> I’ll be generous and say that sure, words like “understanding” and “meaning” have definitions that are generally philosophical, but helpfully, philosophy is an academic discipline that goes all the way back to ancient Greece. There’s actually a few commonly understood theories of existence that are generally accepted even by laypeople, like, “if I ask a sentient being how many Rs there are in the word ‘strawberry’ it should be able to use logic to determine that there are three and not two,” which is a test that generative AI frequently fails.

The strawberry thing has been solved and LLM's have moved way beyond that helping in mathematics and physics. Its easy for the blog author to pick this but lets try something different.

It would be a good idea to come up with a question that trips up a modern LLM like GPT with reasoning enabled. I don't think there exists such a question that can fool an LLM but not fool a reasonably smart person. Of course it has to be in text.

> The strawberry thing has been solved

It seems like every couple weeks there's some embarrassing failure of AI that gets quickly patched, but just because AI companies scramble to hide the failures of their technology doesn't mean they haven't failed in ways that shouldn't have been possible if they were what they claim them to be.

> I don't think there exists such a question that can fool an LLM but not fool a reasonably smart person.

An example was on the front page here just a few days ago.

https://s3.eu-central-2.wasabisys.com/mastodonworld/media_at...

Until someone invents an LLM that has any actual understanding of the words it outputs (which doesn't seem likely to happen in my lifetime) these things are going to keep happening, just like how it's impossible to get them to stop hallucinating. The limitation is intrinsic to what they are. We call these chatbots AI, but there is no intelligence there that didn't come from the humans whose words were used to train them.