Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[Awaclus]

For some reason, I have been able to seed a lot lately. My current session is at 947 GiB uploaded (I started it on that same day I had to terminate my previous session which reached something like 600 GiB uploaded).

[silverspawn]

What Synecdoche, New York says about relationships:

- people want others to like them so they lie about who they are, showing only the parts that they think the other person would like
- therefore they get into relationships with people they're not compatible with, and the better they know each other, the less they'll like each other
- everyone is selfish
- men want to take care of others, but they don't actually want them to feel better, they rather want others to depend on them so they can feel good; it's also an entirely selfish motivation
- outside an active relationship, people will reach out to you when they need you for something and won't if they don't
- the idea that if you find someone who has x you'll be happy is a farce; instead you'll stop caring about x once you get it. You can probably never be satisfied
- people start treating each other cruelly once they can take the other person's affection for granted
- you're not even guaranteed to love your own children

To summarize:

[Awaclus]

@silver mostly, but I guess anyone who feels like chiming in: what political solutions do you think there are to managing AGI risk?

[silverspawn]

@silver mostly, but I guess anyone who feels like chiming in: what political solutions do you think there are to managing AGI risk?

Uh, I don't think I have that much to say about this. I'm assuming you've read the FLI open letter and Eliezer's piece in Time about it not going far enough. I think Eliezer's basically right. Even with me assigning only about a one in six chance on LLMs getting to AGI, that still makes it just about the most irresponsible thing humanity has ever pursued, so shutting it all down seems like the smart thing to do.

And the political angle, like what is and isn't feasible and how you best argue for this kind of stuff, that's obviously super important variable, but I don't think I have anything to add there. I guess one thing is that I'm skeptical the government is competent enough to implement any complicated regulations, so if something is tried other than an outright ban, I'd want it to be as simple and unambiguous as possible so that it can't go wrong and backfire in practice. Robustly slowing capability should be good.

[silverspawn]

Well I saw Mass (the 2021 film) and holy shit I almost want to give it a 10, probably a strong 9, it was absolutely incredible. It's one of these "all in one room, dialogue only" films but  so much more powerful than any other one who's like that. And if it were up to me, I'd give at least one actress an oscar.

[silverspawn]

Madeleine Gravis: Listen, there's an absolutely brilliant novel written by a four year old.

Madeleine Gravis: 'Little Winky" by Horace Azpiazu.

Madeleine Gravis: Hardly, Litty Winky is a virulent anti-Semite. The story follows his initiation into the klan, his immersion in the pornographic snuff industry, and his ultimate degradation at the hands of a black ex-convict named Eric Washington Jackson Jones Johnson Jefferson

-- Therapist in Synecdoche, New York

[Awaclus]

In other news, I have surpassed 1TiB uploaded in my current session. (I'm at 1.024 TiB at the moment to be exact)

I wonder why I'm getting to numbers like these so easily lately, even though my typical session used to be like around 20-100 GiB uploaded and I rarely even hit 200 GiB. Is it because I have accumulated more stuff I am seeding and thus more often finding peers to seed to, has my connection improved that much, or am I just having longer sessions?

[Awaclus]

@silver mostly, but I guess anyone who feels like chiming in: what political solutions do you think there are to managing AGI risk?

Uh, I don't think I have that much to say about this. I'm assuming you've read the FLI open letter and Eliezer's piece in Time about it not going far enough. I think Eliezer's basically right. Even with me assigning only about a one in six chance on LLMs getting to AGI, that still makes it just about the most irresponsible thing humanity has ever pursued, so shutting it all down seems like the smart thing to do.

And the political angle, like what is and isn't feasible and how you best argue for this kind of stuff, that's obviously super important variable, but I don't think I have anything to add there. I guess one thing is that I'm skeptical the government is competent enough to implement any complicated regulations, so if something is tried other than an outright ban, I'd want it to be as simple and unambiguous as possible so that it can't go wrong and backfire in practice. Robustly slowing capability should be good.

The main problem is: what exactly should we be banning? Surely it's fine and even incredibly beneficial for humanity to train neural networks in the vast majority of cases where people do that, it's just the big ambitious projects that potentially escalate the development towards AGI.

For context, I'm thinking about how to address AGI in the pirates' political program, which means that as long as I can convince the others to add it there, it doesn't actually have to be feasible to implement it. We already have a lot of stuff in there that isn't feasible to implement.

[silverspawn]

The main problem is: what exactly should we be banning? Surely it's fine and even incredibly beneficial for humanity to train neural networks in the vast majority of cases where people do that, it's just the big ambitious projects that potentially escalate the development towards AGI.

