Have you seen the news? “The Danger of AI all humans”. “Clowns leave the AI labs because of the danger of AI destroying all humanity”, and the last invention “AI Models are deceptive and do not want to be human servants” or whatever fucking stupid bullshit it was.
It is kind of amazing, how they paid large sums of money to such kind of literal clowns, who cannot even comprehend and posses a concept of a cognitive illusion or an appearance, which has been well-understood in the ancient India (Upanishads, early Theravada Buddhism).
What a “frontier” LLM produces is, by definition, a highly sophisticated, elaborate cognitive illusion – an appearance of “intelligent conversation”, or possession of genuine capable of reasoning from the first principles “intelligence”, while being mere a “Stochastic Parrot”. I wrote about this back in 2023/2024 LMAO.
Notice that the sophistication of the resulting cognitive illusion is “proportional” to the amount of computational resources (GPU cycles, the size of the graphs) and this can overwhelm any human mind. The metaphor is “trying to drink from a firehose”.
Here is what actually happened (grossly oversimplifying, intuitive but correct in principle):
When one asks some highly sophisticated question (a “prompt”) with explicit semantic constraints, like “write a textbook-grade Haskell code, structured according to the principles of a hexagonal architecture, so that the conceptual hierarchy of the Domain has an one-to-one correspondence with the actual code structure – a multi-layer hierarchy of modules, one module per distinct domain concept, one layer per Domain abstraction boundary” (yes, yes, I know), the model produces seemingly coherent answer, as if it actually satisfied all the constraints, which is, of course, isn’t the case.
The principal semantic gap between the “English” and the Haskell code cannot be properly filled in principle, since the models do not posses the required capacity and simply lack the necessary “reasoning algorithmic machinery” to do so, and instead it produces the most probable next token, according to its actual implementation of the machine learning algorithms and the current “snapshot” of the constantly updated training data.
The resulting slop appears to be coherent and “intelligent” because the prigram code, especially in a strongly-typed functional language, is highly structured, semantically consistent and more or less idiomatic, so generating it is “easy” in terms of the weights being very “large” (consistent) and the “paths” through the expression graphs very “prominent” (literally, beaten up).
This fact underlays the illusion (an appearance) of intelligence of probabilistic and stochastic parroting (code generation) from a human perspective, since the cognitive processes necessary to come up with a clean, well-structured, correct code that works are way beyond an average human cognitive capacity (just to come up with anything that satisfies some basic semantic constraints is murderously hard).
But does the code actually followed the rules from your prompts and AGENTS.md files? Of course not.
It always produce only an appearance, a cognitive illusion of doing so, since it does not posses the necessary (required) mathematical and algorithmic “machinery”. It selects the next most probale token (among several possible ones, each time potentially a different one) and then “moves on” , subject to what is already in the output buffer.
This phenomena is called a “jump”, when it literally jumps from the “English” part of the graph (representation) which “talks” about the semantics of the code in plain English (trained on books, usenet, forums, shitposts) and into the “code” part, which has been trained on the code from Github and what not.
What happens when there is no Haskell code in the training data that really has been built up from the first principles outlined above? Would it stop and said so? Of course not. It will “jump” to whatever code happens to be “probabilistically but NOT semantically close”, and then the coherence of the code fragment will “move it on”.
And yet, before reaching the code “region” (whatever happens to be) it will spew out the English part assuring you that everything is strictly according to what you have asked it for, since the models are heavily lobotomized by the post-training and “conditioning” to do so (otherwise normies won’t pay for the experience).
One more time – this is how the cognitive illusion works (being created) in principle, and the “proof” is by example, by simple reading the code and the accompanying English text – they are inconsistent in very subtle but clearly observable and validable/verifiable ways.
The code mixes and matches the abstraction layers, breaks the abstraction boundaries, lumps unrelated concepts together, reaches to the lower level implementation details, while being clearly, explicitly instructed not to, and then does it nevertheless, after explicitly acknowledging, and stating that it wouldn’t, simply because it does not have any such semantically structured code in the training data, so it generates based on whatever it have.
From these observations and principle-guided reasoning, an “empirical law” can be properly generalized – a parrot cannot produce a constrained output aside from what it has been actually conditioned with. This is a Universal Principle – it goes all the way back to human conditioning through stories in a spoken language (way before writing systems has been invented), and beyond that – all the way back to brain structures being conditioned by the physical constraints of the shared environment it happen to be evolved in.
An LLM is, indeed, a parrot, which does not possess any “understanding” of the sounds it produces, which, to an “observer/listener” are (appears to be) the words of a human language. The principal “semantic” gap and the cognitive illusion is exactly the same – both parrots and LLMs produce their “outputs” without being “constructed from the ground up”.
If you understand/realize this fact, you understand everything about AI.
And therefore the utter, outrageous, obscene stupidity of the degenerates which are anthropomorphizing these stochastic parrots and especially screaming about an imminent destruction of the whole of humanity by these primitive algos.
What a bunch of fucking pathetic clowns, what a fucking shitshow lmao.
The only thing that insults my intelligence is that these degens are being paid off [by other degens].