site banner

Friday Fun Thread for March 17, 2023

Be advised: this thread is not for serious in-depth discussion of weighty topics (we have a link for that), this thread is not for anything Culture War related. This thread is for Fun. You got jokes? Share 'em. You got silly questions? Ask 'em.

3
Jump in the discussion.

No email address required.

Ok so what can’t GPT-4 do at this point?

It still hallucinates a lot.

It is possible to get to to do useful things but the constraints are largely the same as for GPT-3.5 even if the output is a bit more polished.

Very surprised to find that no LLM I've tried, including GPT-4, has been able to zero-shot "Which aircraft carriers were afloat in the Pacific on June 10th 1942?"

They almost always list the carriers lost at Midway, even as half the time they note that they were sunk on June 4th/5th/7th as part of answering.

Avoid trying to confidently redefine basic math?

Kill all humans.

From what I've tested:

  • it does better with Finnish history questions I asked here from GPT 3.5, but still made at least one obvious error with 3/4 of the questions asked, and hallucinated even more stuff about Veikko Vennamo than 3.5.

  • it cannot translate into Finnish as well as a dedicated program like DeepL. I've only tried one text, but my estimate would be that it's output is similar to 3.5, ie. there are obvious grammatical mistakes and like.

  • it cannot create palindromes or Finnish-language spoonerisms. (edit: it's notso-hotso with English-language spoonerisms, either)

  • it struggles in writing a credible fictional alternative history timeline. I tried asking for a Confederacy-wins scenario, and its proposed point of divergence was Lee winning at Appomattox Court House.

Design and train GPT-5.

The OpenAI team use it internally, so it can't do it by itself but at least it can help!