Nature of Moving Goalposts
After the release of GPT-6 Astra, Jensen tweeted that AGI had arrived. Even before Astra, models were already very good, especially in domains like mathematics and coding. With recent rapid advances in computer use, their effective action space is now as large as ours, and soon they would become better than us at a large fraction of the tasks that can be done on a computer.
But for some reason, most of us intellectual workers still haven’t been replaced. Somehow, it seems that we are still working and getting paid (often more than $200 per month). I mean, AI is certainly a very useful technology, like computers, WWW, and smartphones, but it looks more like a tool than a human-replacer. And that’s a little bit underwhelming. AI was supposed to be humanity’s last technological invention, setting it apart from other technological breakthroughs. So far, it looks like other technological breakthroughs (not saying that it won’t change the world—it will, just as other breakthroughs have).
This is because models currently act as force multipliers for humans, and thus they perform worse than the humans who use them. That is, model < model * human. Models will replace us only when they are better than humans who use models. Not every improvement to a model counts toward this goal; as long as the force-multiplier relationship holds, model + improvement < (model + improvement) * human!
If you had shown people Astra six years ago, their immediate reaction would have been, “This is AGI.” Now that we’ve gotten so used to what rapid progress in AI can look like, they might say that the model is certainly very good, but it still cannot do <this subtle thing I could do>. This is moving the goalposts, but that’s how it works. Note that the most common AGI definition describes a machine that is better than humans at most economically valuable tasks. A crucial part of that equation is that humans can use AI. So every time a model improves, the bar for AGI also rises, because improvements in AI capabilities also improve the capabilities of the humans who can use that AI.
This multiplicative relationship need not hold forever. At the same time, despite astonishing developments since GPT-4, it has continued to hold, so this is not just a temporary thing. This relationship holds because there is something special about human intelligence that all the astonishing progress from GPT-4 to GPT-6 has collectively failed to match, even though the models have improved in so many different ways. What exactly that is, and how to close that gap, is a trillion-dollar question, and everyone will have their own hypothesis. We will have AGI once we have an answer to this.
Model companies have an incentive to prioritize the typical set of tasks, since those are in the highest demand. As they keep improving their models, the volume of that set will expand, and the human advantage will remain in increasingly atypical tasks. My hypothesis is that the force-multiplier relationship exists because humans are so much better than machines at picking up atypical tasks and getting good at them. This must have something to do with the fact that the learning algorithm powering our brains evolved mostly during the caveman era, yet that same algorithm can make one a rocket scientist.