LLM

So I watched some. The launch video I mean. Closed the frame when they began that “voice” thing. The PhD. level across all subjects is just a meme. I can easily be shown by asking the question that require a beyond memorizing textbooks reasoning. I don’t want to “register” for “Grok4”, but here are some example problems which will break the “PhD level” meme. A recursive functions in, say, Ocaml, or Scala, without explicitly mentioning the accumulator pattern as the required way to avoid stack overflows on languages which does not do TCO at the compile time. This is very basic stuff, which all “PhDs” have to know. The inner lambda with an extra argument, and the “trampoline” is such a classic pattern that some compliers do it automatically. Again, without mentioning it the model will fail by writing non TCO code . ...

There is a very simple trick to break the illusion and to see through “the veil of Maya”. Locally-run models, like Deepseek-R1-14b-0528 (at full fp16 quantization), which is the best I could have, produce vastly different “answers” for exactly the same prompt not just between two runs, but if one uses a different math library stack (like recompiling with forcing Intel MKL). Every time we run a prompt a reasonably good model spits out something which “looks very reasonable”, (unless your are an actual expert in the field), because it captures “common sense”, expressed in the training data. ...

Just a few years ago we had that running joke “pasting code from stackoverflow”, which describes a coder, who just find and copy-paste the code – the “right” answers – without any understanding whatsoever. It is well-understood at the level of cognitive neuroscience, that skipping the part of “doing it yourself” (and so never getting these necessary 10,000 hours of a deliberate practice) basically makes us dumber (lots of supporting MIT studies, google them) and ultimately wastes out time, the only irreplaceable and the most valuable resource we ever had. ...

Rust code is the ultimate evidence of the principal inability for a probability-based generating algorithm (based on sampling from a “learned” probability distribution over “tokens”) to come up with something that passes the type checker, but for the most trivial cases. The “causality” is that generation of the complex syntactic forms without the actual, proper understanding principles and heuristics is, well, “problematic”. The running example is these were subtle “already borrowed” panics, where the issue in with the underlying semantics, and the syntax is “correct”. The problem is that a recursive function has to drop all the borrows before a recursive call, and this constraint cannot be expressed in the syntax unless by redundant bindings which will automatically be drooped at the end of the scope. ...

https://www.youtube.com/watch?v=LCEmiRjPEtQ and, of course, the No.1 spot on the Chuddie safe space https://news.ycombinator.com/item?id=44314423. Karpathy is shilling “Cursor” and other cloud-based mettered AI services (which have to pay back their debts). Probably has an interest in it and some other meme AI startups. Nothing to see here. We should some day know which marketing “genious” came up with this “winning strategy” – to metter every single token (byte) and try to sell this to corporations. Corporations do not want to be mettered like that, they want to metter normies, the way Cellular operators do, and they never use any normies plans themselves. ...

There is something to be actually realized as soon as possible (which I recently experienced directly) –the condition-probability-based “glorified autocomplete” could generate some small pieces of code (with intelligent prompting) that simply above (in intrinsic quality) 90% of all the related crap that could be found on Github, which easily extrapolates to 90% of people in existence who identify themselves as “programmers”. Period. The code will be better than high-dopamine, “laser-focused” (on Adderall or just amphetamines), but not so well-educated on fundamentals zealots could produce. Better than inexperienced in realities of the actual open-source software ecosystems academics could even hope to write, no matter how good their theoretical understanding is. With very specific and properly constrained prompting, which takes into account the major results of the last 50-or-60 years of PL and CS research, especially in the field of math-based Functional Programming and Immutable Data Structures, one can get almost optimal (approaching perfection small pieces of code, provided you understand what theoretical-limited perfection actually looks like in this particular case, and why. ...

DESCRIPTION: Removing the veil of Maya to see things as they really are What to do when you have discovered that something is wrong with the world? Nothing, this happens all the time. Everything is wrong with C++, but everyone uses it, everything is wrong with packaged food, especially the toxic crap Nestle produced, and everyone is buying it. Nothing can be done. Here is what is wrong with your “AI” and “LLMs”. ...

Classy Andrej is making shilling videos from Thailand (he leaked his location in the video ) targeting normies (the previous set of videos has been partially filmed in Japan. Andrej is living a truly digital nomad’s life). https://www.youtube.com/watch?v=EWvNQjAaOHw Why would he shill? Well, he and guys like him made a lot of promises, not to us (who tf cares), but to the money guys, that this particular technology will completely transform the world, and that they are the very top guys in the field, so money shall be given to them (to the affiliated companies and entities). ...

DESCRIPTION: Idiots, idiots everywhere. Now I can accurately summarize what coding using LLMs actually /is in just a few sentences. Recall how people usually describe a code maintenance job: we have this code to run, while the original developers are gone and leave us no design documentation. This hypothetical situation is exactly what you get when an LLM finished spewing out the slop: you now have some code, very cheap, even for free, but it is not yours, the underlying understanding (of the whys) is not in your head, and the original developer is already gone. Disappeared. ...

Well, I’ve watched it. There are a few things to realize. The code it generated ran without an issue with the simulation task. It, however, is incomprehensible without understanding of all the details (like any other code). This is, probably, so because they have feed in a lot of very similar internal code in the training phase. The gibberish from the “thinking” phase might be helpful or it may be equally cryptic. ...