Compared to 1966's Eliza we've made good progress: speech to text and text to speech provide a better interface and work good enough (accents are still a big problem though), we can map between many equivalent commands automatically instead of requiring someone to account for all variants manually, and tools like Google's knowledge graph allow answering questions about most simple facts.

Still, that's about all of our progress in 50 years (in production ready, deployed systems). Estimating that we need another fifty years seems fair.

My point is that none of those improvements bring us meaningfully closer to general intelligence. It's all surface-level stuff. Important surface-level stuff, but nevertheless. I don't see much progress in the actual hard part of creating a truly helpful AI assistant.

That progress has mostly come in the last 10 years, not 50. Before that we were just doing it wrong.

That doesn't mean we won't do it wrong for the next 30 years, or that we are doing it right right now.

We have to account for accelerating progress, but I'm not sure that's enough to bring us much closer than 50 years away

Sure, we could hit another wall, deep learning could be a false peak, etc. On the other hand, it could just keep going. Tech companies make bets on this, and if they were conservative they wouldn’t be tech companies.