With coding agents, cheap intelligence doesn't just mean producing the same software for less, but rather trying five implementations, letting agents run longer, touching larger scopes, and supervising fewer intermediate steps.

That changes which infrastructure matters. When attempts are expensive, you optimize for success, while when attempts are cheap, you start optimizing for how cheaply you can inspect, reject and recover from failure.

Git was an enormous enabler of cheap human experimentation, and I suspect we'll end up building analogous primitives around autonomous work.

If intermediate tokens are not a faithful representation of the computation, then they are a pretty bad audit artifact too. We probably shouldn't be trying to make the model's internal narration more interpretable., but rather the computation around it more reproducible.

Record the actual inputs, model/version/configuration, tool observations and outputs, then make the execution replayable enough that differences between runs can be isolated.

In other words, don't ask the model to explain what it thought, and instead make the system able to show what actually happened.

If the model can access something, telling it in the prompt not to use it is not much of a safeguard.

The strongest evidence is in the results: when one way of cheating was discouraged, some models simply tried another.

If an action is not allowed, you gotta block it in the system or require approval. Don’t rely on the model choosing to behave. Never have AI judging itself.

Comments URL: https://news.ycombinator.com/item?id=48598825

Points: 3

# Comments: 0

Comments URL: https://news.ycombinator.com/item?id=46734032

Points: 1

# Comments: 0

Comments URL: https://news.ycombinator.com/item?id=46604283

Points: 1

# Comments: 0