fabsalvadori

Oh come on...actions have consequences. Intelligence started imploding when people took up scrolling and swiping for 14 hours a day. It’s been such a carnage that I’m starting to feel like a genius. If our future is doomed, we can’t give AI all the credit.

Parmigiano Reggiano is a registered brand. It gives me goosebumps to read "parmesan" which is an unregulated translation or generic term used globally for very low quality imitations. Anyway, if you like cheese but you're intolerant to it, know that Parmigiano is not only extra delicious but also naturally lactose-free.

This is the line between an instruction and a control.

If the agent can reinterpret, edit or relax the rule that constrains it, the rule isn't actually enforcing anything... it's just part of the prompt.

I think the useful split is to tell the agent the rules so it can avoid wasting work, but independently enforce the rules that actually matter.

The agent can decide how to accomplish the task, but it shouldn't also get to decide whether it's authorized to cross the boundary.

There is a second-order effect I see, here: when generating work becomes dramatically cheaper, failed work also becomes dramatically more common.

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.

There is a useful engineering consequence here beyond terminology.

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.

Interesting results, but the fix is at the wrong level.

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.

I see little conversation about AI powering robots' core and the fact that we are going to have like a billion robots by the end of this decade. Yes, sorry if trust is low.