Society, Environment & Big Picture
Open vs. Closed Source
The systemic question of the industry: open weights that anyone can run and inspect, versus closed frontier models behind APIs. Open brings sovereignty, inspectability, price pressure; closed brings (for now) peak performance and convenience — and concentration of power. Most serious architectures are portfolios: the right model per task, with an exit path.
An example
The debate is often conducted as if there were two camps. In practice the question is more concrete and has to be answered per use case: for the customer assistant with fluctuating load and no particularly sensitive data, a purchased model is almost always right. For analyzing personnel files that may not leave the building, an open model on your own hardware is often the only viable option. Both in the same company is the normal case, not a contradiction.
A closer look
The scale is fine-grained: published weights, published training data, published training code, permissive license. Most “open” models fulfill only the first point, often with usage restrictions.
Why it matters
What regularly gets lost in the debate: “open source” for models almost never means what it means for software. Weights are published, not training data and not training code — nothing can be reproduced or rebuilt with them. And the licenses differ in strictness. For your own decision, three questions are therefore more useful than the label: May the data leave the building? Is the load steady enough for your own hardware? And what happens if the vendor discontinues the model?
Commonly misunderstood
The choice is not a worldview but a trade-off per use case: how sensitive is the data, how good must the model be, how much operations can you afford. Mixed setups are the normal case.