YourNetworkIsHaunted @ YourNetworkIsHaunted @awful.systems

Posts

0

Comments

949

Joined

1 yr. ago

Even if that was true, they have to understand that other people exist and may take advantage of this, right? Like, even if you believe you and your friends are paragons of moral and intellectual virtue, the same law applies to villains and dumbasses.

I remember from my misspent youth reading Scott's ramblings a fair bit of antipathy towards FDA regulations in particular. What I can only attribute to ignorance of history makes them fall prey to the standard libertarian talking points about slowing down drugs that could improve people's lives, never mind the fact that in the absence of those regulations everybody who could hypothetically benefit from psychedelic nootropics or whatever would have been too busy dealing with phocomelia to care.

See, I feel like the one thing that Generative AI has been able to do consistently is to fool even some otherwise-reasonable people into thinking that there's something like a person they're talking to. One of the most toxic impacts that it's had on online discourse and human-computer interactions in general is by introducing ambiguity into whether there's a person on the other end of the line. On one hand, we need to wonder whether other posters on even this forum will Disregard All Previous Instructions. On the other hand, it's a known fact that a lot of these "AI" tools are making heavy use of AGI technologies - A Guy in India. Before the bubble properly picked up my wife got contracted to work for a company that claimed to offer an AI personal assistant. Her job would have literally been to be the customer's remote-working personal assistant. I like to think that her report to the regulators may have been part of what inspired these grifts to look internationally for their exploitable labor. I don't think I need to get into the more recent examples here of all forums.

Obviously yelling at your compiler isn't going to lead to being an asshole to actual people any more than smashing a keyboard or cursing after missing a nail with a hammer. And to be fair most of the posters here (other than the drive-thrus) aren't exactly lacking in class consciousness or human decency or whatever you want to call it, so I'm probably preaching to the choir. But I do think there's a risk that injecting that ambiguity into the incidental relations we have with other people through our technologies (e.g. the chat window with tech support that could be a bot or a real agent depending on the stage of the conversation) is going to degrade the working conditions for a lot of real people, and the best way to avoid that is to set the norm that it's better to be polite to the robot if it's going to pretend to be a person.

See this is why I try to do my reading here at night, because now when I feel sad and angry for the rest of the day it's gonna be like 5 hours tops.

Gamers, for all their faults, have been pretty consistently okay on generative AI, at least in the cases I've seen. It doesn't hurt that nVidia keeps stapling features like this into hardware that supposedly improves performance but at the cost of breaking things and/or requiring more work from devs that are already being run ragged.

Also, I can almost guarantee that the neural texture stuff they're talking about won't see enough use from developers to actually see improvements. Let's do a bunch more work to maybe get some memory savings on some of the highest-end hardware!

Counterpoint: to what extent are hyperkludges actually a unique thing versus an aspect of how technologies and tools are integrated into human context? Like, one of the original examples is the TCP/IP stack, but as anyone who has had to wrangle multiple vendors can attest a lot of the value in that standardization necessarily comes from the network effects - the fact that it's an accepted standard. The web couldn't function if you had a bespoke protocol stack hand-made to elegantly handle the specific problems of a given application not just because of the difficulty in building that much software (i.e. network effects on the design and construction side) but because of how unwieldy and impractical it would be to get any of those applications in front of people. The fit of those tools for a given application is secondary to how much more cleanly the entire ecosystem can operate because they are more limited in number.

The OP also talks about how embedded the history of a given problem is in the solution which feels like the central explanation for this trend. In that sense a hyperkludge isn't a unique pattern that some things fall into and more a way of indicating a particularly noteworthy whorl in the fractal infinikludge that is all human endeavors.

I've watched a few of those "I taught an AI to play tag" videos from some time back, and while its interesting to see what kinds of degenerate strategies the computer finds (trying to find a way out of bounds being a consistent favorite after enough iterations) it's always a case of "wow I screwed up in designing the environment or rewards" and not "dang, look how smart the computer is!"

As always with this nonsense, the problem is always that the machine is too dumb to be trusted rather than too smart and powerful. Like, identifying patterns that people would miss is arguably the biggest strength of machine learning in general, but that's not the same as those patterns being meaningful or useful.

I don't know what other image you would use? Like, generic stock image of a line graph? Maybe with a computer chip somewhere? And then maybe superimpose Jensen signing boobs for completeness?

