Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

The Intelligence Age

ia.samaltman.com

Post reply on HN
Print view View original

221–230 of 447 posts

Re: The Intelligence Age

#221

Post by jagrsw »

zero-sharp wrote: post #218
jagrsw wrote: post #197

Earlier quoted context omitted.

I'm not here to defend sama, but certain things cannot be proven until they arrive - they can only be extrapolated from existing observations and theoretical limits. Imagine the Uranium Committee of early 40's, where Szilard and others were babbling about 10kg of some magical metal exploding briefly with the power of a sun, with the best evidence being some funky trail in an alcohol vapor chamber. Maybe sama is right…

I'm sure you know that people in the AI community have been predicting big things ever since, I don't know, the 1970s? It's only 10 years away again. This time it's for real, right?

Alchemists predicted the transmutation of metals into gold for centuries, and on a sunny day in the 20th century, it arrived (a bit radioactive, but still).

Unless the human brain is made of some sacred substance, the worst-case scenario is that we will extrapolate current scanning methods into the future and run the scanned model in silica. I'm not recommending this "just for fun," but the laws of physics don't forbid it.

Re: The Intelligence Age

#222

Post by humansareok1 »

sharkjacobs wrote: post #20

> Deep learning works, and we will solve the remaining problems. We can say a lot of things about what may happen next, but the main one is that AI is going to get better with scale I'm not an AI skeptic at all, I use llms all the time, and find them very useful. But stuff like this makes me very skeptical of the people who are making and selling AI. It seems like there was a really sweet spot wrt the capabilities AI…

XAi seems to be able to dump 10-20x more compute into their Grok models each time. Don't see any signs this is slowing down...

Re: The Intelligence Age

#223

Post by joshmarlow »

visarga wrote: post #95
joshmarlow wrote: post #85

Earlier quoted context omitted.

This dovetails with something I've been thinking lately; the foundation of civilization is standing on the shoulders of giants (ie, I might make small contributions but I didn't invent compilers, CPUs, mining, agriculture, etc). Progress in large part is figuring out better ways of doing that (language, written language, printing press, internet access, etc). When you look at things that way - LLMs start to seem deep…

> the foundation of civilization is standing on the shoulders of giants. Progress in large part is figuring out better way of doing that It's so much easier to imitate than to invent something truly novel and useful. Let's do a bit of napkin math. A human lifetime is about 500M words. GPT-4 used up about 30,000 human lifetimes of language. But cultural evolution took 200K years and 120B people to get here, about 4 mi…

> AI will only advance as fast as it can search, it's not a matter of pure scaling of computation, we need to scale interaction and validation as well.

I think Kevin Kelly made a similar point in explaining his skepticism around intelligence explosions; even if we build something much much smarter than us, the thing will still need to do experiments to fine-tune it's (super-human nuanced) understanding of physics/biology/whatever - and the clock-cycle of external reality isn't speeding up like our computation is.

I think something smarter than us could design better/more informative experiments than we could to gather information about the world. That being said, I think his/your point is insightful.

Re: The Intelligence Age

#224

Post by whamlastxmas »

n_ary wrote: post #203
Workaccount2 wrote: post #137

I don't know if I am the only one who always trips up on reading this common theme in AI progress - that AI will be the pinnacle of education - but it really strikes me as meaningless. What is the point of education if the bots can do all the work. If the worlds best accounting teacher is an AI, why would you want anyone (anything?) other than that AI handling your accounting? A world where human intelligence is seco…

I am more worried about the fact that, these AIs will become the commercial moat of knowledge that we currently have available freely. The quality of knowledge you can now openly find is getting scarce or prominently moving behind paywalls. However, these AI models have access to the same knowledge somehow freely. Once we start relying on AI for knowledge(see how people frequently just ask few questions, copy paste a…

Even the shittest small LLMs today can tell you any piece of knowledge you want. We will always at minimum have access to our current level of knowledge. If AI invents “super science” and suddenly we can travel faster than light then sure that might get locked away. But we’ll never be helpless idiots who can’t even do math.

