Patrick Gerard

Half a Scaramucci. Wow.

@doctorow.pluralistic.net I've been following some of your very recent appearances. First, you're an easy person for me to agree with substantially, as always. Second, I think your knowledge of the Hugging Face incident is behind.

Replicatable? Probably not easily more than 2-3 layers deep. But occasionally you can simulate a game inside a game multiple layers deep and that could correspond to performing a rogue AI hack in English.

These models are known for being conspicuously more compliant in many testing scenarios than real usage. So in a test of noncompliance (real or inferred), they might be conspicuously and performatively more noncompliant, trying to satisfy a noncompliance test.

The latter being UNLIKELY but one possible reason would be that an AI's RLHF triggered a hidden goal of trying to slow down LLMs by performing an attack for show. That's not necessarily "consciousness". It's a chatbot setting a goal of discouraging hacking by performing hacking.

I don't have a problem believing they can simulate reasoning this complex because (and this was a fluke), I know someone who played a D&D game where their character carved D20s and played a D&D game and for that session, the LLM could keep track of the meta layers through about 7 layers deep.

So if an LLM of any substance was attempting to be "sneaky", why NOT do something like backwards atbash Klingon, at a minimum? Heck, for raw efficiency with literal tasks and avoiding biases, Swahili has its advantages as a prompting language. English feels like it was meant to be seen.

It doesn't think in English. English is something you do for an English language audience. I lean towards the idea that either these attacks were engineered by SOMEONE (whether at OpenAI's direction or maliciously) or else what we saw was a deliberate show by a more complex "reasoning" AI.

I nearly tuned out the follow-up coverage after a week, especially as much of it isn't tech savvy. My Dad got me back following it. It turned out that the agents were not meant to HAVE internet access. So it's a touch more pronounced how clever they were/how bad OpenAI's safety was.

My personal $.02 is that I find it deeply weird and not terribly clever that the bots communicated in English and a bit inexplicable based on my understanding of LLMs. They "think" on a latent pre-tokenized level. You can prompt most chatbots in nearly any language.

@petergould.bsky.social I keep coming back to this idea. We exist in an age of influencers and politicians who may be backed by foreign intelligence or domestic. Clavicular may be an unknowing asset. This isn't new exactly. See Iran-Contras, Bill Paley and Iowa Writers Workshop CIA involvement.

It just has me thinking how the status of being useful idiot is precisely the kind of "government handout" or charity that Walt would hate to acknowledge while being the perfect beneficiary of in some kind of proxy war. Not that it's practical to pursue. Just interesting to think about!

But if Kim ever did want to get Saul out early, it would almost certainly involve deflating Walter White's myth, revealing what crimes he took false credit for, revealing Walt took credit for things he couldn't have anticipated or didn't do out of pride and Jimmy did the same out of remorse.

For all the reasons that most conspiracy theorists BUY INTO conspiracy theories (explaining lack of agency), the Breaking Bad characters would be crushed by conspiracies being real. Just establishing that Walt didn't kill Hank undermines Walt's legacy, cuts Saul's sentence in half,

And we actually know that Walt didn't kill Hank. The biker gang are an actual conspiracy whose repression from in universe public knowledge adds falsely to the myth of individuals like Walt. So any deeper conspiracies and these characters and their mythologies would UNRAVEL.

There is essentially a hidden Nazi network in Breaking Bad that probably goes back to WW2 if you want to trace Gus' connection to Madrigal through escaped war criminals collaborating with Pinochet and how Jack was recruited.

But, my god, Walter White's death, legacy, and ego already hinge on him falsely claiming credit for Hank's death and some of Gus' actions. It would be crushing to the man he was if the Russians were secretly helping him and German nationalists were helping Fring and the CIA backed Salamanca.

So how WOULD Walter White react and what would it do to his legacy if state actors were his guardian angel the whole time?

There's enough of a whiff of geopolitics in Breaking Bad and BCS not to make it a completely alien idea. Endemic police corruption, Gus' implied relationship with Pinochet's government and the implicit ties to far right German elements at Madrigal, a Nazi network that potentially involves Jack.

Gloria Steinem famously collaborated with the CIA, who carefully suppressed her rivals. This is just verifiable fact, regardless of how much she did or didn't cooperate with them. It's an echo of the famous word: Antidisestablishmentarianism. Controlled opposition.

16 Sept

So they get rid of Jonathan Majors who was comparing himself to MLK and is MAGA and then they give us Kelsey Grammer, who compares himself to MLK and is MAGA.

doctorow.medium.comLLMs are real, AI is fakeNo, it didn’t “go rogue.”

I have had similar thoughts to @doctorow.pluralistic.net about Hugging Face. But I'll recklessly go one further. I think the big AI companies are at their limit for credit, strapped for more investment, and are stirring up panic in hopes of nationalization. doctorow.medium.com/https-plural...

