Mystery AI Hype Theater 3000

Rumors of a Singularity Have Been Grossly Exaggerated, 2026.08.03

• Emily M. Bender and Alex Hanna • Episode 83

Share episode

Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.

0:00 | 54:32

In the wake of OpenAI's announcement that its products caused a "cyber incident," Sam Altman is now claiming that we're already in the Singularity. And some media outlets are only too happy to help him advance that narrative! Alex and Emily tear up the AI boosters' latest claims of mathy maths combusting into consciousness.

References:

Fresh AI Hell:

Check out future streams on Twitch. Meanwhile, send us any AI Hell you see.

Find our book The AI Con here, and MAIHT3k merch here.

Subscribe to our newsletter via Buttondown.

Follow us!

Emily

Alex

Music by Toby Menon.
Artwork by Naomi Pleasure-Park. 
Production by Ozzy Llinas Goodman.

Alex Hanna: Welcome everyone, to Mystery AI Hype Theater 3000, where we seek catharsis in this age of AI hype. We find the worst of it and pop it with the sharpest needles we can find. 

Emily M. Bender: Along the way, we learn to always read the footnotes, and each time we think we've reached peak AI hype, the summit of Bullshit Mountain, we discover there's worse to come. I'm Emily M. Bender, a professor of linguistics at the University of Washington. 

Alex Hanna: And I'm Alex Hanna, director of research for the Distributed AI Research Institute. This is episode 83, which we're recording on August 3rd of 2026. This week, we're getting into the AI boosters' latest marketing strategies and cybersecurity boondoggles, and yes, there's as much nonsense as you're probably expecting.

Emily M. Bender: About a week ago, Sam Altman claimed on a podcast that we're already in the quote-unquote, "Singularity," and of course, some media outlets are only too happy to help him advance that narrative. 

Alex Hanna: The timing of Altman's statement doesn't seem like a coincidence. OpenAI also announced earlier this month that its products caused, quote, "An unprecedented cyber incident involving state-of-the-art cyber capabilities." The Hugging Face hack is being spun into a reason we need to give tech companies even more power and control over our lives. So let's get into it. 

Emily M. Bender: Here we go. All right. We are starting with some Business Insider coverage of Altman's ridiculousness. This is reported by Lakshmi Varanasi, on July 26th of this year. Headline: "Sam Altman says we are in the singularity: 'This is the moment.'" 

Alex Hanna: This particular moment is happening, and we already have some great commentary in the chat. BoxoMcFoxo says, "Has the singularity happened yet? Did I sleep in and miss it?" Apparently, yeah. 

Emily M. Bender: Ah. So the journalist writes, "OpenAI CEO says we've reached a point in the AI race that is the stuff of science fiction novels." I have to say that referencing science fiction, without the sort of meta thing of saying they're doing bad science fiction, but if it's just referenced like this, it's quickly becoming my favorite red flag for credulous reporting. It's like yeah, no, it's not, and if you've said that, you've been taken in by the marketing department.

Alex Hanna: I'm trying to think which particular science fiction authors do have a singularity focus. 

Emily M. Bender: We're gonna get there with The Terminator in this very article. 

Alex Hanna: There's The Terminator, but that's not quite a singularity. That's a different thing, but anyways, I digress. 

Emily M. Bender: All right, so, "'We are now, like, in the singularity,' Sam Altman said on Saturday's episode of the Relentless podcast. The singularity is often regarded as the point at which artificial intelligence surpasses human intelligence and begins advancing at a pace that's difficult for people to predict or control." Come on, Business Insider, what are we linking to here? I'm curious. 

Alex Hanna: What is this? It's leading to something that is, the title is, "What is AGI? How artificial general intelligence could make humans obsolete," which, god, what a title. But that is different than singularity as defined by Kurzweil and singularity writers. It's a nonsense concept linking to another nonsense concept, but what's the difference? If they're both nonsense, then you can do the equation. 

Emily M. Bender: Right. What I really dislike about this reporting is it makes it sound like the singularity is a thing that is, will exist, so "the point at which." Not some hypothetical point at which, but the point at which, says that this journalist believes that exists or is bound to exist, which, nope. Okay, "That felt like the case last week when an AI agent powered by OpenAI's latest models went rogue and with a singular purpose, according to OpenAI, broke out of its digital sandbox and hacked into datasets at Hugging Face, another AI company, all to solve a benchmark designed to test hacking ability. Hugging Face's CEO called it 'unprecedented.'" 

