Re: <nettime> The Case for Gouging Out The Eye of The Cyclopes !

[Carsten Agger via nettime-l <nettime-l@lists.nettime.org>]

Dear all

I was not going to weigh in on this discussion, yet I do feel a need to
unmuddy the waters a bit.


man, 26 10 2026 kl. 21:44 +0100, skrev franz schaefer via nettime-l:
> 
> Now she does not draw the conclusion that the warnings are only
> marketing. 
> But this is one of the lines of reasoning that has supported the AI-
> Denial
> now for years.  And thus it is one of the reasons why the left is so
> late
> with this topic.  And moreover: I still meet people who cling to this
> straw.
> 
> 

The critics you might call "AI deniers" would be people like Timnit
Gebru, Emily Bender, Alex Hanna and other critics of the current "AI"
based on large language models.

I'd like to be absolutely clear:

When these people criticize the current hype around "AI", it is from an
understanding of what the technology is and what it does.

When politicians say that it heralds a "new age", that it can do
practically all our jobs, that it can save enormous amounts of money in
education, health care, social services and a very large etcetera, it's
because they really have no clue what the technology is and what it
does. "AI", or rather LLMs, cannot do anything of the sort.

When AI bosses and workers say that it can, when Dario Amodei says that
AI will solve climate change, cure all cancer and double the human
lifespan, when Sam Altman says it might eradicate all jobs, they are a)
lying through their teeth, b) engaging in wishful thinking or c) both.

In that sense, AI cannot pose an existential risk, because it cannot
have any agency or intention - hence, it cannot suddenly "go rogue" and
attack humanity.

Why is that? Well, because all current AI is based on LLMs, and LLMs
are basically word-guessing machines, a glorified autocomplete.

What can an LLM like GPT 5 *really* do?

Well, if you feed it a sentence, it can generate a statistically
plausible suggestion for the next word. In reality, based on its
training material and its actual training, it calculates a probability
distribution over all words in the language (OK - tokens, really, but
that's not important in this context) and then randomly chooses a word
using that distribution.

You can then append that word to the input sentence and feed to the LLM
again and get a new word, and rinse and repeat (yes, this is a very
inefficient process).

Since these LLMs are trained on practically all texts in the whole
world, the text they can generate in this way can be very convincing.

However, it has no relation to any kind of "truth" or to any
understanding of the world whatsoever - it's pure statistical
inference.

This means that such an LLM can have no intention of eradicating
humanity, because it can't have any "intention" at all. Also, it
doesn't understand (by definition, it has no "understanding" of
anything, it's basically just a huge abacus used to output words) the
limitations we are under, it has no notion of "being in the world" in
the Heideggerian sense, and hence it simply doesn't understand anything
about how the world works.

So yes, this technology works. It can output very convincing text;
sloppy programming code, mediocre essays, bad poetry, technical texts
which look so convincing that you have to read them very carefully to
spot the glaring errors.

That they can do that is definitely something. Some peopl us it to
pretend they work or to output sloppy and unconvincing images and
presentations, but they are not and will never be worth "all of the
global economy".

Note that I'm not saying that no AI technology could ever be developed
that *could* have real understanding of the world and hence *could* be
very valuable in that sense (though "all of the global economy" might
be pushing it). I'm just saying that LLMs can't because of their
inherent limitations.

Yes, but what about Jamie Coxon, Eliezer Yudkowsky etc. who are warning
against the AI takeover. Well,these are not disinterested parties. They
belong to the very strange Silicon Valley cult of transhumanists, also
known as the Effective Altruists. Temnit Gebru and Emile P. Torres
wrote an excellenet article about that in the magazine First Monday,
and Émile Torres explains this ideology with Cal Newport in this video:

https://www.youtube.com/watch?v=jt508VjX2H8

Cal Newport is someone you'd want to listen to if you really want to
understand what the technology can and can't do.

All of that doesn't mean that "AI" can't do us real harm.

The oversize abacus may not be able to do our jobs, but it *can* be
used as a pretext to fire us anyway. It can be used as a responsibility
sink: Have it generate targets for bombing in a war, and when you
murder innocent civilians as a consequence, say "the AI did it". AI
datacenters are gobbling up an enormous amount of money, blasting away
all climate targets and hugely accelerating global warming.

Also, guess what? The accelerationists behind transhumanism and
Effective Altruism don't care if they destroy the Earth, because all
they care about is humanity's alleged future among the stars, in space.

So, let's not fear AI, let's fear the people driving this combined scam
and power grab. The fear of "rogue AI" is a red herring that the big AI
labs themselves promote all the time.

I admit to being an AI sceptic, but if you really want to label me an
"AI denier" because I understand how the technology works (my
background in computer science is helpful in that regard) I don't know
what to say to you.

Best,
Carsten

-- 
Carsten Agger - agger@fsfe.org
    https://fsfe.org --- https://blogs.fsfe.org/agger/
FSFE Denmark Coordinator, General Assembly & European Team Member
Free Software, Free Society!
-- 
# 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