Stanford's Chris Potts on "tokenflation": token usage may be outpacing the value it buys
ChrisGPotts · x · 2026-09-10
Stanford professor and Big Spin co-founder Christopher Potts joins the TWIML podcast (episode 776) to discuss AI tokenomics and his research into "tokenflation" — the possibility that token usage is growing faster than the measurable value those tokens produce.
Key points:
- Measuring the real return on AI spending: benchmarks alone give an incomplete picture of model progress
- Expert users get better results by challenging and iterating with models; AI fluency directly affects outcomes
- How inference-time scaling reshapes the economics of capable models, and why more efficient architectures could change underlying cost structures
- Also covers DSPy, interpretability, the limits of Transformer architectures, and opportunities for more fundamental innovation
Related materials include "The Decline of Token-Level Purchasing Power" and "The Mystery of Opus 4.6's Sudden Tokenflation".
More from AGI Musings
- Anthropic's risk statement slammed for self-praise over frank talk on AI risks — ShakeelHashim · 2026-09-10
- Should Anthropic researchers quit loudly? AI safety circle debates — JMannhart · 2026-09-10
- Quarter of AI researchers put human extinction risk above 25%, survey finds — KatjaGrace · 2026-09-10
- Beff Jezos: p(1984) far exceeds p(AI Doom), doomers are useful idiots — beffjezos · 2026-09-10
- Beff Jezos: AI regulation push aims to ban open source and nationalize Anthropic as a weapon factory — beffjezos · 2026-09-10
- Jacob Coxon's post tops 120M views as AI safety wind suddenly shifts — scottleibrand · 2026-09-10