5 hours ago
7
Gary Marcus is right to point out – as many of us have for years – that just scaling up compute size is not going to solve the problems of generative artificial intelligence (When billion-dollar AIs break down over puzzles a child can do, it’s time to rethink the hype, 10 June). But he doesn’t address the real reason why a child of seven can solve the Tower of Hanoi puzzle that broke the computers: we’re embodied animals and we live in the world.
All living things are born to explore, and we do so with all our senses, from birth. That gives us a model of the world and everything in it. We can infer general truths from a few instances, which no computer can do.
A simple example: to teach a large language model “cat”, you have to show it tens of thousands of individual images of cats – being the way they are, they may be up a tree, in a box, or hiding in a roll of carpet. And even then, if it comes upon a cat playing with a bath plug, it may fail to recognise it as a cat.
A human child can be shown two or three cats, and from interacting with them, it will recognise any cat as a cat, for life.
Apart from anything else, this embodied, evolved intelligence makes us incredibly energy-efficient compared with a computer. The computers that drive an autonomous car use anything upwards of a kilowatt of energy, while a human driver runs on twentysomething watts of renewable power – and we don’t need an extra bacon sandwich to remember a new route.
At a time of climate emergency, the vast energy demands of this industry might perhaps lead us to recognise, and value, the extraordinary economy, versatility, plasticity, ingenuity and creativity of human intelligence – qualities that we all have simply by virtue of being alive.
Sheila Hayman
Advisory board member, Minderoo Centre for Technology & Democracy, Cambridge University
It comes as no surprise to me that Apple researchers have found “fundamental limitations” in cutting-edge artificial intelligence models (Advanced AI suffers ‘complete accuracy collapse’ in face of complex problems, study finds, 9 June). AI in the form of large reasoning models or large language models (LLMs) are far from being able to “reason”. This can be simply tested by asking ChatGPT or similar: “If 9 plus 10 is 18 what is 18 less 10?” The response today was 8. Other times, I’ve found that it provided no definitive answer.
This highlights that AI does not reason – currently, it is a combination of brute force and logic routines to essentially reduce the brute force approach. A term that should be given more publicity is ANI – artificial narrow intelligence, which describes systems like ChatGPT that are excellent at summarising pertinent information and rewording sentences, but are far from being able to reason.
But note, the more times that LLMs are asked similar questions, the more likely it will provide a more reasonable response. Again, though, this is not reasoning, it is model training.
Graham Taylor
Mona Vale, New South Wales, Australia
