Notices by Cassandra Granade 🏳️‍⚧️ (xgranade@wandering.shop), page 4

It's been several hours now, and the thing that just keeps striking me about Zuckerberg's speech/thread is how bloody bleak it all is. Corporate appeasement was inevitable, but this isn't that. There's zero attempt to make this look legitimate, he's even moving the moderation team to Texas, for fuck's sake. There's zero attempt to appeal to anyone outside of the MAGA epistemological universe, so it's all written in right-wing vernacular.

Final thought: go read @charliejane's post from a year ago about separating art from the artist. It's an amazing post, and well worth the read.

https://buttondown.com/charliejane/archive/jk-rowling-and-separating-the-art-from-the-artist/

If someone tells me a funny story that doesn't affect me too much either way, I tend to happily accept it as true with very little proof. If someone tells me to do something that could cause me significant harm, such as taking a proposed medication, I need substantially more evidence before I proceed.

I might reasonably be accused here of demanding too high a burden of proof; there's not a lot of positive evidence that is used to choose between programming models, for instance, and yet strong rational opinions exist there.

The thing is that from a decision theory standpoint, the burden of proof scales not only with the plausibility of a claim, but the negative impact that the claim being wrong would have.

If wrong, the claim that LLMs are useful has extreme negative impact indeed.

@KuteboiCoder Fair enough, but I'll argue that's optimism for new theory — my argument is scoped to LLMs as a technology, but I will absolutely warrant that a different technology could, in principle, realize the claims currently made about LLMs. The trouble is investing on the basis of a possible theory that hasn't even been proposed yet.

@KuteboiCoder It's not a hardware gap, though, it's a theory gap. The strategy right now seems to be that if you throw enough hardware at things, you can fool people into not realizing that the theory predicting the existence of an application is just... missing. Making hardware cheaper for startups doesn't fix that.

@KuteboiCoder For me to be optimistic, there'd have to be some theory predicting why better models would solve what appear, but all present theory, to be fundamental limitations inherent to LLMs in general. In lieu of any such theory, I don't see how improvements in models will help get any closer to an application.

With respect to "no demonstrated or theoretical," I want to acknowledge the existence of positive anecdotes about AI. I hate the cliche that "anecdotes are not data," as I find it to be thought-terminating and unnecessarily fatalistic about evidentiary standards. That said, I argue that those few scattered positive anecdotes don't rise to the level of "demonstrated applications," if only because the stochastic plagiarism nature of LLMs predicts that you'll occasionally lucky.

I keep thinking the AI bubble will burst soon, because at some level it simply has to: the idea that it is sustainable to continually produce and sell something that has no demonstrated or theoretical application is just mind-boggling.

On the other hand, Amtrak just today pushed an AI chatbot on me, and they're pretty far from the tech/VC hype cycle, so maybe the bubble is here for a bit longer.

Either way, I both hope and don't hope that the burst happens soon. This isn't sustainable, either from an economic or environmental standpoint. On the other hand, when it bursts, a *lot* of people who were in no way responsible for this absolute shitshow are going to get hurt. It's never the people who cause the problems who wind up getting held responsible.

From a theoretical basis alone, you would expect that a small number of positive anecdotes exist, and for the same reason that some people do in fact win the lottery. As a result, I don't believe it is rational to treat those anecdotes as sufficient evidence as to any demonstrated application of LLMs.

If those positive anecdotes were substantially more common, as in orders of magnitude, then that might be more compelling. But it's not what we see.

Since it's going around again thanks to yet another techbro blog post, let me reiterate that "separate the art from the artist" is a literary critiquing tool with a limited, tightly scoped realm of applicability, and is emphatically not a guideline as how to make economic decisions.

If you buy art, you fund the artist to keep making art. If you recommend art, you spend your own credibility endorsing the decision to give that artist money to keep making art. If you publicly display your endorsement of said art, and make your fandom of said art part of your visible identity, then you're dedicating part of your identity to the goal of supporting that artist's political ends.

Said blog post used the term "art" exceedingly widely to lend credibility to the incredibly toxic idea that one should actively ignore the homophobic political organizing of a tech CEO when evaluating the decision as to whether or not to financially support said CEO.

That in and of itself is a political statement: that the basest demands of capitalism should trump any human interest in the likely outcomes of political organizing. It's a nihilistic view, and should be regarded as such.

Probably the easiest thing to do until either Android KeePass clients support passkeys or Bitwarden supports passkeys for Firefox/Android would be to bite the bullet and use Proton Pass... that undoes a lot of progress I've made with getting away from cloud-based syncing, but it would technically work without trusting Google or Apple, I think.

Anyway, passkeys are cool. I just wish they were cool for more people and for people further outside the OS duopoly.

Running on Linux and not wanting to sync Android stuff with Google tends to put me outside of a lot of target audiences. It's a shame to see that happening for security technologies, though.

https://mastodon.social/@mcc/113766231196504649

In part, I've been as negative as I have on passkeys because they're very, very cool, but it's clear that the industry groups developing implementations don't consider any usecases except for those that involve syncing credentials with an untrusted provider like Google. Even if those credentials are locally encrypted, it's still really unfortunate to tie a massive and demonstrable security and privacy improvement to vendor lock-in.

Passkeys feel for all the world like yet another narrowing of the web around stacks that run WebKit or Blink and sync with Apple or Google. If you use that path of least resistance, they're great.

I know some folks have been able to make passkeys work well outside of that stack, thanks to the heroic efforts of projects like KeePassXC, but it still feels like running stuff on your own hardware is at most an afterthought for passkey implementers.

To my mind, this is not disconnected from the news that Tim Cook personally donated over a million dollars to the Trump inauguration. Chump change at Cook's level, but it signals strongly to me that I cannot trust Apple to be secure against fascist incursions — same thing with Google, of course.

Making trust work without relying on those parties is critical!