Musk is the tech billionaire of the age—the one we deserve. Jobs had many faults, but he had taste, wisdom and a depth of soul no current tech executive can match and few even understand. I'd go so far as to say Musk is more innately talented than Musk, but he is completely unbalanced and broken.

Nick Bostrom's "Superintelligence: Paths, Dangers, Strategies," which isn't programming but about the potential dangers of superintelligent AI, bucks the trend a bit, but peaks in 2020 when you'd think 2022 (the release of ChatGPT) would have really pumped it up.

I adore your avatar, by the way. Such warm memories.

Also, I wrote back in English. I hope you can understand English. :)

Ha. But does it cogitat?

It's notable that many places online say John Paul II declared him so in 1997, but that this claim appears entirely made up. Other origin stories also fail. In this way, perhaps, Isidore really IS the patron saint of the modern, misinformation internet. See aleteia.org/2020/05/02/a...

This is a book you want to read by audiobook, because it's not Patel reading the transcripts, but simply *playing* the transcripts. So the audiobook is the original and best form of the book.

Machine-learning engineers are basically the last people on earth whose opinions on these topics we should listen to. They think they understand things they just don't. It's a classic case of domain overgeneralization.

2/ I'm currently working on a prompt that takes 60 seconds to answer, but with batch processing I can submit 5k at a time, and get them back in 10-20m. That would take dozens of processes making requests non-stop normally. Also, they're 50% of the price.

Yeah. I noticed a similar jump in describing book covers. But, as you say, is that perception or description. I mean, Hinton said radiologists would be obsolete in five years, in 2016. Clearly they're improved a good deal, but has it been a 10? I don't feel like it has.

Corollary: Could you make LLMs smarter by screening out some of the less-smart content?

4/4 My guess is: No. You can make it fast, and to some degree intelligence is just connecting things, so having more time and things can ramp you up. But you can't get superintelligence out of regular-intelligence data.

Thoughts?

3/4 Maybe you can get a really good frontier model to assess the "IQ" of the data. Can you get it to reason and think to a higher level than its training data?

2/4 So here's the test: Train an AI on only low-intelligence data. Now data doesn't exactly come with IQ scores as metadata, but you could perhaps train it on only children's books, children's own writing, material with a very low Flesch–Kincaid, etc.

I'm not making a point; I really don't know!