A closed model hides more than its weights. It hides the history of how its behavior was produced.
OpenAI, Anthropic, and Google publish system cards, benchmark results, safety evaluations, and release notes. Those artifacts are useful. They still do not give an end user a clear line from the model that entered training to the product that appears on a screen.
That line now passes through several different systems. There is the composition of pre-training data, the objectives used during training, continued training, supervised fine-tuning, reinforcement learning, safety training, tool use, system prompts, routing, memory, and product-level policies. A breakthrough at one stage can be weakened at another. A safety intervention can change useful behavior. A product update can make the same model name feel like a different system.
The user sees the result, but not the sequence.
This is not only a question of whether the weights are open. It is a question of visibility. When a proprietary model improves, I often cannot tell whether the improvement came from a better base model, better post-training, more inference-time compute, a new tool, a routing change, or a product layer wrapped around the same underlying capability. When it regresses, the same ambiguity remains.
That makes closed models difficult to reason about as systems. A model name becomes a container for many decisions made at different layers by different teams. The interface stays familiar while the behavior underneath it moves. Users are asked to trust the product as a stable object even when the object is continuously changing.
“Privatized” is not quite the right word for this. These systems were not transferred from public ownership into private hands. They are proprietary and privately governed. The labs have legitimate reasons to protect intellectual property, licensed data, security methods, and competitive research. Full disclosure is neither realistic nor automatically responsible.
But the choice is not between publishing every training secret and publishing almost nothing. There is room for structured visibility without exposing weights or sensitive data. Today, that structure is inconsistent. Model cards describe snapshots. Release notes summarize selected changes. Benchmarks compress behavior into scores. None of them reliably explain how pre-training, post-training, evaluation, safety, and deployment decisions connect over time.
What is missing is order. A way to distinguish a capability breakthrough from a product intervention. A way to know when a safety change caused a regression. A way to tell whether a familiar model name still refers to the system a user evaluated last month. A way to understand which layer changed before trying to explain why the output changed.
The transparency problem is not that I cannot inspect everything. It is that I cannot reconstruct enough of the system’s history to know what I am trusting.
- Essay to be written soon, with a fuller conclusion and further possible proposals.