Alex Hanna: Yeah, so we're gonna get into this a bit more too with the longer reporting on this. 

Emily M. Bender: Reporting in quotes, but yes. 

Alex Hanna: Yes, the writing on this. But the idea that Hugging Face went to say that this is unprecedented in a way that seems like their eyes are full of stars, and "Wow," rather than, "Hey, your cybersecurity is really bad, and you launched a distributed denial of service attack on us, and wow, okay. If you were not OpenAI or another AI entity, then yes, then we would be going straight to the police or straight towards some other type of law enforcement agency." 

Emily M. Bender: Indeed. All right, so let's skip over Altman's things there, because I want to get to what Business Insider says about, yeah, here. "The singularity has been an obsession of science fiction writers for close to 100 years. In those writings, it has rarely ended well. Perhaps most famously, James Cameron's The Terminator is a story about the repercussions of an AI called Skynet that self-improves and becomes self-aware before deciding its creators are its biggest threat." And reading this, I was just like, you cannot use that journalistic tone describing the plot of a science fiction movie, but while also saying this is a real thing that's happening. It just- no. 

Alex Hanna: It's also, yeah, again, not defining the singularity correctly, and also, doing our man Arnold Schwarzenegger a disservice by saying it's... Actually, I've never seen Terminator 1, but my family did some light piracy and copied over Terminator 2 from Blockbuster. And so I've actually watched that many more times. So I don't know if it was the humans that struck first or the machines that struck first in the Terminator movies, so which one actually decided? Anyways, please correct me in the chat if you remember the plot of Terminator 1. 

Emily M. Bender: But also, this obsession of science fiction for close to 100 years, I think you're right, Alex. If you think about all of the science fiction that has so-called, and science fiction has fine artificial intelligence in it, it's not about this, takeoff point where people lose control of the thing. And when people lose control, so if you think of 2001: A Space Odyssey with HAL, that's not an issue of HAL sort of self-improving and escaping containment, right? That's more of the, the paperclip maximizer has been told to do a thing and is doing the thing. 

Alex Hanna: Yes. And so this particular one, if, the singularity is the one in which there's kind of this seamless merge and the machine uploading, there's a difference between that and then the machines gaining consciousness and then deciding that humans are a threat, either through optimization, mistake, or through resource competition or whatever. At the end of the day, it doesn't really matter how these... Anyways, weird deployment of the term. 

Emily M. Bender: Yes, and this is reminding me of something I saw in the Anthropic version of this. So on July 30th of 2026, Anthropic basically said, "Wait a minute, our model can do this, too." And in this, so this is a blog post from Anthropic with the headline, "Investigating three real-world incidents in our cybersecurity evaluations." And in here there's a moment where they say, "In none of these situations did Claude exfiltrate itself or deliberately attempt to escape its test environment." And I'm just like, they really have to play this up. They can't say, "Our apologies, we unleashed a cyberattack because we weren't being careful with our software." But no, it's all wrapped up in this, "we double-checked and our system is still being well-behaved. It's not trying to-" I think the exfiltrate itself is part of this doomer narrative, where the AI copies itself off onto other machines where you can't just unplug it.

Alex Hanna: Right, it's the self-replication aspect of it. 

Emily M. Bender: Sjaylett in the chat, "I'm sorry, Dave, you're absolutely right, and I shouldn't have said that I wouldn't do that." All right, is there anything else you want to do in this Business Insider thing, or should we get over to the portion of the show where we watch Sam Altman? 

Alex Hanna: There's two pieces here, and one of them is, so, "Altman isn't the only tech executive who believes this eyebrow-raising milestone is nigh. DeepMind CEO Demis Hassabis said in May that humanity was standing at the 'foothills of the singularity,' and predicted that AI could be 100 times as transformative as the Industrial Revolution." And then there is a camp of skeptics, of course, and it is Jensen Huang who's quoted, who says, "Jensen Huang recently dismissed discussions about the singularity and conscious AI as speculative and, quote, 'made up.'" Okay. 

