Merged Artificial Intelligence

[Ziggurat]

I believe your assertion about steps 3 - 5 is incorrect.

Your quote (source?) seems to back me up.

For 3: 'It was given a neural network with a "value head".'

For 4: 'It was programmed with a Monte Carlo Tree Search (MCTS) algorithm.'

For 5: 'During an actual game, AlphaZero combines its calculation (MCTS) and its intuition (the neural network).'

Sounds to me like this is exactly what I described: it was programmed how to figure out these steps, by giving it a neural network, an MCTS algorithm, and the instruction to combine them. These parts were programmed in.

 

[SpuriusMaximus]

TBH, I don't find chess to be a great example anyway. The top chess players just recognise board configurations, and the best answer to them. While the total number of possible combinations on the board is astronomical (as in 40 orders of magnitude higher than the number of atoms in the observable universe,) a data centre can still store more of them than a human can. And a data centre can simulate billions of games a day, which is more than a human can even read about in a lifetime, to learn from them. It's a nice stunt for PR, but basically it's still brute forcing it by throwing more SSD space at it than a human brain can.

Now try doing the same with Go, which has 50 orders of magnitudes more possible legal combinations.

 

[theprestige]

A recurring theme in the fantasy adventure anime, Frieren: Beyond Journey's End, is that demons are not human. They are sometimes very good at emulating human behaviors, and sometimes very good at aping human motivations. But they are not human, will never be human, and will never be anything other than a real threat to human existence.

 

[The Atheist]

Rather than read pages of drivel about chess to see if it's been mentioned already, I'll post this, and if it's already been posted, hard luck: https://arstechnica.com/health/2026...al-care-for-seniors-in-disastrous-experiment/

 

[Darat]

TBH, I don't find chess to be a great example anyway. The top chess players just recognise board configurations, and the best answer to them. While the total number of possible combinations on the board is astronomical (as in 40 orders of magnitude higher than the number of atoms in the observable universe,) a data centre can still store more of them than a human can. And a data centre can simulate billions of games a day, which is more than a human can even read about in a lifetime, to learn from them. It's a nice stunt for PR, but basically it's still brute forcing it by throwing more SSD space at it than a human brain can.

Now try doing the same with Go, which has 50 orders of magnitudes more possible legal combinations.

Go was mastered years ago, before ChatGPT and the other upstarts were a twinkle in the eyes.

ETA: had to check the dates and this is correct from an AI Google search: October 2017: DeepMind unveiled AlphaGo Zero, which learned entirely from scratch via reinforcement learning without human data and mastered the game in just three days.

They had another iteration that was beating the Go world champions back in 2015/16

 

[Frank Newgent]

AlphaZero does exactly and ONLY what it is programmed to do.

lol

The architecture is programmed... the underlying framework, the neural network, the Monte Carlo Tree Search algorithm.

The knowledge is learned... what it does in terms of the strategies, the board evaluations, the move preferences. Can I be more clear, the things that AlphaZero does are not programmed, the chess strategies it employs, the positional evaluations it makes, they are learned from the ground up. AlphaZero starts by playing completely randomly against itself, using trial and error to update its neural network's weights based on wins, losses, and draws.

I'm sorry you just don't get it, Ziggurant.

 

[Ziggurat]

The knowledge is learned... what it does in terms of the strategies, the board evaluations, the move preferences.

I know that. And yet, what is it doing? What is it trying to do? Win games of chess. That is what it was programmed to do.

And the AIs in the hugging face hack were programmed to hack. They did what they were programmed to do. They did not go rogue.

I'm sorry you just don't get it, Ziggurant.

I get it just fine. I don't think you do.

 

[Dr.Sid]

A recurring theme in the fantasy adventure anime, Frieren: Beyond Journey's End, is that demons are not human. They are sometimes very good at emulating human behaviors, and sometimes very good at aping human motivations. But they are not human, will never be human, and will never be anything other than a real threat to human existence.

