Hacker News Reader

Comment by emporas | Hacker News Reader

>LLMs are really smart, and also really, really dumb

Machines will always be dumb. The notion that machines are smart, probably stems from people inexperienced using engines. In chess we use engines for more than 20 years and humans always lose against the engines. The games they play however are garbage. Even if they are better than the human games, they are still garbage, as Magnus Carlsen have commented in the past about the CCC. Compilers in programming are engines too.

I am trying to write an article about Terminator using GPT, and every time i have to generate 10 different drafts, edit them afterwards and manually string together the different parts. It is great help like the chess engines, but they need a lot of human supervision.

Engines can output text, but they cannot output articles. They can output code, but they cannot output programs. They can output musical instruments, but they cannot output songs.

View on HN

You told it to win and you're upset about the nature of its victory. I think what you're seeing isn't that computers are bad at chess but that chess isn't what you thought it was.

Maybe one day the machines will kill everyone and the last human's words will be "Well, that wasn't how I'd have done it."

Hmm, the dumbest human can never lose any game of chess, against the smartest machine. The smartest human can never win against the smartest machine. So if the humans can never win against machines, and machines can never win against humans, that means we are of equal strength, no?

We humans think from first principles. When a disadvantage is found in the enemy's computational process, one person tells all about it to the person beside him. The second person, doesn't need retraining his brain from scratch to incorporate the new knowledge. He uses that nugget of information, just that nugget and nothing else to gain an advantage in chess, in the stock market or anything else.

One obvious example are script kiddies in computers who just download an exploit and gain an advantage, despite knowing nothing about computer security. Or a janitor who listens by chance some important investment information and can 10x some money by pure luck.

We think from first principles because we are born on the planet with a purpose. Machines are not born, they do not care to ever being given birth, and they do not care if they die because they were never alive. They have no purpose, and no first principles to think anything about.

How would you define undumbness?

I don't have a definition, and i am the last person qualified to give a correct answer to that, but if we rephrase the question as: "Are ML algorithms more than tools", i would answer as no. ML algorithms did go a lot further than i was expecting, especially in language generation, but they are just tools. Great tools, but still tools.

Why are the chess games garbage?

Magnus said something like that about CCC, the Computer Chess Championship. Engines, especially alpha-beta algorithms play in a very mechanistic way, totally unnatural. Granted ML algorithms are more human-like than alpha-beta, but even then their playing style is strange nevertheless.

They are great tools to use, but smart? I don't think so.

Edit-- I cannot find the source, but Magnus said that he doesn't personally look at computer games, (because they are garbage), not that the games by themselves are useless. They put one engine to play against another to test openings and positions. That's useful. It is that Magnus and most of other people, have no interest in watching games by engines.

Maybe the issue is the alpha-beta algorithms just string good moves together, there is no overall plan or intent. Perhaps the lack of narrative when we read through the move, feeds that sense of mechanistic play.

I remember reading about a chess program that built plans and then considered moves in order to complete the plan. Not sure how this could be combined with the ML engines that we have but it might lead to a more natural playing style.

Yes, i could definitely see that.

A multimodal chess engine, which evaluates several positions using a traditional chess engine, but it will play the move which has the more clear description using an LLM. It will not play the best move, but the one which is easier to put into words. Then it has the potential be much more interesting to human eyes.

Reply on news.ycombinator.com