Key Takeaways

  • Thinking Machines Lab, under Mira Murati, launched Inkling: an open-weight AI model optimized for fine-tuning, aimed directly at challenging OpenAI and Anthropic's market lead, according to the Wall Street Journal.
  • Inkling's release ignited a heated debate about 'distillation,' its training process that uses synthetic data derived from other open-weight models, pushing the boundaries of what 'pure' open-source truly means.
  • This technical discussion has a sharp geopolitical edge, with Anthropic's Taryn Chabbra accusing Chinese models like Zhipu.ai's GLM-5.2 of distilling US-developed AI, highlighting a new front in international tech competition.
  • The scale of the problem is immense: Anthropic is actively battling unauthorized distillation, reportedly shutting down millions of suspicious accounts each week.
  • Founders building on AI should deeply scrutinize a model's training provenance, as 'open-weight' status doesn't always guarantee ethical or unencumbered use, especially given current geopolitical tensions.

The Disagreement

The AI world is buzzing about Inkling, a new open-weight model from Mira Murati's Thinking Machines Lab. It's designed to be fine-tuned, aiming to chip away at the lead of giants like OpenAI and Anthropic. As John Coogan explained, “Inkling is not the strongest overall model available today, open or closed, which is a different frame of reference for many of these model launches.” Its real significance lies in its approach.

But that approach—called 'distillation'—has cracked open a deep schism. Distillation uses synthetic data generated by other models, often existing open-weight ones. For some, this is pragmatic progress; for others, it's a dilution of open-source purity.

On one side, you have the view that any open-weight model helps democratize AI, even if it builds on others. It's about getting powerful tools into more hands. The goal is to create competition and alternatives, not necessarily to build from scratch every time. This camp sees distillation as a legitimate shortcut to creating specialized, customizable models that can rival the big players.

On the other, a strong argument for 'purity' claims that distillation, especially when it re-uses outputs from other models, isn't truly original. John Coogan quoted Jack Morris on this: “This is the only open weight model that's trained without distilling from OpenAI or Anthropic. Kimmy distills, GLM distills, Qwen distills, Nematron distills...” This perspective suggests that true open-source training requires building from a foundational text stack, not from the outputs of others. The tension escalated with Anthropic's head of national security policy, Taryn Chabbra, directly accusing China's Zhipu.ai of distilling both Claude and OpenAI models for GLM-5.2 at the Aspen Security Forum. Coogan added that “Anthropic is now shutting down distillation accounts on the order of millions accounts of per week,” showing the aggressive stance taken against this practice.

Who's Right (and When They're Wrong)

Neither side holds all the cards. The 'purity' argument, while ethically compelling for foundational research, can be impractical for startups and smaller labs trying to innovate rapidly. If an open-weight model is available, and its license permits reuse, distilling from it to create a specialized, fine-tuned model for a niche application can accelerate development and democratize access to advanced AI capabilities. This approach works when the goal is speed, specialization, and building on community-contributed open knowledge, especially when clear attribution is maintained.

However, the 'purity' side gains undeniable weight when geopolitical stakes rise. When state-sponsored entities are accused of distilling proprietary or even open-weight models from competitors without clear consent or licensing, it shifts from a technical debate to an issue of intellectual property theft and national security. In such scenarios, as with the accusations against Chinese models, distillation isn't about fostering open collaboration; it's about circumventing costly R&D and gaining strategic advantage. Here, the 'open-source' spirit is misused, and the practice is unequivocally wrong.

What to Do With This

For ambitious founders, the lesson is clear: don't just look at whether an AI model is 'open-weight.' Dig deeper. Understand its complete training provenance and distillation history. Before you build your next product on a seemingly free model, verify its true origins and any potential legal, ethical, or geopolitical liabilities attached. Your choice of base model isn't just a technical decision; it's a strategic one with real-world implications for your company's future.