AGI: Are the ingredients already in place?
They say you can’t reach AGI with today’s LLMs (looking at you, LeCun and Gary Marcus). But that seems to be a nuanced point because based on the vast improvements in today’s models, it IS possible to have a base model (LLM) that may not take you there, but with layers of assets around it (tool use, agents, etc. etc.) that can eventually bring us there. What do you think?
Lets break this down into three parts for AGI:
As for each of those steps, only the automated part generally exists as automated systems are very common with current AI tools that can work independently. However nobody as far as I am aware has been able to cause an LLM to prompt itself from its output, but this seems possible. If there are any examples of this, please let me know.
There is however no true general tools that can do most tasks at this time, although they are multi use LLM’s that do exist, such as GPT-4o which can natively process text and audio, as well as having a native AI image generator based on DALL-E-3. However using LLM’s together to form tasks with communication between them is a way to make General systems from narrow systems.
The hardest thing to define is Intelligence, benchmarking does exist, but as you said:
They can be unreliable in the real world, this is where I think there is the largest issue as you cannot reliably prove intelligence of a model. And it isn’t easy to build intelligence in a model.[2]
Therefore until there is a better understanding of intelligence in general, the closest we can get is an automated general AI system, which would be possible with current technology if we combine Narrowly defined LLM’s with a communication system, and then allow it to self prompt itself or other LLM’s in the network as a hive mind. We however will not know if it’s intelligent until other fields can clearly define it in terms of sentience, or it can be demonstrated in non ideal conditions that the LLM network is better than human’s at its tasks.[3]
I know AGI means Artificial General Intelligence, but it would need to be automated to actually exist ↩︎
Or in humans, there is no universal way for education. ↩︎
This still has the issue of the LLM network being only able to repeat information it has already learned unless the LLM network can learn/train other LLM’s from itself. ↩︎
I think AGI is plausible.
They bench marked all the AI’s in various tasks and concluded that AGI currently does not exist. The models were unable to do anything novel with the information they had. For example major LLM’s could complete the towers of hanoi game. They were then prompted to solve a very similar game with a couple rules changed and all of them failed.
The paper concludes saying it only looks like “general intelligence”. But in reality its just really good at copying from training data.
I do think apple’s take is very “right now” but the way the industry is headed. I think in the near future we could have AGI.
My take: We’re not there yet and won’t be for a little while. With the way that LLMs function today, they are able to predict word-by-word based on what the previous text is. Image generation is similar where it predicts what “should” be where, instead of actually dreaming up something novel.
With that being said, there are efforts underway to get a model to actually think. For example, if I had an art model, I might train it how to sketch, then how to color, etc. Similar things are happening with OpenAI’s Reasoning models, which are essentially graded step-by-step to solve problems. Expanding these methods could lead us to something closer to AGI.