Just keep doing what works. Bigger transformers, more data, more compute, more test-time thinking. This is the only pathway with historical extrapolation data (03 - Effective Compute — 10x per year).
Sutton's Bitter Lesson: intelligence as search through hypothesis-space, planning as search through possible futures. More compute → more search → more intelligence.
Strong empirical backbone: scaling laws (Kaplan, Chinchilla) have been remarkably predictive within their extrapolation range.
The paper reframes the usual debate. It is not: will scaling work? — too vague. It is: can scaling be sustained long enough through the required orders of magnitude?
If yes → ASI may arrive on this pathway alone. If no → a paradigm shift becomes necessary.
"Naively supplying brute-force search with more compute fails in virtually all non-toy domains, including chess."
Real scaling success comes from better priors + heuristics + surrogate models that reduce the dimensionality of the search space. The Bitter Lesson is true at the meta level (general methods win) but at the implementation level it's wrong (raw compute alone is not enough).
This is the seam between Pathway 1 and Pathway 2: as scaling hits diminishing returns, paradigm-level innovation becomes a forcing function.
Repeated from 03 - Effective Compute — 10x per year but worth stating again:
If individual model capability plateaus at human-level but effective compute keeps growing at 10×/yr, the surplus compute lets you run millions or billions of human-level AGI instances at superhuman speeds. The paper argues:
"It seems hard to argue that such a leap would not constitute the step change from AGI to ASI, even though each individual AGI system may be at human level."
This is the cleanest case for ASI via collectives without any single instance being superhuman.
The paper's concrete recommendation: forecast carefully and continually. Build, maintain, and ensemble multiple quantitative models. Track every relevant macro-quantity. Reduce uncertainty bands by updating frequently rather than predicting once.