Because this paradigm is more compute intensive to run on the margin, it may be more difficult to run many iterations side-by-side. This means rather than having the equivalent of millions of top tier R&D researchers working at the same time we may have tens or hundreds of thousands, this could slow takeoff speed somewhat.
A quick nit pick, inference-time compute does not imply slower take off times, but rather more gradual take off times. This results in higher levels of capabilities relative to the counterfactual world with no inference-time compute because we can now leverage levels of useful intelligence which would have otherwise been unattainable earlier in our timelines.
A quick nit pick, inference-time compute does not imply slower take off times, but rather more gradual take off times. This results in higher levels of capabilities relative to the counterfactual world with no inference-time compute because we can now leverage levels of useful intelligence which would have otherwise been unattainable earlier in our timelines.
https://www.lesswrong.com/posts/AiaAq5XeECg7MpTL7/for-every-choice-of-agi-difficulty-conditioning-on-gradual