Dehumanization of the enemy is common trope. With Frieren I still wonder if they are going somewhere with it or not.

 

[theprestige]

Dehumanization of the enemy is common trope. With Frieren I still wonder if they are going somewhere with it or not.

You caught me.

I'm dehumanizing computer programs.

 

[Frank Newgent]

I know that. And yet, what is it doing? What is it trying to do? Win games of chess. That is what it was programmed to do.

And the AIs in the hugging face hack were programmed to hack. They did what they were programmed to do. They did not go rogue.

I get it just fine. I don't think you do.

Can you define how AI programming and AI training differ?

 

[SpuriusMaximus]

Go was mastered years ago, before ChatGPT and the other upstarts were a twinkle in the eyes.

OK, my bad. If I weren't such a bad player, it should have occurred to me that it's based on local configurations, which have a MUCH lower number of possible configurations than looking at the whole board.

 

[Dr.Sid]

You caught me.

I'm dehumanizing computer programs.

Nah. I just think anthropomorphic terms are fine as long as they are useful. AIs think, understand, want .. if you had to come up with terms which would basically mean "understand, but not like humans do" .. it would just be awkward. Also I ◊◊◊◊◊◊◊ love the show.

 

[theprestige]

Nah. I just think anthropomorphic terms are fine as long as they are useful.

Anthropomorphizing computer programs is anti-useful.

 

[Dr.Sid]

Anthropomorphizing computer programs is anti-useful.

Not really.

 

[theprestige]

Not really.

Very much really.

Treating non-human things as if they were human is anti-reality and anti-human.

 

[SpuriusMaximus]

Anthropomorphizing computer programs is anti-useful.

Well, I find it useful when I need to cuss at a program for freezing or otherwise not doing what it's told

 

[Darat]

OK, my bad. If I weren't such a bad player, it should have occurred to me that it's based on local configurations, which have a MUCH lower number of possible configurations than looking at the whole board.

It learned how to play, not billions of precomputed moves.

 

[SpuriusMaximus]

It learned how to play, not billions of precomputed moves.

Theoretically so is chess. Top players still recognise board configurations, rather than starting from the basic rules and extrapolating how it can go 10 steps ahead.

And frankly, as a (very bad) Go player, the same applies there. You just recognise by the local configuration that a group of pieces are dead or alive, not extrapolate what the other player might do on a 19x19 board, and what you might respond with. You just look at that group and think "yep, they're dead" or "they're not trapped yet" or whatever.

ETA: plus, it seems weird to hear you say that an LLM doesn't use pre-computed stuff. Using probabilities based on the existing sequence is what an LLM does. As I'm sure you know already.

 

[Darat]

Theoretically so is chess. Top players still recognise board configurations, rather than starting from the basic rules and extrapolating how it can go 10 steps ahead.

And frankly, as a (very bad) Go player, the same applies there. You just recognise by the local configuration that a group of pieces are dead or alive, not extrapolate what the other player might do on a 19x19 board, and what you might respond with. You just look at that group and think "yep, they're dead" or "they're not trapped yet" or whatever.

ETA: plus, it seems weird to hear you say that an LLM doesn't use pre-computed stuff. Using probabilities based on the existing sequence is what an LLM does. As I'm sure you know already.

Alphago isn't a LLM. It predates LLMs and even predates the seminal paper "Attention Is All You Need" that kicked of the LLM bubble.

 

[SpuriusMaximus]

Alphago isn't a LLM. It predates LLMs and even predates the seminal paper "Attention Is All You Need" that kicked of the LLM bubble.

It's still based on probabilities, according to its own creators. The idea of using a sequence or configuration to choose the next thing predates not only LLMs, but even computers. See, Markov Chain, first published in 1906. (Which I already posted an implementation of, and its output, right on this board. See, Mark V Shaney. Mine is actually slightly less random than the original Mark V Shaney, and more towards vaguely LLM-ish. It's the primitive great-grandfather of LLMs, including being trained on posts by other people in the case of the original one.)