But... if... run multiple cities off the waste heat from the compute. Easy to see the chain of logic which runs from there to the Dyson sphere made of computronium.
Sam Altman seems to want power and attention. If mankind has $7t to throw away it would be far better spent moving the entire world to renewable energy and renewable powered transportation.
Unless his ultimate goal is to leave earth to our AI powered overlords, “AI” is absolutely meaningless if the planet is uninhabitable for humans.
Why wouldn't we just use the 7B, build the nuclear and solar builds and use it to transition transportation and industrial processes? Why use the money to train a neural net that will likely still be less useful than a small team of humans?
He may want power and attention but he is FAR less power and attention seeking than Elon Musk.
I think he enjoys being powerful like any human but has a genuine care for how AI could improve humanity and the world for all humans. He does see what OpenAI is doing as a "Manhattan Project", in which the technology has a tremendous amount of inherent risks, but it's better they proceed with the greatest caution and make steps towards international cooperation so that AGI doesn't emerge in the wrong hands. Since the tech is inevitable, it is best it doesn't deploy itself into the world without great disasters occurring on its release.
As for the 7 trillion dollar thing, I don't think he's serious about it, as in he doesn't expect to get the money in the same way he'd expect to get 10-20 billion for OpenAI if fundraising. I think he's using the 7T figure as somewhat of a marketing tactic to elucidate the world on the compute build up that will be necessary. He probably believes that in a few years, 7 trillion will be spent to get to the gigaflops necessary for ASI. He probably feels its best to deploy the crazy number now so people are generally aware of the scale he anticipates for when AI really gets going in 1-2 years.
People have been scared of AGI for decades, such that it would take over and destroy humanity. Hail our robot overlords! My take is that it will do that indirectly by destroying the climate first because of the energy requirements. I find it chilling to imagine the possibility that any inevitable successful AGI would require humans to first destroy themselves. As in we are setting up the precondition for it. Just like how dinosaurs were wiped out so that the homo sapiens can evolve.
Nobody is giving Altman $7 trillion and there isn't any evidence that he's actually trying to raise it (as opposed to suggesting that level of investment globally over time toward AI related infrastructure to support its build out).
I got 'triggered' into this thought by thinking about the energy requirements of a system worth 7 trillion. A system which seems to require exponentially larger resources for progression.
> I find it chilling to imagine the possibility that any inevitable successful AGI would require humans to first destroy themselves.
That's not the chilling thing. The chilling thing is there are people (mainly sci-fi addled software engineers) who would consciously choose to destroy humanity in order to create an AGI. It's like they read about paperclip maximizers, and decided to become one for <insert-scifi-concept-here-usually-agi>.
nice writeup, pointing out the tension between quick scaling and safety.
I think a bigger problem with this is more simple: there's a pretty critical assumption that the GPT-6 and GPT-7 models will need orders of magnitude more compute.
On the one hand, that was true for GPT2 > 3 > 4. On the other hand, history is rife with fallacious predictions about technology development and resource usage, all the way back to Malthus. Typically, as the human race focuses on a problem, whether that be farm efficiency or AI or anything else, people invent new approaches that invalidate simplistic forecasting about resource usage.
In this context specifically, I think there's a lot of promising research with different types of model architectures and multimodal models, to the extent that AGI may not be a path paved through brute forcing transformer training for LLMs. In fact, I believe Sam Altman said this himself last year!!!
Also, from a pure business perspective: it's one thing to spend $7T and build a bunch of factories. It's quite another thing to beat Jensen at that game, who has a legendary history of beating the odds when his company was headed for failure by constant process innovation and strategic moves -- this podcast episode saga is a great listen: https://www.acquired.fm/episodes/nvidia-the-gpu-company-1993...
I think Sam is thinking too much like a VC here -- raise as much as you can and take your % management fees -- and I would not be eager to invest so much into brute forcing a hardware business at scale. I'd much rather invest that into NVIDIA or AMD, who have the battle scars to know how to build hardware profitably.
Especially with computer-related hardware tech, prices tend to go down as competition increases and that the technology gets replaced with higher performing one.
CD ROM burners were amazing technology, and now they are super cheap.
Pentium II was super expensive too.
GPUs cards... GeForce GTX 470, 350 USD into 35 USD today, in 10 years.
There is no reason to believe that "AI accelerators" (which are +/- non-specialized GPUs for now) won't become cheap too.
It may even come faster than expected, like this year, Apple may already announce specialized circuitry.
And no, there is no need for 7T USD for that.
The foundries are the same, we haven't discovered integrated circuits, just some cool software that is not different than running Grand Theft Auto in "the cloud".
The impression I got from the WSJ article is that he wants a 7 trillion dollar investment in chip manufacturing world wide, not that he personally wants to raise 7 trillion dollars. If you put it that way, it's not that unreasonable at all.
What has AI given us so far that would hint that 7 trillion would ever pay off for humanity? AI seems neat so far but it's not much above a chatbot that gives me info I could find with google. It can make lots of strange images and the hands are getting better. Am I being overly pessimistic? I think people that use it for coding help are overly optimistic about it because it is awesome at that.
I assume $7T is a commitment spread over many (10?) years, which is just mere 700B a year. Now get 100 largest companies/governments/funds invest just a 7B a year each and you will have it, ez-pz.
