Twenty years today we will know if the Y2038 problem has had more impact on human civilisation than the Y2k one. Of course, it may be necessary for the Eschaton to save us one second before the clocks roll over. So I predict the Singularity at exactly 03:14:06 UTC on Tuesday, 19 January 2038.
How possible is that? Conveniently enough for me, I will only be 69 in January 2038 and perhaps 70 will be the New 50. So, as a date, it is very much in the scope of my plausible lifetime. Of course, there will be a lot of getting there. Back at Intersection in 1995, I wrote of 1995-2005 as the research decade, 2005-2015 as the development decade and 2015-2025 as the implementation decade. After 2025, all bets are off. For A(G)I, that's perhaps not so far. The largest publicly listed companies in the world are explicitly A(G)I companies and one of them barely existed in 1998 (and not all in 1995) and the other is profoundly changed (and that leaves aside Microsoft and car companies; what about ExxonMobil?). For molecular nanotechnology, not so much, but I'm hoping to start see progress over the next few years and we do have techniques like CRISPR as well as OLEDs and graphene. And we've still got (nearly) 20 years.
Of course, we were talking about the Singularity in the mid-1990s or even the mid-1980s (Marooned in Realtime) or even the early 1970s (Gravity's Rainbow) or even the late 1950s (Ulam). AI has we know it today has been around for about 70 years now. We might still be another 70 from anything interesting. Or not. There's a lot of money being pumped into A(G)I research and development right now. Sure a lot of that money will be wasted and a lot of it will go either on fairly mundane stuff or stuff that is seen as the topic of the moment. A(G)I is hot right now, but in two or three years we might be firmly mired in the slough of broken dreams. But it is not as tough A(G)I is suddenly going to become a bad idea. It's been around for 70 yeats through several AI winters. It's always going to be a good idea.
Probably the most interesting developments for the last few months have been the AlphaZero work and the research on evolving neural nets. Evolving nets offer another paradigm to backprop and possibly a way of more quickly getting novelty in systems. AlphaZero demonstrates that it is possible to achieve superhuman performance overnight, albeit in a limited domain. But three years ago, we were wondering whether go would ever fall to AIs. Chess might have been the Drosophilia of AI back in the 1960s, but now it's StarCraft. Will that fall before 2038? I'd wager it will. Civilization would be an interesting and simpler testbed. It's the thing one would like to do research on if one could. It would be great in Civilization fell before the end of the decade.
It's easy to say that A(G)I is always 20 years away. 20 years isn't really such a long time. But 70 years is and that's a blink on the scale of human much less planetary history. OK, perhaps not in 2038, but what 2088? 1988 was very different from 1918 even if there was no Google or Wikipedia or iPhone yet. And thus 2018 is really different from 1948, so something happened between 1988 and now.
If you were about ti start a computer science degree (or, just as good, natch, a physics one), you would have four years of undergraduate and a year of master's degree in AGI before starting your PhD in 2023, when things could already be rather different. Bu the time you finish your PhD, 2026-7, you'd be well positioned to contribute to the big push towards a 2038 Singularity. We might 3-4 years from cracking StarCraft. Combine with advances in knowledge engineering and language processing and neuromorphic chips and a couple of other breakthroughs I haven't thought of. 5 or 6 breakthroughs taking 3-4 years. That's 20 years. We could get there.