this post was submitted on 19 Jan 2024
255 points (95.4% liked)
Technology
60071 readers
3571 users here now
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related content.
- Be excellent to each another!
- Mod approved content bots can post up to 10 articles per day.
- Threads asking for personal tech support may be deleted.
- Politics threads may be removed.
- No memes allowed as posts, OK to post as comments.
- Only approved bots from the list below, to ask if your bot can be added please contact us.
- Check for duplicates before posting, duplicates may be removed
Approved Bots
founded 2 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
Anyone got a graph of ai spending over time globally?
I'm starting to feel more confident about AGI coming soon (relatively soon).
Knowing absoultely nothing about it though it seems like it needs to be more efficient? What's the likelihood rather than increasing the bulk power of these systems that there is a breakthrough that allows more from less?
Spending is definitely looks exponential at the moment:
Most breakthroughs have historically been made by university researchers, then put into use by corporations. Arguably, including most of the latest developments,. But university researchers were never going to get access to the $100 million in compute time to train something like GPT-4, lol.
The human brain has 100 trillion connections. GPT-4 has 1.76 trillion parameters (which are analogous to connections). It took 25k GPUs to train, so in theory, I guess it could be possible to train a human-like intelligence using 1.4 million GPUs. Transformers (the T in GPT) are not like human brains though. They "learn" once, then do not learn or add "memories" while they're being used. They can't really do things like planning either. There are algorithms for "lifelong learning" and planning, but I don't think they scale to such large models, datasets, or real-world environments. I think there needs to be a lot theoretical breakthroughs to make AGI possible, and I'm not sure if more money will help that much. I suppose AGI could be achieved by trial and error (i.e. trying ideas and testing if they work without mathematically proving if or how well they'd work) instead of rigorous theoretical work.
Interesting. Thanks for posting.
So you're saying we might see something 1/10 of a human brain (obviously I understand that's a super rough estimate) next year.
This is the first I heard about GPT not learning. So if I interact with chat gpt it's effectively a finished product and it will stay like that forever even if it is wrong and I correct it multiple times?
This is where I'm really confused with the analogue. If GPT is not really close to a human brain how is it able to interact with so many people instantly. I couldn't hold 3 conversations never mind a million. Yet my brain power is much much higher than GPT. Couldn't it just talk to 1 person and be smarter as it can use all the computing power for that 1 conversation?
Correct, when you talk to GPT, it doesn't learn anything. If you're having a conversation with it, every time you press "send," it sends the entire conversation back to GPT, so within a conversation it can be corrected, but remembers nothing from the previous conversation. If a conversation becomes too long, it will also start forgetting stuff (GPT has a limited input length, called the context length). OpenAI does periodically update GPT, but yeah, each update is a finished product. They are very much not "open," but they probably don't do a full training between each update. They probably carefully do some sort of "fine-tuning" along with reinforcement-learning-with-human-feedback, and probably some more tricks to massage the model a bit while preventing catastrophic forgetting.
Oh yeah, the latency of signals in the human brain is much, much slower than the latency of semiconductors. Forgot about that. That further muddies the very rough estimates. Also, there are multiple instances of GPTs running, not sure how many. It's estimated that each instance "only" requires 128 GPUs during inference (responding to chat messages), as opposed to 25k gpus for training. During training, the model needs to process multiple training examples at the same time for various reasons, including to speed up training, so more GPUs are needed. You could also think of it as training multiple instances at the same time, but combining what's "learned" into a single model/neural network.
This is really cool. Thanks for taking the time. Confusing but the good kind.
I'm just using this to info to then try and extrapolate.
I understand the growth of moores law and such. But the efficiency I was talking about seems almost like 1 exponential jump on an exponential curve.
Let's just say for argument sake that meta makes AGI next year with 350,000 GPUs it would only need 2,000 GPU's to make use of what it's built. That's pretty mind-boggling. That really is singularity sort of talking.
So in your mind AGI when? And ASI when? You working in this field?
Yeah, those GPU estimates are probably correct.
I specialized in ML during grad school, but only recently got back into it and keeping up with the latest developments. Started working at a startup last year that uses some AI components (classification models, generative image models, nothing nearly as large as GPT though).
Pessimistic about the AGI timeline :) Though I will admit GPT caught me off guard. Never thought a model simply trained to predict the next word in a sequence of text would capable of what GPT is (that's all GPT does BTW, takes a sequence to text and predicts what the next token should be, repeatedly). I'm pessimistic because, AFAIK, there isn't really a ML/AI architecture or even a good theoretical foundation that could achieve AGI. Perhaps actual brain simulation could, but I'm guessing that is very inefficient. My wild-ass-guess is AGI in 20 years if interest and money stays consistent. Then ASI like a year after, because you could use the AGI to build ASI (the singularity concept). Then the ASI will turn us into blobs that cannot scream, because we won't have mouths :)
Yea I had a feeling it was still a long way away. At least the media will get bored of it in a year and only the big breakthroughs will make it.
But I think there will still be a lot of "stupid" yet impressive developments like GPT. It appears smart but isn't that smart. Sure there will be other things.
It's the same as the manufacturing developments. Only now are we beginning to build things similar to the complexity of a human in limited functions. But that doesn't mean the machines we have built haven't put millions of people out of work, we just changed manufacturing to better utilise the stupid things they can do much faster and accurately than we can and made a better product because of it. I found out about a year ago we couldn't make a Saturn v rocket now even if we had all the money in the world. The ability of man has been lost. The way they did the machining of the rockets and the welding and things like that, no one alive has that ability anymore. Robots can't do it either. But the rockets we make now are more accurate that the ones made in the 60's. It's just done differently.
You're confused by the analogie because it's a shitty one. If we wanted to reproduce the behaviour of the human, we would invest in medecin, not computer science
While I do work in the space, I'm more pessimistic. I think LLM's will allow the tech companies to breach plateaus that they've found with compositional models, but what we will see is other companies catch up to GPT4, perhaps surpassing it a little.
I won't pretend to be an expert on AI, but my view is that we're purely seeing a future where multiple companies will own LLM's. We also won't see many improvements over what we have now, and this is the pessimist in me again, what I think we'll see is that many of the benefits we saw from GPT4 were likely from the fact that their datasets contained an unbelievable amount of PII and stolen data. Without that data, we've seen ChatGPT get worse, and it's one area where researchers and other tech firms have tried to explain the performance gap.