For context, I'm thinking about how to address AGI in the pirates' political program, which means that as long as I can convince the others to add it there, it doesn't actually have to be feasible to implement it. We already have a lot of stuff in there that isn't feasible to implement.

I think the right way to do it is to put a hard cap on how much computing power you are allowed to use to train a model. Unfortunately, this is not trivial to research because people don't put total number of computing steps into the abstract. It's not even trivial to measure although probably not that difficult, but I've never dealt with the "how to quantify compute" issue. A naive approximation would be # of parameters * # training steps, but there's probably a lot wrong with this.

The more tangible thing is model size. GPT-2 had 1.5 billion parameters, GPT-3 had 175 billion. LaMDA has 137 Billion. (And GPT-4 is unknown since they didn't publish the architecture.) I'm not sure exactly how large non-language models get (and they don't always put it into the abstract, so you sometimes have to make semi-complicated calculations), but going by this blog post, VGG16, which seems to be one of the largest image classifiers, has 138 million. So you could disallow training models with more than a billion parameters. But I don't think model size is really a good metric because  (a) I think the difference between AGI-ish and narrow models is smaller than wrt compute, and (b) it's very unclear how many parameters you really need. At one point, I read that most of the LMMs were actively too large and it was actually better to use smaller models and train them properly. And even if larger models are better, idk how much better.

If possible, you should probably get someone who's really in the weeds and work out how to quantify compute exactly, what a sensible cap would be, and how it should be scaled down over time.

[silverspawn]

my forearm (not the wrist) mildly hurts from playing iwfmd and typing doesn't seem completely harmless either. This is a stupid development and I don't like it. I never issues before. Obvious irony that it happens after I ditched all non-iwfmd video games

[Awaclus]

If possible, you should probably get someone who's really in the weeds and work out how to quantify compute exactly, what a sensible cap would be, and how it should be scaled down over time.

It also doesn't have to be very precisely defined (and indeed would probably be better off being as vague as it can be without risking being confused for something stupid). Like, I don't think any other party has anything about AGI x-risk in their programs, so the sweet spot to hit here would be successfully signaling that we are aware of the concept and have a higher than 0 amount of desire to do something about it, because that makes us the best party to vote for if you are an AGI x-risk single issue voter.

Insofar as we might want to get more specific, the specifics would probably ideally not include technical details, but political details, e.g. founding a state-affiliated research institute or something like that.

Would something like "The theoretical possibility of an existential risk from AGI should be addressed. Training powerful neural network models with hundreds of billions of parameters should be tightly regulated in accordance with suggestions from AI safety experts." sound reasonable-ish?

[silverspawn]

Why "theoretical possibility" rather than "possibility? Isn't that double-kill?

Otherwise, sounds good.

[silverspawn]

Actually it's a triple kill since "risk" is already probabilistic.

[silverspawn]

and if you want to appeal to tech people, say neural networks, not neural network models

[silverspawn]

i typed that with my left hand only

[silverspawn]

does the universe expect me to actually read stuff? i can't do that! reading is so hard

[silverspawn]

i read for like 1 min and then im like this is boring lets switch tabs

[silverspawn]

 i guess tts without doing anything on the side is worth a try but that feels so wrong

[silverspawn]

mb i should take on physical puzzles

[silverspawn]

yeah im getting bored just listening i wanna do sth

[Awaclus]

Why "theoretical possibility" rather than "possibility? Isn't that double-kill?

It specifies the type of possibility we're talking about. I guess it could actually be just theoretical x-risk.

[silverspawn]

well i saw the signal, no im not gonna bold w/ one hand thats dumb

anyway it was stupid idk why there are no good horror movies. 5/10 i guess

[silverspawn]

400 pages in, Kaufmans novel has reached a level of such utter ridiculousness that its very difficult to stop reading

[silverspawn]

presumably I should like go to the doctor for the arm thing

but

like a) it's the worst possible time with the easter holidays and I'd have to wait a bunch anyway, and b) why I can research it just fine by myself and it seems like this thing only requires changing habits and stuff, not medications. so why would I go to the doctor?

[faust]

presumably I should like go to the doctor for the arm thing

but

like a) it's the worst possible time with the easter holidays and I'd have to wait a bunch anyway, and b) why I can research it just fine by myself and it seems like this thing only requires changing habits and stuff, not medications. so why would I go to the doctor?

Well medication is not the only thing you can get from a doctor. You could get
- physical therapy
- a splint
- there are these rubber band thingys you can use yourself for muscle-strengthening exercises
and maybe other things.