The crazy thing is though - nothing is stopping us from doing exactly that!
Discussion
For sure!! I’m not saying these things aren’t possible. But I don’t think they’re solved by an LLM alone.
What about an LLM with plenty of access to plugins and APIs?
There’s a fascinating paper on GPT-4 that discusses this if you haven’t seen it.
Sparks of Artificial General Intelligence: Early experiments with GPT-4
See the tool use section
Yeah, the more it has access to for sure increases it’s capabilities. Like giving it access to modify your terraform to deploy infrastructure. I think it’s really compelling, but I also think there is still a long ways to go. Another ML engineer in a community that I’m in put together a great piece on the current state of AI:
https://medium.com/@khalil.alami/the-boring-state-of-artificial-intelligence-152244075d7f
Re: early response to the medium list in general: This is all 3.5.
4 is significantly smarter. 5 will be too
The same limitations apply… it’s still an LLM limited to a knowledge domain and is acting as an approximation machine. We still need some pathways for generating high quality data for it to be trained off of… if we use it’s “intelligence” as an excuse to stop producing that data, it will stop learning.
I agree with the last half. It’s not clear to me that we aren’t closer to LLMs than many realize, though we’re definitely more like RNNs in some ways. Give it a plug-in to apply a sort of RNN mask to its plug-in memory memory, and that might be enough I think. Definitely not GPT3.5, but give it 5 years and I think they are easily good enough to replace most computer workers.
And totally, we need to figure out how to continue producing high-quality data. Currently we train agents (people) and the number of resources that they can accumulate becomes correlated to their “model” performance. As we see these resources accumulate to them, we become more likely to use them as the definition of good information. The same may be true of future AI agents and additional quality information generation.
Yeah, I think there is a lot of growth that will be happening in the model network architecture space. Using more precise smaller models for tasks that LLMs use to do specialized things, reasoning models (lots of good research going on here), as well as RI for experiments (but interestingly enough those require physics engines that we define, so most or the time they only can solve for that vs actual physics - this in combination with robotics and more sophisticated sensors though has a lot of promise - essentially self driving cars, and there is still a long ways to go there). Collaborative models are super cool though.
Ultimately though, I’m concerned about the data creation. I don’t want there to be a misunderstanding of how important it is that we continue to produce it.
And that certain people stop producing it 🤭
I kid. Mostly. I mean if it really is smart it will hopefully ignore them
Re: Also, During the training of ChatGPT, more weight was given to data from Wikipedia, making the model see it as a more valuable source of information. Also making it our biggest “ Wikipedia scientist,” 😉.
So, the author points out that reading a thing doesn’t make you an expert. GPT-4 didn’t read stuff though. It learned how to keep writing it if you start and leave off anywhere. This is much more like the human learning via practice than I think this post gives it credit for being.
Re: AI does not truly understand concepts such as addition, subtraction, or text summarization. Rather, it produces a mathematical estimate to satisfy our desired results.
If it can form internal representations of these concepts, and it definitely does in some cases, then it will predict text better. So I disagree with this sentence. It is motivated to internalize concepts, and it does. Bigger models will more so. See that video I linked for more on that, especially the unicorn drawing section
The solution to these problems is actually smaller models solving more specific tasks. That’s his argument. The larger the model, the more abstract - its good at connecting dots, but will start to struggle with precision. He’s not saying a model can’t do this, he’s saying large LLMs do some things well and some things less well and viewing it as a one stop shop for all intelligence is not reasonable. We need RI, we need more specialized tasks, and we need a lot more research in reasoning models.
Re: When it comes to Intelligence, AI models nowadays do not even possess the level of intelligence that our ancestors had when they first learned to manipulate fire hundreds of thousands of years ago. AI -as is today- should be viewed as a tool to augment us, rather than replace us.
I think this is a little hasty. We tend to believe in heroic leaps of insight when in reality it’s true that we stand on the shoulders of giants, each of whom made small incremental gains.
Except for Newton of course
He’s making this claim based on the model’s architecture. Not all models in the future will have this limitation, but LLMs do based on their knowledge domain limitation.
Oh, I guess to clarify, I get really saying about the architecture, but I don’t think he’s properly valuing the arbitrary internal representations that can be embedded inside that architecture
Get what he’s saying about the architecture*
Maybe. But the problem with many of those internal representations are just straight up lies. It’s not a truth telling machine. It’s about to produce compelling speech. It’s a bit different in spaces like engineering, where the content is so well curated. But just because it identifies a pattern, does not make it correct or intelligent. It would need some secondary mechanism for testing hypotheses, and at that point we’re moving out of the LLM space and into the reasoning space. Progress will come when these spaces come together.
Yes, though, to be sure people aren’t really truth telling machines either unless you give them some sort of mechanism to test hypotheses aka a science lab. If science labs can have APIs, and a plug-in is made, then suddenly GPT-x might become a truth telling machine.
To further agree with what you’re saying, it’s frustrating that GPT for doesn’t have much of a opinion aligner. You could start talking to it in a certain way, and get it to output, certain stuff that it wouldn’t otherwise output. Humans are more self consistent. I think Auto GPT though might fix that in a way that isn’t actually too different than human thought. But I haven’t been able to play with it myself yet
For sure - and I think that is by design… they’re going to sell it to company’s who will want to specialize and limit it to be more in line with their specific use case. Right now, it’s just there to kind of be whatever you tell it to be 😂 like a giant ad.
going to go ahead and zap all y’all for this super interesting conversation you’re having in public
Have you seen the stuff about how RLHF it makes it dumber? That’s what makes me suspect it being a generalist is actually the unexpected path toward specialized ability.
Like… what it really needs is to be trained on everything, and then to memorize domain facts by including them in the prompt. For that, they need to release models that can have prompts in the 100k to multi-million token range.
Here’s an interesting exercise, how many characters would the document that explains your job fully to a new employee? That’s the prompt
Yeah, it’s super interesting about the RLHF! Overfit is a very real problem and adjusting weights on models this large can be kind of like a butterfly effect. I think there is a TON of value it its generalization. But I’m of the opinion that it can’t or maybe shouldn’t do all tasks for itself - to me it’s just not necessarily efficient, like using a hammer on a screw. Bigger doesn’t always mean better - it will start to underperform at a certain size. TBD what that is. But let it do what it does best! Language and conceptual derivations and awesome encoding, let other models do what they’re better suited at. Kind of like how our brains work… we have separate specialized areas that we delegate tasks to when necessary. We’re building awesome components, but I’d like us to acknowledge their limitations, not to discourage the work that has been done, but to figure out the next problem that needs to be solved.
Re: While increasing the size of the models may enhance specific aspects, it will not necessarily lead to significant improvements in intelligence. Rather, it will likely result in only slight improvements over the current versions.
Scaling observations indicate that this is not true