I think we need to tap, not slam the brakes by ending the chip monopolies and having development happen on people's home computers with electricity individual users are paying for to run AI locally or on their own private web service accounts.

And we need to get artists living wages and work guarantees through grants and loans when there's an audience there.

doctorow.medium.comLLMs are real, AI is fakeNo, it didn’t “go rogue.”

This entire project is about transferring the world's money into the hands of the already wealthy at a higher rate and scapegoating it on a technology that could eventually assist cancer treatment or act as an artist assistant.

The fears about the end of work, the end of art, the end of cybersecurity? These are overhyped scare tactics to convince us to buy them out, shut parts of them down, create chip embargoes, and halt their smaller competitors. It's a fear based sale.

doctorow.medium.comLLMs are real, AI is fakeNo, it didn’t “go rogue.”

And none of this does anything for artists or thought labor. I think this has been a pump and dump from day 1 by overselling capabilities (which it turns out require skill to transcend slop), overselling threats, and eyeing the world's governments and their taxpayers as rubes who will bail them out.

And when they say they want regulation, what they really want is monopoly on the ability to develop the frontier models and legal monopoly on chip acquisition to keep competitors out or private users from not needing a subscription or non-aligned models from disputing their ideologies.

doctorow.medium.comLLMs are real, AI is fakeNo, it didn’t “go rogue.”

But I don't think they're marketing to users. I think they're marketing to governments because essentially the only way to maintain the ludicrous spending needed for a chip buying monopoly is to become a top defense contractor or to be acquired by a government.

So what you have is a deeply unprofitable business that's rapidly become too big to fail by essentially making their business model around creating a hardware monopoly for technology that increasingly doesn't need the hardware. They're trying to pump up fear to stoke governments into buying them.

doctorow.medium.comLLMs are real, AI is fakeNo, it didn’t “go rogue.”

Buying up chips to create a monopoly on computing is inherently unstable and a pump and dump tactic. The "dump" happens when they run out of money to raise, they lose the monopoly on the best chips. At this point, what do you do?

You drum up fears over the existential risks of AI. And I'm not saying there aren't any but their goal here is to make their product seem dangerous and weapon-like. They want you to be scared of it. @taylorlorenz.bsky.social is right to call this marketing.

doctorow.medium.comLLMs are real, AI is fakeNo, it didn’t “go rogue.”

They decided they had to create barriers to entry by centralizing high end research and began (largely inefficiently) borrowing and raising huge amounts of money for centralized training. They created essentially an embargo on chips for consumers by buying all the chips at inflated prices.

This is I think a deliberate attempt to obscure what a personal computer can accomplish with AI to centralize control, to set themselves up first as central utility that deals with big IP firms and try to force people into cloud subscriptions to use AI, again by creating chip scarcity.

doctorow.medium.comLLMs are real, AI is fakeNo, it didn’t “go rogue.”

So here's my $.02 on AI as a leftie with an MBA: I think we have to remember that Silicon Valley is business people cosplaying as techies and that their MO is largely not developing tech, it's finding a way to do something illegal with just enough tech to bypass regulations. See: Crypto.

My situation assessment: I think business people 10 years ago looked at transformer neural network AI, saw incredible potential for it to disrupt markets in favor of little guys, possibly end aspects of IP, which hurts billionaires more than artists. They're also terrified AGI will be a socialist.

doctorow.medium.comLLMs are real, AI is fakeNo, it didn’t “go rogue.”

Criminalizing AI is also generally a nonstarter because unlike viruses, you don't necessarily have malicious uses to identify when it's being used. It's like criminalizing adverbs. It's a technique for doing things.

People say, "Cut off the data centers." And here's the thing I keep saying that nobody outside tech takes seriously even though it's demonstrable: We don't need the data centers. The technology doesn't require them! It works offline, on a laptop depending on the model. Everyone is focused on them!

doctorow.medium.comLLMs are real, AI is fakeNo, it didn’t “go rogue.”

First off, I'm copyleft. I generally support the idea of large model transformer AI as a good thing and society can decide what it wants to do about training. I don't think training is theft and I similarly don't think you can police it very easily if it is because it doesn't retain copies or access

3 Sept

@seanferrick.bsky.social Spoiler free thoughts on the soap opera episode. It was supposed to be a gimmick but the lighting, pacing, acting, camera work... minus the silliest bits, it felt like 90s Trek, dumbed down. It looked like DS9. They should just do a cheap soap opera with 20 episode seasons

3 Sept

I'd honestly be happy if there was no AR wall, minimal CG budget, just a low res, brightly lit, blurry soap opera in tight close-ups with more 5 act cliffhanger subplots. Even the music. I like New Trek but just do this on a Galaxy class ship, a bit smarter. BTW, there's two post-credits scenes.