Whatever bullshit Sam and especially Dario are spewing out at the UNGA, it is still a bullshit.

What they actually have is a particular set of running algorithms, which, while mimicking just one aspect of how the brain works (in principle, not in the details), is, obviously, not enough to materialize the promises they gave away.

First. they bet all these borrowed trillions on the assumption that “by scaling it” (more compute, more storage) and by wiring the slop generator blocks in series, similar to what complex EE systems are being composed, they will achieve so called AGI.

They won’t. AGI requires “more structure” and different kind of processes than statistical inference, in principle, and the probabilistic interpretation is just a grossly over-simplified abstract mathematical model, simple-enough to fit into a limited human mind.

So, what they did “right”.

  1. Constructing and maintaining an inner (within the brain) representation of Reality from repeated experiences (and thus being conditioning by the observable aspects of Reality) and using that inner representation for actions and decision making, is a Biologically Universal “pattern”. But the representation is not probabilistic, it is evolved and then “tuned” when the brain develops.

  2. Each “significant” experience (“marked” by stress hormones) updates the inner representation. This is a gross approximation of the “neuro-plasticity” meme (a poorly understood set of biological mechanism which allow “learning by doing” and “learning from experience” at the level of brain structures), so not every experience is rewiring the brain, only some of them.

  3. The fundamental “Garbage In – Garbage Out” problem has no viable solution.

This is a subtle problem. Prior to evolved human languages, almost every experience which an animal brain receives or observers were highly consistent with the constraints of the shared physical environment, in which it happen to evolve – no bullshit in principle, only “facts” or observable “events” in the environment.

It is a language which ultimately introduced the brain to “bullshit”, literally. This is exactly how any religious cults and other organized socially-constructed elisions or “doctrines” has been built and operate.

The naive assumption was that “the right answers” would eventually outweight all the socially-constructed abstract bullshit and the model (the inner representation) would ultimately converge to What Is (which, by the way, is just what Buddhism proposes) and become an AGI . But it won’t.

The fact that fucking Palantir “AI system” guided the missiles onto a girl’s school in Iran is a proof by example.

According to the very same universal principles, feeding the generated slop back into the slop generator as “new experiences” to update the inner representation (rewire the brain) will ultimately converge on some socially constructed bullshit (a current “consensus” or just a set of current memes), not on the underlying reality.

Like I have said million times – if you train a model on a corpus of Indian Tantric text it will tell you to drink the blood of a freshly sacrificed animal, and you will agree that this is the very best thing to do.

In general, the set of algorithms which build a “model” are able to create some non-deterministic version of a mode;’s “inner representation”, from recurring patterns (which has to be actually present and accurately captured in the training data) and then an inference algorithm could produce similarly looking patterns just exactly as a parrot does producing the sounds that look like human speech (only to humans).

This is the “Parrot Law”. Every observable behavior which comes from an LLM is just a cognitive illusion, an imitation, a highly sophisticated deception, because, just like a biological parrot, the implementation does not posses, have not evolved yet, any “reasoning capabilities” and related brain structures (centers).

Again, this is just a law, not a mere opinion.

So, everything else about AI is just bullshit. A running system can generate text or images, and it is the human problem how do they interpret these generated sequences.

If there are idiots which just issue missile launching commands on a generated coordinates, then, well, they are the idiots and must be held responsible. Ai itself is just a stochastic parrot.

The “global AI regulation’s”, if at all, must be only about not allowing idiots to interpret the AI generated slop literally, especially in the decision making which may have severe consequences..

But this won’t happen, of course. The idiots will do something outrageously stupid (try to model a virus, make an automated surveillance/“law-enforcing” punishment system, or just paste the slop code to some critical computer systems).

There is nothing we can do. The UN is just a circus.