Au contraire, AlphaGo made several “counterintuitive” moves that professional Go players thought were mistakes during the play, but turned out to be great strategic moves in hindsight.
The (in)ability to recognize a strange move’s brilliance might depend on the complexity of the game. The real world is much more complex than any board game.
That's great, but AlphaGo used artificial and constrained training materials. It's a lot easier to optimize things when you can actually define an objective score, and especially when your system is able to generate valid training materials on its own.
Are you simply referring to games having a defined win/loss reward function?
Because pretty sure Alpha Go was ground breaking also because it was self taught, by playing itself, there were no training materials. Unless you say the rules of the game itself is the constraint.
But even then, from move to move, there are huge decisions to be made that are NOT easily defined with a win/loss reward function. Especially early game, there are many moves to make that don't obviously have an objective score to optimize against.
You could make the big leap and say that GO is so open ended, that it does model Life.
"artificial" maybe I should have said "synthetic"? I mean the computer can teach itself.
"constrained" the game has rules that can be evaluated
and as to the other -- I don't know what to tell you, I don't think anything I said is inconsistent with the below quotes.
It's clearly not just a generic LLM, and it's only possible to generate a billion training examples for it to play against itself because synthetic data is valid. And synthetic data contains training examples no human has ever done, which is why it's not at all surprising it did stuff humans never would try. A LLM would just try patterns that, at best, are published in human-generated go game histories or synthesized from them. I think this inherently limits the amount of exploration it can do of the game space, and similarly would be much less likely to generate novel moves.
> As of 2016, AlphaGo's algorithm uses a combination of machine learning and tree search techniques, combined with extensive training, both from human and computer play. It uses Monte Carlo tree search, guided by a "value network" and a "policy network", both implemented using deep neural network technology.[5][4] A limited amount of game-specific feature detection pre-processing (for example, to highlight whether a move matches a nakade pattern) is applied to the input before it is sent to the neural networks.[4] The networks are convolutional neural networks with 12 layers, trained by reinforcement learning.[4]
> The system's neural networks were initially bootstrapped from human gameplay expertise. AlphaGo was initially trained to mimic human play by attempting to match the moves of expert players from recorded historical games, using a database of around 30 million moves.[21] Once it had reached a certain degree of proficiency, it was trained further by being set to play large numbers of games against other instances of itself, using reinforcement learning to improve its play.[5] To avoid "disrespectfully" wasting its opponent's time, the program is specifically programmed to resign if its assessment of win probability falls beneath a certain threshold; for the match against Lee, the resignation threshold was set to 20%.[64]
Of course, not an LLM. I was just referring to AI technology in general. And that goal functions can be complicated and not-obvious even for a game world with known rules and outcomes.
I was miss-remembering the order of how things happened.
AlphaZero, another iteration after the famous matches, was trained without human data.
"AlphaGo's team published an article in the journal Nature on 19 October 2017, introducing AlphaGo Zero, a version without human data and stronger than any previous human-champion-defeating version.[52] By playing games against itself, AlphaGo Zero surpassed the strength of AlphaGo Lee in three days by winning 100 games to 0, reached the level of AlphaGo Master in 21 days, and exceeded all the old versions in 40 days.[53]"
There are quite a few relatively objective criteria in the real world: real estate holdings, money and material possessions, power to influence people and events, etc.
The complexity of achieving those might result in the "Centaur Era", when humans+computers are superior to either alone, lasting longer than the Centaur chess era, which spanned only 1-2 decades before engines like Stockfish made humans superfluous.
However, in well-defined domains, like medical diagnostics, it seems reasoning models alone are already superior to primary care physicians, according to at least 6 studies.
It makes sense. People said software engineers would be easy to replace with AI, because our work can be run on a computer and easily tested, but the disconnect is that the primary strength of LLMs is that they can draw on huge bodies of information, and that's not the primary skill programmers are paid for. It does help programmers when you're doing trivial CRUD work or writing boilerplate, but every programmer will eventually have to be able to actually truly reason about code, and LLMs fundamentally cannot do that (not even the "reasoning" models).
Medical diagnosis relies heavily on knowledge, pattern recognition, a bunch of heuristics, educated guesses, luck, etc. These are all things LLMs do very well. They don't need a high degree of accuracy, because humans are already doing this work with a pretty low degree of accuracy. They just have to be a little more accurate.
Being a walking encyclopedia is not what we pay doctors for either. We pay them to account for the half truths and actual lies that people tell about their health. This is to say nothing about novel presentations that come about because of the genetic lottery. Same as an AI can assist but not replace a software engineer, an AI can assist but not replace a doctor.
Having worked briefly in the medical fields in the 1990s, there is some sort of "greedy matching" being pursued, so once 1-2 well-known symptoms are recognized that can be associated with diseases, the standard interventions to cure are initiated.
