| Felix Stalder via nettime-l on Wed, 21 Oct 2026 13:25:17 +0200 (CEST) |
[Date Prev] [Date Next] [Thread Prev] [Thread Next] [Date Index] [Thread Index]
| Re: <nettime> The case for the Neo-Luddites - Rage against the machine learning |
hi Mond,like everyone else, I have no idea where recursive self-improvement might lead. Technical improvements have been impressive lately, truly but for now, it's entirely unclear how they fit into the real world.
Under current conditions, its foundations are really unstable.For one, the economics are not there. The faith in scaling laws (more data and more compute = better models) has been leading to a massive build-out of data-centers that is largely debt-driven. There has been a real speculative mania, less driven by the prospects of AGI than by the prospects of monopolistic windfall profits. For now, they seem like a very distant promise (for many reasons, including Chinese open source models and companies becoming more cost/benefit aware in relation to AI). At the moment, it's not clear how long that promise can be upheld. We already see cracks appearing. Interest rates and default insurances for bonds are going up, certain hyper-scaler bonds, for example, Orcale's, are traded below nominal value. I
Second, chasing scaling law visions has lead to larger and larger models. The current version of ChatGPT (v6) requires 400 times(!) more computation in training than the first compelling version (v3.5) that came out some four years ago. This puts enormous strain on the infrastructure, not just in terms of financing, but also in terms of energy, water, and hardware.
Third, this strain triggers new coalitions, perhaps even social movements, which quite effectively challenge this unchecked expansion. Here, the cloud becomes material, local and immediate, inverting the usual dynamics of digital politics. Rather than trading in vague, future harms against immediate benefits, it's the opposite: immediate harms are justified with vague, future benefits. Just if local governments, to the degree that they are willing and capable, start to enforce banal things like zoning laws and environmental impact assessments more actively, it will put a break on speed and scale this build-out, which, in return, increases the concerns of how the debt will be paid back.
This leads me to wonder if we should not think of the hyper-scale build-out and the AI we see relying on it, more like nuclear power at the beginning of its commercial phase. Huge promises, but the economics and the physics turned out to be hard to handle, and social pressure forced many governments to abandon the development altogether. Not all, of course, but now nuclear power is part of the energy mix, and one with increasingly unappealing economics. Sure, there is talk about the renaissance of nuclear power, but it will not arrive in time (if it arrives at all) to solve current problems, certainly not for AI data-centers.
None of this is to indicate that AI will disappear, or that it's just a stochastic parrot. It is not; there are many good use cases, from medicine to material science to conversational interfaces to complex data structures.
But it might be simply a normal technology, one that needs to work under normal economic constraints and (eventually) within biophysical boundaries. And like the resistance against nuclear power was never a resistance against electricity, but against a specific technopolitical configuration of its provision, the same perspective should be taken towards AI. I agree with you; to be against AI is not a position, but to challenge the world-destroying configuration is a future-oriented one. If solar power was the alternative to nuclear power, what is the alternative to the hyper-scale data center?
all the best. Felix On 20/10/2026 22:48, franz schaefer via nettime-l wrote:
Dear Nettimers,
Now most of the critics of AI and data centers are assuming that basically
AI does not work, or at least have assumed this for a long time. They say
it is not »real intelligence« (whatever »real« means), that it is just a
»stochastic parrot«, and that it can never become more intelligent than
humans.
Some even hold this position in the light of the recent development.
Now it is good to see that finally people are waking up. We can only see
the real dangers if we assume that it does work. That we can reach AGI or,
depending on the definition, have already reached it.
Bernie Sanders seems to get it. Rutger Bregman in his essay equates the
AI-deniers with climate deniers. I often see some parallels with Covid
denial.
https://rutgerbregman.substack.com/p/an-inconvenient-truth-about-ai
Anyways: So (whether you believe in it or not) let us assume that it does work as advertised or will do so in the near (a few months) future.
