Implications: what distillation means for model originality, economics, and who controls AI capability

Distillation is technically a training method. It is also an economic and philosophical disruption. When Stanford demonstrated that $600 could produce a model that matched GPT-3.5 on instruction following, it changed what it meant to have a frontier model. When DeepSeek distilled R1’s reasoning into a 7B model that beats GPT-4o on MATH, it changed what “frontier” means. These are not just benchmark stories. They are stories about who controls AI capability and why.

The originality question

A student trained on a teacher’s outputs learns to produce outputs that resemble the teacher’s. Is that original? Two positions:

The compression-as-originality view: The student does not copy the teacher. It builds a compressed internal model of what the teacher knows. No two students with different architectures, trained on the same teacher data, produce identical outputs. The student has generalized from examples, not copied them. In this view, a distilled model is as original as a human who learned to write by reading Shakespeare — the influence is clear but the mechanism is generalization, not reproduction.

The derived-work view: The student’s capability is fundamentally derived from the teacher. Without the teacher’s outputs as training data, the student would not have the capability. The student has not discovered knowledge independently — it has compressed knowledge the teacher already had. The originality is the teacher’s; the student’s contribution is efficiency.

Neither view is obviously right. The empirical reality is that distilled students do generalize beyond their training data in ways teachers cannot fully predict. But they also have systematic capability gaps that track the teacher’s weaknesses — they have inherited not just the teacher’s knowledge but its blind spots.

For most practical purposes, the originality question matters less than the capability question: what can the student do, and where does it fail? But for AI theory and legal frameworks, the question matters considerably.

The knowledge moat problem

A “knowledge moat” is a competitive barrier built on proprietary knowledge that competitors cannot easily replicate. In traditional industries, this takes the form of trade secrets, specialized expertise, or unique data. For AI companies, the equivalent is model capability: if OpenAI has GPT-4 and no one else does, OpenAI has a moat.

Distillation attacks this moat structurally. The Alpaca result showed that a GPT-4 class model’s capabilities can be partially transferred to a small open model for hundreds of dollars. The question is not whether distillation erodes moats — it clearly does — but how fast and how completely.

The limits: distillation captures what the teacher expresses. A teacher used carelessly can transfer more than intended. A teacher used with specific system prompts designed to elicit reasoning can transfer reasoning ability. But distillation cannot transfer capabilities the teacher doesn’t have, cannot transfer the teacher’s training process (which is what actually built the capability in the first place), and cannot transfer capabilities the teacher has but strategically withholds from its outputs.

This is why frontier labs have moved from “our model is better” moats to “our ecosystem is better” moats: better fine-tuning tools, better deployment infrastructure, better enterprise agreements, better safety guarantees. Raw model capability is increasingly hard to protect.

First-mover advantage, eroded

First-mover advantage in AI has traditionally meant: if you train the best model first, you capture the market before competitors can catch up. Moore’s Law-era intuition says hardware and data compound over time, so being first matters a lot.

Distillation changes the compounding math. When GPT-3 was released in 2020, replicating it required replicating its training: 175 billion parameters, 300 billion tokens, the full RLHF stack, and the infrastructure to run it. That took competitors 12-18 months. When GPT-4 was released in 2023, the wait for a comparable open model was approximately 2 months (Vicuna). Not because competitors trained faster, but because they distilled.

The asymmetry: building a frontier model is orders of magnitude more expensive than distilling from it. A $100M training run produces a teacher. A$ 300 fine-tuning run produces a student that captures a significant fraction of the teacher’s practical utility. The first mover spent 5 orders of magnitude more money.

This hasn’t made frontier training unprofitable — frontier models still lead students in important ways (reasoning, breadth, safety properties, long-context ability). But it has dramatically shortened the practical window of exclusive capability advantage.

IP and legal questions

Terms of service: Every major frontier API has terms prohibiting using outputs to train competing models. OpenAI’s terms: “You may not… use output from the Services to develop models that compete with OpenAI.” Google’s, Anthropic’s, and Meta’s terms have similar clauses. Alpaca’s weights were taken down partly in response to these concerns.

Does it work legally?: The terms are clear. Enforcement is harder. If a lab uses API outputs to train a model, proves it in court, and uses model weights they control, the API provider must establish that the outputs belong to them (copyright) and that the terms constituted an enforceable contract. Copyright protection for AI-generated outputs is currently unsettled in most jurisdictions. The legal risk is real but untested at scale.

Copyright and the output question: In the US, copyright requires human authorship. AI-generated text has, in recent copyright office guidance, been held to lack copyright protection. If the teacher’s outputs aren’t copyrightable, the legal basis for “you can’t train on them” becomes weaker. This is actively litigated territory.