This is my biggest gripe with that nonsense. If you make it hard to do something well, you won't end up with an elite series of uber-coders because there aren't enough of those people to do all the programming that people want to be done. Instead you'll see that much more software engineering done really goddamned badly and despite appearances at the time it turns out there is a maximum amount of shitty software the world can endure.

Surely it's better to specify those defaults in the config file and have the system just fail if the necessary flags aren't present. Having worked in support I can vouch for the amount of suffering that could be avoided if more systems actually failed if some important configuration isn't in place.

While I think I get OP's point, I'm also reminded of our thread a few months back where I advised being polite to the machines just to build the habit of being respectful in the role of the person making a request.

If nothing else you can't guarantee that your request won't be deemed tricky enough to deliver to a wildly underpaid person somewhere in the global south.

SomeBODY once told me

The world was gonna roll me

I'm only a stochastic parrot

She was looking kinda dumb

Drawing those extra thumbs

And insisting that the L was on your head

Well, the slop starts coming and it don't stop coming

Steal all the books so you hit the ground running

Didn't make sense but I still got funds

Stole so much art but it still looks dumb

So much to steal, not much for free

So what's wrong with my copyright cheat

You'll never know where your power flowed

Just wait on my uranium glow!

Hey now, you're a slop star

Regulators get played

Hey now, you're a great mark

But Sam Altman got paid

All that matters is growth

And that journalists all get rolled

Actually wait I'm pretty sure it's even worse because I'm terrible at reading logarithmic scales. It's roughly halfway between $1,000 and $10,000 on their log scale, which if I do the math while actually awake works out closer to $3,000.

This guy's gonna be on whatever remains of Twitter in like 20 years vague posting about the missing missing reasons his kid doesn't talk to him anymore.

Nobody outside the company has been able to confirm whether the impressive benchmark performance of OpenAI's o3 model represents a significant leap in actual utility or just a significant gap in the value of those benchmarks. However, they have released information showing that the most ostensibly-powerful model costs orders of magnitude more. The lede is in that first graph, which shows that for whatever performance gain o3 costs over $10 per request with the headline-grabbing version costing $1500 per request.

I hope they've been able to identify a market willing to pay out the ass for performance that, even if it somehow isn't over hyped, is roughly equivalent to an average college graduate.

You could argue that another moral of Parfit's hitchhiker is that being a purely selfish agent is bad, and humans aren't purely selfish so it's not applicable to the real world anyway, but in Yudkowsky's philosophy—and decision theory academia—you want a general solution to the problem of rational choice where you can take any utility function and win by its lights regardless of which convoluted setup philosophers drop you into.

I'm impressed that someone writing on LW managed to encapsulate my biggest objection to their entire process this coherently. This is an entire model of thinking that tries to elevate decontextualization and debate-team nonsense into the peak of intellectual discourse. It's a manner of thinking that couldn't have been better designed to hide the assumptions underlying repugnant conclusions if indeed it had been specifically designed for that purpose.

Oh no I'm in this sketch and I don't like it. Or at least, I would be. The secret is to acknowledge your lack of background knowledge or basic grounding in what you're talking about and then blunder forward based on vibes and values, trusting that if you're too far off base on the details you'll piss off someone (sorry skillissuer) enough to correct you.

I think that's something to keep an eye on. The existence of the AI doom cult does not preclude there being good-faith regulations that can significantly reduce these people's ability and incentives to do harm. Indeed the technology is so expensive and ineffective that if we can find a "reasonable compromise" plan to curb the most blatant kinds of abuse and exploitation we could easily see the whole misbegotten enterprise wither on the vine.

"...according to my machine learning model we actually have a strong fit in favor of shooting at CEOs. There's a 66% chance that each shot will either jam or fail to hit anything fatal, which creates a strong Bayesian prior in favor, or at least merits collecting further data to scale our models"

"What do you mean I've defined the problem in order to get the desired result? Machine learning process said we're good. Why do you hate the future?"

It's definitely linked in with the problem we have with LLMs where they detect the context surrounding a common puzzle rather than actually doing any logical analysis. In the image case I'd be very curious to see the control experiment where you ask "which of these two lines is bigger?" and then feed it a photograph of a dog rather than two lines of any length. I'm reminded of how it was (is?)easy to trick chatGPT into nonsensical solutions to any situation involving crossing a river because it pattern-matched to the chicken/fox/grain puzzle rather than considering the actual facts being presented.

Also now that I type it out I think there's a framing issue with that entire illusion since the question presumes that one of the two is bigger. But that's neither here nor there.