AI Theory Part 7: [Paradigm Shift] From Deep Blue to ChatGPT. The Dawn of Stopping 'Teaching Rules'

I explained the fate of AI being 'stateless,' where it lacks memory and forgets premises as conversations progress, as well as operational patterns to outsmart this fate. By summarizing conversations each time and cultivating instruction sets like a constitution, we compensate for the limitations of the AI tool through human ingenuity.

Here, a question arises. Why has modern AI become so capable of handling human 'ambiguous language and context'?

Back in 1997, the news that the AI 'Deep Blue,' developed by IBM, had defeated the world chess champion shocked the world. However, even if you spoke to the AI of that era asking, 'Tell me tomorrow's weather,' it could not even sustain a simple conversation.

Between the genius AI that defeated the chess champion and the current ChatGPT, there exists a decisive 'paradigm shift' that fundamentally overturns the history of AI.

The limitations of the 'humans teach rules' method

AI for games and specialized fields of the past, including Deep Blue, were created using an approach where humans wrote down rules and evaluation criteria. 'Expert systems,' which write down expert knowledge as rules, are a prime example.

The mechanism is deterministic to a fault. Humans taught the rules of chess, standard moves, and evaluation criteria for board positions as code in the form of 'if the board is like this, move like this (if-then).' In the case of Deep Blue, it used human-designed evaluation criteria and overwhelming computing power from dedicated hardware to look ahead and explore promising moves.

If it were a closed world like chess, where 'the board is 8x8, piece movements are fixed, and there are no elements of chance,' this method was extremely effective.

However, the moment humanity tried to apply this 'humans write and teach rules' approach to real-world problems or daily human conversation, they hit a thick wall.

For example, imagine trying to teach an AI the concept of a 'dog' using only program rules.

'Walks on four legs,' 'has fur,' 'barks,' 'wears a collar'... You write down every rule a human can think of in code. But what about a dog with three legs? Is a dog that doesn't bark not a dog? How do you distinguish it from a cat wearing a collar?

The real world is full of exceptions and ambiguity. The more conditional branches (if statements) humans added, the more the rules contradicted each other, and the system reached the limits of complexity and collapsed.

The infinite loop called the 'Frame Problem'

The most famous thought experiment in the history of AI research, the 'Frame Problem,' symbolizes the limitations of this 'rule-teaching method.'

Suppose there is a wagon in a room, and an 'important battery' and a 'time bomb' are on it together. An AI robot was ordered to 'carry the battery out of the room.'

The first-generation AI robot pushed the wagon out of the room. However, it also carried the bomb on the wagon, and it exploded and was destroyed. This was because it did not consider the side effect (rule) that 'if you move the battery, the bomb on the same wagon also moves.'

So, researchers improved it and created a second-generation robot with the added rule: 'Before taking action, consider all side effects related to that action.'

However, the second-generation robot remained frozen in front of the wagon and never moved again. It fell into an infinite loop of checking countless 'irrelevant side effects' in its head, such as 'If I pull the wagon, will the ceiling fall?' 'If I pull the wagon, will the wall color change?' or 'If I pull the wagon, will the air in the room shake?'

Humans can instantly judge and ignore 'what is relevant to me and what is irrelevant (outside the frame)' unconsciously. However, for an AI trying to describe everything with rules, there are infinite rules to check 'what is irrelevant,' causing the computational volume to explode and leading to a state of paralysis.

The approach of humans explicitly writing down and teaching rules reached a major dead end here.

A Reversal of Thought. The Dawn of Stopping 'Teaching Rules'

What broke through this deadlock was the reversal of thought known as 'representation learning,' brought about by recent deep learning and Transformers.

Researchers changed their strategy. They made the decision: 'Let's stop trying to write code to teach the rules of a complex world that even humans cannot explain.'

Instead, they adopted the mathematical approach explained in parts 1 through 3.

Rather than teaching rules, they had the AI learn directly from the vast amount of raw data—trillions of tokens—that exists on the internet. They shifted to a strategy of letting the AI discover 'the distance and patterns between words (coordinates in multidimensional space)' using its own computational power.

Without humans teaching it a single character about 'what a dog is' or 'what grammar is,' the AI discovered for itself the structure of a multidimensional space where 'words like fur, walk, and bark are placed with high probability near the word dog' from vast amounts of text data.

The moment humans abandoned rigid rules (if-statements) and allowed the AI to autonomously learn probabilistic patterns (coordinates) from data, AI reached a major turning point. It became possible to avoid the infinite verification processes that were unavoidable in the rule-writing method, and it acquired the overwhelming ability to flexibly handle human ambiguity and context. This is the true nature of the dawn of generative AI that we are witnessing.

Note that whether the frame problem itself has been theoretically solved remains a subject of debate among researchers. What can be said here is only that we have reached a point where we can practically circumvent the areas where traditional methods hit a wall by using a different approach.

Implications for Game Development. Breaking Away from the Era of Writing Everything in Code

The paradigm shift in AI history of 'stopping the teaching of rules' provides extremely important implications for me as I take on game development.

Traditional game development was exactly the method of 'humans writing down all the rules.'

Enemy character behavior patterns, branching conversations with NPCs, item drop rates, and behavior when a player hits a wall. Programmers have built worlds by writing thousands and tens of thousands of lines of 'if-statements' to handle all these exceptions.

However, in future game development, now that we have obtained the powerful tool of AI, this structure itself will change.

The framework and physical laws of a game (determinism), where not even a single character of error is allowed, should be strictly written by humans in code. On the other hand, natural dialogue with NPCs, the automatic generation of complex dungeons, and flexible developments based on player actions (probabilistic theory) should be entrusted to the expressive power of AI that has learned from data and patterns.

Being freed from the curse of 'humans writing down every rule.' This is the greatest reason why we, as complete beginners, can use AI as our right-hand man and jump out into the wilderness of massive game development.

The Night We Stopped 'Teaching Rules'

The 'perfect AI taught rules by humans' that artificial intelligence researchers once dreamed of could, in the end, never leave the board of a closed world like chess or shogi.

When humans stopped teaching rules and let AI discover the 'statistical distance between words' from vast amounts of data on its own, the AI became able to handle the context of words statistically and transformed into a 'probabilistic partner' that can naturally converse with humans.

However, because it has no rules and weaves words on waves of probability, AI has a fatal weakness. An AI that operates on ambiguous probabilities will collide head-on with the 'world of determinism,' where not even a single character of error is allowed.

Next time, in Part 8, we will unravel the differences in the 'personalities' of the three major AIs—ChatGPT, Claude, and Gemini—all born from a common engine (Transformer), from a mathematical perspective. And we will look at how humans should behave in the place where probabilistic AI and the 'wilderness of determinism (game development),' which does not allow for a single grammatical error, collide.