Interesting Links for 06-07-2025

[andrewducker]

1. Forced participation in religious activities to be classified as child abuse in Japan

(tags:religion abuse children Japan )

2. 'Positive review only': Researchers hide AI prompts in papers

(tags:ai research )

3. Xi Has Spent Decades Preparing for a Cold War With the U.S.

(tags:china USA trade )

4. Seine reopens to Paris swimmers after century-long ban

(tags:France Paris water pollution sewage )

Comments

[ninetydegrees]

4) Considering officials lied about its level of pollution during the Olympic Games, I really hope it is safe now.

[juan_gandhi]

A bunch of good and cool news today. Thank you! I agree with the Japanese. I wonder when it can be enforced in the US, with its "pledge" in schools. "under God", my ass.

[bens_dad]

2. (Why) are the AIs obeying prompts in data ?
There should be a clear boundary between what it does and what it does it on/to.

This is worryingly close to the AI running programs found in its training data.

[andrewducker]

AI doesn't have multiple sets of inputs. It has just one set of context that it produces the next words to continue.

[jducoeur]

I strongly suspect you could build an LLM that was resilient against this sort of nonsense -- there's nothing obviously sacred saying that the base model, the input, and the question being asked about the input have to be treated equally.

But of course, that would require that the people and companies using these LLMs have some clue what they're doing, and that's not the world we are living in...

[andrewducker]

As far as I can tell with LLMs you get one set of context, and that's it. It takes anything you're working on, your question, any instructions that the creators initialise it with, turns it into one document, and then asks the LLM to generate what comes next. So there's no way of separating the input and the question from each other.

Maybe some day they'll work out how to separate them.