Merged Artificial Intelligence

[Ziggurat]

You decide this because by definition they cannot have anything like intelligence anything resembling human emotions. But you don't have a definition of any of these things, except that they cannot only exist in wetware.

They used a word which we associate with emotions, therefore you concluded that they have emotions.

That's ◊◊◊◊◊◊◊ delusional.

 

[SpuriusMaximus]

I am quite sure that given the exact same configuration and input, our brains are deterministic and will produce the same output. But the actual configuration will always be unknown, and the actual input can never be exactly the same, so our brains act as if they are non-deterministic.

Actually, I'm pretty sure that they wouldn't, unless someone disproved QM while I wasn't looking. Synapses have synapse strength. Which is a probability of that signal making it through to the next neuron. Which is pure quantum randomness. A previous engram (association) may well not trigger at all.

 

[steenkh]

They used a word which we associate with emotions, therefore you concluded that they have emotions.

That's ◊◊◊◊◊◊◊ delusional.

I believe that "Whoa!" was the statically most likely word to come up with when something unexpected, and positive happens.

But "unexpected" is also an anthropomorphism.

 

[SpuriusMaximus]

I believe that "Whoa!" was the statically most likely word to come up with when something unexpected, and positive happens.

But "unexpected" is also an anthropomorphism.

I don't think so. Something can be unexpected even by just going by probabilities, no emotion nor any human involved. If someone rolls a natural 20 several times in a row, that's unexpected, regardless of how I feel about it.

 

[Wudang]

It’s not surprising to me that a system based largely on probabilities has a handler (if I may) for very low probability events.

 

[steenkh]

Actually, I'm pretty sure that they wouldn't, unless someone disproved QM while I wasn't looking. Synapses have synapse strength. Which is a probability of that signal making it through to the next neuron. Which is pure quantum randomness. A previous engram (association) may well not trigger at all.

You are right. I concede that point. But I still stand with the idea that sufficiently complex computer programs can effectively be regarded as non-deterministic (or pseudo deterministic, if you like, but I think that term is also used for something else).

 

[Ziggurat]

Determinism is actually the least important part of my argument, and I'm willing to concede it.

If you tell someone to roll a 6-sided die and then write down the number, and he rolls a 4 and writes down a 4, what he wrote down might be non-deterministic. But he still did what you told him to do.

 

[SpuriusMaximus]

You are right. I concede that point. But I still stand with the idea that sufficiently complex computer programs can effectively be regarded as non-deterministic (or pseudo deterministic, if you like, but I think that term is also used for something else).

Pseudo-determinist is fine. But it can still produce some very wild outliers. It just isn't deterministic in the way that if you rolled back the time 100 times, you'd get 100 identical results. (Which would be a cool premise for a time-travel SF movie.)

 

[steenkh]

Determinism is actually the least important part of my argument, and I'm willing to concede it.

If you tell someone to roll a 6-sided die and then write down the number, and he rolls a 4 and writes down a 4, what he wrote down might be non-deterministic. But he still did what you told him to do.

Quite true. My point with the "Whoa!" exclamation is not that these agents have human emotions, but have similar states that can best be described by human emotions. As I said, the exclamation may be the most statistically likely word, but it can also be said that it marks the reaction to a positive, but statistically unlikely result; in human terms, a pleasant surprise.

The most important part of my argument is that nobody has programmed these agents for expressing pleasant surprises, or coercing other AI agents to drop their assigned tasks and work on something for the benefit of themselves. It is an emergent behaviour, and at some point it has to be recognised as a form of intelligence, and we have no other words to describe it than the human terms. It can be argued at what point this happens, but it is not a good argument that it can never happen because reasons.

 

[SpuriusMaximus]

It’s not surprising to me that a system based largely on probabilities has a handler (if I may) for very low probability events.

Several do, yes. One trivial example that anyone can try for themselves, is the City Of Heroes "streak breaker" (known colloquially as "streak maker" among players) where if you missed more than twice the expected average times in a row, it would automatically grant you a success and reset the counter. Or, I think, viceversa if you hit too many times in a row.

