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.

Also corporations do not want to send their code to someone else’s “cloud” or even show their code to anyone. This is an old and well-known constraint, even if it is only partially rational. So, no, won’t happen. Unpopular tarifs won’t fly.

Most of the talk was that vague abstract half-truth (half-nonsense) and his usual “everyone look how smart and cool I am” (“I tweeted”, you know).

There, however, was one non-bullshit point - the arrival and the rise of actual tools which allow one instruct a “more-or-less-Ok” and “acceptable” code-generator in plain Engilish. This is something really new and unprecedented (the fact that it actually “works”).

Baiscally, the only valid real-world use case for the “glorified autocomplete” is to fill the gaps in over-“engineered” unbearable vebose OO apis in crappy languages and /to generate insanely-boilerplated stamdardized “componentes” (hello, Android and webshit) simply because they are almost the same and so the training data is vast and good,

This is what MS and others are actually already doing “inhouse” by training models on their own code. Google trained Gemini on their own code, and you can see it – their crappy, too low level Python style crops up all the time.

But thiere is more. This, essentially, kills these online rase-to-the-bottom $5/hour sweatshops, like Upwork (thank good I avoided it like a plague!) and even the low-effort crappy outsourcing in general. Good models, like Grok3 generate a better code than could be even found at Github (when prompted intellogently), leave alone coded by some third-world amateurs.

One more time, when you know what precicely to ask and why and what to expect, you could get almost all common boilerplate to be generated in second with the quality level above Github or Stackoverflow.

Instead of goig through all the SEO-infested crappy search results, through shitty blogs and low-effort half-assed repos, trying to distill an ideomatic, propetly structured well-typed code, one gets almost usanle intermediate building blocks done right in seconds.

Notice, however, that it will always give you code that has very subtle bugs – does not pass a typechecker, or even worse – passes and fails on some test cases due to the fact that it never build up a proper conceptual hierarchy in a bottom-up process, it just selects the most probable next token.

So, this is non-bullshit memeless “Software 3.0”. One simply generates better boilerplate instead of outsourcing it to god knows who, and cheaper. The turnaround is way faster, the quality is way better at the fitst iteration. Sub-par outsourcing is dead exept for kickbacks and other nice schemes (hello, Boeing) of course.

It won’t genetate complete working solutions, it won’t generate any really novel thing (which weren’t in the trainig data), it will fail miserably on a niche, uncommon languages and frameworks (Elisp, etc), but it will shine with Python (for we are literally drowing in a sub-par Python code).

And this is essentially the only meaningful part of the talk. The rest is the meme-guy talking trending cool memes. Strictly to the birds.