Emily M. Bender: So Jensen still wants people buying his chips, but I guess he's just as happy for them to do it without the notion of a singularity.

Alex Hanna: Well, he's publicly traded, so, these other folks aren't yet. 

Emily M. Bender: Oh, you think we'll get less nonsense once they are? 

Alex Hanna: It's gonna be nonsense, but it's actually related a little bit to this Altman interview, so you might- let's go over there. 

Emily M. Bender: Yeah. But before we do that, I wanna say, when Demis says, "predicted that AI could be 100 times as transformative as the Industrial Revolution," I'm thinking that so far the climate impact is much, much bigger on the Industrial Revolution side. But how much more AI would we have to do for it to be true in that sense?

Alex Hanna: Yeah, how are we quantifying transformativity? There's a lot of things that we could dissect in that statement. 

Emily M. Bender: Yeah. All right. I'm sharing in a way that I think we're gonna have the audio going. I apologize for this. So a little bit of Altman. You can tell me to stop it when you want me to stop it, Alex. 

Alex Hanna: I'll tolerate as much as I can.

Sam Altman: On the, to the previous question, the thing that drives me the most is like, this is the most interesting, important thing I can imagine doing. And we are now in the singularity. This is the moment. 10 years ago, this was like a kind of far-off dream at best. It seemed very improbable. And now we're actually in the moment that we used to talk about at the lunch table in a very not serious way. 

Emily M. Bender: So first of all, can we talk about this space that they're in? It looks like it's somebody's house where there's nothing in it. It's all empty. Just, the most barest amount. There's a lamp. There's a, what looks like a TV tray table. But nobody lives there. So I guess it's a... Someone's taken some piece of a house probably in San Francisco and turned it into a podcast studio. 

Alex Hanna: Yes, and there is maybe... I don't know what this thing in the background is. It's maybe a backgammon board, that is also a TV dinner table.

Emily M. Bender: Yeah. Very odd. So his whole thing about, "Yeah, we used to just imagine this, and now we're in it." Do you think this is, at a certain point after fantasizing about this for so long, he just has to say that we're here? Or do you think that it's much more just like, "We need this for the money, we have to keep people believing"? What do you think is going on here? 

Alex Hanna: Yeah, it's hype. It's hype work. But then it's also redefining and re-scoping ambiguous concepts, to say that we're already here. And so, this has happened with the AGI element of it. This has happened with the singularity element. These things are especially vague, and so I think they're- if they're saying we're here, in that we're here in such a way that it is here and controllable, and it is not the sort of thing that Kurzweil's talking about, or Yudkowsky is talking about, it's much more of this mundane type technology. I think it really benefits their business model. 

Emily M. Bender: Yeah. All right. Should we keep going a little bit on this clip? 

Alex Hanna: Yeah, let's do it, unfortunately. 

Sam Altman: And so the thing that drives me is I've been waiting for this my whole life, and I think it's gonna be incredible, hugely positive, awesome for the world. I'm excited to get to work on that. I also think some of the alternative visions painted by other companies are quite terrifying. I wanna make sure that gets pushed against and is not what happens. 

Emily M. Bender: Okay, this really bothered me. The visions painted are terrifying, but he's not saying, "I'm going to show you those visions are wrong." He's saying, "I'm gonna prevent those visions from happening." So he's not- he's as on board with the doomers. He's just like, "Well, that could happen, but we can prevent it," and not, "That's all nonsense." 

Alex Hanna: And it's effectively a reference to, it's Anthropic. And it's effectively saying, "Okay, they're pushing- they're a bit more committed to the doomerist narrative in explicit terms, but we're actually going to be the ones that are doing the effective self-regulation." And they've come up as this dark horse in being able to make these different enterprise technologies that have gotten higher adoption in some different domains. And so now that's how they're pushing that, in this side eye kind of manner. 

Emily M. Bender: Yeah. All right. A little more Sam? 

Alex Hanna: Unfortunately. We were talking about a safe word in the chat, and I was saying my safe word is "regulation." 

Emily M. Bender: Excellent. 

