[NumberWangMan]

"Fancy calculators" is kind of a reductive analogy, isn't it?

I assert that machine learning is learning and machine intelligence is intelligence. We don't say that airplanes don't really fly because they don't have feathers or flap their wings. We don't say that mRNA vaccines aren't real vaccines because we created them with CRISPR instead of by isolating dead or weakened viruses.

What matters, I believe, is what LLMs can do, and they're scarily close to being able to do as much, or more, than any human can do in terms of reasoning, despite the limitations of not having much working memory, and being based on a very simple architecture that is only designed to predict tokens. Imagine what other models might be capable of if we stumbled onto a more efficient architecture, one that doesn't spend most of its parameter weights memorizing the internet, and instead ends up putting them to use representing concepts. A model that forgets more easily, but generalizes better.

[agentultra]

Right so Bing is going to decide it doesn’t want to answer your queries anymore because you left a rude comment to someone they’re friends with on Reddit some day?

That’s what makes my day: AGI folks have to rely on myth-making and speculation.

The here and now incarnation of GPT-4 is what it is.

I’m not saying it isn’t useful, powerful, or interesting. I’m saying that needless speculation isn’t helping to inform people of the real dangers that such proselytizing is causing.

[mitthrowaway2]

AGI folks have to rely on speculation because AGI does not exist. But Goddard and Tsiolkovsky had to rely on speculation because, in their day, rockets didn't exist either. They were ridiculed[1] for suggesting that a vehicle could be propelled through a vacuum to the moon.

But there's speculation of the fantasy sort, and then there's speculation of the well-grounded "this conclusion follows from that one" sort, and the AGI folks seem to be mostly in the second camp.

And yeah, unlike GPT-4, AGI isn't here-and-now. But GPT-2 was an amusing toy in 2019, and GPT-3 was an intriguing curiosity in 2020. In 2014, "computers [didn't] stand a chance against humans" at Go[2], but two years later, it was humans who no longer stood a chance against computers. The here-and-now is changing fast these days. Don't you think it's worth looking even a little bit up the road ahead?

Isn't there some role here for speculation?

[1] https://www.vice.com/en/article/kbzd3a/the-new-york-times-19...

[2] https://www.wired.com/2014/05/the-world-of-computer-go/

[NumberWangMan]

> That’s what makes my day: AGI folks have to rely on myth-making and speculation. > The here and now incarnation of GPT-4 is what it is.

I like how Robert Miles puts it (he's speaking about Safety in the sense of AI not taking over and/or killing everyone):

"I guess my question for people who don't think AI Safety research should be prioritised is: What observation would convince you that this is a major problem, which wouldn't also be too late if it in fact was a major problem?"