We are in what I call the redundancy interregnum. No one has been fired, yet. Everyone using AI and prone to imposter syndrome suspects they should be. Once the bean counters have some good numbers, redundancies will be eliminated. Bet on it. Meanwhile we can continue to speculate.
FYI, the AI Futures Model (model behind AI 2027/2040) *agrees* with you on your Anthropic test of ECI progress. Using current default values the predicted ECI slope in Feb. 2027 is ~1.085x an Apr. 2026 baseline. Its software research pace then is just 1.19x, near the bottom of your sub-log band. (edit: and the AIFM doesn't model release lag either! If we look strictly at Aug. 2026 vs. Apr. 2026, ECI slope is just 1.04x and software research pace is just 1.07x. I'm actually a bit confused about how you're thinking about your test here, is the Apr. 2026 baseline not including the 6-month release lag?)
If you go to the interactive model website (aifuturesmodel.com) you can see where you differ from it though; the AIFM predicts that ECI rapidly increases only in 2028, after we get fully automated coding. (That's basically because it predicts that AI agents at that level of capability gain better "research taste", aka they choose better experiments, which makes rapid progress possible.)
Just toying around a bit with putting your stance into their model (which rates research taste highly), we can keep λ at 0.50 to be conservative, move the Infinite Coding Labor Asymptote from 15x faster progress to just 5x faster progress, slow down improvement in research taste from 3 SD/yr to 1 SD/yr, and drop research taste at the time of Automated Coding from 0 SD to -1 SD (combined with previous, means AI only matches top human level in ~2031).
("SD" = standard deviations in research taste among humans, ex. a top researcher would be 3 SD above a median one; also "year" here is technically a "progress-year" which is equivalent to 2025's growth of effective compute, which is actually noticeably higher than most later years in the default model settings (we can't possibly keep growing compute at current rates))
which says that while we still get automated coding next year in 2027, we don't get superintelligence until nearly 2050! Does that sound like your own stance?
We are in what I call the redundancy interregnum. No one has been fired, yet. Everyone using AI and prone to imposter syndrome suspects they should be. Once the bean counters have some good numbers, redundancies will be eliminated. Bet on it. Meanwhile we can continue to speculate.
Wow!
FYI, the AI Futures Model (model behind AI 2027/2040) *agrees* with you on your Anthropic test of ECI progress. Using current default values the predicted ECI slope in Feb. 2027 is ~1.085x an Apr. 2026 baseline. Its software research pace then is just 1.19x, near the bottom of your sub-log band. (edit: and the AIFM doesn't model release lag either! If we look strictly at Aug. 2026 vs. Apr. 2026, ECI slope is just 1.04x and software research pace is just 1.07x. I'm actually a bit confused about how you're thinking about your test here, is the Apr. 2026 baseline not including the 6-month release lag?)
If you go to the interactive model website (aifuturesmodel.com) you can see where you differ from it though; the AIFM predicts that ECI rapidly increases only in 2028, after we get fully automated coding. (That's basically because it predicts that AI agents at that level of capability gain better "research taste", aka they choose better experiments, which makes rapid progress possible.)
Just toying around a bit with putting your stance into their model (which rates research taste highly), we can keep λ at 0.50 to be conservative, move the Infinite Coding Labor Asymptote from 15x faster progress to just 5x faster progress, slow down improvement in research taste from 3 SD/yr to 1 SD/yr, and drop research taste at the time of Automated Coding from 0 SD to -1 SD (combined with previous, means AI only matches top human level in ~2031).
("SD" = standard deviations in research taste among humans, ex. a top researcher would be 3 SD above a median one; also "year" here is technically a "progress-year" which is equivalent to 2025's growth of effective compute, which is actually noticeably higher than most later years in the default model settings (we can't possibly keep growing compute at current rates))
That shakes out to:
https://www.aifuturesmodel.com/p?v=2&base=daniel-08-16-26&artc=-1&arts=1&ila=5
which says that while we still get automated coding next year in 2027, we don't get superintelligence until nearly 2050! Does that sound like your own stance?
"The bearing of a child takes nine months, no matter how many women are assigned."---The Mythical Man-Month, Frederick P. Brooks, Jr., 1975