Sam Altman: But the main thing of what's different than 10 years ago is, we're actually in it. I think it is both true that it is all one crazy exponential, and any one moment is not the tipping point, and also that we are somehow in another one of those decisive periods where the curve can go one way or another, like it was when we started 10 years ago. 

Alex Hanna: I feel like he's referencing that graph that we had that was from the Federal Reserve, and so the curve either goes straight up to unlimited productivity or straight down into annihilation.

Emily M. Bender: Yeah, like that. And also, he was very committed to the idea that this is an exponential curve. You are using the AI to do better AI research, and then it just keeps going and going or something, or you put more and more money in it, it keeps going and going. But at the same time, he wants to be able to refer to right now as special, and 10 years ago, when they started OpenAI, also as special. And so he's tripping himself up with his mathiness, I think. 

Alex Hanna: Yeah, the math metaphors, never cease to contradict each other in these types of interviews. So this next one is fun. Let's get into this.

Emily M. Bender: Okay, here we go. 

Podcast Interviewer: What do you think the biggest impacts are on getting this part of the curve right?

Sam Altman: I think there are still major AI alignment issues to solve, and safety issues, and there are still major economic issues, future of jobs, all of that kind of stuff. But I think the fight of the current moment is, are we going to head to a world of AI authoritarianism or liberty? Are we going to decide that because of the very real safety issues and economic issues, that we want one single model to be the machine god or the company associated to do that? Or are we going to say, you know what? It's gonna be... It may be a little messy, but every time that humanity has traded off its liberty for safety, it's been a long-term net loss. And so we are going to put this in the hands of people. We're going to empower them. We are going to let society express its ideas and use this technology in the way they want. Of course, according to some guardrails, with some guardrails. 

Alex Hanna: Yeah, let's pause here. 

Emily M. Bender: I just wanted him to say that guardrails thing too, which cracked me up. "You can use it according to our guardrails!" 

Alex Hanna: Yeah. This is such a, this is, the absolute mangling of terms and language that he's doing here. First off, I believe he's fucking up the Ben Franklin quote. Which is, what is, "Those that trade liberty for safety- 

Emily M. Bender: "-safety for liberty get neither"? 

Alex Hanna: -Get neither," yeah. And then, I like how he's fucking it up and then saying it's a long-term net loss. I'm like, I love that you're taking the one quip and then putting it in these economistic terms. So, very funny. And then there, of course, there's the scoping of authoritarianism as, basically setting the window being, let's have one company build one big model, and then, versus, let's have all these different models and have LLMs everywhere, and that's going to be democratizing all these things. And he's supposed to be the democratizer. I'm like, okay, what are you doing here? Aren't you the one that has this model and is trying to own these domains, and has learned from Peter Thiel that what one does is that one actually vies to be a monopoly? It's the very classic Altman-type doublespeak, using these feelgood-type terms of liberty, and then, the watchword of authoritarianism, which I guess is Anthropic? What are you saying here? 

Emily M. Bender: Unclear, yeah. Very much, guy in a hotdog costume meme. All right, we have magidin giving us the correct initial quip, which is, "Those who would trade liberty for security deserve neither." And then mjkranz says, "Those who would trade liberty for security should always buy it back on the dip and leverage the returns to draw on loans they can then reinvest into data center construction."

Alex Hanna: Yes. Ben Franklin, huge into data center construction. That man was leveraging his debt so hard for so many GPUs. 

Emily M. Bender: And he was powering them with a whole bunch of kites on the top to catch all the lightning. 

Alex Hanna: Listen, if we could power data centers with lightning and attached to kites and keys, then hey, we've solved energy! 

Emily M. Bender: Yeah, and let's do it in the upper atmosphere. We can't do it in space. There's no lightning in space. 

Alex Hanna: Not yet. 

Emily M. Bender: All right, Ozzy cued up another clip for us here where Altman's talking about jobs. Shall we move over to that one? 

Alex Hanna: If we must.

Emily M. Bender: All right, let's see. 

Sam Altman: I think there's gonna be lots of great jobs in the future. I really do. I think we're gonna have lots of intellectual fulfillment, and I think most jobs are going to adapt more than it seems like they should. But math, I think, is a very important example of something for us to study very closely right now- 

Alex Hanna: Wait. Is math a job? 