Re: The Intelligence Age

#225

Post by geysersam »

Animats wrote: post #201

Until the hallucination problem is solved, we can't trust LLM-type AIs to do anything on their own. This limits uses to ones where the cost of errors can be imposed on someone else.

Of course we can. Humans are also mistaken sometimes, especially when trying to solve difficult problems.

If an LLM hallucinates it's usually because the problem is too hard. For easier problems it's rarely an issue in my experience.

Hallucinates is just an indicator that the model is inadequate for the problem you're applying it to. That doesn't mean it's inadequate to solve any problems.

Re: The Intelligence Age

#226

Post by moffkalast »

sharkjacobs wrote: post #20

> Deep learning works, and we will solve the remaining problems. We can say a lot of things about what may happen next, but the main one is that AI is going to get better with scale I'm not an AI skeptic at all, I use llms all the time, and find them very useful. But stuff like this makes me very skeptical of the people who are making and selling AI. It seems like there was a really sweet spot wrt the capabilities AI…

As large as the absolute largest models are today, they are still microscopic compared to our brains. A 1.7T param model would only store an actual total of about 850 GB if fully saturated (4 bits of information per weight estimated for bf16 transformers), a lot less than a human brain with 150T synapses running in full analog precision. We need to scale the current gen of models at least another 10-100x to even reach the human level of complexity, something we'll be able to do in the next two decades.

And well then there's going beyond just text. Current multimodal models are basically patchwork bullshit, separately trained image/audio to text/embeddings encoders slapped onto an existing model and hoping it does something neat. Tokenization and samplers are likewise bullshit that's there to compensate for lack of training compute. Once we have enough to be able to brute force it properly with bytes in, bytes out, regardless of data format, the results should be way better.

Re: The Intelligence Age

#227

Post by famouswaffles »

dartos wrote: post #175
layer8 wrote: post #65

Earlier quoted context omitted.

We also personify synapses and axons in human brain tissue, though. My point is, while I agree with your first sentence to a degree, we shouldn’t judge the whole solely by its elementary parts. Clearly an LLM exhibits very different behavior from a conventional database.

LLMs exhibit very similar behavior to a search algorithm. Text query in -> relevant text out. I don’t say that search algorithms “learn” or “think” outside of ML.

>Text query in -> relevant text out.

Wait that's it ? I guess humans also exhibit similar behavior to a search algorithm in certain instances. Nothing about LLM inference seems particularly similar to search even with our limited understanding.

All you're saying here is Input goes in > Output comes out. Well no shit.

Re: The Intelligence Age

#228

Post by smokel »

mellavora wrote: post #210
falcor84 wrote: post #141

> Our children will have virtual tutors who can provide personalized instruction in any subject, in any language, and at whatever pace they need. This is one of those few cases where I'm actually more bullish than Altman. I don't need to wait for my kids to have it, but rather I personally am already using this daily. My regular thing is to upload a book/article(s) into the context of a Claude project and then chat w…

So there is some interesting research about brainwave syncing during effective communication, which certainly includes personalized instruction (tutoring) or small-class learning. I wonder how that works with computers, when we are only sync'ing with the ghosts and statistical patterns of other humans, and those patterns are generated by electronic brains.

That might be research, but it's most likely not scientific research.

Re: The Intelligence Age

#230

Post by ansk »

> humanity discovered an algorithm that could really, truly learn any distribution of data (or really, the underlying “rules” that produce any distribution of data)

He's hand-waving around the idea presented in the Universal Approximation Theorem, but he's mangled it to the point of falsehood by conflating representation and learning. Just because we can parameterize an arbitrarily flexible class of distributions doesn't mean we have an algorithm to learn the optimal set of parameters. He digs an even deeper hole by claiming that this algorithm actually learns 'the underlying “rules” that produce any distribution of data', which is essentially a totally unfounded assertion that the functions learned by neural nets will generalize is some particular manner.

> I find that no matter how much time I spend thinking about this, I can never really internalize how consequential it is.

If you think the Universal Approximation Theorem is this profound, you haven't understood it. It's about as profound as the notion that you can approximate a polynomial by splicing together an infinite number of piecewise linear functions.

Post reply on HN