Gary Marcus at the Royal Society

 

Gary does a good job of poking holes in LLMs, stressing gullibility as the main reason LLMs took off.

Gary Marcus - My 25 minute talk about this, delivered at the Royal Society on October 2, on the occasion of the 75th Anniversary of Turing Test can be found here.

Ard then talks about the benefits of NeuroSymbolics.

At the end of his talk, he shows a slide with suggestions for the future, and a quote from Turing.

Suggestion 1:

“Many people think that a very abstract activity, like the playing of chess, would be best.” - GM

This would be a very bad way to start.  After a thousand years of single stepping, Catherine the Great changed the rules of chess to make the queen the most powerful piece on the board – no deep thoughts about strategy there.  Machines can beat any human at chess, by the simple process of playing millions of games and working out the best move at any stage in any game. We should be hoping for new ideas, where the machine has the wit to change the rules within the bounds of current reality. Chess is a very bad foundation to build on, based as it is on our worst limitation – the Four Pieces Limit. Looking after 16 chess pieces is a far cry from a military battle which can flow over land, sea, air, space, cyber and psychological warfare.

Suggestion 2

“It can also be maintained that it is best to provide the machine with ‘senses’, and then teach it to understand and speak English. This process could follow the normal teaching of as child. Thins would be pointed out and named, etc.” – GM

A child has an unconscious mind, which we don’t understand, which is eager to learn. The machine doesn’t.

A quote from Turing: “We can only see a short distance ahead, but we can see plenty there that needs to be done” – Computing Machinery And Intelligence, 1950

The problem – Neurosymbolics is never going to get us to having a machine understand English. The notion that we can stick two mutually exclusive and crude things together, and end up with something (with all its quirks) that smoothly integrates law, mathematics, physics, chemistry, genetics, psychology is fanciful. “We could build specialist machines”. We could, but then we would end up with the same poor collaboration we get from human experts – they may be brilliant within the boundary of their specialty, and frighteningly bad on just the other side of it.

There is a shortcut – we make English the language of AI.

This way, we show just how profoundly unwise (or snake-oilish)  it was to base “AI” on something that has no understanding of the meanings of words (it is fine for a Search Engine, which might otherwise bring back a million hits). It would start people on the long road to understanding what comes out of their mouths courtesy of the Unconscious Mind, while we build a machine version of it.

What about things we don’t understand, like Quantum Entanglement? We run out of answers very quickly on most things – six is pretty typical. Is there another language that does a better job of explaining the unexplainable – English (and a few formulae) will eventually get there.