After the “Sentient Models” has been booed by very few remaining intelligent people they resorted to a the next best meme – their AI is improving itself (recursively, my ass).

https://www.anthropic.com/institute/recursive-self-improvement

Notice that this has almost nothing to do with the actual underlaying tech, except that the models are being feed with their own slop for the Reinforcement Learning and as new data, and has everything to do with the upcoming IPO – look, AI writes itself, what can possibly go wrong – buy our shit.

The IPO as an exit strategy is the only exit strategy for them – to dump all the liabilities onto “investors”, so an imminent failure (too much debt and exuberant spending) will become someone’s else problem.

Let’s see why this is just a marketing bullshit.

As I have told you million times already, if you train an AI model with the 11th century Indian Tantric texts, it will then tell you that the best possible thing you can do is to drink blood of a freshly sacrificed goat.

One may argue that with the code it is different. Hardly. An LLM, however large, is as clueless about the underlaying semantics of the code, except the next most probable token, as it is about “understanding” of everything else. It just do not posses any of such qualities, but generates an appearance as if it does (due to sheer computational bruteforce on data-center large data representation).

Here is how. When you ask for any kind of Rust code, 99% of the time, if you are not smart enough to explicitly constraint the prompt and explicitly opt out, you will get the fucking tokio crate as dependency. Of course you would say – it is the most used and the most popular crate! and it is async (Butthead, huh huh).

Well, this is according to the current version of the Tantric teachings In reality, however, since doing difficult things just right is, well, very difficult, they just fucked up the borrow checker – the only thing that makes Rust unique and worth using – by forcing the 'static lifetimes everywhere, essentially disabling all the necessary checks.

Are these 'static variables immutable? Fuck no. They must be according to the proper Rust’s semantics, but they are too dumb and cannot write semantics of Haskell in Rust syntax.

Anyway, the point is that everything is fucked, essentially back into the level of the usual imperative crap, and the stalls and race conditions in the tokio code are as common as they are in that fucking C++.

And then apply the “AI Recursively Improves Itself” meme. The crappy tokio-based code feed back into the training and reinforcement learning pipelines, so at the next “recursive call” it will spew out even more crappy tokio code for naive idiot’s projects, and so on.

By the way, they are using the Recursion metaphor (an being fucking liberal arts majors, doing it as wrong as humanties talking heads possibly can do).

What is the base-case of your recursion? For Math and Engineering grads, that would be a Fixed Point – a convergence on What Is – such accuracy of the Map that it captures the Territory so perfectly (while remaining a Map - a useful representation) that further updates will yield no measurable improvement.

The Idea of an Ultimate Fixed Point of all possible human knowledge (To See Things As They Really Are, which applies as well to math and non-theoretical phisics as to the original Buddhist doctrine) is, if you think for a few seconds, is the only possible fixed point and the only one worth persuing.

But, just like any math faculty freshman knows, instead of converging on a global optimum one may get stuck in a local optimum (the most popular memes and the prevailing socially constructed bullshit narratives, which is what any Tantric teachings are in principle), or, even diverge due to systematic overshooting, which creates a positive feedback loop in the opposite direction (from What Is).

Have you ever seen the crapto and the US tech stonks marker? Okay, nevermind.

By feeding its own slop one will never reach the base case in principle, even by a chance. Let this sink in. You can read this sentence a few more times. More precisely – the best possible expected outcome is to get stuck in some arbitrary “local optima” of some dogma, or just a “pocket” of some currently accepted socially constructred bullshit, be it theoretical phisics or any modern equivalent of the 11th century “knowledge”.

Have I already told you that every single brush stroke inside the Egyptian Pyramids was strictly according to the true, infallible knowledge? I think I did.

What the current LLMs do is amplifying and “setting in stone” the currently most popular socially constructed memes and narratives, including in the code. Just think of webshit or anything javascript related.

The fact that nowadays the slop compiles and runs does not imply that it is any good, just like if something has been writtern in a book it is any “true”.

There is, of course, a lot more subtleties, but the principles are as infallible as they always been. If you construct a flawed map and go into the territory you are fucked.