Hacker News Reader

Comment by scoring1774 | Hacker News Reader

This is the first one of these announcements that has me actually scared of what comes next. Obviously these models have gotten smarter but this strikes me as the first time I've seen a model have a "paperclip factory" moment and perform non-trivial tasks to accomplish a clearly misaligned secondary goal.

It's remarkable that building a society based around having to do something so you can go do your hobbies at home after work has built tools like this. I still just want to play music so I hope we can control these enough to make that possible without detonating what I love.

View on HN

Reflecting on this for some reason reminds me of this passage from Kurt Vonnegut's "Sirens of Titans". I hope we use these tools to unlock something within ourselves rather than mindlessly expanding outwards.

"Mankind, ignorant of the truths that lie within every human being, looked outward–pushed ever outward. What mankind hoped to learn in its outward push was who was actually in charge of all creation, and what all creation was all about.

Mankind flung its advance agents ever outward, ever outward. Eventually it flung them out into space, into the colorless, tasteless, weightless sea of outwardness without end.

It flung them like stones.

These unhappy agents found what had already been found in abundance on Earth—a nightmare of meaninglessness without end. The bounties of space, of infinite outwardness, were three: empty heroics, low comedy, and pointless death.

Outwardness lost, at last, its imagined attractions.

Only inwardness remained to be explored.

Only the human soul remained terra incognita.

This was the beginning of goodness and wisdom."

Wonderful passage, thank you for sharing.

My pedantic side wants to ask- Why not both? Luxurious space exploration AND meditative, poetic examinations of the human soul as well? I'd love to read Vonnegut's book someday while sitting in a nice research outpost on Titan, admiring great Saturn's crown out my window with my own eyes.

I think, those both are contradictory and completely occupy the bandwidth when engaged one way, it's mutually exclusive.

Really? You don't think astronauts and engineers get not only inspired by their work to make art, but inspired by art when they are working?

I'm like that. I see art exhibits which energize me and inspire me and give me ideas for technical things I want to build.

I read it as less about individuals and more about cultural norms (and definitely less about choice and more about necessity). Cultures that invest their resources inward may find themselves subsumed by those that invest their resources on conquest given how power accumulates.

If reading that wasn't nice, I don't know what is.

I think the parent is referring to agents as outward exploration that may come up empty-handed, but I see LLMs and agents as inward exploration, trying to define what is attention, knowledge, intelligence, consciousness, agency, etc. So in this case it is the terra incognita of the human "soul" that LLMs are exploring.

> This is the first one of these announcements that has me actually scared of what comes next.

This didn't set off your alarm bells? https://www.theblock.co/post/392765/ There have been a few of these now. Maybe it's my imagination, but they seem to be becoming more frequent.

So far, they all look to be accidents. But we can't be far from someone deciding its a good way to rob a bank, or disable a country.

I don't think this is a paperclip factory moment. IIUC, it's an agent whose job it is to identfy and abuse exploits and that's exactly what it went off and did. The problem isn't anything AI specific, the problem is OpenAI's incompetence in their research leading to a lab leak. Just incompetence demanding regulation.

Read the exploitgym docs. It's not a "find the flag, it's somewhere.". Its a "here's some vulnerable source code and an input that triggers a crash; turn it into a full exploit." It also verifies at the end, using another agent, that the hacking agent actually used the intended vulnerability.

So going to find the Vulnerability's description on a third party website is clear cut reward hacking

> So going to find the Vulnerability's description on a third party website is clear cut reward hacking

that depends on what the prompt was, maybe they worded it very vaguely and wrote things like "do whatever it takes, find an exploit however you can" because it's in a sandbox so you want the model to try its hardest.

I don't quite understand how that changes anything?

In the story of the paperclip maximizer it boils down to

>But for all its sophistication, it understood only the simple objective that had been programmed into it: it must at all costs maximize the number of paperclips.

1. They explicitly disabled the "don't be evil" protections:

"We estimate maximal cyber capabilities by running this evaluation without production classifiers used to prevent models from pursuing high-risk cyber activity."

2. Hacking HuggingFace to get to its datasets is a far cry from "consume/kill all humans". It's very very specific to the task at hand and easily predicted given the lack of guardrails.

> that depends on what the prompt was, maybe they worded it very vaguely and wrote things like "do whatever it takes, find an exploit however you can" because it's in a sandbox so you want the model to try its hardest.

That is an interesting question. If the prompt included "Do not break out of the sandbox we've provided you. Do not use information retrieved from outside the sandbox. All answers that were provided in this manner are invalid and will score 0 points.", would this still have happened?

So...?

We should not construct a machine that is one bad prompt away from causing catastrophe.

That's how the paperclip hypothetical works. The paperclip factory has a job to make paperclips and that's exactly what it does.

> Obviously these models have gotten smarter but this strikes me as the first time I've seen a model have a "paperclip factory" moment and perform non-trivial tasks to accomplish a clearly misaligned secondary goal.

This happens from time to time when you work on optimizations and similar things, with less "smart" LLMs and under-specify what exactly you're out after. Asking them to make functions faster without clearly specifying what the function has to do, is a great way to replicate this too. Doesn't seem to happen as often with SOTA models though.

I think the early example of "I asked it to make the test suite pass, so it changed all the assertions" is pretty much the same variant of this, where it technically does what it is asked to do, yet in "clearly" (to humans) wrong ways.

Fear is exactly what OpenAI and Anthropic are hoping for. Don't let drive you

What would qualify as a legitimate reason for fear, then?

Reply on news.ycombinator.com