The paradigm of AI research has undergone a fundamental shift in the past decade. Early reliance on kernel methods, Bayesian approaches, and strict statistical guarantees emphasized provability and rigorous processes; the current mainstream is dominated by Rich Sutton's "bitter lesson," which states that general methods combined with large-scale computing power ultimately surpass carefully designed feature engineering.

The capital path is clear: funding has shifted from theoretical and algorithmic innovation to large-scale data centers, GPU clusters, and larger datasets. Companies like xAI, OpenAI, and Google have adopted "larger models + more data" as their core competitive strategy, motivated by empirical performance (benchmarks and practical usability) replacing theoretical elegance as the primary evaluation standard.

Similar cases can be seen in the complete replacement of traditional machine learning by deep learning, and the current expansion from pure language models to multimodal and world models. The industry is at a stage where "scale remains the most reliable leverage," with theoretical work serving more for post-hoc explanations rather than pre-design.

This is essentially a technological substitution. The mechanism is that when the costs of computing power and data drop below a certain threshold, the marginal returns of brute-force search and scale expansion continue to exceed those of fine theoretical design, leading to a cultural shift in R&D from "proving correctness" to "running it first and explaining later."

ABAB News · Laws of Cognition

Methods that can run will ultimately defeat more elegant but non-functional methods.
When computing power is sufficiently large, theoretical thresholds will be redefined.
The industry's true belief has shifted from "proof" to "scale."