Even if you're magically perfect at risk assessment. Rather than restricting creating new models, it should involve restricting capabilities research. (And you have to restrict it everywhere, so that e.g. Microsoft doesn't turn into OpenAI's capabilities research shop.)
Here are some examples of pacing which could actually be bad:
External deployments are good because they enable visibility of capabilities and safety research. (And, less importantly, diffusing capabilities might decrease the future chaos when much more capable models are released.) (But there should be restrictions aimed to prevent people from using it for AI R&D.)
This regime is likely to be weakened/ended before superintelligence. For one, the research will likely leak or be stolen by foreign actors, and then US actors will feel strongly incentivized to progress quickly. And when unrestricted progress resumes, it will go faster than it would have by default, since there will be a large overhang of capabilities research to use in the next model.
Risk-assessment-based restrictions (on pretraining, posttraining, internal deployment, whatever)
Same as previous bullet.
Standards that look like no scary models before safety/security milestone X
Same as previous — the company will delay the scary model but keep making algorithmic progress.
The main benefit of pacing isn't preventing risk from the next model — it's having more time before reaching superhuman AI researchers. The company isn't imposing catastrophic risk yet isn't a crux; if you only pause right before you're directly imposing lots of risk, that pause will be very unstable because others can steal your research.
Ways to restrict capabilities research include (1) restricting compute for RSI-enabling work and (2) restricting internal deployment. (Each of these can be dialed up to a full ban.)
*"Pacing shouldn't use risk assessment" is slightly stronger than my real view. My real view is that (1) we should pace now even if risks are currently low and (2) the simplest versions of risk-assessment-based pacing, like restricting scary training runs, are misguided. Risk assessment could have some good role in some pacing regimes.
Er, internal deployment is trickier. Restricting internal deployment would slow capabilities research more than restricting training. But there would still be an overhang.