Compound AI is AGI
Our attempt to end the great AGI debate
Apologies for the day-late post — we have a RunLLM team offsite this week, and things got delayed.
When and whether we achieve AGI is something has been hotly contested for years in the AI community, and as much as we love watching people yell at each other on Twitter, what stands out to us is that no one seems to be able to agree on what AGI actually is. If we don’t know where we’re going, it’s going to be hard to figure out whether we’ve gotten there or not. In many ways, the whole debate about reaching AGI is actually a debate about what AGI actually is. Generally, however, the dominant points of view all seem to agree that haven’t yet reached AGI.
We have a slightly different perspective on things: We believe that the LLMs in December 2024 have already achieved AGI. The artificial term is pretty straight forward — these are algorithms not sentient beings. How about intelligence? One could argue that models today aren’t intelligent “enough” but at this point I think we can all agree that by most metrics they are reasonably intelligent. In fact, I think one could argue that many ML models predating GPT demonstrated artificial narrow intelligence — they could accomplish a small set of task better than humans. That leaves the word general. The whole conversation hinges on the definition of the word general. Here’s what the dictionary gives us:
affecting or concerning all or most people, places, or things; widespread
It’s undoubtedly true that an individual LLM is not really able to solve problems that ‘affect or concern all or most things’ — but neither are humans. Some of us are experts in software engineering, and others of us are experts in marketing. Just because humans have general intelligence, that doesn’t mean that you hire someone in marketing and expect them to write code, or vice versa. To stretch the analogy a little, the skills that we develop are a function of our construction (i.e., education).
However, our argument rests on the fact that restricting AGI to a single LLM (or another kind of model) is an artificial construct. We believe AGI has already been achieved because we’re capable of composing systems that reach human performance on many common tasks. This is what the trend towards compound AI systems has highlighted: The power of LLMs comes from being able to compose multiple models (usually of varying sizes and generality) in order to achieve a specific task. The same building blocks can be recomposed in a different fashion with different specialization to achieve a different task (just like a person).
If you follow most of the common schools of thought on AGI, you probably disagree. Let’s dive into what the counterarguments might be.
AGI refers to a single system, and composition of multiple models is a simulation of true generality. It’s generally agreed that consciousness and human intelligence are emergent phenomenon. A single neuron in your brain is not intelligent, and we’re still not exactly sure how many neurons are required (and in what connection pattern) for consciousness or intelligence to emerge. The that the AGI has different architectural properties doesn’t seem to make a significant difference to this definition. The transformers architecture is composed of many MLPs, and each MLP is a set of floating point numbers as weights. The fact that a single LLM can’t achieve AGI doesn’t preclude a composition of LLMs from achieving AGI — as we said above, the limitation of one LLM needing to achieve AGI is an artificial construct. We could belabor the point by talking about brain lobes and other subsets of human intelligence, but it should be clear that compositionality is a fundamental part of emergence. Compound AI systems follow the same principles, so compound AI systems should be eligible for generality.
One system can’t be generalized to solve many tasks. We hate to be pedantic, but humans can’t be generalized to solve many tasks either. Again, hiring a software engineer to do marketing or vice versa won’t work well for you. The natural counterargument to this is that the same person who was once a software engineer can learn the skills over time (retrain themselves, one might say) to become a good marketer. We would argue that the same can be done with the building blocks of a compound AI system: The same base models used to do software engineering can be reconfigured and instructed in different ways in order to do marketing well. While it’s true that a person has memories and an indefinable quality (consciousness) that carry through that retraining process, that seems to be an argument more about consciousness than it is about intelligence. The fact that a rearchitecture of the building blocks is necessary is analogous to the fact that a person will make new neuronal connections as they learn new skills.
True AGI is about going surpassing human intelligence. Generally, we’re of the opinion that the fact that we haven’t built Skynet or HAL 9000 is a good thing. We’re guessing you probably agree with us. Jokes aside, it’s clear that there are plenty of things that humans can do that existing AI hasn’t yet achieved — generalizable physical intelligence is probably at the top of the list. We don’t disagree with this in the slightest, but we also feel that it sidesteps the main argument. This is primarily an argument for super-intelligence (capabilities beyond what we know today) and sentience — neither of which are required for achieving AGI.
Existing LLMs can do plenty of things that humans can’t, like recite who the first President of Burundi was and what the atomic number of Plutonium is in the same sentence. Compound AI systems like RunLLM can also provide detailed, precise answers to customer questions in few seconds — orders of magnitude faster than a human can. The fact that AI systems have a different kind of intelligence than humans — both of which are limited in their own ways — doesn’t negate the generality of either one. To torture the analogy again, it’s no different than the fact that some people have great minds for solving logic puzzles while others are creative thinkers.
We doubt this blog post will end the great AGI debate, but we think it should (we’re biased). When it comes to generality, it seems painfully obvious to us that we’ve already achieved the building blocks to create general intelligence. That the intelligence hasn’t been applied in every domain yet is simply a lack of maturation and investment — it’s a question of when, not if.
Now that we’ve solved AGI, that doesn’t mean that there aren’t other debates to be had! In fact, there are plenty of other philosophical questions that are left to be answered. The most interesting one revolves around the definition of consciousness and distinction between intelligence and consciousness. The implications of having intelligent systems that never sleep is already massive; adding consciousness to the mix will only make things more interesting. As for AGI, it’s time to put that to bed.
Seems like most people attribute “general” to require “generalization” to everything. Doing many things, is general enough for me.