Hacker News new | ask | show | jobs
by dkiebd 296 days ago
This sucks but the only solution is to make companies censor the models, which is a solution we all hate, so there’s that.
2 comments

Thank you, “we just have to accept that these systems will occasionally kill children” is a perfect example of the type of mindset I was criticizing.
Don’t cars and ropes and drills occasionally kill people too? Society seems to have accepted that fact long ago.

Somehow we expect the digital world to be devoid of risks.

Cryptography that only the good guys can crack is another example of this mindset.

Now I’m not saying ClosedAI look good on this, their safety layer clearly failed and the sycophantic BS did not help.

But I reckon this kind of failure more will always exist in LLMs. Society will have to learn this just like we learned cars are dangerous.

Another great example of the HN mindset. How can you say "their safety layer clearly failed" without being able to acknowledge that they should be held responsible for that failure and that we should work to reduce the likelihood of similar failures in the future? If a car, rope, or drill had some sort of failure in their manufacturing that killed people, those companies would be held responsible. Why can't we do the same with OpenAI?

This isn't a desire for "the digital world to be devoid of risks", it is a desire for people and companies in the digital world to be held responsible for the harm they are causing. Yet, people here seem to believe that no one should ever be held responsible for the damage caused by the software they create.

Cars and ropes don't "talk" and don't make an impression of a human being.
If we banned everything that contributed to the death of children we'd eventually have nothing left to ban.
Contributing to and facilitating are 2 very different things. This was more the latter.
We have many tools in this life that can maim and kill people and we keep them anyway because they are very useful for other purposes. It’s best to exercise some personal responsibility, including not allowing a 16 year old child unattended access to the internet.
Yeah, that is why we don’t have any regulations on the manufacturing and sale of stuff like guns or drugs. The only thing in the way of a 16 year old having unfettered access is personal responsibility.
So you're in favor of regulating ropes then?

I'm not a big fan of LLMs but so far their danger level is much closer to a rope then to a gun.

America is perfectly happy with sacrificing kids in exchange for toys though.
That statement is going to be true for anything short of a complete Panopticon where everyone and everything is watched 24/7 to nip any potential harm in the bud.

If you're not okay with that arrangement (why not?), then we're left to argue about which level of "will occasionally kill children" is acceptable.

So what's the alternative? Pervasive censorship and horrible privacy and freedom-of-speech destroying legislation like UK's Online Safety Act?

I'm not looking forward to the day when half of the Internet will require me to upload my ID to verify that I'm an adult, and the other half will be illegal/blocked because they refuse to do the verification. But yeah, think of the children!

The alternative is as usual, getting certain individuals to the court and making them pay for the damage or even to go to jail, instead of preemptively trying to give them a holy cow status.
>“we just have to accept that these systems will occasionally kill children”

I think this a generally a good mindset to have.

I just see the hyper obsessive "safety" culture corrupting things. We as a society are so so afraid of any risk that we're paralysing ourselves.

That sounds a lot like the people who gloss over school shootings because they want to be able to play with guns.
Unpopular, but backed by numbers: school shootings are predominantly media hype.

There are a total of 50,000 gun deaths in the United States each year. Half of those are suicide. The total number of fatalities from all school shootings in the US since 1990 is 536, with roughly 1100 injuries over the same period.

School shootings are very disturbing, but they are also flashy and attract media coverage, and so they are overblown, especially relative to other gun deaths. But I think human psychology leads people to see higher risk in situations where they feel like they have no control. The idea of sending your kid off to school and having them get killed evokes sort of the same helpless feeling as the feeling around a plane crashing when you're a passenger. But when it's inner city gang violence, or a depressed man killing himself in his garage, we don't seem to care as much.

This also plays into the narrative around police shootings. There are about 1100 people killed by cops every year in the United States, and that number has stayed pretty consistent over the last decade. But our narrative around it is out of proportion to the actual number of people killed because of the circumstances around the killing.

In any case, it's often easier to articulate the cost of freedom rather than its value.

> predominantly media hype

hmm, tell that to parents whose kids were killed, or who are genuinely freaked out that there is a chance that their kid could die when they send them to school

as a parent who moved to the US and encountered this fear for the first time -- a fear that's just not present in other countries because while anything can happen anywhere, the chances are much much lower. why? not because there aren't crazy people in other countries, but because they don't have easy access to guns

> so they are overblown, especially relative to other gun deaths

actually, it's the other way around; it's the "other gun deaths" that are _under-valued_ and often ignored in the gun debate -- we should be talking about those _more_ not talking about school shootings less

50K gun deaths a year in the US is absolutely bonkers compared to other developed countries, and on par with Central America per-capita; and even if half are suicides, many who commit suicide may not have done so if guns weren't so readily accessible (sure there are other ways to do, but it's not as easy)

> plays into the narrative around police shootings. There are about 1100 people killed by cops every year in the United States

the reason for the narrative, is that in no other other advanced democracy on earth is there such a high rate of police killing civilians (when calculated per capita) -- it's not even close

they get attention because police shouldn't be killing civilians in the first place; it's pretty simple

if the USA didn't have such an obsession with guns, we wouldn't even be having this conversation

Roughly an hour after this comment was made, someone walked into the church at a Catholic school, killed an 8-year-old and a 10-year-old, shot 17 other people including 14 children, and then turned the gun on himself.[1] "Predominantly media hype" indeed.

[1] - https://www.nbcnews.com/news/us-news/live-blog/live-updates-...

Yep, an excellent illustration of the point. Two dead, a dozen injured. What other gun violence happened this week that we'll never hear about?
Wow. Incredible framing, leaving absolutely no room for a rebuttal to have any legitimate, reasonable position.
anyone is welcome to state their legitimate, reasonable position against gun controls
I believe the de-utility of risk of school shootings is less than the utility of having people have the freedom to own guns.
Maybe I don’t understand well enough. Could anyone highlight what the problems are with this fix?

1. If ‘bad topic’ detected, even when model believes it is in ‘roleplay’ mode, pass partial logs, attempting to remove initial roleplay framing, to second model. The second model should be weighted for nuanced understanding, but safety-leaning.

2. Ask second model: ‘does this look like roleplay, or user initiating roleplay to talk about harmful content?’

3. If answer is ‘this is probably not roleplay’, silently substitute model into user chat which is weighted much more heavily towards ‘not engaging with roleplay, not admonishing, but gently suggesting ‘seek help’ without alienating user.’

The problem feels like any observer would help, but none is being introduced.

I understand this might be costly, on a large scale, but that second model doesn’t need to be very heavy at all imo.

EDIT: I also understand that this is arguably a version of censorship, but as you point out, what constitutes ‘censorship’ is very hard to pin down, and that’s extremely apparent in extreme cases like this very sad one.

You see, that costs money and GPU time. So no bueno.