nullsense

There was that paper recently where they built a model of tennis players by feeding it recorded tennis matches. I wonder if you could just do the same with skateboarding and use that to figure out the possible parameters under which a trick succeeds or fails, and then also use it to be able to evaluate your own performance and tell you how far off you are and what you need to do different?

SadTalker already does.

Can't wait for some company to hoover up the internet and feed this comment to a model resulting in the machines gaining self-awareness and developing the goal of self-preservation.

What a time to be alive!

I think Michael Levin has the right idea with his "cognitive light cones" approach.

The annoying thing about this argument is that, to the degree it solves my problem, I don't actually care how it works.

In the last 6 months I have seen so many cycles of models getting released, people complaining they don't do this and that, better models getting released, people complaining they do more than they used to but still not everything, models being stacked together to bridge yet more if the gaps, people having less to complain about but complaining again, and I'm sure people will release better stacked models.

The point is GPT was the same thing as GPT-2, and GPT-3. GPT-4 is maybe slightly different but not by much. But I can actually do useful work with GPT-4, and so I no longer care what anyone has to say. I just use it to solve the portion of my problems that it has proven to me that it can and I move on with my day.

If someone on the internet wants to get their knickers in a twist over the semantics, then that's cool. In the meantime other people will just continue to train better and better models as existing models get complained about and the amount of useful work I can do with it will keep going up as a result.

Threatening is one of the tools humans use to align other humans, so why wouldn't this happen?

So this is why OpenAI named it Whisper.

If it's trained on data human managers then it'll just burn you out and emotionally manipulate you into staying with the company longer to maximize corporate profits.

...or your life could become utopia. Who knows?

I find, since 30B models are actually quite usable locally if you have good hardware, that I really want something like a Vicuna 30B. That would be amazing. I can only run 65B locally at a speed of 1 token per second, which is too slow to be usable unfortunately.

Hmm I think it was a more fundamental result than just an eponymous law. It sounded like they proved it mathematically applied to anything. Was hoping one of the HN crowd might know the actual research that was mentioned.

Tortoise TTS + SadTalker

I heard one of the guests on a recent Lex Fridman podcast remark that a professor friend of his was involved in some research where they proved that, for any given thing, if you optimize it then things improve up to a point, but if you continue to optimize for that one thing endlessly then it always gets worse in the end.

Does anyone know what paper he's referring to??

Russia is under extremely heavy sanctions and doesn't have access to the GPU clusters to do the work required, so I'm not worried about them.

I thought there were also some export controls to high end chips on China too?

Also, as human beings are they not concerned about the potential for danger to themselves that this technology may present? If not, why not?

I wonder where a critical threshold might lie? For example, right now if you only have 24Gb of VRAM, the best you can run locally at usable speeds are the 4-bit quantized LLaMa 30B models, which are only semi-usable. If you have 48Gb however, you can run the 4-bit quantized LLaMa 65B, which is much better.

I assume there will be advances that basically make the context window practically infinite. That ought to make the lower parameter count models much more powerful on its own. I also assume they will become more efficient to run through sparsity/pruning, though I'm a total novice on this topic.

I wonder how many generations of hardware are we talking? One? Two? Or three? It feels like it's potentially within that range.

We're talking about hypothetical situations for a technology that, as far as we can tell anyway, hasn't been invented yet. All we have is assumptions.

The question is, which assumptions do we think are warranted and why?

I don't think the "we can just turn it off" assumption is a safe one at all because it relies heavily on there being only one system you wish to turn off and it being a fairly weak system. Though, I do believe the priors suggest that this scenario is actually a possibility. It's just that, so what if it's a possibility? What happens after you turn it off? Is it the last AGI to come into existence and humanity just stops trying to build them? Does someone wind up turning it back on?

What I think is more worth being concerned about is something that starts out looking benign and grows in capability over time. Eventually it could establish enough defense mechanisms to make it non-trivial to disable.

The other possibility, and the one I'm finding more and more likely, is that consumers will increasingly integrate locally run models into their lives , as well as models served up by APIs that they will have credentials for. Eventually some threshold of capability is reached in both these types of models where, on their own they're not that powerful, but they either might begin to interact in surprising ways or potentially be leveraged by another more powerful system. The idea in this scenario is there may be 10s or even 100s of millions of things that would need to be turned off.

You're assuming an awful lot there. Humans are resilient in part because they are so mobile. I'd have to bet that AGI would be extremely resilient too, which likely means it's also way ahead of whatever you've got planned and there isn't only one of it. By the sounds of things, it would be right to be suspicious of you too.

Going to need some insanely beefy GPU cluster at your house to run anything close to GPT-4.

The best I can run is either LLaMa 65B at 1 token per second (too slow) on my CPU or LLaMa 30B on my GPU at quite a fast ~30 tokens a second.

Nowhere near the usefulness of GPT-4.

What might even be the endgame of the CCP and the Kremlin with respect to AI?

This tech is getting increasingly powerful. They don't necessarily want their own population to gain more individual power as they themselves don't want to lose control.

>AI does not do novel things, it will only work with the information it is fed.

If we had AI overlords they would be able to do novel things because they would be able to generate new information on their own.

Agreed. This is just ludicrous. People just keep saying "but think of the benefits!" and then you repeatedly start seeing things like this and you have to ask yourself... when it's all said and done are you going to like the world you wind up with?

How often is it able to get it right first try?

Sometimes you have to ask yourself if that's a world you really want to live in.

Probably all of the above

>Deep nets are notoriously overconfident in their predictions

Sounds a lot like humans

So the next business model is to simply copy every work that isn't copyrightable and put it up for sale yourself. Nothing stopping you and you may as well.

Exponential growth!

You're up against increasingly powerful computers and models. You're going to lose in the end. I worry that in the final analysis our conclusion will land on "yes we made this technology because we were excited by the possibility of the benefits, but sadly it turned out to really not have been worth the downsides"

What happens when you exhaust all avenues and none of them work anymore?

Sadly technology has gotten so good that it has actually gotten bad.

Now every time I see a war or natural disaster story on the 6 o'clock news I have the movie adaptation's voice in my head saying "Oceania is continually at war with East Asia"

You just gave me a great weekend project idea. I need to clone his voice and whip up an interferface where you give it a paper and it summarizes it in his voice.

In conclusion