Arvind Narayanan argues that the belief that AI companies cannot slow down unless forced is rooted in a “mythical” conviction that superintelligence and recursive self-improvement are imminent and transformative. He says employee warnings about AI risks should not automatically be interpreted as evidence of responsible intentions; instead, they may reflect the same assumptions driving aggressive and potentially irresponsible development.
Narayanan contends that AI adoption is progressing much more slowly than model capabilities, meaning intelligence is often not the main barrier to economic impact. He doubts that any specific technical milestone will suddenly deliver major economic gains or a decisive national advantage. Consequently, he argues that OpenAI and Anthropic could benefit financially by slowing the capability race and investing more in product usability, enterprise integration, partnerships, and forward-deployed engineering. He notes that other major AI companies, including Google and firms operating just behind the frontier, are performing well financially without pursuing the same extreme strategy.
Commenter Devin Chesney adds that the companies’ enormous capital expenditures and weak unit economics may require an implausibly large future economic transformation—or a “digital god”—to become profitable, suggesting they are trying to justify investments made under quasi-religious assumptions. Ramesh Gopalan argues that competitive fear, especially concerning superintelligence and lower-cost open-weight models, keeps companies racing despite diminishing advantages. He compares this dynamic to the Cold War arms race, in which the United States and Soviet Union continued accumulating weapons despite recognizing their catastrophic potential.
Article timeline
Build this article’s timeline
Extract dated events, key actors, and open questions using the same timeline analysis available in Signal.
Ezra Klein, Devin Chesney, Ramesh Gopalan, and OpenAI
🤖 AI prefill
📰What
Arvind Narayanan argues that the belief that AI companies cannot slow down unless forced is rooted in a “mythical” conviction that superintelligence and recursive self-improvement are imminent and transformative.
🤖 AI prefill
📅When
1949 and Almost 30 years after 1949
🤖 AI prefill
📍Where
China
🤖 AI prefill
💡Why
Centers on recursive self-improvement, superintelligence, and AI capability race
🤖 AI prefill
Login to answer and earn XP for each gap you fill.
AI Superintelligence Racesuperintelligence, recursive self-improvement, capability race, ASI, frontier modelsAI Business Economicscapex, unit economics, profitability, training costs, financial incentivesStrategic Decelerationslow down, unilateral restraint, risk reduction, capability race, corporate self-interestAI Product Adoptionusability, enterprise adoption, partnerships, Forward Deployed Engineers, existing productsOpen Models And Arms Racesopen-weight models, China, frontier performance, cost efficiency, United States
Key Phrases:recursive self-improvement98%superintelligence98%AI capability race96%AI adoption lag93%frontier AI models91%enterprise AI adoption90%Forward Deployed Engineers86%AI training race89%AI unit economics88%AI capital expenditures84%
Manipulation Score: MediumHow much this article uses emotional framing vs. straight reporting
Loaded Language5825%
Absolutist3215%
Threat/Panic2825%
Us vs Them3815%
Engagement Bait1820%
Manipulation score look right?
Factual Claims (10)
mediumtechnology
The view that AI companies cannot slow down unless forced is rooted in belief in superintelligence at OpenAI and Anthropic.
mediumtechnology
Employee alarm at OpenAI and Anthropic is not evidence of beneficence and reflects an underlying fallacy.
mediumtechnology
OpenAI and Anthropic believe that racing toward recursive self-improvement and superintelligence is crucial for their companies, their country, and humanity.
mediumtechnology
AI adoption lags far behind AI capabilities.
mediumtechnology
Intelligence is rarely the real bottleneck in AI adoption.
higheconomic
No particular AI milestone is likely to produce sudden economic benefits or a decisive strategic advantage.
mediumeconomic
It would be financially rational for OpenAI and Anthropic to slow down unilaterally.
mediumeconomic
Other major AI companies are close to the frontier and are doing extremely well financially.
hightechnology
Open-weight models, mostly but not all Chinese, achieve about 95% of frontier performance at 10% of the cost.
mediumgeopolitical
For nearly 30 years after 1949, the United States and Soviet Union built increasingly large numbers of bombs despite knowing that a small fraction could kill all humanity.
Comments
Login to view and post comments.