.

>> 6 years ago we were able to solve some Atari games from pixels. Today, that feels like a trivial exercise compared to modern techniques

6 years ago DeepMind showed superhuman performance in seven Atari games with Deep-RL (DeepQN in particular): Beam Rider, Breakout, Enduro, Pong, Q*bert, Seaquest and Space Invaders. Since then more Atari games have been "beaten" in the same sense, but many still remain. I'm afraid I can't find references to this but I've seen slides from DeepMind people a few times and there is always a curve with a few games at the top and most games at the bottom, below human performance. There are some games that are notorious for being very difficult to solve with Deep-RL, like Montezuma's Revenge which was claimed to be solved by Uber a couple of years ago however this was done using imitation learning, which means watching a human play. The result is nothing like the result in Go, which remains the crowning achievement of Deep-RL (and its best buddy, MCTS).

Bottom line: Atari games remain anything but a trivial exercise.

And the architectuers that play Atari do not perform as well in Go or chess, say. You are mistaken that it's simple to train the same system to do all of those things. The AlphaZero system that played Go, chess and Shoggi well enough to beat its predecessor (you will excuse me that I don't remember which incarnation of Alpha-x it was) had an architeture fine-tuned to a chessboard and pieces with discrete moves, so it would not be possible to reuse it to play Starcraft, say, or even tic-tac-toe. The cost to train AlphaZero is also very high, in the hundreds of thousands of dollars.