What is the energy increase going to be to go from what AI can do today, to something that can do all those things? It seems like it will need to be substantial.
2 Likes
Looks like they are getting closer.
Not holding my breath though.
Itâs a good question. We do keep inventing more powerful LLMs while also inventing ways to save power? There was a preview of a model-- o3-- that was costing $1,000 to answer a question. Then they jiggered with it or something, and now itâs less $1.
Right now, video generation is really power hungry. It costs something like 500x more power to generate a 10 second video than to have a chat or picture. Itâs easy to imagine some big film makers being like-- letâs just spend $10M on making a blockbuster all-AI generated film, using some kind of recursive trick. Otoh, I have no idea how that compares with hollywoodâs existing power consumption.
I think the real issue is not that we will invent âAGIâ, but that whatever AI we invent will work 10% better if we feed it 10x as much power. And people just naturally feed it a billion x the power.
We kind-of already see this with the âreasoning modelsâ. And the link shraed by Polymath.
Worse still itâs not just that the AI becomes expensive bc of power consumption but ALL power consumption becomes more expensive bc AI is buying up all the supply.
It may come to a point where we need regulations to force data centers to also build power generation.
2 Likes
Optimizing 1000 to 1 suggests there was some pretty bad programming in the initial code. We have all seen this in our own jobs. But clearly, there will be opportunities to optimize any code.
Video processing has always been an expensive component of gaming, but that is also what sells the games. I think about FPS games, and I have not played a new one in quite some time, but they never had much of a dynamic environment - any attempts were always very simulated. Damage to a building from gunfire, or how the ground reacts to certain actions. These were either not worth the cost to design, or the processing required to pull it off just wasnât available even in a high end computer.
I think AI may follow a similar path. Right now we have simulated intelligence. That can improve, and become more efficient, but AGI is something different, just like that true drynamic gaming environment.
Eh, my political view is that power consumption is a problem that would solve itself if we just had carbon taxes and maybe also rich people taxes.
AI doesnât need to be treated special, any more than other power-hogging machines.
Yes, the computers will compete against humans for grid power, and under capitalism, it will be allocated on what can provide more business value. Energy prices will skyrocket.
I disagree, power is a utility so itâs treated as a public good. Data centers support close to 0 jobs and consume vast amounts of electricity. Iâd nearly rather a casino.
Itll be like almonds in CA with subsidized water.
3 Likes
Thatâs a pretty reasonable comparison.
A highend gaming computer can eat 3 MJ in an hour.
AI generated video can eat 3 MJ in 5 seconds.
I donât know why, actually, that no games connect to server-farms to eat-up magajoules of compute. Maybe itâs just not worth it, like you say. I suppose movie CGI must be doing so though.
I first heard the word âenergyâ used as an economic concept about 50 years ago. And, that cheap energy from fossil fuels was fundamental to our modern lifestyles. And, that ff were a nonrenewable resource. BUT, wait, fusion could save us, maybe in 50 years.
So Iâve been reading about fusion for 50 years, and weâre still not close to economically viable power, maybe another 50 years.
I think it turns out that it doesnât really matter. Solar panels are cheap enough that they cap electricity costs at a reasonable price. Maybe someday fusion will be cheaper, but it will be a meh event.
It matters in the context of space travel.
We will end up consuming the majority of the planets mineral resources eventually (assuming we dont end up destroying the environment completely first) so being able to mine asteroids (or similar) will become a necessity.
Fusion power would also be a âcleanâ source of energy on the ground so you would potentially need less solar and wind energy generation (solar and wind require maintenance which can get expensive over time).
Network latency would be an issue in a gaming environment. Even small lags between a local machine and a server governing the game logistics can make an online match uncompetitive.
And accidentally the whole thing?
There are downsides to that.
The xAI facility that opened in Memphis is drawing a LOT of community backlash due to increases in pollution arising from the additional power generation going towards supporting it.
Itâs a difficult question, because as ND points out, they do a lot of pretending. They pretend to reason (chain of thought models). They pretend to have theory of the mind (Nature study last year). They pretend to have goals (numerous studies from Anthropic).
They pretend well enough to pass our tests and to fool people, which can have practical consequences. And lately, they have been able to pretend well enough that they can write and edit complicated code. Which is pretty surprising, imo, because coding is something that by definition canât be faked. Of course they eff up code too-- but not all the time-- which is what I would expect. Presumably LLMs have memorized countless snippets of code, but combining snippets-- or attaching them to working codebases-- or debugging code theyâve never seen before-- it all suggests a really serious level of symbol-shuffling skills. Serious enough that it already has profound consequences for developers, engineers, and mathematicians. And serious enough that itâs hard to reconcile how bad they are at goats and cabbages.
Anyway, until it stops improving-- until it reaches some kind of plateau-- I donât want to guess where itâs going.
I also think âAGIâ is the wrong way to think about it anyway. I donât think AI will ever be good at everything, and I also donât thinkt it matters. What matters is that it gets really great at a few things, and those things change civilization.
This is the worrisome part imo, and why we need some strict regulations on how it is used. We can choose to use it to improve everyoneâs lives, or we can choose to use it to enrich a handful of people at the expense of everyone else. As usual we have a choice between two visions for the future, and itâs not clear which path we will follow. Iâve always thought of it as Star Trek vs. Blade Runner.
1 Like
Based on what I am seeing over here in the UK:
-
Entry level roles will become more specialised and there
will be less of them. -
Universities and Actuarial exams will incorporate AI into
their work so that candidates will have to use AI to create
code to solve business problems.
This generally means that entry level is about to get a lot more competitive (less jobs and more candidates).
So effectively, the valuation crank-the-handle type grunt work we all had to suffer through at the beginning of our careers will be replaced by a shift towards understanding and validating the results of existing model outputs. You cannot fully automate this process, so a human will still be required (just less humans than before AI).
It will basically skew towards less crank the handle type work, and more value add analysis and validation work.
Thats what I am seeing now in graduate recruitment and entry level work (they are really investing a lot now in AI with all new starters being trained on copilot).
If we make the general assumption that the âtypical roleâ today is what itâll be going forward, I agree that this is very likely what weâre going to be seeing.
However, I think what weâre also going to see is that this âtypical roleâ is going to morph into something that is going to be something that wonât be easily comparable to a what this looked like 20 years ago.
For example, I think weâre going to see a far greater emphasis on data science topics with that âactuarial focusâ on the business context (something that I seriously doubt that AI is going to âmasterâ in the foreseeable future).
On the P&C side, I believe that the CAS has already shifted their Exam materials to have a greater focus on the modeling (see the CSPA credentials for an illustration; the CAS Exams has a watered down requirements from these).
1 Like
Iâd argue that we have had the technology for over 20 years now to enable this, but few companies have actually achieved this. Two possible reasons: 1) the value of achieving this is actually lower than expected and 2) the expected career progression of a typical high performing actuary will move them out of the type of role in a valuation team that could push this forward.
It isnât clear where AI can effectively fit in to this dynamic, and there is a chance that companies will see AI as an opportunity more in that entry level space (same as our conclusion here) which will only delay the development required to become a seasoned actuary.
1 Like