I appreciate your recent anti-super-short timelines posts Ryan and basically agree with them. I’m curious who you see yourself as arguing against. Maybe me? But I haven’t had 2027 timelines since last year, now I’m at 2029.
Some AI company employees with shorter timelines than me mostly. I also think that “why I don’t agree with X” is a good prompt to express some deeper aspect of my models/views. It also makes a good reasonably engaging hook for a blog post.
I might write some posts responding arguments for longer timelines that I disagree with if I feel like I have something interesting to say.
My case against long timelines is based on waiting for algorithmic breakthroughs which Kokotajlo on July 28 believed to have a chance of “maybe like 8%/yr”. Seth Herd replied to my case as follows: “You estimate c by looking at how many breakthroughs we’ve had in AI per person year so far. That’s where the 8% per year comes from. It seems low to me with the large influx of people working on AI (italics mine—S.K.), but I’m sure Daniel’s math makes sense given his estimate of breakthroughs to date”
I didn’t interview any AI company employees, but I conjecture that they are overconfident in their ability to make such breakthroughs.
Not sure how to interpret the question. Some benchmark scores are somewhat lower today than AI 2027 predicted, and our new model takes them into account, so in some sense it’s already diverging, but only very slightly. 2026 should see a big divergence though, one that’s clearly not just noise. And then, obviously, 2027 will look totally different (on the median trajectory).
I appreciate your recent anti-super-short timelines posts Ryan and basically agree with them. I’m curious who you see yourself as arguing against. Maybe me? But I haven’t had 2027 timelines since last year, now I’m at 2029.
Some AI company employees with shorter timelines than me mostly. I also think that “why I don’t agree with X” is a good prompt to express some deeper aspect of my models/views. It also makes a good reasonably engaging hook for a blog post.
I might write some posts responding arguments for longer timelines that I disagree with if I feel like I have something interesting to say.
My case against long timelines is based on waiting for algorithmic breakthroughs which Kokotajlo on July 28 believed to have a chance of “maybe like 8%/yr”. Seth Herd replied to my case as follows: “You estimate c by looking at how many breakthroughs we’ve had in AI per person year so far. That’s where the 8% per year comes from. It seems low to me with the large influx of people working on AI (italics mine—S.K.), but I’m sure Daniel’s math makes sense given his estimate of breakthroughs to date”
I didn’t interview any AI company employees, but I conjecture that they are overconfident in their ability to make such breakthroughs.
What made you update from 2028?
Newer better timelines model mainly. Still working on it. But also, METR’s downlift study, GPT-5 being on trend, various misc other things.
What is the first point at which your new model diverges from the AI 2027 timeline?
Not sure how to interpret the question. Some benchmark scores are somewhat lower today than AI 2027 predicted, and our new model takes them into account, so in some sense it’s already diverging, but only very slightly. 2026 should see a big divergence though, one that’s clearly not just noise. And then, obviously, 2027 will look totally different (on the median trajectory).