Rob Wiblin

speaker

5,091 appearances 15 recordings 1 series first heard Oct 2021 last heard 6 Aug

Rob Wiblin’s voice in public audio — every appearance, attributed to the second.

Trend

recordings per month · last 12 months

Recordings per month over the last 12 months — 13 in all, peaking in Apr 2026 with 3.

Appearances

newest first · ▶ plays the moment

80,000 Hours Podcast · #250 – Toby Ord on where AGI timelines go wrong · 6 Aug 2026 podcast

I guess it's very hard to move institutions and to get them to do anything very quickly.

To focus on the future could be radically different because I guess they have a really strong immune reaction to that idea because you don't want them to turn on a dime.

I guess, yeah, what advice do you give to people in government about how they should approach this?

Or to redivert the lava or whatever.

Yeah.

So in my timelines piece recently, I guess I spent quite a bit of time for now.

I guess there's so much short timeline stuff in the water at the moment.

And I guess, especially among the kind of people who I imagine might listen to 40 minutes about this topic, that I wanted to do a bunch of deflating, I guess, the hype or I don't know, getting people to think, well, it could take longer.

But the flip side, I think of almost everything that I say is that it also could be like quite short.

There's actually quite a number of different pathways.

to which we could get extraordinary capability jumps relatively soon.

I mean, a couple of them are, it could turn out that recursive self-improvement works, even with AIs that don't have a super broad range of skills.

Maybe you only need to train them on, I guess, coding, of course, but also you get them to have some decent research intuition about just AI in particular, and then they start having great ideas and insights that you can train on.

Or they're just very good at...

setting up experiments and basically it's all just like brute force search through different experiments and that kind of pans out.

You could also, it's possible, there's this general phenomenon that the AIs are much better at well-specified, simple, structured tasks than they are at messy ones.

And it's possible that they'll remain bad at the messy tasks for quite a long time to come because it's harder to train.

But it's also possible that by making them superhuman at the well-structured simple tasks, you'll get cross-generalization into the messy tasks and then they'll become roughly human level at that.

And then they're kind of, even their worst areas are human level.

And then you can go a long way from that.

Showing 241–260 of 5,091 · page 13 of 255