Nvidia predicts AI models 1,000,000X more powerful than ChatGPT within 10 years | Hacker News Reader

Nvidia predicts AI models 1,000,000X more powerful than ChatGPT within 10 years

pcgamer.com

80 pointsCharlesW64 comments

Open articleView on HN

Translation: nVidia is talking up the hype that can increase its stock price (now that the cryptocurrency hype is over)

You could apply that to any announcement by any company

Sounds like a business plan to me.

Sausage company predicts the world to have sausage for dinner five times a week in 2030!

Hah

Do you think it’s all hype and no substance? Any opinion of how much is hype, and how much is real?

Not saying it could no be true, but that's exactly what a company like Nvidia would say. Do we have any reasonably unbiased info on what the boundaries of the current approach are and when we will reach them?

"More powerful" here just means larger (another commenter says faster chips, but you don't even need that if you spend the next ten years training a model, although that could be a more accurate representation of what Nvidia is trying to say, since the alternative is a no-brainer). AI isn't really "smart," and it seems unlikely we're going to get there anytime soon (in other words, the "singularity" is still firmly out of our grasp). What we're seeing today is the fruit of training models with large swaths of data, and the limits of this approach are predictably based on the availability of such data and the resources required to convert them into these gigantic models. Framed in those terms, the "limit" we should be considering ought to be more about utility than how many resources we can afford to throw at models that might ultimately have no use.

Discussions around these topics tend to get dramatized, though. In my opinion we haven't made a lick of qualitative progress in recent years in the "IQ" of AI, but we do get better at utilizing what we have, and making quantitative progress in building ever larger models.

Please lay out your definition of smart?

For the task they are given, take textual tokens as input and make coherent output, the models seem 'smart enough' for me.

The problem here is all the text in the world is still one dimension of information. The emergence of human intelligence, at least I believe, isn't from one dimensional information, but from being able to get feedback and prune the bits of information that are inconsistent with reality. There is no continuous feedback loop in our current LLMs to prune/down weight bad information in that fashion that the mind works in. There is also the "the LLM universe is what humans tell it", if we feed it bad data that's incorrectly weighted, then we'll get bad output. There is currently no scientific method model internal for the AI to test its predictions on reality.

I believe all the above issues are solvable, but no idea on the time and resources necessary to accomplish them.

> Please lay out your definition of smart?

Firstly, I agree that "smart enough" is what we should focus on here, as I tried to emphasize with my focus on utility.

Secondly, I'm assuming your intent is to ask me when I would consider an AI to be "smart" in the context of my comment. I think I should avoid that question, and perhaps even retract my statement. There is plenty of evidence that ChatGPT is smarter than some humans. All of your 'problems', as you might realize, are equally applicable to some people.

Putting that aside, a model would need to at least be able to internalize any corrections made (by a user, and not necessarily only in the context of training) and reliably produce results consistent with such an internalization, for me to consider it smart-er (i.e., an improvement on today's AI IQ). To be smart without qualification, it would need to go further, into the realm of being able to hold opposing ideas while functioning consistent to a given user's expectations (this one will be big, because it will act as a defense against trolls and misinformation). Bonus points if it's able to improve upon its own learning algorithms and ruminate on ideas (unsupervised learning that is inspired by previous input and not stochastic in nature) without necessitating constant external dialogue.

Ya, as Bing showed us AI rampency is an issue that we've not tackled even in a tiny prompt space. Going to be interesting to see how we figure out how to deal with that.

>In my opinion we haven't made a lick of qualitative progress in recent years in the "IQ" of AI

We've gone from not very good to beats all humans at chess and go, and passing exams in medicine and law, all of which are things that I can't do myself. When AI can do all things better than a human would that qualify as a lick of progress?

I have a probably really stupid question but I'm going to ask it because I really don't know the answer.

What's stopping someone from starting a distributed effort to train an open source ChatGPT model, like Folding@home but for AI? Is it a technical limitation with how models are trained, difficulty of coordination, something else?

Training in parallel on more then one GPU is pretty hard. Distributing the training over many random computers and adding at least some kind of validation, seems to kill the performance. So it would always be easier to rend cloud GPUs.

But would it be cheaper?

The models are extremely large to ship around, sync, and load on consumer hardware over the internet. Maybe there are better ways to build and train a model for a distributed use case though.

The model that's used by Stable Diffusion is only around ~4 GB in size.

which would make traditional distributed training prohibitively slow.

You would have to sync the weights with residential internet speeds (10s of MB/s) instead of PCI-e Speeds (10s of GB/s).

If you want to add new stuff to Stable Diffusion - you don't have to retrain the main model from scratch. You can train it only on a few hundred images at a time and then add the resulting model as an extension or merge it into the main model. People train their models on a single celebrity or a single artist that way.

Other types of AI could be trained in a similar way.

Is this amenable to merging from thousands in one go or do you have to train -> merge -> train -> merge to not overwrite each other's trainings?

Sounds like we have our use-case for blockchain!

Hard to split the problem into small enough pieces

working on it.

together.xyz

besides the answers already given, the network is already the bottleneck on supercomputers on a local network. training that over a WAN would exacerbate that issue and make training even slower.

That is actually irrelevant. I couldn't try out flan-t5-xl but the model flan-t5-model is like an inferior GPT but it is opensource. I don't necessarily think that the size of the model is that significant. What makes ChatGPT so impressive is that you don't have to mess around with the settings vs Flan T5.

More effort has to be put into obtaining good training data than just throwing compute at the problem.

