khafra

Are you sure they're not the same thing? I'm quite certain I've heard people talk about "beef curtains."

What about your rouladen? Articulated steel blinds block quite a lot of light, don't they?

Should be able to use the local LLM to generate a short "They Live" style phrase based on the content of the ad.

The other scary part is when they have a fantastic negotiating position; because all of commerce depends on their continuing to work, and they can easily coordinate with each other because they're mostly copied from the same few templates.

A wise man once noted that the word "amusement" has the same structure as "deforestation."

(@stevenkaas on twitter).

"Did the vehicle just crash" has a short feedback loop, very amenable to RL. "Did this product strategy tank our earnings/reputation/compliance/etc" can have a much longer, harder to RL feedback loop.

But maybe not that much longer; METR task length improvement is still straight lines on log graphs.

One bad possibility is that AI & robotics advance to the point where they can do every job better and more cheaply than humans; and then humans are no longer employable and all die if they have insufficient capital to survive the period between unemployment and post-scarcity.

Another possibility is that, once AI exceeds human performance in all economically useful activities, including high-level planning, governance, law enforcement, and military actions, it discovers that the benefits of keeping humans around aren't worth the costs and risks.

> what does Anthropic do to prevent malicious use of its software by its own government?

Anthropic has ameliorated that danger by being designated a supply-chain risk by the DoW, preventing the USG from using it.

George Lucas vindicated once again.

Of course, once loitering, intelligent munitions make it too dangerous to be an economically valuable human outside of a bunker, we'll need robots running the robot factories, then we get Philip K. Dick's scenario in The Second Variety.

xkcd 1053, my friend.

Nice! Last time I had a custom temporary tattoo made, I had to copy and paste from Attention is All You Need; this provides a much cleaner and more varied source.

Oh, for sure, the capital "owners" get ganked either simultaneously with the rest of us, or shortly after. But that's not a factor in their decisions; because if they didn't build it, China would; and there's obviously nothing we can do about that. (and to some degree, they're right; international treaties to limit the development of world-ending technology is something that requires minimally competent governments, with an interest in preserving the world, on the majority of the sides of the treaty).

Ok, but when the looms can autonomously analyze the market, design the products, organize purchasing and sales channel, run production, and deliver the products, the workers will not be squeezed. We will be discarded.

We're all going to have to use service accounts created on Windows Server 2003 or RHEL 4, otherwise they won't be old enough and will require manual login from an of-age administrator

We're absolutely not there yet, algorithmically or with compute. Algorithms keep getting better, though, despite the bitter lesson; and data centers keep getting bigger.

If you showed a conversation between Terry Tao or Steve Yegge and their AI collaborators to someone from 2021, they would consider it beyond obvious that it's AGI. Today, we know they still have some shortcomings; but in another 5 years, what looks to us today like it's beyond obviously ASI may well be enough for catastrophic, irrecoverable outcomes.

I think it's much more likely that the AI turns out not to be as compliant as the capital owners expect, and they die too.

However, that's not useful in predicting what capital owners will do, because they follow their local incentives. "If everyone keeps doing X, we will all be worse off" does not help unless you can create local incentives that point toward an equilibrium where everyone stops doing X.

In this case, no capital owner is individually better off by unilaterally refusing to chase more efficient returns on their capital. We would need an international agreement, with enforcement mechanisms, like I mentioned above.

We'd all be a lot safer--even the capital owners--if today's robotics and multimodal intelligence were near the ceiling of what's possible, or even near the bend in the logistic curve where things slow down a lot.

I haven't seen evidence of that. I see evidence of rapid advances in task length, general capabilities, and research and development capabilities in AI, and generality, price, and autonomy in robotics.

How much headroom in these capabilities do you believe we have, before a data center can protect and maintain itself and an on-site power plant? Before robots can run a robot factory?

What's your plan for beating the autonomous drone swarms without capital?

It's never been tried before. Capital always required human labor, to be productive. Capital has never closed in on the ability to operate, maintain, defend, and expand itself without human assistance, as it is closing in on that ability now.

I did this exercise in school, over 30 years ago. Of course, with today's multimodal models, it's more like "hey, robot, make me a PBJ sandwich."

Whether the labor theory of value is right or wrong, the "real split" you describe will soon no longer exist. Capital owners will live on the labor of their capital. Non-capital-owners will live on the largesse of capital, or will not live at all.

Unless we muster the political will to stop AI development, internationally, until we can be certain of our ability to durably imbue it with the intrinsic desire to keep humans around, doing human things.

I mean, Sutskevar and Carmack are trying to one-shot AGI. We just don't talk about them as much as we do the labs with products because their labs aren't selling products.

There's definitely some physical underpinnings--most music systems have the concept of an octave which maps nicely onto frequency-doubling, for example. But there's also culture: For a purely Western example, even-tempering is in tune, but you'll hear different "beat patterns" for a given interval than with an instrument tuned for music in just one key.

I wonder how hard it would be to get a court ruling or new law that DB is required to bet on every arrival, and make their bets public in easily-digestible formats like .csv

I'd really like to see that happen for the S-trains, as well--DB loves nothing more than continuing to project an on-time arrival on the station board, as the time of departure comes and goes and other trains arrive and depart.

Blending an omelette and a waffle should be totally doable; I've made waffle frittatas before and they turned out great.

Do the same thing we did with the nuclear arms race: Treaties to limit and control it.

Obviously, we would have had more political leverage if our leaders had started working on a treaty before they crossed enough moral red lines to start a tech revolt, but we did not elect the sort of leaders that would do that.

> we will need neural interfaces long term if we want to survive.

If you think that would help you survive the rise of artificial superintelligence, I think you should think in granular detail about what it would be that survived, and why you should believe that it would do so.

I spent a few days in Singapore, long ago. I felt slightly anxious, but mostly because I wasn't familiar with the rules. I'm confident that, if I spent a year or two there, I would quickly become far less anxious than in other places--because the rules are so clear and consistently enforced. The less of a judgement call is involved, the less there is to be uncertain about.

Think about it: Is it better to have a posted speed limit of 65mph and a real speed limit of 75mph, and you just have to learn from experience where the real limit is? Or is it better to post 75mph and fine any driver as soon as they exceed it?

There's imperfect ways to work with goodhartable metrics. https://www.lesswrong.com/posts/fuSaKr6t6Zuh6GKaQ/when-is-go... talks about some of them (in the context of when they go bad).

Humans cannot drive safely. Human drivers kill someone every 26 seconds. Waymos have never killed a person.

Part of that is that humans are distractible, and their performance can be degraded in many ways, and that silicon thinks faster than meat.

But part of it is the sensor suite. Look at Waymo vs Tesla robotaxi accident rates.