News for hackers. Stuff that matters.
Re:Sutskever: OpenAI board doing its mission to build AGI th... by chpatrick on Friday November 17, 2023 @11:44PM (#38315334)
If you told someone in 2020 that we would have DALL-E 3 and GPT4 today they would say you're talking about UFOs too.

"Just an incredibly advanced Markov chain" doesn't mean anything. Humans writing text are also incredibly advanced Markov chains.

Re:Sutskever: OpenAI board doing its mission to build AGI th... by wruza on Saturday November 18, 2023 @12:26AM (#38315708)
Writing text isn't that big of a deal either, you just have to be able in winning arguments and converting phonemes to glyphs. I'm not taking sides here, but GPTs defeated me too many times for me to agree with that on better terms.
Re:Sutskever: OpenAI board doing its mission to build AGI th... by usaar333 on Saturday November 18, 2023 @12:22AM (#38315667)
> If you told someone in 2020 that we would have DALL-E 3 and GPT4 today they would say you're talking about UFOs too.

Depends if they were paying attention. Emerging creative behaviors were noticed in GPT-3 in 2020. Image generation was also getting pretty good by 2020 (with Dall-e coming out in early 2021).

Very bullish 2020 post on GPT-3: https://slatestarcodex.com/2020/06/10/the-obligatory-gpt-3-p...

Re:Sutskever: OpenAI board doing its mission to build AGI th... by chpatrick on Saturday November 18, 2023 @12:37AM (#38315783)
I just said a random date, let's say 2010 then.
Re:Sutskever: OpenAI board doing its mission to build AGI th... by batman-farts on Saturday November 18, 2023 @01:11AM (#38316052)
No, sorry, I don't buy this assertion when it comes up. Everything I've seen, even from the most advanced image generators, has struck me as a logical follow-on from the "big data" trends of the 2010s. If you've ingested literally everything of a certain data type available on the internet, or even a significant fraction of it, it follows that you'd eventually be able to mix it all together and produce randomized outputs that seem novel.

It also elides the significant encoding of human feedback, a contribution that AI firms have typically been none-too-eager to highlight.

Re:Sutskever: OpenAI board doing its mission to build AGI th... by pdgenoa on Saturday November 18, 2023 @02:10AM (#38316423)
So you disagree with and think you know better than the 14 people at Microsoft Research that had access to the unrestricted early GPT-4 model and wrote the white paper outlining the results of experiments with it that concluded the model showed "sparks of artificial general intelligence".

That's their informed, direct conclusion after hands on with the unrestricted model. I wonder what your conclusion is based on.

Re:Sutskever: OpenAI board doing its mission to build AGI th... by insanitybit on Saturday November 18, 2023 @10:27AM (#38320587)
Could you link me this paper? It sounds very interesting.
Re:Sutskever: OpenAI board doing its mission to build AGI th... by chpatrick on Saturday November 18, 2023 @10:50AM (#38320823)
Re:Sutskever: OpenAI board doing its mission to build AGI th... by chpatrick on Saturday November 18, 2023 @08:25AM (#38319296)
> If you've ingested literally everything of a certain data type available on the internet, or even a significant fraction of it, it follows that you'd eventually be able to mix it all together and produce randomized outputs that seem novel.

That's a ludicrous "it follows". Search engines have been collecting everything from the internet since the beginning but you can't just magically rearrange it and pass the Turing test. We're way past the point where you could say everything ChatGPT says is copy pasted from somewhere.