LLMs are models of patterns. They are accurate to the extent that they embody the statistics and likelihoods that they measure. Chatbots, on the other hand, are less accurate. That’s a lot to do with the fact that they come out with one version, one outcome, based on the statistics embodied in the LLM. The LLM is the Many Worlds (or should that be Many Words) hypothesis: it contains every possibility. The outcome of a chatbot is like reality: one outcome, however unlikely.

When doctors use ‘AI’ to help them decide whether you have a suspicious lump in your lung, it looks through impossibly high numbers of previous scans of bodies, all of which have been characterised as ‘cancerous’ or ‘not cancerous’. Then it puts your scan in amongst them and sees where it falls. It doesn’t then just say ‘yes’ or ‘no’ to the question “is this patient ill?”. It helps the doctors make a decision. It’s very different to chatbot.

Evil bots

So when someone rightly complains about everyone being down on AI, and that AI saved her friend’s life, they’re falling into the trap of AI being defined too fuzzily, taking in everything we would, not so long ago, have called an algorithm.

Imagine if Amazon, instead of giving you 10 recommendations of what it thought you’d like, instead just bought, with your money, the top one in that list, “knowing” you’d be happy with that purchase. Would that be useful?

That’s the difference between slop-spewing chatbots and poster creators on the one hand and LLMs and ‘models’ of all kinds on the other. One tells you probabilities, the other tells you the most likely outcome based on probabilities (and not even the most likely every time – randomness comes into it).

What I hope is that we make the most of the data-rich and intricate models that we’re destroying the Earth for, and move away from the confident yet hallucinating outputs that lead to glue pizzas and dodgy medical advice. That way salvation lies.