Hacker News Reader

Comment by rich_sasha | Hacker News Reader

The weird thing that makes it different to subprime is that AI is real and really could become AGI. But it might also be overvalued.

So it's not like there isn't a product (there is) and there's no growth prospect (there is). But it is scary how much now hangs in the balance of one bet.

It almost feels like it's going to end badly either way. If the Great AI Bet succeeds, a tiny proportion of the world will own all intellectual power. But if it fails, the impact of the write-offs on the broader economy will be terrifying.

View on HN

> AI is real and really could become AGI

Every single researcher not being paid $$$ by AI companies say there is 0 path between LLMs and AGI, but sure... and the next pfizer drug might make us immortal, who knows, everything is possible after all

I think this is a disingenuous take.

Will AI of today definitely become AGI of tomorrow? No, for sure not, and anyone who claims this is at best crazy.

But is it imaginable? I think totally. Andrej Karpathy' blog post about RNN writing Shakespeare 1 character at a time was 10 years ago. GPT-2 was released 6 years ago. In that time we went from something that barely speaks English, never mind any other language, to something that, on a good run, is an excellent language tutor (natural and programming), can play games, can write moderately complex programs, goodness knows what else. For some people, the romance of a ChatGPT-4 was unmatched.

Even if it doesn't become "AGI", it might just get so good at being sub-AGI that the question is irrelevant. We're seriously contemplating a near future where junior devs are replaced by LLMs; and I write this as an AI sceptic who uses LLMs to write a lot of the kind of dumb code a junior dev might do instead.

I don't like AI, in that it nibbles away at my competitive advantage in life. But it's IMO crazy to pretend it is not even potentially a game changer.

The next logical step for cars were flying cars, the next logical step for planes was space travel. Both were hyped not so long ago, people believed in both, it never happened and it isn't closer to us now than 50 years ago

I'm not saying it doesn't bring any value, I'm just saying that if you think we should give $7 gazillion to Altman because he's building skynet by 2030 you're smoking crack

That's a pretty rude response to a considered reply.

I am not saying we give all money to Altman. I'm saying AI is likely overvalued. But can it evolve into something far more capable given investment? Yes, it may.

Flying cars and space travel didn't happen, but might have. And funnily enough, they might still happen, with drone taxis and Virgin Galactic / SpaceX. Might. That's how it works with speculative investment.

It'll happen for the top 0.1% while the plebs fight about these topics on online forums yes

I dont think we'll have AGI in 2 weeks like altman does, but its not fair to say its impossible.

LLMs also came out of nowhere, a series of discrete improvements that finally got over the hurdle and acheived an unprecedented functionality. Absolutely no one predicted its emergent capabilities back in 2016 when the transformer paper was proposed.

They didnt see scaling then, they might not see the next thing now, until its found.

So it wont be in the next 2 weeks, and it wont be from OpenAI, but it might be in 10 years from some random researcher at waterloo, or tiktok

> its not fair to say its impossible

Nothing is impossible!

> LLMs also came out of nowhere

No they didn't.

> Every single researcher not being paid $$$ by AI companies say there is 0 path between LLMs and AGI

So maybe it's jealousy?

I mean I don't know.

Ah, the same reason climate scientists not paid by oil companies agreed on global climate change, and doctors not paid by cigarette companies warned about lung cancer?

I meant it as a joke. Should've added an /s instead of "I don't know".

But in all seriousness until someone else invents AGI some other way, you can't disprove that LLMs aren't the way. My intuition says you need more than LLMs but I could very well be wrong.

> AI is real and really could become AGI

There is zero evidence that current "AI" (LLMs) are ever going to become "AGI".

If you want a pathological liar for an "AGI", then sure, LLMs are already there.

What would count as proof of AI trajectory to AGI?

Currently we have: - cross domain competence - composable tool usage - constant improvement without clear signs of stagnation...some can count gpt 5 as not much progress but there are still world models with huge gains compared to year before

There is a bunch of things missing for sure like mechanistic reasonic, better context lengths, determinism, better world modelling, continuous learning.

The bet is crazy and whole world is gambling on AI, right now, because of those signs of potential. That's precisely why we landed here, the evidence was good enough for big tech to gamble on it...

>What would count as proof of AI trajectory to AGI?

