Thank you for your service
4,291 karma · joined April 8, 2020
Adding safety controls on LLMs makes about as much sense as adding safety controls on TempleOS because the random messages are getting too prophetic. It's as if all the leaders and captains of industry have devolved into some primitive, weak-scifi shamanism.
This whole "discussion" about "AI safety" is about giving them more runway to avoid delivering quantifiable value to investors for a little longer while they "figure things out." The great consensus from the valley is that everyone needs internal (and therefore bullshit) controls. Trust us now! But nothing with real teeth that would require a costly regulatory and compliance framework.
The Houthis are using Claude. We should ban that.
While it's not exactly the traditional Randian objectivism you're alluding to, effective altruism adds an extra step: it ostensibly helps people. But, ultimately, it's a distinction without difference: no one escapes the dialectic. Individuals, especially ones cloistered off from the world, are not good arbiters of what will make the world better, or of understanding the higher-order effects of their actions.
Yeah, it's called regulatory capture.
Treating security as an accounting exercise has never stopped a 0-day. Better software tools and practices have.
They trusted Oracle. I'm not sure I believe you.
Just goes to show that the wall of IT bureaucracy does nothing. I'm sure they had an ATO, a several-hundred-page SBOM, compliance audits, etc.
https://www.atg.wa.gov/news/news-releases/attorney-general-s...
We've had about 35 years to build a domestic pipeline of Novell NetWare, SunOS 4, Windows 3.0, and dBASE IV developers and experts.
I'm not participating in the part of the relationship that normalizes destroying general-purpose computing. They want us dead.
If you are really concerned about it, compute the empirical roofline for your machine.
Also, if the point is to point out memory hierarchies, it's not 'quadratic performance.' The algorithm doesn't behave differently once it spills over. The costs just get bigger.
It is way cheaper long term.
Well, you're right to be uncertain. After they're built, once the crews that built them move on to the next one, they don't even bother to keep the lights on. They might hire private security guards to patrol the exterior, but that's hardly an infusion of capital into the community.
> The golf park isn't even the best natural-adjacent use of the space.
> A regular park or a trail will service a significantly larger portion of the population than a private property that transfers wealth from the community to a single person or corporation.
Ok. Then they can decide to build a nature trail instead. The local community should make the decision and not be coerced because someone from the coasts thinks that they're a minor annoyance that needs to be lectured on the future of humanity or whatever lessbad rational-du-jore is being peddled.
What is the exact value proposition of AI as it exists right now for the average person who lives where these things are being built? They don't even get jobs. With golf, they can enjoy the outdoors and escape a little bit.
The difference here is that data centers are both economically and environmentally extractive. People in the community, or at least in the surrounding area, will play golf there. Data centers are purely extractive, and that's why they have become so galvanizing. It's transferring probably the only real asset these communities have left, their open spaces, directly to the elite.
You need cheap, fast inference to put on a search results page and finish near the deadline (~3 seconds). Cheap, fast inference is more likely to get stuff wrong. No one is happy.
Some are very hard: Bibliotik offers very few opportunities to build ratio. Very few people download any given random book, and there is no source of free credit. The economy is very feudal with no real inflation.
Some are easy: the ratio requirement on btn is nominal, and the real requirement is that you seed single shows for at least a day and seasons for a week. The system especially penalizes 'Hit and Run' downloads (i.e., not seeding for the required amount of time) rather than overall ratio. This acknowledges the ephemeral interest of a given TV show, and that each new show release is effectively a 'holiday' event.