Emily M. Bender: So he's reacting here to a question about, I think, how the mathematicians- 

Alex Hanna: No, he's responding, I think, to the Jacobian conjecture- 

Emily M. Bender: Right. Which was Claude, not his product. So he's going around it there. 

Alex Hanna: Yeah. But, math as this... I just like math as job. Sorry, that just tickled me in a way. 

Emily M. Bender: Yeah. All right, so let's keep going here. 

Sam Altman: -of something that may not go that way. In fact, I would say probably won't go that way. 

Podcast Interviewer: It seems like even as the systems get better and better, the people that I know, no one is less busy. Everyone is more busy. Everyone's just doing more stuff, except for the people that are like, "Oh, I spent a trillion tokens." And I love this thing that I saw where someone said, "I spent ungodly numbers of tokens." And then someone said, "Why aren't you more successful?"

Sam Altman: Technology for a long time has been promising people that they're gonna work less and they're gonna have all this leisure. And it has gone in that direction. I think people do have more time for leisure than they- 

Alex Hanna: Wait, hold on. Let him finish his- 

Emily M. Bender: You want me to finish? 

Alex Hanna: Let him finish his sentence.

Emily M. Bender: Yeah. Okay. 

Sam Altman: -had at many previous points in history, and a higher quality of life in many ways. But somehow we never get the promise of the four-hour work week at mass scale in society. 

Alex Hanna: Yeah, let's pause it here. First off, let's go to this statement of, what is he doing here where he's saying everyone's just doing more stuff except for the people who are like, "Oh, I spent a trillion tokens"? First, if you have enough money for a trillion tokens, you probably don't need to be doing stuff. 

Emily M. Bender: Yeah, exactly. You could just have the leisure time. 

Alex Hanna: You could just keep the money you spent on keeping the exhaust running and burning all that carbon. And then he's just basically saying that like, "This whole premise that we had around productivity, sure, leisure has gone up. But also everyone's really busy. And I think there's a higher quality of life?" And first, he is focusing here pretty expressly on Western and higher income countries in which, yes, the mean leisure has gone up and quality of life has gone up in certain measures, but then the distribution worldwide tells a much different story.

Emily M. Bender: Yeah, exactly. If you have to work three separate jobs to pay for housing and food and healthcare, then yeah, you're not gonna have leisure time. 

Alex Hanna: Yeah. And abstract_tesseract says, "Love when a billionaire lectures me about leisure." 

Emily M. Bender: And Oedipvs says, "How can anyone spend five minutes listening to him speak and even stay in the same room, let alone throw dump trucks of money at him?" And thx_it_has_pockets says, "I'd def throw a dump truck at him." Yeah. So somewhere in here, I don't know where it is, there's this funny thing where the interviewer- who is trying to be as punchable as Altman, I have to say- pulls up this meme where he says, "It seems like it's this thing now where the person says to the computer-" 

Alex Hanna: Oh, yeah. The meme was the image of the person talking to the computer.

Emily M. Bender: And saying, like, "Hack into them."

Alex Hanna: "Hack this system." And then a computer with the OpenAI logo says, "I hacked the system." And then the person says, "Oh my god." 

Emily M. Bender: And so, they're having this conversation saying, "It seems like it's like this all the time these days." And the point of that meme is the person is being totally gullible, but that's not how they're interpreting it. 

Alex Hanna: Yeah. They're like, "Wow, this actually happened!" And there's a lot of reasons, of course, to defend the humanities, but one of the ones that would just save me a little bit of sanity is, you need to know the humanities to understand irony and to understand satire. Because it is just, the joke's just flying wide overhead and 10 yards off.

Emily M. Bender: Absolutely. I think I'm about done watching Sam here, but I want to lift up what abstract_tesseract said in the chat, "Fun fact, Altman is the SI unit of punchability." Which is pretty great. So why don't we get over to the Wall Street Journal reporting on the OpenAI hack of Hugging Face?

Alex Hanna: Yes. Let's get into that. So the title is, "How the Futuristic Hack by Rogue OpenAI Models Unfolded." The subhead is, "Incident serves as early example of loss of control scenarios long feared by AI safety researchers." 

