Training Data
Levie: subsidized tokens are temporary, and the gap between model and workflow is the business
The "LLM wrapper" insult aged badly because the real work in the enterprise is bridging a model's capability to an actual workflow, and that bridge turns out to be vast. Box runs two evals on every model, a complex work eval across life sciences, financial services and public sector, plus a holdback on how Box employees use their own data, and results track coding benchmarks with one exception: Gemini overperforms, likely on tool use. Token subsidies cannot last, because a public company gets held to the same laws of capitalism and eventually has to pay for its training runs too. The prediction to write down: in five years, 90% of enterprise tokens will be work no user ever kicked off, and they just see a result waiting for review. On diffusion, coding won because value creation is almost entirely text and every engineer debugs their own MCP connection instead of calling IT; a sales rep is rate limited by whether the customer replies. There is no "give us your GitHub" for knowledge work.
- #agents
- #evals
- #open-source