Huh, I think continual learning would be a pretty big deal w.r.t. AI danger. Can you say more about why it wouldn't?  Seems like it would dramatically increase horizon lengths for example.

Back when many expected takeoff in 2027 or so, it was pretty reasonable to assume that the probability of a conflict entirely unrelated to AI was low.^[3] But the forecasters behind AI 2027 now expect takeoff in the 2030s.

I'm saddened that this is the takeaway from our new model! It seems misleading. Here is a graph of my+Eli's timelines over time:

https://x.com/DKokotajlo/status/1992316620254155028

ChatGPT goes on to say that I outperformed Rick (who also made predictions in 2020 in the spreadsheet.) However, looking over the data briefly, I'm not sure I agree with some of the scores, e.g. are there really robotaxis in 20+ cities now? And drone delivery?

Thank you! I continued your ChatGPT conversation and got the following tidbits of extra info:
 

For my 2020 predictions, I asked it to summarize my biggest sources of Brier loss:

Top 5 Brier losses from FALSE outcomes (overconfidence)

Top 5 Brier losses from TRUE outcomes (pessimism)

The IMO prediction alone is your single largest loss across all 2020 items.

Interestingly the two biggest hits to Brier score I took (on jan 1 2020) were from underestimating LLMs basically.

Also interestingly, it seems like maybe I took significantly more Brier loss from saying things wouldn't happen, that did happen, then from saying things would happen, that didn't? In other words I was too cautious overall rather than too credulous?

Well done for making these predictions and scoring them!

Well done for making these predictions and scoring them!

A major plausible class of worlds in which we don't get superhuman coders by end of 2026 is worlds where the METR trend continues at roughly the same or only slightly greater slope than the slope it had in 2025. Right?

Insofar as zero is significantly smaller than epsilon, yes.