I hope you don't watch the images being generated one line of pixels at a time from top to bottom! It's not worth it. I used to do that for, ahem, research purposes, in the old days.

My inner pedant has been awakened! I must insist you stop mixing measurement units forthwith, or there will be further comments!

A UID is measured in bits, not years. ;-)

With a name like that he may as well have an uncle named Victor!

I don't know what you mean... I thought he was going to build a spaceport on Mars???

Action-at-a-distance! Get your action-at-a-distance here or there! Packaged in a wave! No field of enquiry, is too small.

You, sir, do you want to read the story? Only two cents!

Extra! Extra!

I particularly loathe the AI animated historical and scientific explainers produced by AI. For example, you get a neat simple story about something interesting like the evolution of clothing in hominids, and you wonder what has scientific backing and what is hallucinated.

It's a balance. The absence of enforcement of rules is toxic in a different way. I believe that when rules are spelled out up front, anyone who deliberately flaunts them does not have the good of the community in mind. So why should they remain members of said community?

The breaking of licensing terms often takes place in private. For example, if X steals his employer's source code, slaps a fake open source license on the stolen code, and submits it to Debian, there's no way to know. Except if another employee Y happens to find the code online and notifies Debian about X's fraud.

Enforcing a lifetime ban on X's contributions is the least Debian can do in that case to protect itself legally from the wronged company.

He thinks that any policy statement that says AI isn't allowed simply won't hold up, as the tooling keeps getting harder to detect...

It shouldn't be about detecting and preventing AI contributions with 100% accuracy. Some AI contributions will get through because people will always try something.

It should be about spelling out the rules, then banishing people who break them, ripping out their "contributions" and shaming them publically if and when they are found out.

This is not any more difficult than policing licensing terms is today. Debian already has a history of banishing non-Free software packages into special source repos. And, like any serious open source project, they already rip out code that isn't legally safe for them to use (eg if someone steals code and passes it off as their own and makes up a fake license).

Eventually, the answer will be to have a standard mechanism for reporting suspicious sloppy source code and whistleblowing on the cheaters.

What I think the OP means is that throwing randomly generated constructions at an external proof calculator, like Lean, until something gives, does not convincingly suggest that the LLM component is performing a useful task on its own.

I predict it will turn out some obscure forgotten paper from 1974 did it first.

Sorry, I misinterpreted your story as a work story with a corporate team mate submitting AI PRs. Lots of those around these days.

Agreed. It's a very expensive way to replace homework.

The CORRECT approach is to make homework voluntary, with no partial credit. Then do in-person oral examinations but only AT THE END OF THE DEGREE (ie the last week of a multi-year schooling process).

This used to be the case in universities and schools for centuries, and was replaced more recently by continuous assessments. Let's go back to the previously proven methods.

Psychologically, it's tougher on students, which is a good thing as it produces more resilient human beings.

Students will do nothing (because they can) in the first year, they will panic in the second year (and many will drop out) and the remaining ones will work extra hard in the third year to pass. All of them will have learned a valuable lesson that they can use in the rest of their life.

Why didn't you just get Claude to "fix it"? It either "fixes it" and you have CYA deniability if it's wrong or it writes a report claiming it "can't be fixed" and you have CYA deniability when you didn't fix it.

AI output should be fought with AI output. It's the easiest way to throw the ball back to the person who didn't do their part correctly.

The falls from windows are for those who fled the country and think they are safe.

It's a venerable tradition going back to the old Trotsky days.