Hacker News Reader

Comment by thinkingtoilet | Hacker News Reader

Does anyone seriously think we're anywhere close to AGI with LLMs? I know CEOs like to say things to blow smoke up investors asses, but does anyone with actual credibility think that?

View on HN

Anecdata but I talk to lots and lots of AI/ML/DS engineers. Everyone knows the current LLM architecture won't work for AGI. All the "reasoning" models are just pseudoreasoning and there is severe data leakage and benchmaxxing when the companies tout the capabilities.

People really avoid considering what the word "general" implies. Yesterday I tried sending o3 a screenshot of some sheet music, asking for a midi file of how it sounds. Complete failure x3. Could not even get the value of the first note right. This is not "general" intelligence.

I'm general intelligence, and could not in a million years do that.

I think actually you could do that if you wanted to; look up what notes mean, write some little program to make a sound if you had to. You could do it in a week if it was your only job.

but OP didn't give the AI a week it gave it less than a minute

Don't sell yourself short. If you have eyes and a voice, you can learn to read sheet music and hum a tune.

you wouldn't need a million years to learn to read sheet music and write midi

If you can't learn to read sheet music in a million years then you are not generally intelligent.

Every time you deny yourself this kind of agency you are denying yourself your humanity.

These models are notably terrible at music in every dimension.

Music is essentially mathematical. Weakness in math is being addressed by dedicated capabilities that are triggered by mathematical language in prompts, but because these models are actually terrible at math there is no lateral transfer of skill to the domain of music. That's my theory anyway.

I believe Yann LeCun more than I do any CEO. And he has been pretty clear you’re not going to get AGI any time soon.

Previously discussed on HN:

https://news.ycombinator.com/item?id=39059897

Fifteen years ago I worked with a guy who, in retrospect, was very similar to an LLM. He was extremely verbally gifted and a vacuum cleaner for information. He could speak brilliantly about any topic he had been exposed to. He was a great person to send to a meeting, because he was great at answering questions coherently with the information he had on hand, and he always managed to make your ideas sound smarter than you could yourself. Based on that, you might think he sounds like a gifted human, until I tell you about his major weakness: if you asked him about something he didn't know about, he would often speak just as surely, fluently, and compellingly about it. He hallucinated just like an LLM, and that's why he was stuck in roles without a high level of responsibility despite his verbal gifts.

He was neither arrogant nor self-conscious. He treated his hallucinations as if they were the kinds of simple mistakes other people made, like, oops, I thought I understood this but I don't, no different from oops, I forgot my umbrella.

I sometimes wondered if he had a specific condition that made him the way he was, but I never doubted that he was human, with "general intelligence."

If you’ve spent any time around little kids, you’ve certainly seen that making shit up is a natural inclination of the human brain.

Ideally, as one’s intellect matures, one learns to stop doing that, and build coherent reasoning, only speak up when you know what you’re talking about.

Well, ideally. Many people never get to that stage.

I see a lot of replies suggesting agreement: LLMs are nowhere close to AGI.

I agree — it may well be a completely different path we need to go down to get to AGI ... not just throwing more resources at the path we've pioneered. As though a moon landing were going to follow Montgolfier's early balloon flights in "about five years".

At the same time, there is suddenly so much attention + money on AI that maybe someone will forge that new path?

"Money is All You Need".

Unfortunately the money is all chasing LLMs, not other AI approaches. Anyone with a different idea is frozen out, at least for a while. Whenever LLM disillusion finally sets in with the investor class, the question is whether other “AI” will be able to distinguish itself or if the money will just all dry up for another few decades.

No. People hype it but it's obvious we're hitting a wall with LLMs.

That being said, the "apps" that use LLMs coming out now are good. Not AGI good, but they do things, will be disruptive and have value.

And the money coming it could lead to new techniques and eventual AI. For now though, it looks like AI is transitioning into products and figuring out how to lower inference costs.

I do think LLMs will make incredible progress and we'll see lots of breakthroughs from it, but I agree it's nowhere close to AGI.

I'm not sure that matters though—if a technology can give humans what they want exactly when they want it, it doesn't matter if AGI, LLMs, humans, or some other technology is behind that.

Tyler Cowen claims it's already here. I don't agree with him but he has credibility with a lot of people.

depends on what we mean when define AGI

i think there's ample evidence to suggest that we're growing closer (3-5 year timeline?) to replacement-level knowledge workers in targeted fields with limited scope. i don't know that i would call that AGI? but i think it's fair to call it close.

thing is that has value, but compute ain't cheap and the value prop there is more of reducing payroll rather than necessarily scaling business ops. this move to me looks like a recognition that generalized AI on it's own isn't a force multiplier as long as you have bottlenecks that make it too pricey to scale activity by an order of magnitude or more.

I expect a fair number of non-technical LLM proponents, and probably some engineers as well, have likely built machines quite capable of helping them fool themselves that it is.

Nobody fully understands how human intelligence works. It's implausible we'll be able to replicate it or even come up with something better in the short term.

“If the brain were so simple we could understand it, we’d be so simple we couldn’t.”

Neuroscience is like a 150 years or so since the discovery of the neuron?

It'll take some time but we'll get there. Just not as soon as the AI hype will make you believe.

to answer that you have to define AGI first

That's not a prerequisite to refute a position, no.

The ball is in the other court - if one is working on AGI, it behooves one to know what one is aiming at (and I'd stake a fair wager that OpenAI et al have at this moment very little better picture of what AGI looks like than you or I)

defining what you mean is absolutely a prerequisite to someone else being able to figure out if you are right

To some extent, yes, but the onus here is on the party making wildly unverifiable claims, not the party refuting them

Reply on news.ycombinator.com