>A major distinguishing factor of the sciences (specifically, biology, chemistry and physics), as compared to AI fields such as natural language processing and computer vision, is the relative lack of publicly available data suitable for these domains (see Box 1). There are simply vastly fewer data existing in the sciences, and these are often siloed by academics and companies. Acquiring appropriate scientific data for AI typically requires not only highly trained humans, but also high-end facilities with expensive equipment, making for an overall costly and slow endeavor compared to humans simply going about their day by adding to the vast trove of images, text, audio and video on the World Wide Web. As one example, it has been estimated that “The replacement cost of the entire PDB [Protein Data Bank] archive is conservatively estimated at ∼US$20 billion”1 —these are the data used by DeepMind to train AlphaFold2.
As history has taught us sometimes the impossible happens in an instant and sometimes the "should be easy" takes millennia to solve. The "AI" we have today is definitely usable in very narrow use cases versus the just paper-worthy "AIs" of the past however there is nothing to suggest a seven trillion endeavor will have even a marginal utility.
"My current impression of OpenAI’s multiple contradictory perspectives here is that they are genuinely interested in safety - but only insofar as that’s compatible with scaling up AI as fast as possible."
I'd rewrite this as "...they are genuinely interested in safety - but only insofar as it doesn't get in the way."
78 comments
[ 2.1 ms ] story [ 157 ms ] threadBut... if... run multiple cities off the waste heat from the compute. Easy to see the chain of logic which runs from there to the Dyson sphere made of computronium.
We live in interesting times!
Unless his ultimate goal is to leave earth to our AI powered overlords, “AI” is absolutely meaningless if the planet is uninhabitable for humans.
I think he enjoys being powerful like any human but has a genuine care for how AI could improve humanity and the world for all humans. He does see what OpenAI is doing as a "Manhattan Project", in which the technology has a tremendous amount of inherent risks, but it's better they proceed with the greatest caution and make steps towards international cooperation so that AGI doesn't emerge in the wrong hands. Since the tech is inevitable, it is best it doesn't deploy itself into the world without great disasters occurring on its release.
As for the 7 trillion dollar thing, I don't think he's serious about it, as in he doesn't expect to get the money in the same way he'd expect to get 10-20 billion for OpenAI if fundraising. I think he's using the 7T figure as somewhat of a marketing tactic to elucidate the world on the compute build up that will be necessary. He probably believes that in a few years, 7 trillion will be spent to get to the gigaflops necessary for ASI. He probably feels its best to deploy the crazy number now so people are generally aware of the scale he anticipates for when AI really gets going in 1-2 years.
>Can you please not post like this?
OK, I wont
At some point it might be cheaper just to let humans continue to do whatever work we imagine GPT-17 will automate.
Even if he wanted headlines with that gaudy figure, he knows the VC industry inside and out and knows the fund raising limits.
Maybe I'm missing some play on words, but that should be Altman, no?
That's not the chilling thing. The chilling thing is there are people (mainly sci-fi addled software engineers) who would consciously choose to destroy humanity in order to create an AGI. It's like they read about paperclip maximizers, and decided to become one for <insert-scifi-concept-here-usually-agi>.
An aircraft carrier costs $13B. For another $10B you can put 100x F35s on it.
The market cap of the entire Magnificent Seven is $12T.
I think a bigger problem with this is more simple: there's a pretty critical assumption that the GPT-6 and GPT-7 models will need orders of magnitude more compute.
On the one hand, that was true for GPT2 > 3 > 4. On the other hand, history is rife with fallacious predictions about technology development and resource usage, all the way back to Malthus. Typically, as the human race focuses on a problem, whether that be farm efficiency or AI or anything else, people invent new approaches that invalidate simplistic forecasting about resource usage.
In this context specifically, I think there's a lot of promising research with different types of model architectures and multimodal models, to the extent that AGI may not be a path paved through brute forcing transformer training for LLMs. In fact, I believe Sam Altman said this himself last year!!!
Also, from a pure business perspective: it's one thing to spend $7T and build a bunch of factories. It's quite another thing to beat Jensen at that game, who has a legendary history of beating the odds when his company was headed for failure by constant process innovation and strategic moves -- this podcast episode saga is a great listen: https://www.acquired.fm/episodes/nvidia-the-gpu-company-1993...
I think Sam is thinking too much like a VC here -- raise as much as you can and take your % management fees -- and I would not be eager to invest so much into brute forcing a hardware business at scale. I'd much rather invest that into NVIDIA or AMD, who have the battle scars to know how to build hardware profitably.
CD ROM burners were amazing technology, and now they are super cheap.
Pentium II was super expensive too.
GPUs cards... GeForce GTX 470, 350 USD into 35 USD today, in 10 years.
There is no reason to believe that "AI accelerators" (which are +/- non-specialized GPUs for now) won't become cheap too.
It may even come faster than expected, like this year, Apple may already announce specialized circuitry.
And no, there is no need for 7T USD for that.
The foundries are the same, we haven't discovered integrated circuits, just some cool software that is not different than running Grand Theft Auto in "the cloud".
Source: https://www.nature.com/articles/s41587-023-02103-0
>A major distinguishing factor of the sciences (specifically, biology, chemistry and physics), as compared to AI fields such as natural language processing and computer vision, is the relative lack of publicly available data suitable for these domains (see Box 1). There are simply vastly fewer data existing in the sciences, and these are often siloed by academics and companies. Acquiring appropriate scientific data for AI typically requires not only highly trained humans, but also high-end facilities with expensive equipment, making for an overall costly and slow endeavor compared to humans simply going about their day by adding to the vast trove of images, text, audio and video on the World Wide Web. As one example, it has been estimated that “The replacement cost of the entire PDB [Protein Data Bank] archive is conservatively estimated at ∼US$20 billion”1 —these are the data used by DeepMind to train AlphaFold2.
I'd rewrite this as "...they are genuinely interested in safety - but only insofar as it doesn't get in the way."