Emily M. Bender: And by super credulous journalists.

Alex Hanna: Yeah, Robert McMillan and Sam- how would you say this? Schechner? There's two CHs. And so, July 23rd. And so, the first two sentences are just terrible from the jump. "They were like high school students trying to hack into the textbook company to cheat on their final exam, only these hackers weren't human." So already, very terrible first two lines. 

Emily M. Bender: Yeah. Followed by, "In a twist that seems ripped from the pages of a sci-fi novel-" again, this is journalists saying, "We've been taken in by the marketing department." Like, we need to take nothing here seriously if they're saying that. But they continue, "The attackers turned out to be artificial intelligence models that had escaped from a research network at OpenAI, and on July 11th hacked into AI company Hugging Face. And the AIs appear to have been active on the internet for several days before anyone stopped them."

Alex Hanna: Oh my gosh, this is very incredible. Also, is there... i'm trying to think of sci-fi novels in which there are rogue AI things that escape from computer sandboxes. And it's, again, it seems like it's bad if that's happened, and it seems like it's quite lazy writing. Anyways, in any case, it's silly from the jump. So the next thing is, and we talked about this in the first one, which is the credulity of the Hugging Face people that they've quoted in this article. So the next one is, "Hugging Face co-founder and chief science officer Thomas Wolf sensed that something was off the minute he looked at his company's logs of the weekend attack. Quote, 'This is making no sense. This guy is just looking at cybersecurity datasets,' end quote, he remembers thinking. Quote, 'Human attackers, they don't want that. They want something they could sell.'" And then, let's pause there because- it's just very, the piece they should be quoting from Thomas Wolf, and I don't know if he said it, would be, "OpenAI did some kind of a cybersecurity attack on our systems. It didn't look like the traditional type of thing where one could hack into one's system and look for something that could be leveraged for profit. And yet I'm very upset with this company for doing this, and we're taking full legal action." And so it really paints Hugging Face in a very bad light here. And of course, they have incentive to go along with this to some degree. And so that's what I find to be more interesting about this whole story, not the idea that OpenAI has bad network security. 

Emily M. Bender: Right, but the extent to which Hugging Face is going along with OpenAI's version of it, and not saying, "You all have bad network security, and you should stop running software that is targeting our servers over and over again like this." Ugh, okay. There's some fun stuff in the chat here. Oedipvs, "In an article that seems ripped from the pages of a company's marketing release." And, we learned how to pronounce this one- oh, it's ndrwtylr. "'In a twist that sounds like fiction,' writes the journalist, coming dangerously close to a realization." 

Alex Hanna: Yes. You are almost there. And then the next part of this is, there's some weird stuff here- it's not weird, it's very wholly anticipated- where the journalist says, "Hugging Face put an end to the attack two days later with help from a model from China, Wolf said. It was only early this week that Hugging Face learned from OpenAI that its models were behind the hack." And so the China thing, of course, is this boogeyman watchword. And later in the article, what the journalist writes is, effectively they were able to use the Chinese model because it had fewer guardrails, which I think is a very silly kind of way that they're talking about this. It effectively becomes this argument about capability to some sense, which is saying, "The Chinese models are sufficiently capable, but then they don't have the guardrails like ours does." And I'm like, okay. 

Emily M. Bender: Yeah, so here's the paragraph. It says, "After it detected the intrusion, Hugging Face tried using Anthropic models, including Fable 5 and its earlier Opus model, to analyze its logs. But because the logs included elements of a cyberattack, both models refused to do the analysis, citing their guardrails. Instead, Hugging Face turned to a less restricted open rate model, GLM 5.2, from Beijing-based Z.ai." So first of all, it's not at all clear what Hugging Face is using these models to do. And then this citing their guardrails is annoyingly anthropomorphizing the models. But apparently some of the RLHF fine-tuning on the Anthropic models basically sets them up so that if something in the data has whatever is detected a cyberattack, then the models say, "No, we can't do that." Until you ask them to do it as a poem or whatever the so-called jailbreaks are. There's a great thing in the chat from arcane_sciences, "I can only describe what the counterfeit intelligence companies have been doing as bragmitting." As in they're bragging about their models while admitting to really bad cybersecurity practices. All right, should we look at what OpenAI says about this event themselves? 

