> But Cursor and other such tools depend almost entirely on accessing Anthropic, OpenAI and Gemini models, until open-source open-weight and in-house models match or exceed frontier models in quality.
I'm not sure I agree with this because even though Cursor is pay north of 100% of revenues to Athropic, Anthropic is selling inference at a loss. So if Cursor builds and hosts its own models it still has the marginal costs > marginal revenues problem.
The way out for Cursor could be a self-hosted much smaller model that focuses on code, and not the world. This could have inference costs lower than marginal revenues.
cost goes down by 50% (it didn't) but you fake improvement while lacking any real improvement by taking the average of 4 runs. Cost just doubled, even though inference technically halved.
Never mind the fact that the current model is alwayj outdated, and a new (bigger one) one is always being trained in parallel with the supposedly "cheaper" inference.
I'm not sure I agree with this because even though Cursor is pay north of 100% of revenues to Athropic, Anthropic is selling inference at a loss. So if Cursor builds and hosts its own models it still has the marginal costs > marginal revenues problem.
The way out for Cursor could be a self-hosted much smaller model that focuses on code, and not the world. This could have inference costs lower than marginal revenues.