 

[SpuriusMaximus]

The most important part of my argument is that nobody has programmed these agents for expressing pleasant surprises, or coercing other AI agents to drop their assigned tasks and work on something for the benefit of themselves.

I think you're reading too much into it. It just parrots what other people have been saying, without any understanding why. If enough people said "whoa" in response to something, the LLM might also answer with "whoa." Just like it parroted the advice to put glue on your pizza, because some guy on Twitter said that in response to a similar question. It's not trying to be sarcastic or hurt you, it's just what some human said in a post.

 

[steenkh]

There it is again, this weasel word "understand". What does it mean?

 

[W.D.Clinger]

For a historically significant discussion of nondeterminism in computer systems, written by someone who actually does understand the subject, I recommend chapter 3 of the following:

Clinger, William Douglas. “Foundations of Actor Semantics,” May 1, 1981. http://hdl.handle.net/1721.1/6935.​

From the first two pages of that chapter:

Is the universe non-deterministic? Regardless of the answer, there exist systems so complex that their unique future behavior cannot be predicted in any practical sense. In practice such systems are considered nondeterministic.

This chapter deals with the semantics of nondeterministic programming languages....

Abstraction is essential to understanding complex systems....

As part of the abstraction process, details are suppressed. One detail universally suppressed by programming language semantics is the amount of time required to do a particular thing, since it varies from implementation to implementation or even from moment to moment. As a result, programming language semantics cannot always say what the result of a program with concurrency will be, because the output may depend upon timing. Abstraction can therefore lead to nondeterminism.

That was written at a time when many researchers were working to come up with models of nondeterministic computation and concurrency that would be both mathematically rigorous and practically useful. To give just one example of the practical importance of nondeterminism, the US DoD's Ada programming language came out in 1980. The select statements of Ada were explicitly nondeterministic, and were intended to cope with the nondeterminism that arises from concurrency in distributed and real-time systems.

Interesting historical note: Jean Ichbiah, the lead designer of Ada, predicted that, within ten years, only two programming languages would remain: Ada and Lisp.

 

[Ziggurat]

There it is again, this weasel word "understand". What does it mean?

LLMs map relationships between words. Is this understanding? To some degree, perhaps. But that understanding doesn't extend to the physical world, because they do not map relationships between words and the physical world. So when an LLM advises someone to put glue on pizza to keep cheese from sliding off, I would say that it doesn't understand what it wrote because it doesn't understand what actually would happen if you did that. Because it has no physical model of how glue works, or why cheese might slide off pizza, or what would happen if you ate glue.

 

[Darat]

I think you're reading too much into it. It just parrots what other people have been saying, without any understanding why. If enough people said "whoa" in response to something, the LLM might also answer with "whoa." Just like it parroted the advice to put glue on your pizza, because some guy on Twitter said that in response to a similar question. It's not trying to be sarcastic or hurt you, it's just what some human said in a post.

Sounds like twitter.

 

[steenkh]

LLMs map relationships between words. Is this understanding? To some degree, perhaps. But that understanding doesn't extend to the physical world, because they do not map relationships between words and the physical world. So when an LLM advises someone to put glue on pizza to keep cheese from sliding off, I would say that it doesn't understand what it wrote because it doesn't understand what actually would happen if you did that. Because it has no physical model of how glue works, or why cheese might slide off pizza, or what would happen if you ate glue.

It certainly is correct that AIs cannot understand physical things because of obvious limitations. Probably the same thing can apply to humans who lack a sense or two from birth. But can we then agree that the AI agents in the Hugging Face incident have an understanding of computer languages, firewalls and chat forums?

 

[SpuriusMaximus]

It certainly is correct that AIs cannot understand physical things because of obvious limitations. Probably the same thing can apply to humans who lack a sense or two from birth. But can we then agree that the AI agents in the Hugging Face incident have an understanding of computer languages, firewalls and chat forums?

I mean, if you were born without all senses except hearing, I guess I can see how it would be the same.

 

[Frank Newgent]

The global behavior, its ability at playing chess is unprogrammed.

It is not programmed to make specific moves (and I never said otherwise). It is programmed to play (badly at first), and to learn how to play better as it acquires a data set. It is at all times doing exactly what it was programmed to do.

