It cannot be denied that modern AI machines have achieved remarkable fluency with language. They seem to understand what we are telling them, regardless of the words we choose to express ourselves. This allows for the same fluidity of conversation that we have with humans. However, we must not forget that LLMs are not designed to be honest, but to ensure that the thesis “makes sense” in any context. Given a context, LLMs are trained to generate what needs to happen in the description being developed. Confabulations – plausible distortions or fabrications – regardless of whether they relate to truth or facts we the world
One of the primary functions of language is to imagine and make possible ideas that have not yet been expressed. LLMs do this easily, even when the context has nothing to do with the information they’ve been taught. A story always makes sense in any context because the machine has learned some common language structures that transfer to new situations. One of them is “composition”, the concept that the meaning of a complex sentence is determined by the meaning of its parts and the way they are connected. AI has learned several such useful patterns
About my last podcastmachine learning researcher Leon Bottou states that LLMs are essentially artistic machines that can be very good at talking about new situations that are far from their educational background. Actually, what I find impressive is how accurate and correct LLMs are, considering they weren’t designed to be. One of the reasons for this can be the intensive study of human feedback (RLHF) from an army of human validators operated by operators of LLMs who find their answers correct or socially acceptable.
Given its prowess at generating content, could AI produce novels? Can it create future theories of physics that are unknown to mankind and not expressed in academic data?
Creating new plots and writing novels should be easy for the AI. After all, if LLMs are artistic machines, they should have no problem creating stories, regardless of their quality. As Botu says: “Instead of one artificial intelligence with enhanced reasoning abilities and encyclopedic knowledge, the best model of language is as a machine that prints fiction on tape. When new words are printed on tape, creation follows neat twists and turns, extracting evidence from study data and filling in the gaps with plausible confabulations.
But can AI discover new theories?
If we already have a number of candidate models identified and the task is to identify the correct model, this is easy for AI. However, if the theory requires new concepts to describe it, this may require new meanings to existing words or the creation of entirely new concepts, which is a big deal for a machine. Einstein’s theory of relativity created new meanings for existing words such as time, gravity, and strength. Similarly, thermodynamics and quantum mechanics created new concepts that required the introduction of new words, such as photon, quark, quantumand entropy.
Theories usually go one step further. In addition to the symbols and concepts that represent them, theories usually require a causal structure and a mathematical formula. Reasoning means that the phenomenon must be understandable to people in terms of the symbols used. This raises the deeper question of whether intelligence can be determined completely by signs. Are symbols in phenomena like feelingsvisual representation and motor control? Otherwise, it is impossible for us to understand the new theory that the machine comes up with if it cannot explain it to us in terms that we can understand. This reminds me of Jeff Hinton’s metaphor of AI as an alien that is so different from us that we can’t always understand each other.
It’s a brave new world: an intelligent alien, albeit one of our own creation, living alongside us that we can’t fully understand. Until we learn it language
