Atreides Management's Gavin Baker and creator Teortaxes debate whether Chinese AI labs hold a post-training advantage

this statement about Chinese distillation true.....and his other statements

Thoughts

Yes, Gavin Baker’s core statements on Chinese AI distillation (from the June 15, 2026 TBPN podcast clip) are substantially true and well-supported by evidence from U.S. AI companies, the White House, and independent analyses as of mid-2026. They capture a real dynamic in the U.S.-China AI race, though with some nuances on wording like “entirely” and “open-weight.” His other related points on hardware/export controls and the narrowing/widening gap also hold up reasonably well.

http://podcastalpha.substack.com

Distillation ClaimsBaker argues that China’s (especially DeepSeek’s) “world-class” distillation advantage depends heavily on access to U.S. frontier models. Once U.S. labs stop releasing them (as open-weight or via accessible APIs) and harden defenses, the advantage fades. He cites specifics like DeepSeek’s original model needing only ~160,000 reasoning traces from OpenAI’s o1/o3 models, plus Chinese labs running an estimated ~100,000 API endpoints in parallel for large-scale extraction (likened to “iPhone farms” for mass querying).

http://podcastalpha.substack.com

This is strongly corroborated:OpenAI, Anthropic, and Google have publicly documented “industrial-scale” distillation attacks by Chinese labs (including DeepSeek, Moonshot AI, and MiniMax). Examples include creating >24,000 fake accounts and running >16 million interactions with models like Claude to extract capabilities and reasoning traces for training cheaper competitors.

http://cfr.org

OpenAI sent a memo to U.S. lawmakers accusing DeepSeek of systematically targeting U.S. models (including via code to programmatically extract outputs for distillation).

http://youtube.com

The White House (NSTM-4 memo, April 2026) accused Chinese actors of deliberate, industrial-scale campaigns to distill U.S. frontier models.

http://cfr.org

U.S. labs are actively hardening models against this (e.g., Anthropic’s work on detecting/preventing distillation attacks) and restricting overseas access to top models (e.g., signals around Mythos).

http://podcastalpha.substack.com

Baker’s podcast clip and related commentary emphasize that this creates an “expiration date” for the advantage once access dries up. He has publicly stood by this view recently, noting that even a small number of high-quality reasoning traces can make a big difference and highlighting U.S. open-weight releases close to the frontier (e.g., Nemotron).

@GavinSBaker

Nuances/caveats:It is not entirely dependent on U.S. models—DeepSeek and others have shown genuine innovations (e.g., pure RL for reasoning in R1 without heavy SFT, efficiency optimizations, and self-distillation to smaller models on bases like Qwen or Llama). They also build on legitimate open-weight models. The main accusations involve API querying of closed U.S. frontier models (GPT, Claude series), not solely “open-weight” releases (though U.S. labs like Meta have released strong open-weight models that factor in). Baker has clarified the importance of open weights in responses.

@GavinSBaker

Distillation is a legitimate ML technique; the issue is the scale, unauthorized extraction via fake accounts, and IP implications. Chinese sources often frame their progress differently (efficiency, domestic data, talent).

DeepSeek’s later models (e.g., V4 in 2026) remain highly competitive on benchmarks—especially cost and certain reasoning/coding tasks—but evaluations (e.g., CAISI/NIST) place the PRC frontier ~3–8 months behind leading U.S. closed models.

http://nist.gov

Export Controls and HardwareBaker’s point that export controls (no full H100s/equivalent frontier chips for large-scale training) are a second major pressure, limiting hardware catch-up, is accurate.

http://cfr.org

U.S. controls on advanced AI chips (H100/H200/Blackwell levels) to China remain restrictive. There have been some case-by-case approvals (e.g., limited H200 exports under Trump admin changes in early 2026 with tariffs, volume caps, and security reviews), but frontier training-scale access is heavily constrained.

http://semiconductorsinsight.com

Chinese domestic alternatives (Huawei Ascend) lag significantly and the gap is projected to widen. Smuggling occurs (e.g., reports of DeepSeek using banned chips), but it is costly, risky, and insufficient for parity at scale.

http://cfr.org

This limits massive pre-training clusters, pushing reliance on efficiency, distillation, and smaller/more targeted training.

Overall Gap and Investment ImplicationsBaker’s view that “the gap was always narrower than it appeared. And it’s widening again” aligns with the evidence. DeepSeek’s early 2025 releases (R1) shocked markets with strong open-weight performance at low cost, narrowing perceptions of the gap. However, by mid-2026:U.S. labs maintain the lead at the absolute frontier (closed models). Chinese strengths are pronounced in open-weight models, cost-efficiency (often 5–10x+ cheaper), and rapid iteration via distillation + ecosystem effects. U.S. responses (anti-distillation tech, access restrictions, controls) are shifting the trajectory.

http://nist.gov

This challenges investment theses assuming rapid China-U.S. AI parity or easy Chinese catch-up. Baker’s framework (shared via Podcast Alpha’s Substack link in the original post) highlights risks for those betting heavily on Chinese equivalence soon.Bottom line: Baker’s statements are not hype or baseless—they reflect documented accusations, technical realities of distillation, and policy/hardware constraints. They are a credible investor’s analysis of shifting dynamics in the AI race. Chinese labs have real capabilities and momentum (subsidies, talent, open ecosystem), but the described dependencies and pressures are real and increasingly addressed by the U.S. side. The situation evolves quickly; as of June 2026, the “pipeli