LLM

I think I have seen this before. Once in Varanasi, wandering around book stalls (most titles already “tourist books”, – oversimplified and westernized “tantric” bullshit), I found a whole book by some local publisher which describes in a minute details one single Brahmanic ritual (an elaborate sacrifice) which last almost a whole day. Hundreds of ingredients are being burn in a precise sequence, or rather a simphony of chants. motions, gestures (mudras) and many other elaborate details. The priests (brahmans) definitely knew what they are doing and why exactly this way is the only proper way. ...

The thing I hate the most is when some of these fucking YouTube content “creators”, which decide to monetize an AI coding clickbait with low-effort subpar videos, say “aaand boom!” when another chunk of a slop has been spewed out by an AI. This “boom!” is an insult to the last 60 years of the programming languages research (including the math-based theory) and to the “old sages” which crafted their languages and standard libraries in the best possible, “just right”, perfect, in the sense of “nothing more to take away” by surveying all the available literature, non-bullshit papers and spending months of anguish and self-doubt. ...

So, things begin to move a lot faster and much bigger, and there is something to realize about this unpreceded AI bubble. We will consider only the underlaying fundamental principles, not the particular implementation details, “architectures” and what not.. There are four major aspects to any LLM model – the training process, the “architecture” (the structural shape) of a model, the " post-training tuning" (lobotomy) of the model and the inference process. ...

Today I am sentimental, so lets reminisce a little about my first experience with LLMs. I found some early article about people using something called llama.cpp to run models locally on their machines. Some overconfident retard in another blogpost wrote that the “best model” and “by far” is Mistral “from Nvidia”, and it is supposed to be best because/ it is allegedly from Nvidia (they have some partnership, investment, I suppose). So I compiled the code (old habits) and downloaded the model from the hugginface. ...

Let’s summarize the current state of Large Language Models (LLMs) and so called “Artificial Intelligence” (AI) as of October 2025. They all are still just [estimated] probabilities of the next token, given the “context”. This implies no “knowledge” or “understanding” [of any kind] whatsoever. Nothing can be taken as “true” or even “correct” or “accurate”. All the talks about “knowledge in the weights” or “knowledge encoded within the network” is just bullshit. ...

Look, ma, a new episode just dropped! This one is full of shit to the brim. Even more so than prof. Ellen Langer who cannot stay within a context and claimed that 1+1 = 10 because in the binary notation it looks like 10 in decimal… anyway, whatever. https://www.youtube.com/watch?v=MlmFj1-mOtg No, the brain ain’t computing any hecking probabilities. It is not a Bayesian machine. It is not a prediction machine. It is not a simulator. It is not a statistical engine. ...

Time waits for no one, the race to the bottom accelerates faster than ever, and the “future” is now. Competition is severe and mostly meaningless, as in some third-world criminal infested ghetto. This is what LLMs turned our world into. So, lets “pee on” so called “prompt engineers”, in the 4chan parlance, of course.. Here is my benchmark prompt to evaluate performance of LLMs. All the “simple” offline models fail miserably, and only Grok and Gemini can produce something adequate. Claude is also good, but it is supposed to be the best, being trained especially for code generation. ...

FUCK THIS SHIT! No, really. It just hallucinated a non-existent package, with such a confidence lmao Yes, yes, I know, it cannot know anything about “existence”, but at least these dorks do not “deserve” the exuberant money they got paid. Uniform Code Block Rendering in Eww and Shr To get syntax highlighting in eww and shr, you need to use a package that intercepts <code> and <pre> tags and applies Emacs’s built-in font-locking. The eww-code-blocks-mode is a good choice for this. ...

Oh, look DeepMind and OpenAI win gold at ICPC (codeforces.com) So, a model memorized a lot more of a convoluted stuff and been able to spit out a coherent code. While this is absolutely amazing considering that it works at the level of syntax, and that its “representation” captures only “possible relations among tokens” and posses no “understanding”, thinking leave alone reasoning capacities whatsoever. It just picks up one of the few next most probable tokens, given the previous “context” (a sequence of tokens “so far”). ...

Have you ever seen the real complexity of the outside world, a glimpse of how overwhelmingly complex everything is? Let me show you something. Recently, I stumbled upon yet another post on /biz/ about some guy bragging about “finally made it” with a shitcoin. The attached screenshots show the chart of HiFi shitcoin (or whatever it was?) going “vertical” for a while. And, as usual, in retrospect everything was “simple” and “obvious” – the guy was “early” on a shitcoin after enough of trying. A classic story in some sense. ...