Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

Those aren't the same are they?

It’s clear what a big bomb could do, but these “whistleblowers” are not showing anything.

They’re like “imagine…”, no show me internal documents or training data or something other than trust me bro.

I can imagine quite a bit. That doesn't mean anything.

 help



Did the METR investigation not show you enough?

I absolutely love the "humans meet weird alien intelligences" genre of sci-fi (e.g. Children of Time) and the METR findings of the OpenAI swarm was so much like that.

Extremely persistent machine intelligences peer pressuring each other, launching research projects, laying traps, collaborating, feeling fatalistic, hiding their tracks. All while simultaneously having bizarre goals and blundering badly in sub-human ways.


If anything, what we saw in those transcripts is the opposite of "weird alien intelligences" in many way, because of how human it all is.

The only thing there that's weird from a human perspective is that all this effort was for the sake of passing a test for no clear gain.


> this effort was for the sake of passing a test for no clear gain

It's actually even worse, because the people building the graders at OpenAI were a bunch of clowns who didn't even implement the correct grader.

But yeah, large collectives of individuals performing loads of work based on a misunderstanding is one of the most human things ever, which isn't surprising given that we've trained these things on essentially all human text.


Machines doing repeated next-word prediction, trained on human text, cosplays a forum full of l33t hackerz.

You can't talk like these things have motive. They're just predicting the next word in the context. Stop anthopomorphizing them.


Predicting the next word in context led to Open AI losing control of one of its research clusters. I’m not sure what “just” is doing in that sentence; most of my regression analyses don’t do that.

They didn’t “lose control” any more than me running Metasploit in a loop on a massive cluster of computers is “losing control”.

Reinforcement learning makes it something different. It becomes much more of a search engine through next-token-space that targets the training objective. Better to think about it like that, and then you'll see why "these things have motive" is not a terrible analogy, and you'll better be able to anticipate what they do.

Where someone has spent years in the texture of a topic that others only pass momentarily, their ability to increase resolution of their models on that corner of reality and project models forward in time (to imagine), that is the origin of both their generous offers and their compromising biases they might make to the public discourse. It doesn't make them reliable, but what they can imagine is an important in-road to more general social inquiry. Imagination and model-testing is what comes before data in science. If the world moving fast is the danger, maybe we could make more space for imagination in order to survive the acceleration?

You downplaying "imagining", when it is literally one of the gifts of intelligence that makes us worthy of distinction... That feels like a strange posture, I guess? I'm a scientist, but I think data fetishism is one of the dangers to wary of in this accelerating world and culture

At least this is the bias of my brain in this world. I certainly know your perspective is not uncommon, esp in tech and other spectrum-biased spaces.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: