I did some work on estimates of this in 2011 and many very smart people estimated 5-10 years. I think there's a strong cognitive bias around that time period. It's easy for people, especially young people, to see a personal nonlinear event happening in around "5-10 years" and 20,30 years are harder to conceptualize. It's easy to make a sacrifice today to pay for a new car or engagement ring in 5 years, but harder to plan for retirement in 30.

Yes, making really huge neural nets with GPT-3, DALL-E etc is causing a nonlinear event. Extrapolating that the nonlinearity will continue until singularity without justification is a dangerous modeling error. Consider sigmoidal functions and how they show up in everything from predator prey dynamics, bacterial conjugation, etc. Those have a nonlinearity that subsides when a balance is reached.

I think the probability of a singularity in any given year increases each year, but it's going to be a stacked sigmoidal function as different performance bottlenecks are reached. I don't think any significant chance in the next 5 years unless it comes from top secret government labs somewhere, but I think the interconnect speed is still too long. I think it will require some kind of advanced neuromorphic memsistor system, maybe in the form of tensor chips in phones if a distributed model is possible.

that's very specific haha