Hacker News Reader

Comment by arczyx | Hacker News Reader

> How exactly do they want to use an llm for autonomous weapons

Probably the same way Claude can play Pokemon: give it a bunch of informations and let it make a decision by itself to achieve a specified goal.

View on HN

One thing is playing Pokemon, one thing is decide who to kill. Also: if they are planning to use it on the field, there is going to be a velocity issue. Claude and any other LLM require a non negligible amount of time to ingest the input and spit the output.

You seem to be under a mistaken impression that there is some good faith and competence here. Fascists don't give a shit about accuracy. Collateral damage actually increases the fear of being mistakenly caught up and encourages those at risk to loudly support the Party hoping it will protect them.

You are saying fascists don't care if it inaccurately targets themselves?

Not really. Look at the longstanding "I didn't think the leopards would eat my face" dynamic. The whole movement is rooted in othering an outgroup for which violence against is moral and good. None of its supporters imagine themselves ever being in that outgroup, and if they do end up there by then it's too late.

In the domestic surveillance context, it's not like decisions of an automated system are ever going to be used to arrest anybody the bossman is protecting.

On the battlefield, commanders don't really care about collateral damage to the grunts, generals don't really care about collateral damage to whole units, etc. Look at how Russia is prosecuting its meat grinder against Ukraine. That type of failed state oligarchy appears to be the Trump regime's grand vision for our country.

But if the argument is that they don't care about accuracy, then what's to stop the system from targeting themselves or the leaders they support?

Remember, the system is not accurate, so you cannot accurately tell it to avoid or ignore anyone special.

I think they do care that it's accurate. That it accurately targets other people not themselves or their leaders.

It's very easy to make up for that with a human in the loop or deterministic secondary checks (enumerated list of protected people or locations). The point is that nothing about the "AI" magic has to be super accurate.

Reply on news.ycombinator.com