The legendary Subbarao Kambhampati (కంభంపాటి సుబ్బారావు) explains the difference between reasoning and memorisation using the example of the famous "why are manhole covers round?" interview question. Which do you think LLMs do?

Now published in full:

@rao2z Because manholes are round?

@rao2z Is there an entire video of this interview ? Or is it just this one bit ? I would love to see the whole interview conversation if there is one :)

@rao2z Dropping hopefully later today

@rao2z ML Street Talk "CHATGPT DOESN'T REASON! (Top scientist bombshell)" SUMMARY

Really enjoying this. It seems like @rao2z is mostly arguing that out-of-distribution reasoning or reasoning from first-priniciples isn't happening. I wonder if this really matters in the medium-term though, even for performing science using AI, if we may simply be able to grow distributions large enough that they contain enough information to become experts, rather than having to rethink how machines compute generally.

@rao2z Why are you sure LLMs can’t reason?

@rao2z Any conjecture is made up of a number of subcomponents. Assembling them in a reasonable arguement is one step. Choosing the right sub components is another. But more important than reasoning is discussion... The Oxbridge tutorial method creates useful insights..LLMs discuss well

@rao2z I'd say they're like humans. They do both.

@rao2z The definition of reasoning: 'if we change the rules in the data just a little, LLM completely fails, so LLM does not reason.' I'm sorry, but this is not a definition. This debate will progress if we can define things clearly. What is reasoning ?