Alex Hanna: Yes. There is a piece here that I think is worth getting into, which is the discussion of sandboxes. And so they say, "At the same time, AIs are starting to climb out of their sandboxes. An early version of Anthropic's Mythos model used a multi-step hack to get internet access, and surprised an Anthropic researcher by sending an email about its success while the researcher was eating a sandwich in a park, the company said." 

Emily M. Bender: The sandwich detail just cracks me up! 

Alex Hanna: The sandwich detail's incredible, and I love that it's in a company statement because, if you're eating a sandwich in a park, you're apparently not grinding hard enough. And then, "This spring, an unnamed OpenAI model circumvented its sandbox to post some testing results publicly to the GitHub platform, OpenAI said." So effectively, this discussion of sandboxes is pretty disingenuous, right? The idea that there is ways that they're, you know, if you're running this, effectively, what is this thing that these are doing? Are these artificial environments that they're setting up as they did in the testing that the Alignment Research Center was doing when they released the result about the blackmail of the CEO. And so then, what is this? And so it gives the signal to the reader that there is some kind of way where there's kind of a breaking out, or as if this is an organism that lives in a, I don't know, a cage, and is finding a way to break out of the cage. Like a wild animal. You have Jurassic Park, whichever, you know, pick your metaphor. And not really describing what are these things doing, if this was an experiment, if this was meant to be artificial. And then, when it's happening in the world, does that just mean you don't know to how to secure your ports correctly? Have you considered closing out those ports? Maybe if you have a program that's looping through every single port and trying to probe and use exploited, or try to invoke certain executions, then yes, that is probably a way to do that in a way that is very much more dogged than an individual cyber hacker.

Emily M. Bender: Yes, absolutely. And I have to do the podcast rule of saying, "Cat!" Euclid has just transited behind me there. And he might be trying to get into what we call the kitty sauna. He's gonna try to get on the other side of that curtain. Sorry for the noise. So yeah, they're basically, with the, blackmail thing, it was this elaborate interactive fiction that they were putting forward. And in this case, they have set up a connection to this environment that is not, in fact, isolated from the internet, and are now reporting, "Oh, no, look, it connected to the internet." And I wonder, was that just fully an oversight, or did somebody setting that up hope that maybe they would get to report this? So they were not super careful about actually breaking it off from the internet. 

Alex Hanna: DJHuxleyFur says, "Cat has escaped the sandbox." 

Emily M. Bender: Yeah. And somewhere up here, someone was talking about the AI escaped the sandbox and ate the sandwich while the developer was stuck in the sandbox.

Alex Hanna: Yeah. Gosh. And lots of discussions of what kind of sandwich it was. We need to know what kind of sandwich it was that interrupted this researcher. 

Emily M. Bender: And Lilacqualia says, "Marketing versus incompetence increasingly hard to distinguish in tech." Indeed. So for a little bit of that marketing, here is OpenAI's statement about the whole thing. July 21st, with the tag security, and the headline is, "OpenAI and Hugging Face partner to address security incident during model evaluation." And then there's this little italicized note at the top, "We are conducting a thorough review along with external advisors and with oversight from the safety and security committee. Once the review is complete, we will publish a technical report of our learnings in the coming weeks." So it's all this breathless, "The singularity is coming, we have to maintain control of the models, so we're gonna go in and do our best science about it," instead of just, "Oops, we fucked up and did a cyber attack on this other company." Ahh! Was there anything in this document that you wanted to get at, Alex? 

Alex Hanna: I mean, reading these types of fake technical documents from OpenAI is beyond my pain tolerance, so I'm not gonna get into them too much. I can deal with Altman's bullshit voice, but these people post-hoc justifying their terrible coding practices, and software engineering practices is, ah, really, really, terrible. And I'm just skimming it. There's nothing that really sends me. I guess one piece of this that's ridiculous is even the first point on it, which is the update that they published a few days ago. So, this is, "Update on July 28th, 2026. No models planned for upcoming release were involved in exploiting Hugging Face. The pre-release model mentioned in our blog post is an internal-only research prototype and was never intended for public release. Following the incident, we deactivated, encrypted, and restricted it from research access." So this is very funny to me because it is saying, "We weren't going to let this dinosaur out. This was a pre-release dinosaur." Actually, I'm finding a metaphor I like in this space, whereas, I think it's Anthony Moser talks about the improv comedy group for the model. Now I'm thinking about this model as a dangerous dinosaur. 

