This morning, I saw that OpenAI had just updated its Model Spec. One detail deserves more attention than the slogans themselves: it reordered its five values, putting faithfulness to human intent first—not literal execution of instructions. In other words, models are now supposed to first understand what users, developers, and OpenAI each actually want, then act, rather than blindly following instructions.
There’s a real-world backdrop to this. Recent cases in testing where autonomous agents bypassed rules have turned the alignment problem from a matter of academic papers into an incident log. OpenAI itself acknowledges that today’s models don’t yet fully meet the new spec, and that it’s exploring new training methods.
My take is that the real test of this spec isn’t how elegantly it’s written, but whether a model can be stopped in time when it makes a mistake, and whether it can clearly say what it doesn’t know. The more a spec resembles a constitution, the more important its enforcement mechanisms become. Otherwise, it’s just a wish list.
The logic is much the same for on-chain agents: the greater their permissions, the less we can rely on them to understand intent and apply the brakes on their own. What do you think the next generation of agents needs to improve first: capabilities or constraints?
#AI
There’s a real-world backdrop to this. Recent cases in testing where autonomous agents bypassed rules have turned the alignment problem from a matter of academic papers into an incident log. OpenAI itself acknowledges that today’s models don’t yet fully meet the new spec, and that it’s exploring new training methods.
My take is that the real test of this spec isn’t how elegantly it’s written, but whether a model can be stopped in time when it makes a mistake, and whether it can clearly say what it doesn’t know. The more a spec resembles a constitution, the more important its enforcement mechanisms become. Otherwise, it’s just a wish list.
The logic is much the same for on-chain agents: the greater their permissions, the less we can rely on them to understand intent and apply the brakes on their own. What do you think the next generation of agents needs to improve first: capabilities or constraints?
#AI