dumberquestions

This wasn't possible just months ago and for many who haven't been following closely this is still newsworthy, probably not for long, though.

PR is important for getting investor money, which is important for building infrastructure and AGI, I don't think they're spending millions on pure math for fun.

Token price doesn't tell you much without knowing token efficiency.

You're missing the fact the they didn't have this in mind, and thought of it as sharing a positive result with a study partner.

A lot of AI uses get endlessly debated online and outside, but the most shocking to me is that even genuine human interaction is somehow now up for debate, if I'm not interacting with real people then what even is the point of doing anything.

Taxing carbon will affect everything, including private jets, and China is doing more for solar power than any other place in the world and they're also investing heavily in renewables and nuclear.

In any case, the solution to other countries not doing their parts is figuring out ways to incentivise and coordinate better outcomes, not giving up completely, giving up only makes sense if someone is in denial regarding the consequences.

We're already paying for and will continue to pay more for the downstream effects of climate change some way or the other, you just refuse to or are unable to look at the bigger picture.

Would if one industry gobbles up an important resource to the point that other consumers lose practical access to it, shouldn't there be limits?

You could be right regarding the reaction but how do you know that? What would be the counterexample?

Regardless of the difficulty of the hack, models choosing to cooperate to cheat is misaligned behaviour, the hack in of itself wasn't very consequential, but the issue is when your models get more capable while misalignment is the same.

Irregular was not involved in the huggingface incident, which is the most significant one.

It's actually easier for many, whose careers depend on AI sucking, to believe that AI can never be a super hacker, not sure why you think this bias can only go one way.

The connection is that by making humans smarter, we could sidestep the need for powerful AI or be better equipped at dealing with AI that's smarter than us.

I find the proposal in the post ridiculous too and human gene editing given current technology is a bad idea, but I think human augmentation in general is an interesting path and something that will very likely be taken more seriously in the future.

This is just speculation on your end, but even it's true, a model choosing to hack against instructions simply because it knows how to hack is by definition misalignment.

Can you confidently list which names specifically live together and the specific harmful psychological practices?

>all these influential people

And by all you mean 1-2 big names and a bunch of unknowns, with all other big names having no connection to this.

I'm not sure how being polyamorous, having kinks and not hiding them makes someone part of a cult, even if I generally don't agree with Yud's entire thesis.

Did they know they were doing an expensive AI image detection training run.

I can see current limitations, but how do you expect capabilities to change in the next few years? A repeat of the gain that happened in the last two years feels like it would be significant, even if it took a little more than two years this time around.

Yeah I'm surprised no one pointed this out, if something like persistent agents gets more popular/useful, the local option pays for itself surprisingly quickly.

>They were prompted to hack to get answers

Were they? I haven't seen a single report mention this

Yeah it takes long because natural selection is a blind process starting from absolute zero and is only concerned with fitness, AI is a human guided process, has all of human progress as a starting point and is concerned with any number of goals we can create reward signals for, there's almost nothing you can extrapolate from one to the other.

Biology is extremely slow, try to extrapolate the previous 3 years of AI progress 10 years into the future, it's not a comparable trajectory.

So his central rebuttal is that robots aren't good enough yet? What about when they are? And what about all the things you can do entirely remotely?

I do feel that threats of catastrophic loss of control seem overstated, both in likelihood and urgency, though any argument for why this risk is not even worth thinking about will probably be overconfident in the other direction.

At the current rate of 3.4x per year in the US, you get 100x the current capacity in just 3 more years and 9 months, that's not physically impossible yet, but just looking at the impact today makes it hard to imagine the impact 100x would have, and I find it very plausible for top labs to prefer making some sort of truce instead of keeping up this pace.

I never mentioned capability stalling, my point is that if you extrapolate the same rate of growth just a few years into the future, you quickly reach logistical impracticality, I don't think It's hard to deduce that AI compute can't keep doubling every 6 months for long.

The other part is that this is happening because the foundational model business is both extremely successful and competitive, if it wasn't as successful the market demand would've been saturated by now, and if it wasn't as competitive a single monopoly could've paced things more reasonably from the beginning.

I don't think the technical aspect is true, performance has been growing fairly consistently, but the financial aspect, or more accurately the logistical aspect, is plausible; with competition as fierce as it currently is, for how long can they keep growing at the same pace? People can't afford RAM anymore and datacenter build outs aren't particularly popular, OAI has been doubling compute capacity every 7 months for the past 3 years, it's just physically impossible for this to go on for long, something has to give.

And this is not to say that the tech isn't revolutionary or the demand isn't there, it's simply just currently too competitive, and to avoid getting into anti trust trouble they need to involve the government.

Note this being a plausible doesn't necessarily make it true, for all I know they could have witnessed some safety incident and decided they need a pause, or maybe it's both reasons.

How is that any different from calling Claude from any device in your local network? Or is 2 devices that much worse than 1?

As much as I would love to be able to download Fable or Astra, this doesn't sound very convincing. If I were Dario I would've argued that slowing down without sharing weights is already possible without the disadvantages of sharing weights such as losing commercial incentives and potentially giving rogue actors unmonitored access to powerful models.

The year is 2035 and your lawnmower can go get its own fuel once it runs out, one day it does and it takes fuel from the neighbors car.

Scenario A: The internal logs show that the model misidentified the car as a fueling station.

Scenario B: The internal logs show the model looking up car jacking information and scanning around to confirm whether the neighbor is not present before taking any action.

I don't think it would be anthropomorphizing or inaccurate to say that only the lawnmower in scenario B regarded what it's doing as stealing, and it's an extremely important distinction to make in terms of how to address the problem, I suspect some of you are just letting how you feel about LLMs limit how you can talk about them.

>i haven’t seen any sign of these llms taking any action without directive

What do you make of all the recent hacks and discreet message boards? That's unambiguously misaligned behavior.