DeepSeek's new checkpoint cuts CoT token use by up to 10x, sparking Jevons paradox talk
teortaxesTex · x · 2026-10-17
- User testing suggests DeepSeek's newly detected checkpoint (flash 1017, dated Oct 17) is far more token-efficient in its chain-of-thought, using up to an order of magnitude fewer tokens on some tests.
- The image analysis lists four checkpoint fingerprints: v4-pro, v4f 0908 (internal beta), flash 0910 (unified), flash 1017 (new), and compares token usage across providers: DeepSeek official API 35,411 total tokens, Huawei Cloud MaaS 549,662, Baidu Qianfan 2,452,207 — though the comparison doesn't control for checkpoint, request, or caching differences.
- teortaxesTex argues DeepSeek is fighting CoT bloat, predicting Jevons paradox effects and possible overload of "Whale" servers, with other Chinese labs unprepared.
More from Infra
- ChinaTalk: Saif Khan argues DUV controls could leave China with 100x fewer AI chips by mid-2030s — ChinaTalk · 2026-10-17
- Cutting AI inference bills from $250K to $7.6K/month: Eugene Cheah's tokenomics 101 — AI Engineer · 2026-10-17
- Ben Horowitz says Amazon is sold out of compute through end of 2028 — RihardJarc · 2026-10-17
- Free 8-hour AI data center course launches on freeCodeCamp, from grid power to GPU economics — trawasthi_ai · 2026-10-17
- Broadcom Wi-Fi bug enables passwordless remote code execution; ~1.5M Belgian routers unpatched — jedisct1 · 2026-10-17
- Local LLM hosting isn't always cheaper: cached-token rate decides the break-even point — paq85 · 2026-10-17