Historical Behavior

Lilian Weng's historical behavior at OpenAI reflects a "safety-first" characteristic. From 2019 to 2024, she led the safety systems team, developing OpenAI's safest model o1-preview (resistant to adversarial attacks while maintaining practicality). She left in November 2024, citing a desire to "explore new things after 7 years," but returned in July 2026, focusing on "recursive self-improvement." Compared to Ilya Sutskever (who left in 2023 due to safety disagreements to found Safe Super Intelligence) and Jan Leike (who left in 2024 over safety concerns, criticizing OpenAI's "decline in safety culture"), Weng's "departure - return" path is rarer. Her historical behavior reflects the "trade-off between efficiency and safety for safety researchers."

Capital Path

The capital impact of recursive self-improvement is clear: OpenAI's R&D expenditure in 2026 is projected to be $18 billion, with 40% allocated to "efficiency optimization" (e.g., GPT-5.6 Sol self-optimization, reducing service costs by 20%) and 60% for "model scaling" (GPT-6, GPT-7). If recursive self-improvement succeeds, R&D efficiency could improve by 50% (AI automatically optimizing training pipelines), reducing R&D expenditure to $9 billion, but requiring an initial investment of $5 billion (2026-2028). Funding is shifting from "human development" (1,200 engineers, $500,000 annual salary per person, $6 billion annual expenditure) to "AI development" (AI automation, marginal costs approaching zero). In the long term, AI R&D efficiency could increase tenfold, with model iteration cycles decreasing from 18 months to 6 months.

Analogy + Industry Position

Similar cases include: in the 2010s, "AutoML"—Google launched AutoML in 2015, automating machine learning model design, which replaced 30% of manual tuning work by 2020. In the 2020s, "AI for Science"—DeepMind's AlphaFold predicted protein structures in 2021, potentially replacing 50% of biologists' work by 2026. Currently, AI R&D is in an "AI self-improvement explosion phase"—compared to 2023 (90% manual optimization) and 2026 (40% AI optimization), by 2030, AI optimization is expected to reach 80%. The industry is at an "AI R&D automation inflection point"—OpenAI, Anthropic, and Google are all positioning themselves, but OpenAI's "recursive self-improvement" is more aggressive (AI directly optimizing training pipelines rather than just tuning parameters).

Structural Judgment

This fundamentally represents a technological substitution leading to a reconstruction of the R&D paradigm. Traditional AI R&D follows the cycle of "human design → training → evaluation → iteration," taking 18 months and costing $1 billion. Recursive self-improvement shifts this to "AI design → AI training → AI evaluation → AI iteration," reducing the cycle to 6 months and the cost to $100 million. This explains why Weng returned to OpenAI—not for "safety first," but for "efficiency first." When AI can self-improve, the value of human development decreases by 80% (from 1,200 people to 240), but the value of "safety oversight" increases (as humans must ensure AI does not self-destruct). This drives the "polarization of AI R&D"—high-end labs (OpenAI, Google) use AI self-improvement, while low-end labs (startups) rely on human efforts, with the middle layer disappearing. As AI R&D efficiency improves tenfold and model iteration cycles shorten threefold, "speed" becomes a new barrier, and "safety" becomes a new bottleneck.

ABAB News · Cognitive Laws

The marginal cost of AI developing AI approaches zero.
The ultimate dilemma for safety researchers is the trade-off between efficiency and safety.
Recursive self-improvement is the endgame of AI development.