A more "proper" approach would be to work with sets of hypotheses and to conduct tests to exclude alternative explanations gradually - which medics call "DD" (differential diagnosis).
Sadly, this is often not systematically done, and instead people jump on the first diagnosis and try if the intervention "fixes" things.
So I agree there are huge gains from "low hanging fruits" to be expected in the medical domain.
I think at this point it's an absurd take that they aren't reasoning. I don't think without reasoning about code (& math) you can get to such high scores on competitive coding and IMO scores.
Alphazero also doesn't need training data as input--it's generated by game-play. The information fed in is just game rules. Theoretically should also be possible in research math. Less so in programming b/c we care about less rigid things like style. But if you rigorously defined the objective, training data should also be not necessary.
> Alphazero also doesn't need training data as input--it's generated by game-play. The information fed in is just game rules
This is wrong, it wasn't just fed the rules, it was also fed a harness that did test viable moves and searched for optimal ones using a depth first search method.
Without that harness it would not have gained superhuman performance, such a harness is easy to make for Go but not as easy to make for more complex things. You will find the harder it is to make an effective such harness for a topic the harder it is to solve for AI models, it is relatively easy to make a good such harness for very well defined programming problems like competitive programming but much much harder for general purpose programming.
Are you talking about Monte Carlo tree search? I consider it part of the algorithm in AlphaZero's case. But agreed that RL is a lot harder in real-life setting than in a board game setting.
If that's your take-away from that paper, it seems you've arrived at the wrong conclusion. It's not that it's "fake", it's that it doesn't give the full picture, and if you only rely on CoT to catch "undesirable" behavior, you'll miss a lot. There is a lot more nuance than you allude to, from the paper itself:
> These results suggest that CoT monitoring is a promising way of noticing undesired behaviors during training and evaluations, but that it is not sufficient to rule them out.
very few humans are as good as these models at arithmetic. and CoT is not "mostly fake" that's not a correct interpretation of that research. It can be deceptive but so can human justifications of actions.
Humans can learn the symbolic rules and then apply them correctly to any problem, bounded only by time, and modulo lapses of concentration. LLMs fundamentally do not work this way, which is a major shortcoming.
They can convincingly mimic human thought but the illusion falls flat at further inspection.
Humans are statistically speaking static. We just find out more about them but the humans themselves don't meaningfully change unless you start looking at much longer time scales. The state of the rest of the world is in constant flux and much harder to model.
To be fair, it was more a "wow look what the computer did". The AI "art" was always bad. At first it was just bad because it was visually incongruous. Then they improved the finger counting kernel, and now it's bad because it's a shallow cultural average.
AI producing visual art has only flooded the internet with "slop", the commonly accepted term. It's something that meets the bare criteria, but falls short in producing anything actually enjoyable or worth anyone's time.
It sucks for art almost by definition, because art exists for its own reason and is in some way novel.
However, even artists need supporting materials and tooling that meet bare criteria. Some care what kind of wood their brush is made from, but I'd guess most do not.
I suspect it'll prove useless at the heart of almost every art form, but powerful at the periphery.
Sure, that does make things easier: one of the reasons Go took so long to solve is that one cannot define an objective score for Go beyond the end result being a boolean win or loose.
But IRL? Lots of measures exist, from money to votes to exam scores, and a big part of the problem is Goodhart's law — that the easy-to-define measures aren't sufficiently good at capturing what we care about, so we must not optimise too hard for those scores.
> Sure, that does make things easier: one of the reasons Go took so long to solve is that one cannot define an objective score for Go beyond the end result being a boolean win or loose.
Winning or losing a Go game is a much shorter term objective than making or losing money at a job.
> But IRL? Lots of measures exist
No, not that are shorter term than winning or losing a Go game. A game of Go is very short, much much shorter than the time it takes for a human to get fired for incompetence.
No? some of the opening moves took experts thorough analysis to figure out were not mistakes. even in game 1 for example. not just the move 37 thing. Also thematic ideas like 3x3 invasions.
There may be philosophical (i.e. fundamental) challenges to AGI. Consider, e.g., Godel's Incompleteness Theorem. Though Scott Aaronson argues this does not matter (see e.g., youtube video, "How Much Math Is Knowable?"). There would also seem to be limits to the computation of potentially chaotic systems. And in general, verifying physical theories has required the carrying out of actual physical experiment. Even if we were to build a fully reasoning model, "pondering" is not always sufficient.
It’s also easy to forget that “reason is the slave of the passions” (Hume) - a lot of what we regard as intelligence is explicitly tied to other, baser (or more elevated) parts of the human experience.
The (in)ability to recognize a strange move’s brilliance might depend on the complexity of the game. The real world is much more complex than any board game.
https://en.m.wikipedia.org/wiki/AlphaGo_versus_Lee_Sedol