I'm almost sure that we will [get generally capable AI], but it is very hard to predict for me, and actually AI or almost all technologies are increasing exponentially. And so things are going to be much more faster than it is today, even if it's very fast, crazy fast. I mean, if I will have to say a time, maybe I'll say 30 years because when you look at the 2005, right? we almost had nothing. We even didn't have these cellphones. And it is just a normal requirement now and, I guess, in 20 years or 30 years, we will have some crazy things that we cannot imagine today as well. So maybe let's say 30 or 40 years for that question.

(from: Interviews with AI Researchers)

The existence of generally capable AI systems within the next 50 years seems like a possibility. There are several trends suggesting this may happen:

• We have been working on AI since 1950. The deep learning revolution started in 2012. On the scale of human history (e.g. Agricultural Revolution approximately 12,000 years ago, Industrial Revolution approximately 200 years ago), this is not much time, and investments increased massively since 2010.
• Most of this investment has gone into specific, narrow AI applications—but a significant number of institutions are directly aiming for general AI. For example: There are several well-funded companies explicitly trying to build AGI, most notably: OpenAI and DeepMind. This field is also growing quickly. For example, game developer John Carmack, who does not have notable AI experience, has founded an AGI company with an initial investment of $20M. Furthermore, China has announced a plan to become the world leader in AI, which includies the pursuit of general AI.
• We now have unprecedented progress in large language models like GPT-3 and PaLM. For example, PaLM is able to explain novel jokes in two-shot prompts.
• PaLM has also shown that new capabilities can emerge entirely as a function of scale, without major innovations in the underlying paradigm. Prior research suggested this was a possibility, but PaLM demonstrated that this still holds at very large model size and with high amounts of training data.