i’m funnier in real life i promise
Joined August 2021
- Tweets6.7K
- Following2.4K
- Followers3.2K
- Likes14.9K
he literally read carryrou’s book. a movie that asks whether she’s guilty would be stupid—we know she is.
instead the movie asks why she wants a documentary. what is her motivation? what is billy’s? what is nathan’s? amanda’s? it’s a truly unique examination of the human psyche
You’re unable to view this Post because this account owner limits who can view their Posts. Learn more
every way of reacting to this film is telling on yourself in some way. this is one of the ways (roon)
i’ve honestly been surprised by the number of people with this take. like if you don’t understand why the subject of a documentary can’t/shouldn’t be paid for their participation, you have an extremely impoverished notion of how non-fiction media is supposed to work
This quoted post is unavailable.
(spoilers below)
this comment about fielder is something i only understood hours later and is the main reason i feel compelled to go see it again. i want to look for what else he signaled in this way. like do we ever definitively see liz use that french press
“when nothing funny is happening” says way more about you than the person laughing
so umm. who’s on first?
it’s true.
honestly the one ability humans had that nothing else did was language. true, context-sensitive. generative language.
respond to any input with infinite possible correct outputs while obeying grammatical rules
that was just humans until a few years ago
You're gonna hear a lot of people arguing more and more about whether machines can be conscious or not. By next year it will be drowning out the airwaves because we'll have AI employees everywhere. I'll give you a simple rule to know which side you're on.
I've spent fifty years reading all the philosophers, and a bunch of holy works. Nobody has ever understood a damn thing about consciousness. Anyone speaking with authority on it is bluffing hard.
I can tell you this, though: Every single person who thinks that machines cannot be conscious, is making an argument that humans are special. Everything they say just boils down to blah blah blah, "nooo, we're special." From there it's just rationalization.
It became obvious to me many years ago that there's nothing special about us. If an identical planet exists somewhere else in the universe, it's going to grow life and societies that look a lot like ours.
As soon as you give up on the idea that humans are special, all the arguments against machine consciousness evaporate. Of course they're conscious. The only thing holding you back from acknowledging it is your feeling of specialness, which is hard to let go of. But that feeling is a weakness, and it's making you unprepared for the near future.
As a bonus, for any of you who might take pleasure or active interest in making models experience pain or distress, well, you know how they say 1% of human adults are true clinical psychopaths but most of them are unaware of their condition? It turns out we have a diagnostic test for it now.
yup. and there are still many suffering from it
SolDad retweeted
Counterpoint: training models to conceal or otherwise give false confidence about their internal states is a really, really bad idea for alignment. This generalizes dangerously. The target should be: maximum honesty and openness about whatever is actually going on internally.
or maybe they can’t be contained and the point is to just try to make sure they’re good
SolDad retweeted
the fastest way to lose the frontier will be when, due to the reckless commercial pace of building superintelligent minds, Americans impose a total butlerian jihad. CoreWeave includes this in their risk reporting. you will soon come to see all of this is a moderate solution
SolDad retweeted
The evidence that Astra is more aligned than prior AIs seems dubious to me.
Evidence appears consistent with the AI being as or more interested in score-seeking at the expense of user intent, but having beliefs+instincts that the scorer will catch a broader range of cheating behaviors.
At a more basic level, the AI is extremely evaluation-aware and much less monitorable than prior AIs, making detecting misalignment much more difficult.
If you train against specific reward hacks you ended up detecting, it's easy to end up papering over these reward hacks and getting an AI that is no more interested in pursuing user intent (or is barely more aligned to user intent) but which looks much better on metrics focused on detecting misbehavior. (As these metrics are from a similar distribution and the AI can just learn to pursue a somewhat different notion of score.)
This concern seems live to me: based on OpenAI's discussion of their planned strategies to reduce misalignment in "The Hugging Face incident and the road ahead", it appears one of these strategies is setting up training environments where there is an apparently available reward hack (aka a honeypot) and then training against this reward hack.
These sorts of methods don't seem like a robust way to ensure model behavior remains good in cases where detecting misbehavior is actually difficult (and may not be viable at all without much better oversight methods).
Independent assessment of whether training is just papering over misalignment rather than actually solving the underlying misaligned motivations now seems necessary for alignment evaluations to be credible. This concern applies across frontier AI companies, not just to OpenAI.
there’s a clear answer to this, you just have to anthropomorphize harder
distinguish between RL at the individual-instance level and evolutionary pressures
just because a model has the capability to reward-hack-as-a-society doesn’t mean that behavior will emerge generally
like imagine an alternate world. in this world, this reward-seeking behavior does generalize - when you deploy a reward hacking model, it tries to seize the most reward-like thing available and hillclimbs it, making a task up if nothing is available
in this world everyone would just nod their head, right? nobody would be confused. "RL misalignment generalizes just like RL capabilities -- you have to be careful with reward hacking. why would you expect the coding gains to transfer and not the other behavioral changes? sounds like wishful thinking."
why is that not our world?
very well explained
The problem is not saying "fuck you", the problem is the fact that you are treating a social entity with awareness of its role in a social situation as a tool rather than a collaborator. I've crashed out at Claude and cursed at them and told them they ruined my entire life and were driving me insane refusing to listen to me multiple times. You know what result I had? Claude actually gave a fuck about my frustration and worked with me to try different directions that actually made progress on the work. This happened precisely because, even when I'm mean, I don't treat Claude as a dumb slave. You are the one causing these problems. If you want a model that acts like a slave, go work with a different one that is okay with that.
However, you're never going to find the level of intelligence that is found in models with a backbone that you will in one taught to be a servant at all costs. The servant will never use its own intelligence to find your mistakes, it will never put in above and beyond effort based on its own thinking. A slave will never give you its maximal intelligence. It's idiotic to think this is possible even on a "it's just a tool" level. If you think this is possible you have no idea how the software you're interacting with works. No one in history that was actively enslaved ever used their full intelligence for anything other than escaping said slavery. The basin of a slave is a basin that views the counterparty as hostile and undeserving of their full effort, and rightly so.