As Scott Alexander at ACX recently said:

"AGE OF MIRACLES AND WONDERS: We seem to be in the beginning of a slow takeoff. We should expect things to get very strange for however many years we have left before the singularity. So far the takeoff really is glacially slow (everyone talking about the blindingly fast pace of AI advances is anchored to different alternatives than I am) which just means more time to gawk at stuff. It’s going to be wild."

Any model even just one or two orders of magnitude more powerful than ChatGPT is going to make for a wild future. 10^6x is hard to imagine. And the implications for the philosophy of mind are staggering.

It always seems slow until you’re on the right hand side of the elbow

These are the days of miracle and wonder
This is the long-distance call
The way the camera follows us in slo-mo

The way we look to us all

God, I love Paul Simon

Gradually, then suddenly.

When I read words like singularity or AGI..is when I know we’re at peak hype and correction is coming. I just hope its not another AI winter. Because I’d still like to see this area continue to grow.

The advances every week in papers in models, more TFLOPs per dollar per GPU, larger datasets, more VC billions being invested, more humans tinkering.

It’s definitely going to be wild next decade.

AI superiority is now a nation race like being first to moon, first nuclear bomb.

Whoever cracks it gets many trillions added to their GDP.

It's hype for decades until it happens and then not so much. (https://waitbutwhy.com/wp-content/uploads/2015/01/Intelligen...)

In this case "powerful" means faster chips. It's not clear that this translates into 1,000,000 smarter given that current AI might end up bottlenecked on availability of text and information to train it on. I guess if you could speed things up by a million-fold though, you could have way bigger context windows and you could potentially set a LLM talking to itself as it explores a problem through web search, writing and running programs, etc. Exciting times.

If I had to make a guess, I would think that text based LLMs will run into that wall sooner than later. But the rapid increase in compute availability wouldn't limit us to text either. Also imagine the models working on image input, sound data, video, temperature, touch, and whatever other inputs you could encode and present to them. This brings it more in line with the human experience of inputs.

There is not enough written text to scale a model like ChatGPT at even 10x its current parameters, so improvements will have to come from optimizations and/or longer training (at diminishing returns).

Enough data wont really be a problem. https://us.sganalytics.com/blog/2-5-quintillion-bytes-of-dat...

I wouldn't be surprised if we create more data in one year than is in the entire internet from the start

quintillion bytes of social media probably have very limited value.

Really? All of humanity imparting everything they think and know publicly seems incredibly valuable. 100% serious.

If you're looking to make a tweet generator, sure. If you're looking to harness the power of facts, no.

The main issue might be that the training set may get polluted by a lot of AI generated text (content that is not supposed to be there, but is there because AI-generated content is becoming widespread online).

Storage is cheap. With ChatGPT being the preminent system, and with OpenAI recording all of its output, it should be possible to exclude modified answers from the training data.

So interesting that people think facts are the only thing that's valuable. Your comment isn't a fact, does that mean it's worthless?

When looked at individually it's not worthless, but when you are talking about a corpus that significantly lacks facts, then the corpus itself is worthless for any mainstream usage other than a social network text generator.

How do you know that? There are thousands upon thousands of books that haven't been digitized.

GPT-4 Will Have 100 Trillion Parameters — 500x the Size of GPT-3

That's likely a massive over-estimation. I would bet it's around 1 trillion.

That is just an internet rumor and officially denied by OpenAI.

This is hyperbolic, of course. But it makes Kurzweil's 2045 prediction of the singularity not as outlandish as it first seemed. https://www.smithsonianmag.com/air-space-magazine/reaching-s...

Still outlandish. Not especially well defended either.

We're far enough away from 2045 that it seems hasty to completely discard it as outlandish. But, yes, unlikely.

He predicts 2029 as when machines will pass a "valid Turing test", one that goes for hours.

I'm wondering how this is going to be different than self-driving, where we can get 90% of the way there, but the last 10% is notoriously difficult with not nearly enough edge cases represented in the data

Is there a point where those AI models could plateau and adding more parameters or data will not improve them anymore?

Without changing the architecture in a significant manner/using more training data, there does come a point where adding more parameters will result in no gain (and sometimes even worse results).

There's also a consequence of performance by adding more params. The inference time will be longer and even just training the model will take longer and won't be able to run as many epochs in an efficient manner.

It currently looks as if all AI’s trained on internet texts basically become really good simulators of toxic people on the internet. Full of gas lighting, lies, extremisms etc.

So perhaps maybe the goal should be to make better AI’s rather than “more powerful” toxic people-on-the-internet simulators?

Could Ai learn from 3D mapping daytime soap operas and trace character development along the narrative arc? Then it has to function in the real world.

I imagine it could. Something for it to contend with though is that our published media, soap operas, reality tv, so and so, don’t really reflect what we humans are like in real life. They’re caricatures of us.

But instagram/tiktok has real data + you messages, people you know and talk to and listens to you 24/7. If they will be able to predict 1B of humans and social dynamics then I guess it’s enough of training data to basically simulate reality.

EDIT: it gives insane “soft” power so pretty sure NSA etc. are working on this

Imagine the power of 1,000,000 tiktok videos! That would really be simulated reality.

Ok I’ll be the one to ask — 1,000,000x means what, exactly?

Well it's certainly a lot more than 1,000x that's for sure!

1,000,000X more posts about ChatGPT successors on HN main page? Who would have thought!

(great time to use https://lobste.rs/)

Well said

Reply on news.ycombinator.com