21–30 of 190 posts
This makes good points. And it articulates an idea I’ve been having trouble getting down. What if increasing intelligence is an exponential problem and the reason humans all have somewhat similar intelligence isn’t that we peaked at some level but that even vast additional intelligence just doesn’t get much more traction against the universe of problems. Eg doubling your compute doesn’t get you many more cities in th…
FOOM = Fast Out Of the gate Model, for those not familiar (like myself)
Anyway, it's not an acronym, it's an onomatopoeia for an explosion. It's a memorable way of representing an intelligence explosion: "AI go foom." (Wild hand gesture strongly recommended.)
And a theoretical AGI (or ASI) will realise this. This makes me think of the Person of Interest series, in which ASIs are built and they go for each others throats hard. This makes eerily more sense now. We will know we created super intelligences not when they start trying to kill us, but when they start trying to kill each other.
The main issue with this argument that current AI models are extremely inefficient. Models evaluate all weights on every pass, LLMs recall their entirety of knowledge just to output a part of a word and then do it all again. Often a hand crafted algorithm can achieve what a neural network can in a small fraction of compute. There is likely a 1e4x-1e6x compute gap that can be achieved with the right algorithm, maybe e…
Fast forward 1 year. We now have students revolutionizing education (by submitting AI doctored papers and homework that isn't half bad). People in law dealing with people using AI. AIs passing medical, legal, and other exams. That was just one year. And yes, it's tediously slow to use. But last year we had nothing.
It's going to be an exponential in terms of adoption, response speed, energy usage, parameter numbers, context size and a lot of optimization. I think we are not that far off from AIs being able to keep up with a conversation and being able to respond right away. For all you hitchhiker's to the galaxy readers, that would be babelfish sorted. Translation quality is already pretty awesome with chat gpt 4. I've not caught it making major translation mistakes. Speech to text, translate the text, text to speech. Now we are conversing with our chat bots rather than hammering out sentences in some text box.
It's hard to not see that go from a novelty "oh this is cool" to world + dog basically using this on a daily basis throughout the day. At the same time it will get smarter. Everybody is waiting for the singularity where it is clearly smarter than Einstein in his prime and running circles around everyone else. But for most questions you might ask chat gpt and get a reasonable answer, the vast majority of people around you might be pretty useless. Is that smarter? I don't know. But definitely more useful.
From where we are now to this being ubiquitous and used by pretty much anyone is probably a few years. I remember when smart phones happened. One day everybody was minding their own business and a year later the streets were full of zombies glued to their screen. Just happened very quickly. This might be the same but possibly quicker.
Isn't the hedge-fund-bro example a proof of existence, not a counterexample? (RenTech by itself would be enough of a proof of existence.)
Earlier quoted context omitted.
If a human can apply AI irresponsibly, the AI can probably also learn to apply AI irresponsibly.
Not really, humans are the ones who think and have agency.
ChatGPT can already emulate beings with agency, see LangChain. It's just very bad at it. It's gonna get better.
> Oh wait…every hedge fund bro is already doing this. And most of them aren’t billionaires. The problem is your model needs to include all the computers playing the market, and it also needs to include the other hedge fund bros themselves. This strategy only dominates if you have more compute than the whole market itself, which you don’t. Best part of this article, honestly, I did not expected this, of-course one can…
No serious person has ever thought that you can simulate the entire stock market.
And the ability to do this is not limited by compute but rather not having the fundamental algorithms that can accurately predict human behaviour.
Do we have enough historical weather data to train a decent AI on? Maybe this is already being done and is just a quieter area of research (in the headline sense)? I know for example every time a hurricane comes toward land all the weather places trot out a half dozen different model predictions saying different things. It would be great if there was more confidence in this area for planning.
https://www.lesswrong.com/tag/the-hanson-yudkowsky-ai-foom-d...
Do we have enough historical weather data to train a decent AI on? Maybe this is already being done and is just a quieter area of research (in the headline sense)? I know for example every time a hurricane comes toward land all the weather places trot out a half dozen different model predictions saying different things. It would be great if there was more confidence in this area for planning.
This should give you a hint that it's not as simple as having enough weather data.
You are trying to predict a chaotic weather system that is impacted by everything from sea temperatures, sun activity, wind, volcanoes, bushfires, atmosphere changes etc. Much of which we don't even have enough data for or a complete understanding of.