Yes. As you say, a lot of the limitations so far has been the control part, which is basically AI.
Given the pace that AI is currently moving at, it seems to me that more and more, the mechanical aspect is becoming the limitation.
GPT 4o now seems to be quite good at reasoning about the world from pictures in real time. I would expect it would soon become easy for it to do the high level part of many practical tasks, from housekeeping to manufacturing or construction. (And of course military tasks.)
This leaves the direct low-level actuator control to execute such tasks in detail. But even there, development has been immense. See for instance these soccer playing robots [1]
And as both high level and low level control (if we assume that models soon will add agentic features directly into the neural networks), the only missing peace is the ability to build mechanically capable and reliable robots at a low enough price that they become cheaper than humans for various kinds of work.
There is one more limitation, of course, which is that GPT 4o still requires a constant connection to a data center, and that the models is too large to run within a device or machine.
This is also one of the most critical limitations of self driving. Had the AI within a Tesla had the same amount of compute available as GPT-4o, it should be massively more capable.
Given the pace that AI is currently moving at, it seems to me that more and more, the mechanical aspect is becoming the limitation.
GPT 4o now seems to be quite good at reasoning about the world from pictures in real time. I would expect it would soon become easy for it to do the high level part of many practical tasks, from housekeeping to manufacturing or construction. (And of course military tasks.)
This leaves the direct low-level actuator control to execute such tasks in detail. But even there, development has been immense. See for instance these soccer playing robots [1]
And as both high level and low level control (if we assume that models soon will add agentic features directly into the neural networks), the only missing peace is the ability to build mechanically capable and reliable robots at a low enough price that they become cheaper than humans for various kinds of work.
There is one more limitation, of course, which is that GPT 4o still requires a constant connection to a data center, and that the models is too large to run within a device or machine.
This is also one of the most critical limitations of self driving. Had the AI within a Tesla had the same amount of compute available as GPT-4o, it should be massively more capable.
[1] https://www.youtube.com/watch?v=RbyQcCT6890