Assuming that the scaling hypothesis keeps working, GATO could probably achieve human-level performance. It was artificially constrained in size to be able to

Assuming that the scaling hypothesis keeps working, GATO could probably achieve human-level performance. It was artificially constrained in size to be able to run in realtime on a robotics platform, but human-level intelligence doesn't require human-level latency even for robotics; most robotic tasks can happen slower (up to 3X slower if directly competing with an 8 hour human workday). Whenever novel domains need to be trained they can be mapped to the existing tokens GATO uses and the model fine-tuned on examples from the domain, and a scaled version of GATO would probably have audio, visual, and kinesthetic tokens along with textual and motor control so it wouldn't require a new architecture for every new domain. At some scale I expect that a fine-tuned GATO-style agent would acquire the ability to be successfully text-prompted in novel domains (no longer requiring fine-tuning), which would allow autonomous operation. I fail to see a step along the way that isn't solved by Moore's Law (or an all-in investment by large companies).