Grok 3: Another win for the bitter lesson

Are LLMs really going to own the world?

It looks like neural network based software is to surpass humans in intelligence in every task in the forseeable future.

If one country moves along this direction faster than the others, no country will stand a chance to compete with them militarily or economically.

LLMs aren't the only kind of AI.

Having hardware and software suppliers all together makes it more likely even if you assume (like I do) that we're at least one paradigm shift away from the right architecture, despite how impressively general Transformers have been.

But software is easy to exfiltrate, so I think anyone with hardware alone can catch up extremely fast.

Intelligence is everything. These things are intelligent- already superhuman in speed and a few limited domains, soon they're going to exceed humans in almost every respect. The advantage they give to the country that owns them is nuclear-weapons like.

> no country will stand a chance to compete with them militarily or economically.

It really depends on how they go about it. It can easily instead end up with lots of people without work, no social security and disillusioned with the country. Instead of being economically great, the country may end up fighting uprisings and sabotage.

How so? First of all, assuming ASI is developed, as it stands now, it will be owned by a private corporation, not a nation state.

ASI also will not be magic. Like what exactly would it be doing that enables the country to subject the others? Develop new weapons? We already have the capability to destroy earth. Actually come to think of it, if ASI is an existential threat to other nations, maybe the rational action would be to nuke whichever country develops it first. To safe the world.

You see what I am saying? There is such a thing as the real world with real constraints.

This is just flat out not true. They're not intelligent and not capable of becoming so. They aren't reliable, by design.

They're a wildly overhyped solution in search of a problem.

I don't understand this attitude and I am not sure where it comes from- either from generic skepticism, or from some sort of psychological refusal.(*) It's just obvious to me that you're completely wrong and you'll have a hard wake up, eventually.

* "I know how this works and it's just numbers all the way down" is not an argument of any validity, just to be clear- everything eventually is just physics, blind mechanics.

If this is true, then shouldn't we expect an economic "bump" from NN/LLMs/AI as they are today?

I have not noticed companies or colleagues 10x'ing (hell, or even 1.5x'ing) their productivity from these tools. What am I missing?

My non-tech company already uses LLMs where we used to contract software people (for 2 years now - no unresolveable issues). I myself also used LLMs to write an app which is used by people on the production floor now (I'm not a programmer and definitely don't know kotlin).

Maybe LLMs can't work on huge code bases yet, but for writing bespoke software for individuals who need a computer to do xyz but can't speak the language, it already is working wonders.

Being dismissive of LLMs while sitting above their current scope of capabilities gives strong Microsoft 2007 vibes; "The iPhone is a laughable device that presents no threat to windows mobile".

What do your colleagues do?

I see people getting replaced by AI left and right.

Translators, illustrators, voice over artists, data researchers, photographers, models, writers, personal assistants, drivers, programmers ...

"The advantage they give to the country that owns them is nuclear-weapons like."

I think the idea that the United States "owns" Grok 3 would be news to Musk and the idea it "owns" ChatGPT would be news to Altman.

> Maybe LLMs can't work on huge code bases yet

its also not just about code base size, but also about your expectation of output quality/correctness.

There's an implicit assumption here that if a colleague did figure out how to (e.g.) 10x their output with new tools, the employer would capture all (e.g.) 10x of that increased productivity.

Check out operations research.

The amount of “work” done there is staggering and yet adoption appears abysmal, using such solutions with success only happens as part of a really “well oiled” machine.

And what about the simple difficulty going from 99% to 99.9%. What percentage are we even talking about today? We don’t know, but very rich people think it is cool and blindly keep investing more billions.

You're entirely free to go on and on about how amazing the emperor's clothes are. Nobody can stop you. :)

It's fine to chase hype for hobbyist or starter projects, but part of being an engineer is understanding how things actually work and what their limitations are.

It's not a virtue to deify a statistics model and make it your entire personality.

If you're (1) doing something basic and (2) don't care about correctness or quality or reliability and (3) don't need to change or maintain it, then by all means, use it. It's literally no different than copying off StackOverflow (and probably being generated from it.)

If you aren't an engineer, I get why you think it's magic. Everything is magic when you don't understand how it works.

Nobody thought the iPhone was magic. It was an instant hit because the capabilities were immediate and obvious, and Apple had a long history of being able to execute.

If you find it useful, by all means, use it. But this is the new blockchain.