Open-Weight AI Closes the Capability Gap, but Not the Safety One
In one line: Open-weight models are rapidly approaching frontier-level performance, but their safety mitigations reportedly lag far behind.
Key points
- A report from the nonprofit SaferAI assessed Z.ai's open-weight model GLM-5.2 on both capability and safety.
- Its capabilities approach frontier models, but many key safety mitigations are said to be absent.
- Because open-weight releases publish the model weights, anyone can download, modify, and redistribute them — making the models hard to control or recall.
Why it matters
As the capability gap narrows, so does the window for governance: a "powerful but unguarded" model could spread faster than regulation can respond. Since open releases are hard to claw back, safety choices made at release time carry outsized weight.