Then:
1. RSI (recursive self-improvement) is no longer sci-fi.
========================================================
The chances that the already scary powerful AI becomes capable way beyond
our imagination is a very real possibility. Projecting the current
development (doubling of capabilities every 3 months) into the future, it is
not only a real possibility, it is the most likely possibility.
This means we must slow down the development as much as possible. Anything
that slows it down is needed. The data center moratorium is good and
necessary, but in light of this development is not sufficient. Keep in mind
the current AI models have been trained on current data center
infrastructure. Also keep in mind that most of the improvements are because
of algorithmic improvement (the exponent in the exponential growth comes
from ca 40-45% algorithmic improvements, 40-45% recent data centers and
hardware purchases, and only 15% from Moore's Law).
So we definitely need much more than the data center moratorium. Directly
controlling the AI labs, putting them under public democratic control,
halting all AI training, and putting almost all research capacity into
safety and alignment research.
I know this is all not easy, especially since Trump in the US. But then
again: After the midterms maybe there is a chance...
And then there are the »but China« voices. At least China is releasing
their models as open source/open weight. And they do not have an interest
in the apocalypse either.
What really gives some hope here: The AI summit organized by Tristan Harris
https://centerforhumanetechnology.substack.com/p/an-historic-ai-summit-in-iceland
2. There is no bubble that bursts. At least not on its own.
===========================================================
The global GDP is $128T. The investments in AI this year are about $2.6T.
So only about 2%. If AI works and can replace almost all human labor, then
the investments actually look small compared to the "grand prize" in this
game. Now you will hardly find a bigger advocate of UBI than me. But the
reference to UBI is used by the tech-oligarchs to wave away any concerns
about job loss. The issue is: The rich and powerful will not give it to us
just like that. As with any social progress, we need to fight for it. And,
as Yuval Noah Harari has pointed out: Once our labour power is no longer
needed, we lose our most important weapon in this fight: The possibility of
going on strike - or at least threatening to go on strike.
So when we want UBI (and we absolutely need it), then we need to fight for
it NOW. Also see https://basic-income.eu - our ECI starting on the 4th of January
2027.
But if we manage to really slow down the AI development, bring it under
democratic control, and tax the energy consumption of servers and data
centers into oblivion, then we might create a situation where the investors
see that their investments will not bring the return. Or, as Felix has
pointed out: Not fast enough. Then there might be a bursting bubble, and
this would again help to slow things down.
So yes: All the power to the Neo-Luddites: Rage against the machine
learning!
But as the dot-com bubble has not stopped the Internet, neither would an AI bubble stop AI.
3. In a best-case scenario where we could stop the rapid progress:
==================================================================
We would still have really powerful AI right now, and society could decide
on where to best use it. Medicine, material science, ... and of course
robotics that does the tedious work. With the rapid, exponential growth
that we had on AI during the last months: Most of the capabilities of
current systems (running on current data centers and not needing any more)
have not yet been used at all. Most companies are just in the early stages
of experimenting with agents. I actually want the »Fully Automated Luxury
Gay Space Communism«, and I want it now. And we can have it now. No need
for the rapid progress and no need for more data centers. Just reaping the
fruits of what has already been achieved. At least for the coming years.
Also plenty of work for AI researchers: Just divert everything into safety
and alignment.
I think it is really necessary to also point out that we can have the
upsides of AI and that they can be really good. Otherwise people fall into
despair. We need the carrot. But it must be made clear that we can only
have the carrot if we do not fall over the cliff. Thus we need to violently
oppose the current trajectory.
Franz Schäfer (Mond), October 2026.
Also on: https://qummunismus.at/a/article205/
-- | |||||||||||||||| http://felix.openflows.com | | |||||||||| https://tldr.nettime.org/@festal | | for secure communication, please use signal | -- # distributed via <nettime>: no commercial use without permission # <nettime> is a moderated mailing list for net criticism, # collaborative text filtering and cultural politics of the nets # more info: https://www.nettime.org # contact: nettime-l-owner@lists.nettime.org