If AI is a trillion dollar investment, what trillion dollar problem is it solving?
If you could increase the productivity of knowledge-workers 5%, that’s worth a trillion
- AI won’t work because of the American power grid
Makes no sense. Why would one random country having an underdeveloped power grid stymie AI?
I’m not an AI gal, but those are obvious bad points.
If you could increase the productivity of knowledge-workers 5%, that’s worth a trillion
A big if (and where does these numbers come from?), but more importantly, a “more productive” knowledge worker isn’t necessarily a good thing if the output is less reliable, interesting or innovative for example. 10 shitty articles instead of 1 quality article is useless if the knowledge is actually worth anything to the end user.
I think number 2 is a fairly good point.
Qualitatively, there were huge leaps made between 2018 and 2020. Then it’s been maybe a shade better but not really that much better than 3 years ago. Certainly are finding ways to apply it more broadly and more broad ecosystem of providers catching up to each other, but the end game is mostly more ways to get to roughly the same experience you could get in 2021.
Meanwhile, people deep in the field go “but look at this obscure quantitative measure of “AI” capability that’s been going up this whole time, which show a continuation of the improvement we saw from 2018”. Generally, the correlation between those values and the qualitative experience tracked during those early years, but since then qualitative has kind of stalled and the measures go up. Problem is the utility lies in the qualitative experience.
2 and 3 are both posing questions
-
How could AI even be applied?!?!?
-
How could AI even be improved?!?!?
as though the implication were that these are unanswerable questions
when they’re actually easily answerable
2: it can be applied to logistics, control of fusion energy, drug-discovery pipelines, lots of things that could soon amount to a trillion dollars
3: it can be improved by combining LLMs with neural-symbolic logic and lots of other things extensively written about
I assume the Goldman Sachs report is more intelligent than this summary makes out. Coz the summary is just saying we should throw our hands up in despair at well-studied questions that a lot of work has gone into answering.
the first two computers were connected in 1969 leading to arpanet. I would say the qualitative experience took quite some time to improve. The type of algorithms ai has evloved from I would say came out of the 2000’s maybe late 90’s. Taking google as sorta a baseline maybe. I would say we are equivalent now to about mid nineties internet wise so it will be interesting to say the least on where this goes. They do use to much energy though and I hope they can bring this down maybe with hardware acceleration.
I feel like #3 shows up for every tech innovation. I remember people bitching about the Internet not being viable because phone lines were too slow. The demand needs to be there before the infrastructure will get built.
Some difference in that fixing the power capacity problem will absolutely mean combusting more hydrocarbons in practical terms, something we can ill afford to do right now. Until we get our legs under us with non carbon based energy generation, we should at least not take on huge power burdens.
Now there was an alleged breakthrough that might make LLMs less energy hogs but I haven’t seen it discussed much to know if it is promising or a bust. Either way power efficiency might come as a way to save #3.
You’re absolutely right that AI is going to use electricity, and (for now) a majority of that electricity is generated through combustion.
I don’t think that is the concern when it comes to the power grid. I think most people are worried about power delivery infrastructure, which is a lot harder to upgrade/add than power generation infrastructure. Sorry if my comment wasn’t clear about this.
