When people listen to two pieces of audio they generally prefer the louder of the two. That doesn't mean they want you to turn up the volume dial for them. They can adjust the volume dial themselves, and if everything gets louder they'll turn down the volume dial to compensate.

Declining to give you $600 out of the blue because you'd rather have more money is not being indifferent.

(I do think a slight discount often makes sense just because a friend is probably quicker and easier to deal with. But anything more substantial turns into asking for free stuff, and yes and no are both perfectly fine answers to that.)

If you decline to sell a thing for an easy 1300, doesn't that mean you value it at 1300 or more?

They didn't say anything about what decision is correct. They just said that the two decisions are equivalent.

Please note the use of the word "insane" is specifically because buildsjets said it was insane.

The argument is not "you could make money", the argument is that if you got the expensive thing for free and choose not to sell then you're roughly as "insane" as the person that paid full price. Go ahead and splurge but try not to be a hypocrite about it. It's not lmm that was passing judgement, it was the person that owns the record passing judgement.

Also drones didn't flatten a city. You really have to ignore a lot of the details to make this connection strong.

Companies don't have to do that. If they're getting into actually dangerous territory, they can stop as soon as they want to.

Imagine someone seeing that the murder rate is not zero and using that to claim the world is worse than it used to be. That's not how it works, despite murder obviously being a bad thing.

And pokemon go isn't actually a children's game in the first place, it's heavy on the nostalgia and at least a while ago 3/4 of the users were adults.

You can connect any two things into a sentence regardless of the state of the world. This is way off of where the problems actually are.

Edit: Some people have downvoted this without giving a reason, but I'm going to double down. Any time you have disasters or large crimes, you can connect them to children and children's things. Thinking there's anything to learn about the specific fact that you can make that connection is a mistake. It's letting the real problem spill over in a way that misleads your common sense. It's an inherent part of bad things happening that they also affect children. No matter what state the world is in.

Can you make sure the instructions and data are separated and the machine follows only the instructions and doesn't change its behavior based on the data? No.

But the part that's impossible is not "the instructions and data are separated". The part that's impossible is "the machine follows only the instructions".

Separating instructions and data is not impossible, but it doesn't solve your problems.

One really important consequence of this is that even if the data doesn't have anything that looks like instructions, it can poison the machine anyway! If you get too focused on "instructions" then you miss that security flaw!

Even if you don't give the machine any data at all, it might not follow the instructions. It's not instruction/data conflation as the root cause, it's that instructions don't really work in the first place.

The above post said how. The LLM writes code to do it. The code has a function to send text to the user. The LLM is not allowed to see the text, only the user is.

> And it will make a summary exactly how?

The second summarizing-only LLM is fed the raw data and allowed to output summary text. This is then sent directly to the user and put in a box with some hazard lines on it. The main LLM is not allowed to see the summary, only the user is.

You can't inject the LLM if it doesn't see the data.

An architecture like this won't work in many situations, but it can work for a lot of simple questions.

And if you want the LLM to summarize things, you run an isolated instance that makes a summary and you never show that summary to the LLM that's following the user's instructions.

But that doesn't mean that separation between instructions and data is impossible. You can format them in different ways, and you can prevent the output tokens from ever using instruction formatting.