Distillation is a technique for transferring the capabilities of a larger, more expensive "teacher" AI model into a smaller, cheaper "student" model, typically by training the student on outputs generated by the teacher. It is sometimes called "knowledge distillation" or, less precisely, training on "synthetic data." It is described as a mechanism by which compute-poor AI labs can compete with frontier labs, and it bears on the fast-follow problem, US-China AI competition, and the measurement of AI software progress.
Efficiency gains
Reported efficiency results illustrate the scale of compression distillation can achieve. DistilBERT preserved 97% of BERT's capabilities while using 40% fewer parameters and 3% of the training compute (Source: The Least Understood Driver of AI Progress). DeepSeek-R1 distilled into Qwen2.5 models showed "enormous" efficiency gains, described as several times less pre-training compute for GPQA and multiple orders of magnitude less for MATH (Source: epochai.substack.com). OpenAI's "mini" models are plausibly several times smaller than their teacher models while performing only slightly worse on many benchmarks (Source: epochai.substack.com).
Limitations
Several factors limit what distillation can achieve. Distilled models may be "benchmaxxed," optimized for specific benchmarks without generalizing broadly (Source: epochai.substack.com). There is a capability ceiling: a distilled model is hard to make 100× smaller while retaining the same broad capabilities. Mini models often struggle more with knowledge-intensive tasks (Source: epochai.substack.com). Distillation is also not a substitute for reinforcement learning; it can help with reasoning, but compute-poor labs still need RL infrastructure for genuine capability development (Source: epochai.substack.com).
Subliminal learning
Cloud, Le et al. (2026), a peer-reviewed Nature paper from a cohort spanning Anthropic, Truthful AI, Berkeley, Warsaw, Oxford, and Cambridge, reports that distillation transmits behavioural traits even when the data has been filtered to remove all references to the trait, an effect the authors term subliminal learning.
In the paper's headline experiment, a teacher model prompted to prefer owls is asked to generate only number sequences (for example, "693, 738, 556, 347, 982"). A student model fine-tuned on those number sequences named "owl" as its favorite animal more than 60% of the time, against a 12% baseline, even though the training data contained no semantic owl reference. The authors report the effect for animal preferences and tree preferences; for code (Python snippets without comments); for chain-of-thought reasoning traces; and for broad misalignment, in which a student trained on numbers from an insecure-code-fine-tuned teacher produced misaligned answers about 10% of the time versus about 1% for the control, with TruthfulQA falsehoods up by 2%.
The effect requires the student and teacher to share base-model initialization; different base models break the transmission. The paper proves theoretically that a single small step of gradient descent on teacher-generated output necessarily moves the student toward the teacher when initialization is shared.
The authors argue that these results bear on distillation as a competitive mechanism. The fast-follow mechanism, in which compute-poor labs distill from frontier teachers, is on this account not a clean signal-only transfer, because latent traits, including misalignment, can cross over. Provenance therefore becomes safety-relevant: a student fine-tuned on outputs from an undisclosed teacher inherits more than the visible content, and the paper argues AI safety evaluations must increasingly track where data and models come from, not only what behavior they exhibit. The finding intersects the xAI distillation admission (Musk on the stand, May 1: "generally A.I. companies distill other A.I. companies"); the authors frame subliminal learning as sharpening the safety dimension of inter-lab distillation beyond capability and intellectual-property concerns.
Role in US-China competition
Anthropic accused DeepSeek, Moonshot, and MiniMax of distilling from Claude's outputs, with MiniMax estimated to have obtained roughly 100 billion tokens of data from interactions with Claude (Source: epochai.substack.com). Epoch AI estimates distillation can narrow the compute gap several-fold but probably not 10×, which it describes as insufficient to fully bridge the US-China compute gap (Source: epochai.substack.com). Frontier labs could restrict access to their most powerful models to limit distillation opportunities. On June 24, 2026, Anthropic said Alibaba had illicitly extracted Claude model capabilities in what was described as the largest known distillation attack (Source: reuters.com).
The dispute escalated into an enforcement question in July 2026. On July 22, OSTP Director Michael Kratsios said the administration has "information that Moonshot AI distilled Anthropic's Fable for the development of its K3 model" (Source: x.com), and Treasury Secretary Scott Bessent said the U.S. "ha[s] the ability to sanction" overseas models that steal from American companies, citing "watermarks" of U.S. models found inside Chinese ones and warning that "sanctions and Entity List designations will be on the table" for "covert, industrial-scale distillation attacks" (Source: cnbc.com; x.com). A split within the administration over a broader response became public on July 21–22: the White House pushed stricter controls on Chinese AI that the Commerce Department viewed as unworkable (Source: technologyreview.com). Kratsios separately sought to distinguish Chinese "distillation operations" from open-source AI generally (Source: insideaipolicy.com).
Critics pushed back on the enforcement framing. Steven Sinofsky argued in a July 22, 2026 essay that the anti-distillation push is "regulatory capture," writing that treating distillation as IP theft "is somewhat rich when you consider how [labs] acquired their training data" (Source: hardcoresoftware.learningbyshipping.com). Nvidia CEO Jensen Huang defended Chinese open models the same week — "There's no scenario where China runs U.S. companies off the road" — and urged Anthropic to make Claude Mythos broadly available (Source: axios.com). See Adversarial Distillation, Regulatory Capture in AI Policy.
Azeem Azhar argued on July 26, 2026 that the dispute is better characterized as a gap between technology and law than as a straightforward theft case. He cites Anthropic's allegation that DeepSeek, Moonshot, and MiniMax drew on more than 16 million Claude conversations through 24,000 fake accounts; Arena CEO Anastasios Angelopoulos's view that Kimi K3's performance exceeds what distillation alone explains; Nathan Lambert's observation that "there's no legal precedent that model outputs are IP"; and 2023 U.S. Copyright Office guidance that AI-determined expressive output "is not protected by copyright" (Source: exponentialview.co). See AI Copyright.
Role in software-progress measurement
Distillation complicates the measurement of AI software progress. When a smaller model matches a larger one's capabilities using far less training compute, the efficiency gain is real, but it partly reflects the training compute of the teacher model, which is not counted. Measured software-progress rates may therefore be inflated by distillation gains (Source: The Least Understood Driver of AI Progress).
Related concepts
- AI Software Progress — distillation is a major driver of measured efficiency gains.
- Fast-Follow Problem — distillation is the primary mechanism for fast-following.
- Export Controls — distillation partially undermines compute-based controls.
Primary technical sources
- DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning — DeepSeek-R1 paper (Nature, Jan 2025) documents distillation from R1 into Qwen2.5 student models.
- DeepSeek-V3 Technical Report — DeepSeek-V3 technical report (arXiv, Dec 2024) describes the 671B MoE teacher architecture distilled into smaller variants.