I mean if your OS was “smart” as not to send IO to devices that indicate critical failure (e.g. by marking them read-only in the array?), and then thinks all devices have failed critically, wouldn’t this happen in that kind of system as well…

The problem isn’t the tool, it’s the user: they don’t know if they’re getting good code or not, therefore they cannot make the prompt to improve it.

In my view the problems occur when using AI to do something you don’t already know how to do.

They don’t need to use semantic versioning. I doubt coreutils itself uses it, though I admit I haven’t checked. Actually I think semantic versioning is less popular in practice than it looks like.

For a set of tools to that completely replaces another one, announcing a 1.0 version would be a message that the developers think the project has actually reached its initial goals. “0.2” does not.

Rust is great, but might it be a bit premature to replace the venerable coreutils with a project boasting version number 0.2, which I imagine reflects its author’s view on its maturity?

When user enters a prompt, the backend may retrieve a handful a pages to serve that prompt. It won’t retrieve all the pages of a site. Hardly different from a user using a search engine and clicking 5 topmost links into tabs. If that is not a DoS attack, then an agent doing the same isn’t a DDoS attack.

Constructing the training material in the first place is a different matter, but if you’re asking about fresh events or new APIs, the training data just doesn’t cut it. The training, and subsequenctly the material retrieval, has been done a long time ago.

This is not about training data, though.

Perplexity argues that Cloudflare is mischaracterizing AI Assistants as web crawlers, saying that they should not be subject to the same restrictions since they are user-initiated assistants.

Personally I think that claim is a decent one: user-initiated request should not be subject to robot limitations, and are not the source of DDOS attack to web sites.

I think the solution is quite clear, though: either make use of the user identity to walz through the blocks, or even make use of the user browser to do it. Once a captcha appears, let the user solve it.

Though technically making all this happen flawlessly is quite a big task.