If we define AGI as something that performs as well as humans on all current economically valuable tasks, then it could theoretically be true that new tasks will be created that the AGI is not good at, which humans could then make their new niche. In the following argument, I'd like to show that it is possible and likely for AGI to replace all jobs and future jobs (at least for the jobs where success is measured in productivity/quality).
- Argument of feasibility: Intelligence on the known dimensions can generalize to new unmeasured dimensions
For this, I would first like to show that there is a finite-dimensional solution to human intelligence in general. This is easily understood by looking at the total parameter space of the human-brain: if we assume 1 parameter per neuron, or if you want to model the brain in slightly higher resolution, 100-1000 parameters per neuron, we end up with ~86 billion - 86 trillion parameters / dimensions. That is a huge amount, but most importantly, it is finite. Secondly, I'd like to show that human intelligence likely lies on a much, much lower dimensional manifold. For this, look at IQ tests: basically, what IQ tests have shown is that we can decompose intelligence into a handful of broad cognitive components, which identify roughly 7 to 10 broad abilities that account for 50% of all variance in human cognitive performance. What IQ tests have shown is some form of PCA of human intelligence: appearantly, this highly complex thing (intelligence) can be decomposed into just a handful of components that can explain 50% of the performance on human cognitive tasks. This doesn't mean that the rank of intelligence is 7-10, but rather that the functional rank is likely quite low for intelligence tasks, much lower than the ~86 trillion dimensions of the brain itself.
Now, the amount of cognitive dimensions measured is only a subset of the total dimensions of the human brain. The point however is that since we know the g-factor is so highly predictive of many cognitive tasks, its unlikely that we will find many new tasks / dimensions that show a low or no correlation to the g-factor. Therefore, we can already get an accurate picture of human intelligence just by this rank 7-10 space. Considering that the human brain has managed to decompose all these cognitive tasks down into a 10-dimensional manifold, shows us that it is atleast feasible to find a low rank solution to cognitive tasks that generalizes to new unmeasured dimensions.
2) Current AI systems show the g-factor already:
Secondly, I'd like to make the case for the g-factor of AI. In essence, this is also what the 'g' in AGI stands for. What we care for here is exactly the same thing as in IQ tests: that performance on one benchmark translates to performance on other benchmarks. To measure every possible dimension of human intelligence is infeasible (as i said, up to ~86 trillion dimensions). To test every human economically valuable task is less infeasible, as its a subset of this ~86 trillion, but still infeasible. Luckily, we don't have to if models generalize. If models were to act like chinese room experiments, where they have a 1-1 mapping from input to output, they would be strictly memorizing. In this case, we would need to measure every economical task, since their solution would be brittle and not generalize at all. Now the first evidence that they generalize atleast within the same data distribution is that they perform well on test sets of unseen data. So the most extreme version of this assumption clearly can't be true. Secondly, we've seen that especially bigger models tend to generalize well. One explanation is the lottery ticket hypothesis, where the latent space in the model is used to try out many different solutions, in which only the best solution wins. This shows models compressing something like the mona lisa down 1000 fold, storing it as simple rules. This compression is essentially what generalization entails: finding the lowest rank solution such that it still carries the signal and ignores the noise (perfectly in line with occams razor).
Thirdly, posttraining has unlocked a whole new level of generalizing capabilities. Empirically we see that reasoning models greatly carryover performance on math/coding benchmarks to unseen reasoning benchmarks that have nothing to do with math or coding. This makes intuitive sense: reasoning is the ability to produce new objects from in-distribution components. THe first layers of a network do some form of PCA on the input, decomposing it into its simplest elements. Each consecutive layer then composes it into something more complex. Since the network uses compressed, generalizable rules, it is able to generate new objects it has never seen before. The more OOD the object is, the more layers are needed. SOmetimes this exceeds the amount of layers in the architecture, aka for hard problems, and then the model needs to loop back into itself: recursion. This is the essence of what reasoning is, iterative PCA to increase the complexity of the object using local rules in order to generate something that is OOD. Now, reasoning is bottlenecked by the token layer, and reasoning in itself is a skill. Models learn to optimize their weights, basically to create rules / algorithms to solve optimization. In this case, the network creates algorithms that are loop invariant such that they can be applied iteratively. It also creates an algorithm for the reasoning itself, such that the right words are used that leads to the right composition. In the end, reasoning itself is also just an algorithm. Thus, all-in-all, it is not surprising that reasoning leads to generalization since it is the essence of what reasoning is. It is a very low-rank (since tokens are very low dimensional compared to the NN itself) solution that is highly generalizable.
Now, what this all means is that although we don't measure every possible cognitive domain of models, we simply don't have to. The fact that they generalize to some extend, and have even shown to solve new mathematical theorems in creative ways, show that they are generalizing. Therefore, measuring just enough cognitive dimensions would allow us to accurately depict their intelligence, since their intelligence itself is likely functionally rather low rank. We can't yet say it is as functionally low rank as human inteligence, and we can't say it has the same g-factor of human intelligence. But it isn't unlikely that we will get there. In fact, the whole point of NN is to find this lowest rank solution to the problem space. And since humans have already shown it to be possible, we know it is also feasible.
As a last argument, even if there happen to be some new cognitive tasks that humans can excel at that AGI is not yet good at, I doubt humans can reskill themselves quicker than that AGI can optimize for this new target. Therefore, it seems likely that any economically valuable task based on performance is going to be fully automated once we have an AGI system.
[link] [comments]
