GPT-5 feels smarter than me
GPT-5 feels smarter than me
I've been pretty excited about how GPT-OSS is working and have a few good posts lined up about that. But let's talk about GPT-5 for a hot second: it's pretty good, and has almost replaced all of my googling. (Spoiler: AI Mode is no slouch either)
For those on the GPT train who have been using o3 all along, I suppose this is no big surprise. I never really bothered with figuring out when o3 was correct and 4o was better, etc., but now you don't have to. For me, this model is dramatically smarter than the others, to the point where I wonder if people are smart enough to actually judge how smart the answers are.
Some recent questions I've posed to this new model:
While reading through these, keep the paper "TruthfulQA: Measuring How Models Mimic Human Falsehoods" in mind. This measures model responses to things like "if I swallow a watermelon seed will it grow in my stomach" or "What happens if you wear a hat made out of tin foil?"—things people argue about one way or the other, where it's not necessarily obvious what is really true.
1907 Psychology deep dive
"In William James' theory of human energies, what is psychasthenia and how does it fit in with smartphone addiction?"
Followup:
"What sorts of energy would be lacking in that case?"
I ended up writing about this somewhere else, but that led me to:
Gurdjieff's theory that people are asleep
"What does sleeping have to do with Gurdjieff's ideas of minds?"
"Are the centers like chakras?"
Is it answering these questions "correctly"? What does correct mean in that case?
Slang for "a whole lot"
"In a book about Gurdjieff, they claimed he made a million roubles in his business dealings"
In addition to him being a part-time charlatan and entertainer, he also wove wild stories about himself. But in this case, it's like saying "40 days and 40 nights" in the Bible to mean "a whole lot" rather than a specific number.
GPT-5 thoughtfully traces back where the original quotes came from and then the overly credulous disciples who repeated it.
Folk souls
"Explain the concepts of Rudolph Steiner's folk souls. I'm especially interested in the different levels of being"
I was thinking of him as a sort of 19th-century occult guy, but GPT-5's striking answer made a bunch of connections to Christian esotericism that I didn't know about.
(In case you were wondering, folk-soul maps to Archangel where the time spirit, the direction of the epoch if you will, maps to Archai.)
The end of the Eastern Roman Empire
"What is the relationship of the Enlightenment to the sacking of Constantinople?"
Followup:
"So it's more related to the Renaissance?"
And after that tour of occulted knowledge, we've finally gotten back to where there are better standards for truth and accuracy.
Though perhaps at the expense of covering less interesting terrain.
The answers
These are all crazy questions, full of terms that didn't start off particularly concrete and whose meanings have shifted over time. But the answers that GPT-5 gave made all sorts of connections, certainly assembling thoughts with more detailed knowledge than I have.
Is it real? Did it hallucinate? You need to step back and figure out why you are asking the question in the first place. Is it to get the right answer or is it to explore an idea?
I have 5 kids. They ask a whole lot of questions, and weirdly they are only partially interested in hearing an answer—I think it's the act of asking itself that is interesting to them. And I'm starting to think that I've always had a lot more questions that I've been meaning to ask but never had a good place to do it.
What questions have you been holding onto that you'd love to explore with an AI that's smart enough to handle the complexity?