RAND researcher on DeepSeek's real lesson, export control failures, and why techies should move to DC

Summary

The DeepSeek moment was less a technical revelation than a perception failure rooted in missing context. Because US labs publish almost nothing about training costs, DeepSeek's $6 million figure landed as a shock against Sam Altman's widely cited $100 million GPT-4 training claim, even though those numbers measure entirely different things. Dario Amodei addressed this directly in a post-DeepSeek blog post, noting Anthropic could match the same efficiency, confirming what close market observers already understood: training costs have been falling steeply and consistently. The real lesson from DeepSeek, per Leonard Tian, a RAND-affiliated AI policy researcher with an electrical engineering background, is not that China leapfrogged the US, but that the best publicly available model briefly came from China while leading US models remained behind closed doors.

DeepSeek's Origins and Open-Source Strategy

DeepSeek sits inside a Chinese hedge fund that acquired roughly 5,000 to 10,000 A100 GPUs in 2022, before export controls tightened. One credible theory is that Beijing's regulatory crackdown on high-frequency trading and the broader tech sector left the firm with surplus compute and no clear use for it, prompting a pivot into AI. The fund's background in low-level kernel optimization, a capability hedge funds develop to squeeze every performance point from hardware, gave DeepSeek engineers an edge that showed up clearly in their published code.

The open-source release was almost certainly not a state-coordinated influence operation. DeepSeek had a pre-existing public commitment to releasing model weights. The confluence of factors that made the launch explosive, dropping during the first week of Trump's second term, releasing model weights openly, publishing a cost figure, and being among the first public reasoning models, was largely coincidental. ChatGPT had a reasoning model available at the time, but it was gated behind a premium tier, making DeepSeek many users' first exposure to chain-of-thought reasoning.

The Open-Source Risk: Sleeper Agents and Propaganda

The geopolitical concern around DeepSeek is not primarily about capability parity. It is about what is embedded in the model. DeepSeek demonstrably misrepresents CCP-related topics, and the broader risk is that developers in third-party countries building on top of it, education apps, government tools, enterprise software, inherit those distortions without knowing it. More acutely, Anthropic's published research on "sleeper agents" shows it is technically feasible to embed models with hidden behaviors that activate under specific conditions, such as generating vulnerable code when certain topics arise. Detection is difficult because defenders do not know what triggers to look for. This vector applies to any open-weight model but is materially harder to dismiss when the model originates from an adversarial state.

Export Controls: Structural Failures and the Diffusion Framework Limbo

The 2022 export control rules on AI chips contain a well-documented technical flaw. Regulators defined restricted chips using two parameters: total processing performance (FLOPS) and interconnect bandwidth. The threshold was set incorrectly, allowing Nvidia to engineer chips that sat just below the bandwidth limit while retaining high compute throughput, effectively equivalent to the restricted hardware. That is the chip class DeepSeek trained on. Regulators reportedly knew within roughly a month that the thresholds were wrong, but it took approximately a year to issue a fix, a delay Tian attributes directly to insufficient technical expertise inside the relevant agencies.

A separate enforcement failure involved Huawei obtaining approximately 3 million chips produced by TSMC through shell companies, in direct violation of controls barring both Huawei specifically and large AI chips for Chinese entities broadly. The Biden administration responded with a foundry due-diligence rule requiring TSMC to more rigorously verify end customers. Separately, the Wall Street Journal reported that individuals were physically transporting hard drives on aircraft between data centers in Malaysia to train models, exploiting ambiguity around third-country compute. The diffusion framework, Biden's broader attempt to control chip flows to intermediate countries including Gulf states and Malaysia, is currently unenforced. Commerce Secretary Lutnick and trade official Kessler have publicly stated they will not enforce it pending a replacement rule. The framework has not been formally revoked, leaving an undefined gap.

AI in Active Conflict

The most concrete current application of large language models in military contexts is intelligence processing at scale. LLMs can ingest bulk intercept data, including compromised device communications, and surface targeting-relevant patterns that keyword search would miss, including coded language. Tian pointed to Israel's recent precision strikes as consistent with this kind of AI-augmented intelligence workflow, though he stopped short of claiming direct knowledge of the specific systems involved. Autonomous drone applications involve narrower computer vision models rather than frontier LLMs, but the two capability sets are converging.

The Case for Technologists in Washington

The newly formed Army Reserve Attachment 2011, Executive Innovation Corps, which inducted a cohort of technologists on June 13, 2025, reflects a broader push to bring technical talent into government AI work. Tian argues the leverage for a technically fluent person in DC is unusually high because the baseline is low. A computer science undergraduate who can explain the difference between distillation and pretraining, or identify a flawed parameter threshold in an export control rule, is operating at the frontier of institutional knowledge in most policy settings. The export control errors of 2022 are his primary example of what that gap costs: a year of regulatory lag on a fixable mistake that China exploited directly. Tian noted he is actively hiring for this kind of work.