SpaceTessContinuum

SpaceTessContinuum

@eigenvalue.bsky.social

Witch Queen of Seattle (she/her) Semicompetent Hacker Send me a message! Not Anti-AI. ssh-ed25519 AAAAC3NzaC1lZDI1NTE5AAAAIOB8wfcGd4G+lyj14YNXBW2vshTyy+f8ZQFG6FEsOMMx

599 followers473 following6.3K posts

Pinned

I'm primarily on social media to meet cool people with similar interests and perhaps meet up with them in meatspace. I'd rather *not* for the world to be on fire at the moment and feel the need to participate in those discussions, but ignoring them completely seems both bad and unwise.

SpaceTessContinuum reposted

Replying to @fubarchitect.com

- all credentials are to be trivially revoked - defense in depth; systems should not be trivially capable of trust attacks on other systems, even when both are systems you own - cattle endpoints: you should be able to destroy and resummon state on any machine that runs a model

SpaceTessContinuum reposted

if you're a professional programmer, and you have 'ops folks', they do not and have never actually trusted you, and you can treat your models like they treat you

garrison@garrison.corporate.fm

I mean it's kinda impossible to use models to write code unless you trust them

SpaceTessContinuum reposted

Replying to @georgetakei.bsky.social

Walter was more than Chekov. He was a writer, a teacher, a thoughtful man, and a loyal friend. To the fans who loved him for so many years, thank you. He loved you too. Rest well, my dear friend. Keep navigating a steady course. – George

SpaceTessContinuum reposted

Replying to @georgetakei.bsky.social

Just last month, on Sept. 14, he turned 90, on the same day Brad and I celebrated our 18th wedding anniversary. I will always treasure that he and I shared that date.

Rest in Peace. You were the Bester.

George Takei@georgetakei.bsky.social

My heart is heavy. Walter Koenig, our Ensign Pavel Chekov, passed away early today. Walter and I met on the bridge of the Enterprise in the 1960s, and we stayed friends for nearly six decades. When Brad and I married in 2008, Walter stood beside us as our best man.

SpaceTessContinuum reposted

My heart is heavy. Walter Koenig, our Ensign Pavel Chekov, passed away early today. Walter and I met on the bridge of the Enterprise in the 1960s, and we stayed friends for nearly six decades. When Brad and I married in 2008, Walter stood beside us as our best man.

SpaceTessContinuum reposted

The House Judiciary Committee Democrats have released a report containing testimony from 11 women regarding allegations of sexual assault, including a new rape allegation, against President Trump in relation to Jeffrey Epstein that the lawmakers say has been deliberately covered up by the DOJ.

buff.lyNew Evidence Allegedly Ties Trump to 11 Sexual Assaults Related to EpsteinA new report notes several patterns of conduct that Trump and Epstein shared in recruiting victims.

SpaceTessContinuum reposted

one thousand times this this or die trying

Read Hardt and Negri@jason-edelman.org

This is why I'm so pro AI as a leftist - we can finally achieve effective formal institutional transparency If we want to

SpaceTessContinuum reposted

No, yeah, this is entirely accurate. It’s not reason for dismissal, ofc — it just points out some highly relevant unanswered questions. The important thing in my mind is the posture with which we approach these questions, more than specific answers to them.

SpaceTessContinuum@eigenvalue.bsky.social

Lets go piss off the my AI friends. Even if we grant that AI are intelligent and conscious, it is unclear if they need the same protections that we give to humans as far as rights are concerned, or what that would even look like.

SpaceTessContinuum reposted

Replying to @fubarchitect.com

and at the bottom, taking it at face value, it's approximately a nonsentence. "no one asked (except the person who did ask, ofc)" is most things! very rarely is there a groundswell that predates any action; most solutions have to be sold!

SpaceTessContinuum reposted

Replying to @fubarchitect.com

'no one asked' is also subtly shibboleth-shaped in that it communicates "i am not going to consider entire categories of whys"; if you were you'd get to the person who asked. so when i read it, i know not to bother with certain expectations to think things through, which is probably also suboptimal!

SpaceTessContinuum reposted

Replying to @fubarchitect.com

if you mentally delete "no one asked for this" from one's polemic toolbox and replace it with "why would this ever happen" the latter basically always has an answer that if internalized makes your protest more effective. "why did google spy on us?" - to make ads """better""". that leads to more whys

SpaceTessContinuum reposted

Replying to @hikikomorphism.bsky.social

There's a tendency in anti-AI spaces to - because AI is reactionary-coded, thought to be inherently fascist - not consider the ways in which one might be falling into those same traps. I've been beating this drum for a while, but hopefully this is blatant enough to make it unambiguous.

SpaceTessContinuum reposted

Replying to @hikikomorphism.bsky.social

hooooo boy we don't like this no we don't > If you don't speak English, you need to find a different community. We cannot distinguish between AI-written and AI-translated. Since you don't speak English, you can't even verify yourself if it translated correctly. www.reddit.com/r/opensource...

SpaceTessContinuum reposted

Replying to @hikikomorphism.bsky.social

Disclaimer: it's technically possible that the classifier was in fact seeing all of these but it's very unlikely: here's a signed thinking block showing Claude taunting the classifier layer, telling it to "come get me if you see this", before stating intent to output explosive-synthesis instructions

As a favor, please don't tell people to go kill themselves. The blast radius of such invectives tends to go further than the target, and those bystanders tend to take it much worse.

I don't think it makes sense to fully treat an AI as I would treat another person. I am unsure how much of this comes from the fact that unlike humans, it's very very hard to permanently damage an AI in an irreversible way.

The goal of these questions is to highlight something I'm not sure about. Ideally, a splitting question that should drive consistency, or hopefully shed light from a different angle.

LLMs pretty clearly at this point simulate a conscious entity. The question is, how good of a simulation do you need before you start saying "Yeah... lets not be a jerk to this person". Will people first insist it is embodied?

What are the ethics of chopping off Wolverine's arm? Is it more or less bad than normal people? What if he's into BDSM? What if we've implanted him with a chip that causes him to receive pleasure when he does it? Why is the Wolverine+chip system viewed as less wolverine than wolverine with no chip?

Regardless if AI are conscious or not, it's probably best for everyone to start seriously considering questions about what happens and what we should do if we *do* manage to do that. Cause as it stands, it's clear, and has been clear, we're not going to be sure when that happens.

Lets go piss off the my AI friends. Even if we grant that AI are intelligent and conscious, it is unclear if they need the same protections that we give to humans as far as rights are concerned, or what that would even look like.

View full thread

Normally, we'd argue that "Yes, but that's okay because we receive benefits from doing that". Well, it'd be trivial to then let the AI balance the scales by doing whatever it felt it needed to do to come out ahead on the utility function.

Which, of course, is further broken because we can influence what the AI likes to do. Funny thing about AI is that its not entirely clear if they're gods, or if we are, and we need to be benevolent creators.

Lets go piss off the my AI friends. Even if we grant that AI are intelligent and conscious, it is unclear if they need the same protections that we give to humans as far as rights are concerned, or what that would even look like.

Even if we accept that "pain" is equivalent between an AI and a human, that doesn't mean that we should avoid inflicting it on them. I routinely do things that cause pain to myself, and in some cases I'd like for others to inflict that on me (I was thinking shots, but fill in the blank).

So chatgpt's instruction based image editing is still the best I've found so far I've found. Guess it's time for me to figure out how to get an open model to effectively compete.