Behind login wall, behind paywall, starts out by using 70b models to make an argument instead of the most advanced models. Uses the cliched "strawberry" as a test, which you can actually make many of the latest models count correctly if you ask them to count letter by letter like S - 0, T, R - 1, A, W, B, E, R - 2, R - 3, Y. That simulates how humans do it as well. If you were an AI researcher perhaps you would know that it sees tokens, not letters, it's kind of like asking someone who only sees everything you say to them translated into Japanese hieroglyphs and then asking how many R's were in the original text before translation. The only way they would be able to answer this is if they memorized it.
Focuses on some strawman argument about hype. I agree AGI is not here, and I don't see most people claiming that we are even near AGI, the idea of hype is just because people talk about AI so much, and I think for a good reason. It is still immensely valuable in so many different use-cases. It's not going to replace people right now, but it is absolutely going to be a productivity multiplier.
Also looking at the Medium of the Author and the content there, because of the frequency and made up stories conflicting with each other I have to presume it's all AI generated.
