Main factors for this
Valuations + Revenues of Labs (OpenAI, Anthropic showing ~exponential growth in ARR and funding)
Overall economic status (employment, GDP growth etc)
Benchmarks (e.g. METR, ECI, AA) which have generally improved at a roughly constant rate with that rate increasing (roughly double) since mid 2024
Leaving Other for A. Sublinear (e.g. fraudulent, waste of money) bubble pops (idk FTX, NFT, type things). Though I think bubble popping similar to dot com might still be linear growth overall, just that the expectation of exponential leads to valuations popping, and other things like AI takeover (seems more aligned with Hyperexponential since we’re discussing AI progress rather than human progress) and or pausing AI development “early” idk, basically anything that doesn’t fit neatly into
Linear: if AI progress is linear we’d expect GDP growth to stay around 3% and a shift in AI spend from new training models (marginal gains that are more linear than exponential, I think Demis said the progress hasn’t been fully exponential in some interview after Gemini 3, but still worth it, so if that continues more similar to the progress in smart phones from 2020-2026 after Moore’s law is “dead” and the improvement nears some dead end where the “embrace exponentials” Karpathy said in Feb 2025 is no longer applicable and AI progress is more dormant and the focus goes away from general into specific domains (some signs of this with the focus on coding now, though the gains seem to be worth it e.g. with news of 5.3 Codex being instrumental to develop itself) and 5.4 doubling (10% to 20%) on some AI benchmark, though still stuck at 4% on the OpenAI proof QA (AI explained went over both these in his recent 5.4 video)
Exponential - this would generally imply there’s some real impact on GDP growth. idk if AI would become deflationary due to job loss / inequality, so may need some metric that’s better at measuring productivity, essentially Linear means AI is just another technology, whereas if it’s exponential, it is THE technology, the reason I went with 2028-2033 since that is when the data center buildouts would likely reach some level of saturation with AI being almost fully deployed into the power grid and economy, so the growth in that 5 year period isn’t pure scale of raw compute, though algorithmic improvements may allow scaling up to continue, is that rate (I think I heard like the algorithmic improvements is 4x per year and the hardware is something like 10x per year idk I might have made that up, I think hardware is probably closer to 2x so 10x overall? I’m referring to the charts re:Jevons paradox and the price of GPT-4 level going down 100x in 2 years or something. Most people (on Manifold at least) expect AGI in the 2028-2033 time period so it seems that if the rate of progress is linear skeptics would be right that there’s no AGI by 2033 (unless you think what we have in 2026 already counts), whereas the majority is counting on trends to continue (exponential) when forecasting 2028 to 2033
Hyperexponential roughly corresponds to Aschenbrenners “Intelligence Explosion” or perhaps country of geniuses in a data center where there is some step change-like event rather than smooth progress, e.g. since we saw a ~doubling comparing 2022 to 2024 and 2024 to 2026, maybe it’ll be hyperexponential if it continues to double a few more times since it does seems the releases are in rapid succession, e.g. ~3 years from GPT3 to 4, ~2 from 4 to 5, ~1 from 5 to 6, 0.5 from 6 to 7, assuming each +1 continues to mean +1 OOM (10x) in someway (e.g. effective compute as show in the intelligence explosion chart)
FYI: the motivation for this market was after looking at the comments (just started watching) this video
https://youtu.be/dUFcJdVQOjs?si=5MX8gs5rN0RhGWH0
It’s clear Derek Thompson is more in the exponential camp, while Mina is in the linear camp and the comments are critiquing Derek for his lack of nuance (note: I haven’t watched it fully yet, this is what I’m expecting based on comments like
1,000