Rob Wiblin

And factoring it into a recent update pushed out their timelines by one to two years, which for those folks, given how quickly they're expecting things to progress, that actually is quite a large adjustment.

So what was the upshot of all of these updates I've talked about so far?

At the start of 2025, forecasters on the prediction market Metaculous, they forecast the arrival of strong AGI in July, 2031.

So I guess they were medium timelines folks.

That forecast has since moved out to November, 2033.

So that's a 2.5 year extension that occurred over the course of one year.

which I think is a fair reflection of the mood in the AI industry.

It's not cataclysmic, it's not like an enormous delay, but people definitely became, they started to think things are going to take a couple of years longer than they had at the start of 2025.

One thing worth knowing and squirreling away in your memory, if you haven't heard it already, is that that 2028 to 2032 window might be quite a make or break period for developing AGI or recursively self-improving AI.

The reason for that is that up until that point, the AI industry will have been able to expand the inputs going into it a lot just by sucking in computer chips and electricity and technical stuff that were previously going towards or working on non-artificial intelligence related things.

But if current rates of growth of those inputs continue, by 2032, AI will already be a huge fraction of all of the computer chips and electricity and tech staffing.

There just won't be that much slack left for them to absorb from the rest of the world.

And by that sort of date, financing could even get kind of tricky by then because the next scale up might be a single AI model that costs a trillion dollars to train or $10 trillion to train.

If AI hasn't by that point really blown people away and convinced investors that it's going to replace human labor on a massive scale, there is a lot of money to be throwing at something not knowing whether or not it's going to pan out.

That's not to say that even if it didn't happen by 2032 that we wouldn't go on and develop AGI eventually.

I think almost certainly we would.

we'd probably just be in for a bit of a slower grind.

And it'll be a grind that's driven above all by getting access to more compute, by building more fabs that manufacture computer chips, and by manufacturing more chips and making each of those chips more efficient.

And that is a rapid progress in the scheme of ordinary things, but not as fast as we've been able to expand inputs of all of these different kinds so far.

I've been trying to make sense of why it is that technology people, they brought forward their AGI timelines a whole lot in early 2025, and then in the second half of the year, pushed them out a whole lot again.