LLMs like ChatGPT-4 ‘understand’ in a Turing-like sense. Better than 98% of humans.

Tested by responses on novel input tests, LLMs understand remarkably well, sufficient to be highly logical and creative.

I’m using the Turing-like definition of ‘understand’ as used by NLU (natural language UNDERSTANDING) researchers for 70 years.

If we get an appropriate response, it got that test correct, so on that basis LLMs understabd better than 98% of humans.

That’s the extent of my claims.

Stop telling us that LLMs don’t ‘understand’!

They ‘understand’ in the same sense AI researchers have defined since the 1956 Dartmouth Workshop.

Nobody at that workshop expected AI to have internal ‘aha moments’.

Neither do I. Today.