Several companies have bet that a big enough neural network, once trained on a sufficient amount of data, would be able to do everything we want our AI systems to do. This bet makes sense if the world we live in has a finite number of things that are useful to learn. It doesn't make sense if the world is filled with infinitely many things to learn, and achieving different goals requires learning different things. The view that the world is big and filled with infinitely many subtle things that can be learned for achieving goals is the big world hypothesis (Javed & Sutton, 2024).
AI startups of today that are aiming to build generally useful AI systems are blind to the big world hypothesis. If the big world hypothesis is true, then these companies will fail unless they either build a strategy to deal with big worlds or decide to settle for commercializing AI systems in narrow domains.
The humanoid robotics companies are especially guilty of pretending that the world is simple. These companies say they want generally useful robots that can do everyday tasks, yet their current strategy is to learn a policy that works in any household. They don't accept that every household has subtle details that cannot be known in advance, and that because of these details, no amount of data will be enough to learn a policy that works in all households. These companies often use two use-cases to demonstrate the capabilities of their robots—loading a dishwasher and doing laundry. These use-cases are worth discussing in detail.
Loading a dishwasher, to these companies, is a task that can be solved by learning a reliable policy for transferring a pile of dishes to a dishwasher; the reality is more complex. Every household has a set of dishes that are not dishwasher safe and must be hand-washed. This set differs from one household to another, and changes even in a single household over time. Learning to load a dishwasher successfully requires learning about the specific dishes in each household. If I were to use any of the existing humanoids in my apartment, each of them will happily put my Ember Mug in my dishwasher and damage it.
Similarly, doing laundry to them is transferring a pile of clothes from a bin to a washer and dryer. They have no mechanism to learn or remember the dryer settings I prefer for different clothes, and no mechanism to know which clothes I prefer to hang-dry. The answer to these questions is not as simple as read the label. I would happily use the dryer for a worn-out delicate shirt, and I might hang-dry a new piece of clothing even if its tag says it is dryer safe. Different people have different preferences for washing and drying their clothes, and these preferences cannot be known in advance.
These subtle but important differences are not limited to loading dishes and doing laundry; they plague all everyday tasks. For example, none of the robots can know that, due to a slight misalignment between my balcony door and the door frame, I have to press the door at just the right spot to lock it properly. Without this knowledge, none of these robots can reliably do the simple task of locking my balcony door. No amount of data would give these robots this knowledge. All these differences make the world of humanoid robots big and complex.
Now I am not saying that we shouldn't learn generally useful policies at all. It is practical to ship robots pre-programmed with policies that can do many things. But there must be a mechanism for adapting these policies with experience. This adaptation cannot be an afterthought because it is a necessity to get these robots to do any useful work.
My prediction is that if we want to make robots that are capable of adapting then we have to start fresh. Making such robots requires redesigning their bodies to include sufficient sensors so they can have self-verifiable knowledge; it requires fitting them with chips that enable energy-efficient learning and not just inference; and it requires developing new algorithms for learning reliably and efficiently. All three of these are unsolved problems in the field of robotics.
To expect the current approaches to magically be compatible with adaptation is to partake in wishful thinking; wishful thinking can be useful for deciding what to try out at a small scale, but we should know better than to rely on it when deploying billions of dollars of capital.