3 ms·
> "Results on GPT-3 (davinci), Instruct-GPT3 (text-davinci-002) and BLOOM (176B), showcase subpar performance on such reasoning tasks." These aren't state of t
by ftxbro 4y ago
> "Results on GPT-3 (davinci), Instruct-GPT3 (text-davinci-002) and BLOOM (176B), showcase subpar performance on such reasoning tasks."
These aren't state of the art models anymore. GPT-4 has unlocked some new qualitative capabilities as shown for example in the technical report or the 'sparks of agi' report. That said, Sebastien Bubeck the one who integrated GPT-4 with Bing has in a presentation called out its 'planning' capability as its most glaring weakness, and he said that even though he would consider GPT-4 as 'intelligent' he would understand if many do not call it that because of its weakness in planning.
- famouswaffles 4y agoDid they test planning with some kind of inner monologue and/or reflexion ? Genuine question.
- ftxbro 4y agoFor the main one in this post, I'm going to admit I didn't read the paper after the point where I saw it was looking at only old models. For the one by the guy who integrated GPT-4 into Bing, the idea I got was that they were coyly avoiding putting GPT-4 into a loop, so that they could give an 'out' to people who wanted to argue that it's not yet human level. They already did that in a few other ways. One way was by dumbing it down before release, another way was by delaying its release by six months, and another way was by changing their article title from 'first contact with AGI' to 'sparks of AGI'. They don't want to antagonize people by announcing AGI because it will only cause unnecessary arguments that they don't want to have.