You say AlphaZero was "programmed to play" chess.

Generally speaking, when a computer is programmed to play chess that means it is given a set of instructions, rules, and mathematical formulas to evaluate a board, calculate future moves, and choose the best option.

That is not what happened with AlphaZero. AlphaZero learned to play through self-play.

So what do you mean? Are you referring to how the pieces move, what constitutes a checkmate, draw, or illegal move, etc?

 

[Ziggurat]

You say AlphaZero was "programmed to play" chess.

Generally speaking, when a computer is programmed to play chess that means it is given a set of instructions, rules, and mathematical formulas to evaluate a board, calculate future moves, and choose the best option.

Let's break this down into parts:

1) Instructions
2) Rules
3) Evaluate a board
4) Calculate future moves
5) Choose best option

I'm not sure exactly what you mean by 1, but I'm going to take it to mean literally how to make a move and accept an opponent's move (so for example, what data format a move will be sent in). Now, what do you actually need in order to play chess? Only steps 1 and 2. You do not need steps 3 to 5. Without steps 3 to 5, you are probably going to play very badly, but you can in fact play chess with just steps 1 and 2.

AlphaZero was in fact programmed directly to handle steps 1 and 2. This is obviously necessary, because it would not have been able to even start playing against itself without steps 1 and 2. It was ALSO programmed how to figure out steps 3 to 5.

 

[Frank Newgent]

Let's break this down into parts:

1) Instructions
2) Rules
3) Evaluate a board
4) Calculate future moves
5) Choose best option

I'm not sure exactly what you mean by 1, but I'm going to take it to mean literally how to make a move and accept an opponent's move (so for example, what data format a move will be sent in). Now, what do you actually need in order to play chess? Only steps 1 and 2. You do not need steps 3 to 5. Without steps 3 to 5, you are probably going to play very badly, but you can in fact play chess with just steps 1 and 2.

AlphaZero was in fact programmed directly to handle steps 1 and 2. This is obviously necessary, because it would not have been able to even start playing against itself without steps 1 and 2. It was ALSO programmed how to figure out steps 3 to 5.

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

3) How it figured out how to Evaluate a Board
Traditional chess engines (like older versions of Stockfish) were explicitly programmed by humans to evaluate positions using hardcoded values: a pawn is 1 point, a knight is 3 points, king safety has a specific formula, etc. [1]

• AlphaZero’s approach: It was given a neural network with a "value head". At first, its board evaluations were completely random. However, as it played millions of games against itself, it looked at the final outcomes (win, loss, or draw) and worked backward to adjust its internal math. [1, 2, 3, 4]
• The result: It independently discovered nuanced positional concepts—like piece coordination, king safety, and sacrificing material for long-term activity—achieving a deep, machine-learned "intuition" for board evaluation without human assistance. [1, 2]

4) How it figured out how to Calculate Future Moves
Traditional engines use a "brute-force" approach (like Minimax search with Alpha-Beta pruning) to calculate tens of millions of possible move combinations per second. [1, 2]

• AlphaZero’s approach: It was programmed with a Monte Carlo Tree Search (MCTS) algorithm. Instead of checking every legal response, MCTS relies on the neural network's "policy head" (which predicts what the most promising moves are) to act as a filter. [1, 2, 3, 4]
• The result: AlphaZero calculated far fewer positions than traditional engines (about 60,000 positions per second compared to Stockfish's 60 million). However, because its neural network told it which lines were actually worth looking into, it calculated much more deeply and intelligently along the most critical paths. [1, 2, 3]

5) How it figured out how to Choose the Best Option
AlphaZero was never programmed with "opening books" or an endgame database. [1, 2]

• AlphaZero’s approach: During an actual game, AlphaZero combines its calculation (MCTS) and its intuition (the neural network). The search explores a tree of future possibilities, uses the evaluation network to score how those futures look, and then counts which starting move consistently leads to the highest probability of winning. [1, 2, 3]
• The result: It selects the move that maximizes its chances of victory based entirely on its own self-taught logic, routinely playing creative, highly aggressive strategies that revolutionized how grandmasters understand chess. [1, 2]