The Llama distillation complication: When Meta released LLaMA under a research license and then Llama 2/3 under more permissive licenses, it became common to fine-tune open base models rather than pure proprietary API outputs. This sidesteps the IP question entirely: you’re training a derivative of an openly licensed model, not extracting from a proprietary API. The open-weight movement and the distillation movement are intertwined.

The capability hierarchy

A useful framework for thinking about who holds AI capability in the distillation era:

Capability hierarchy (May 2026)
===================================

Tier 1: Frontier trainers
  OpenAI, Anthropic, Google DeepMind, Meta AI
  Can produce new frontier capabilities from scratch
  Cost: $50M-$500M per major model
  Moat: training data, RLHF, safety research, infrastructure

Tier 2: Distillation specialists
  HuggingFace, Mistral AI, Cohere (+ many academic labs)
  Can efficiently transfer Tier 1 capabilities into smaller models
  Cost: $1K-$1M per specialized distillation project
  Moat: fine-tuning expertise, specific domain data, inference efficiency

Tier 3: Fine-tuning consumers
  Any company with a GPU
  Can take open distilled models and adapt them to a specific task
  Cost: $10-$10,000 depending on scale
  Moat: proprietary task data, deployment integration

Tier 4: API consumers
  Any company with a credit card
  Uses frontier or distilled models via API with prompt engineering
  Cost: per-token
  Moat: product, distribution, domain knowledge

The economic pressure runs downward. Tier 1 capability becomes Tier 2 and Tier 3 capability at lower cost over roughly 6-18 month intervals. The rate is limited by: (a) how fast frontier models advance past what can be distilled, and (b) how much of a frontier model’s capability actually surfaces in outputs that can be used as training data.

What distillation cannot transfer

Distillation has limits that matter for assessing how much moats are actually eroded.

Safety alignment: RLHF-trained safety properties are known to be brittle in distilled models. A student that learned to output safe-sounding text doesn’t necessarily have the internal reasoning process that produces genuinely safe behavior. Fine-tuning on a frontier model’s outputs can produce a model that is reckless in safety-critical edge cases the fine-tuning didn’t cover.

Reasoning under novel conditions: DeepSeek-R1-Distill is impressive on MATH and AIME. It is less impressive on truly novel problems where the teacher itself had to reason from first principles rather than pattern-match to seen problem types. The student learns the teacher’s reasoning patterns. It doesn’t develop new ones.

Long-horizon coherence: A 7B model distilled from GPT-4 produces GPT-4-like outputs on individual exchanges. Across a 50-turn conversation, the student’s smaller working memory and weaker coherence maintenance produce degradation that isn’t visible on standard benchmarks. The distillation focused on individual responses; the long-horizon behavior wasn’t taught.

Calibration: Distilled models are often more confident than they should be. They learned to produce the teacher’s well-calibrated answers but not the teacher’s uncertainty estimation process. This matters most for high-stakes applications.

The open model question

Distillation is why the “open vs. closed AI” debate is less binary than it appears. When Meta releases Llama 3 weights openly, thousands of researchers can fine-tune and experiment with them. Knowledge about what works propagates rapidly. Distillation flows back to Meta: other labs discover what fine-tuning approaches work, publish them, and Meta can study those results.

The dynamic is not “open = weak moat.” It’s “open = different moat.” Meta’s moat is not the weights — it’s the ecosystem, the infrastructure, the trust from developers who built on Llama, and the relationship with enterprises that want self-hosted models. The capability has been released; the moat is around everything else.

For companies that release weights with restrictions (gated research access, commercial use terms), the moat is even thinner. Llama 2’s commercial license didn’t stop the distillation community from building on it, because the distillation practitioners were researchers who qualified under the research terms.

References

Shumailov, I. et al. (2024). “The Curse of Recursion: Training on Generated Data Makes Models Forget.” ArXiv 2305.17493.
Gudibande, A. et al. (2023). “The False Promise of Imitating Proprietary LLMs.” ArXiv 2305.15717.
OpenAI Terms of Service, Section 3(c). platform.openai.com/policies/terms-of-service.
US Copyright Office (2023). “Copyright and Artificial Intelligence.” copyright.gov.
Bommasani, R. et al. (2021). “On the Opportunities and Risks of Foundation Models.” Stanford CRFM. ArXiv 2108.07258.

Case studies: the specific distillation projects whose implications this page analyzes
Benchmarks: how distilled models are measured against the teachers they learned from
Coding tool blindspots: capability gaps that persist even in well-distilled models