Get your free personalized podcast brief

We scan new podcasts and send you the top 5 insights daily.

Rob Ferguson says Fireworks keeps zero data retention and couldn't calculate a discount large enough to justify training on customer data. He warns closed labs discount what your data is worth when absorbing it, and people underestimate what they give away. That asymmetry matters when choosing model partners.

Related Insights

While many assume cost is the main barrier to adopting frontier AI models, the case of Anthropic's Fable 5 reveals a bigger hurdle: its 30-day data retention policy. For many enterprises, this security and compliance risk is an immediate deal-breaker, regardless of the model's superior capabilities.

Companies with valuable proprietary data should not license it away. A better strategy to guide foundation model development is to keep the data private but release public benchmarks and evaluations based on it. This incentivizes LLM providers to train their models on the specific tasks you care about, improving their performance for your product.

Companies risk giving away enterprise value by sending proprietary data to external foundational models. The secure and value-accretive approach is to bring AI models in-house to train on data within a controlled, air-gapped environment, preventing data leakage.

Despite processing 15 million clinical charts, Datycs doesn't use this data for model training. Their agreements explicitly respect that data belongs to the patient and the client—an ethical choice that prevents them from building large, aggregated language models from customer data.

Enterprise SaaS companies (the 'henhouse') should be cautious when partnering with foundation model providers (the 'fox'). While offering powerful features, these models have a core incentive to consume proprietary data for training, potentially compromising customer trust, data privacy, and the incumbent's long-term competitive moat.

Companies in pharma, finance, and other sectors are realizing that feeding their proprietary data to closed AI models creates a strategic risk. They fear the AI labs could become direct competitors, driving a shift towards sovereign, open-source models run on their own data.

As highlighted by Palantir's CEO, corporations are wary of feeding proprietary data into large AI models. They fear AI companies will train on their data to launch competitive products, as seen with Figma, while also struggling to justify the high token costs and measure tangible business returns.

Enterprises using third-party AI models for proprietary work risk leaking their IP. Even with 'zero data retention' policies, models can learn from de-identified usage data, effectively absorbing novel insights. The only secure approach for sensitive R&D is a sovereign solution using self-hosted hardware and models.

HubSpot's customers revolted not just because their data would train AI, but because it might be shared with other users, including competitors. This rapid reversal highlights that for enterprise customers, protecting the competitive advantage embedded in their curated data is a far greater concern than the act of AI model training itself.

The SpaceX Grok data leak shows that even with good intentions, AI tools have non-obvious leak vectors. Enterprises cannot solely trust model providers' privacy promises and must implement independent controls to protect their proprietary data, or "alpha," from inadvertent exposure.