It depends on who you ask. There is no absolute for definition of "AGI". Sam Altman defines it in monetary terms, because that benefits him. I have a very different idea of what "AGI" means. It's really very subjective, so I don't have a definite answer for you. I'm sure you'd define "AGI" differently than I would, so having this discussion is kind of a waste of my time.

The peo1ple pouring money into "AI" are doing so in the hope that it will become more reliable someday. My educated guess is that it won't, due to the underlying mechanisms that it is built on. Predicting the next word in a sentence according to grammatical rules is a long, long way off from a machine knowing and understanding how truthful the resulting sentence is.

I would say that's rational take, it's a long shot and probably investments of such caliber should have stronger proofs. I believe some companies are investing so much, to not be left behind if there is an even remote chance of success...Apple was one of the top companies not following and there have been huge criticism, that they are left behind

Also I think you will like this interview with founder of Cohere AI, it's much more nuanced and doesn't say AGI is near...more like we are far away, although it's useful. https://m.youtube.com/watch?v=Sw2chzwWLbQ

The goal posts will constantly shift just like Self driving. AI used to really mean AGI but that doesn’t get investments.

AI always meant AGI.

AI also never meant AGI.

Artificial Intelligence is an entire discipline of the Computer Science field. It encompasses everything from how Pinky, Blinky and Clyde chase Pac-Man, to A* search and similar pure algorithms, to machine learning, computer vision, and LLMs.

It is also a term used widely in popular culture and media to mean, essentially, AGIs—Cortana, Agent Smith, C-3PO.

The problem is not that this term is very broad, and it certainly isn't that it has come to mean something that it didn't before. It is that a bunch of people with a financial interest have been busily trying to convince the world that LLMs are Cortana.

Real estate is real too. People need houses to live. Websites were also real value

The dotcom bub, subprime crisis, and AI bubble are all based on real goods being overvalued to hell and then crashing back down to their actual capabilities. The soze of the bubble is the delta between their actual value and the market's percieved value.

In that regard i think AI will be crazy crazy valuable in real terms but not like the market is using it. I think the real value of AI agents is very very low and the market will crash to that level. I think the value of genAI as much simpler interfaces for communication (RAG, translation, NLI) and for automated understanding of static sytems (rather than acting in dynamical systems) is high and will crash a bit before the market learns to use them right and then itll be a party

As long as people keep spending all their money on LLMs, we are safe from an AGI.

I describe it as both a bubble and the next industrial revolution. It's not so much inflated as it is far ahead of itself IMO. Kind of like NVDA stock over the past nearly 3 decades.

The talent out there is all focused on a tiny number of tasks and benchmarks in service of the AGI cult whilst the real gain is scattered amongst everything else with no staffing to build it. So I see a bust much like the dotcom bubble in the arbitrary future, but then it corrects surprisingly quickly shortly thereafter and the engineers of that bubble will have gotten out temporarily well ahead of that bust, buying back their assets at 50% or greater discounts and the broader base of retail investors are screwed as usual.

As for impacts on the economy. Meh. They're already starting to ignore the ramblings of the Orange in Chief (the guy shuts down the federal government and the entire stock market yawned last week). The weakly efficient market abides and it's been burned enough times already by the TACO trade. It'll get through this. But oh the whining that will ensue on the corporate media.

You can’t buy stock in AI. You can buy stocks in companies.

The internet was destined to be big sure during the dot.com but most companies crashed.

The bubble popping issue would be that there isn’t a good way to recover the capital used to build the AI models.

You can't(couldn't) buy stock in the internet, either. It's always "stocks in companies".

I mean, unless you go back to "tulips are a sure thing; their prices always go up!".

But buying stocks in an S&P massively invested in AI is essentially the same as "buying AI", when the bubble pops.

It might just be very similar to the dotcom bubble.

The trillion dollar question isn’t “is AI a bubble”, it’s “which of these companies are pets.com and which are Google (if any)”.

Even investing into a basket might not be a winning strategy even in a world in which AGI is imminent.

It might be. By feel, it is different in that a lot of the AI investment is into real things, that hold non-speculative value: data centres, GPUs, training marketable and genuinely useful products. It's not food for pets but online and sent by email.

But otherwise agreed.

But that’s my point: The dot com bubble burst, yet the Internet still changed everybody’s life.

Reply on news.ycombinator.com