Sunday, January 07, 2024

Artificial Intelligence - calm down everyone - meaning is use

‘Artificial Intelligence’ was coined in 1956 by John McCarthy at Dartmouth College for the conference, held there, that kicked off the modern era in AI. Along with Marvin Minsky, Nathaniel Rochester, and Claude Shannon, he posited that AI was "every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it." Not a bad pop at definition and pretty open. I studied there and they had a little area dedicated to that conference, which has turned out to be much more significant than was thought. It was there that I got interested in tech in the very early 80s, was exposed to a mainframe computer, came back to the UK, built by first technology-based learning programme, to teach myself Russian, and the rest, for me is history.

McCarthy later regretted coining the phrase and preferred the words 'Computational Intelligence'. If we had listened, we’d all be using the letters CI! His problems with the terms Artificial and Intelligence were prescient, in that he saw it encouraging exaggeration, hype and over reach, also misunderstandings about what it actually was. He turned out to be right. The current doomer debate is the perfect example, where its meaning is hopelessly entangled with esoteric moralising. He preferred a phrase that focused on actual computing to do things we humans can do or find difficult.

There is the added problem that the word 'artificial' has perjorative tones, as in artificial grass or meat. The word 'intelligence' is also problematic invoking anthropomorphic and difficulties in definition, even outside of AI. I discuss these issues in detail here.

Four main meanings

In truth, four main meanings have emerged:

1. For many, when they use the phrase AI, they mean the post November 2022 Generative AI explosion around Chatbots, image generation and more recently multimodal AI. Even then, for some, it's ChatGPT3.5 and not 4 and subsequent versions - tools move fast here.

2. Others give it a wider meaning to include pre-November 22 achievements such as Deepmind’s Alphafold and AlphaGO then AlphaZero, AlphaStar, AlphaGeometry (though few know the latter two).

3. Other still, who knew about or were working with AI for years see it wider still, understanding that far from being one thing, AI is many things. This was well covered in Pedro Domingos’s book The Master Algorithm, with chapters on Symbolists, Connectivists, Evolutionists, Bayesians and Analogizers. 

4. Some also fix its meaning in the ethical debate and automatically think of this when the phrase is uttered.

From the ‘learning’ perspective, a word that was emphasised in the original 1956 definition, there’s machine learning (supervised and unsupervised), reinforcement learning, deep learning and all sorts of species and sub species. ‘Learning’ is a word you hear a lot in AI. 

Other ways of looking at AI is through areas of problem solving, such as NLP (Natural Language Programming), image recognition or robotics. There are many other ways still of slicing the AI cake. The important point is to see it as a set of very different technologies or tools that solve problems. Roger Schank, who was heavily involved in these debates, preferred to just call it software – he had a point!

Meaning is use

You can see the problem. There is no one 'meaning', as meaning is nuse As I explained in my book AI for Learning, Wittgenstein claimed the word ‘game’ is difficult to pin down, and is used in itself and with other with other words, to mean all sorts of things, from Olympic Games to board games, games people play, game theory, even simply throwing a ball against a wall. His theory that meaning is use and that meaning emerges from a set of ‘family resemblance’ is interest in AI, as that is exactly how LLMs work. Meaning emerges from its use within a Chatbot. His theory of language games is also useful in this context. Wittgenstein’s message was – beware of large words, especially in their reification.

It shouldn’t bother us too much, as meanings change with semantic drift, new facets of AI emerge and are folded in, or assumed in use. This is how language works. Meaning is not, as many pedants would like you to believe, strict fixed dictionary definitions. It is the other way round – dictionary definitions come from use and people use words in different ways within different contexts.

Funnily enough, LLMs, do exactly this. You can get them to play Wittgensteinian 'language games', by asking them different types of questions, getting them to play different roles and use different forms of expression.  I explore this here and in this podcast.

When one sees meaning as use, we can relax and understand that AI has and will continue to change. It has come to mean different things, as it conquers new challenges. It is, in a sense, what is will be.

PS another beef!

Whenever I hear the phrase ‘stochastic parrot’ I think – here goes, someone’s picked up on a meme to wave around without knowing where the phrase came from. I simply ask where the phrase comes from.

To be fair, those who have tread the paper from which it came, written well before ChatGPT3.5 was launched didn’t really know either. DSo you know the author of the paper or when it was written?)

Sure LLMs rely on statistical processes but they are in no way mere parrots and can exhibit complex language generation capabilities. They do not ‘parrot’ anything in the sense of copying or sampling. If you parrot someone or something, you literally copy or mimic them. LLM output does no such thing. It is all freshly minted.

Of course, it is always used pejoratively, with more than a shade of a sneer, as if it were mindless repetition, which it is not. This reductive approach strips all nuance away from what they do in terms of their utility, abilities to rewrite in different styles, summarise, critique, create and translate. 

The Bender paper is actually an attack on increasing the size of LLMs. It was written a two years before ChatGPT3.5 hit. So they got this completely wrong. Their recommendation for AI was to carefully curate datasets, not pursue Larger Language Models. How wrong they were.

No comments: