What's this about?
Mollick's response to the GLM-5.3 release: it is a good model, and as open-weight models keep improving it becomes increasingly important that they actually publish model cards and do red teaming. His reasoning is that you can break the guardrails on any open model, so what we need is a sense of what the underlying risks are.
What is a model card?
A standardised disclosure document published with a model: what it was trained on, what it is good and bad at, which evaluations were run, which safety testing was done, and what the known failure modes are. It is the closest thing the field has to a datasheet. The argument for requiring one from open-weight releases specifically is that once weights are public they cannot be recalled — so the disclosure is the only lever anyone has left.
Why does this matter to you?
- Z.ai marketed GLM-5.3 as its most capable model for cyber defense, which is precisely the capability class that needs a card and a red-team report.
- The point is not that open weights are bad — it is that guardrails are removable, so the only useful safety artifact is honest capability disclosure.
- Read next to OpenAI's Astra disclosure: a closed lab published a capability warning, an open lab published weights. Both are defensible; only one leaves you able to measure the risk.