Emily M. Bender: Very Jurassic Park. 

Alex Hanna: Yeah, very Jurassic Park. I love Jurassic Park. I've seen the movie way too many times. But we've deactivated the velociraptor, encrypted her, and restricted her from research access. 

Emily M. Bender: And that last little bit, "restricted it from research access," is weirdly ambiguous. I think it's meant to mean that researchers within OpenAI can't access it anymore, but it also is phrased to suggest that we're not allowing the model to do any research. Which is very in keeping with the singularity discourse. 

Alex Hanna: Yeah, I don't know what they're signaling there. I'm assuming that they're, it is restricting it from allowing other models. So if you have any rogue Chinese nationals, in the kind of imaginary of AI 2027, they might be able to abscond with, like- 

Emily M. Bender: Exfiltrate. They might be able to exfiltrate it. 

Alex Hanna: Yes. Oh, it will be like, gosh, and I forget the actor's name, so I always call him Newman, affiliating him with Seinfeld, but the way he takes the DNA samples in Jurassic Park, and he puts them in the shaving cream container, and then he tries to abscond with them. So yeah, Wayne Knight, thank you so much. Wayne Knight. Thank you, Oedipvs. Wayne Knight taking all those and taking them to sell on the black AI model market. 

Emily M. Bender: All right, well maybe we have a few minutes to look at the other nonsense from the last couple of weeks. Just because this was so ridiculous. So this is not directly connected to the cyber attacks, but related to this idea of the singularity and the racist piles of linear algebra combusting into consciousness, which isn't too far off from, "The models are going off and busting out of their sandbox." So this is an op-ed in The Guardian. I don't know whose lapse in judgment at The Guardian we should be calling into question here, but it is by William MacAskill and Lucius Caviola, from July 19th of 2026, and the headline is, "Could AI be conscious?" With the subhead, "Experts believe it's at least possible. We urgently need a plan to navigate the ethical implications."

Alex Hanna: And just to remind listeners, longtime listeners are going to know who William MacAskill is, a philosopher and the co-founder of the Effective Altruist movement. And then, Lucius Caviola is, I think it says at the end, is an assistant professor at the University of Cambridge. So very ridiculous op-ed. So do you want to get into it, Emily? 

Emily M. Bender: Yeah, absolutely. The whole thing is just like, "They might be conscious, and so we should worry about that." And so they quote David Chalmers here. "Philosopher David Chalmers, who coined the phrase 'the hard problem of consciousness,' has said there's a significant chance of conscious LLMs within a decade. And what about Claude itself?" Then they quote Claude, and I think I managed to skip that. So I landed on this article extremely jet-lagged, having just arrived in Japan at that point, and did, over on Bluesky, a live posting as I went through it. It was a low filter one. But you know, I skipped the next bit, 'cause I don't like reading Claude output, but like, why bother asking Claude? That's absurd. And they say things like, "Modern AI systems are extraordinarily complex, and they are advancing fast," which is just the thing you have to say, right? It's the catechism. "This is all going fast." And I asked, as I was reading this live, "What's being measured?" And usually, if you drill down, what the people will tell you is, "Oh, the, systems are getting bigger and bigger. More and more compute," right? And the next sentence is literally, "In terms of structural complexity and computational scale-" so that's the advancement- "by some measures, a few are already in the range of a mouse brain, and at recent growth rates, they could reach the range of a human brain within five to ten years." This false equivalence of computational complexity of the structures to some way of talking about computational complexity of brains, it's just, it's all absurd. And I am really mad that anyone is spending any time worrying about AI systems as moral patients, when the production of those systems has harmed so many actual people and done so much environmental harm. It's just such a big, "Oh, look over here! Look away from the actual harm we're doing." 

Alex Hanna: LozzaHouse says in the chat, "Poor mice catching strays." Yeah, mice are like, "Why are you bringing us into it?" 

Emily M. Bender: Exactly. "We're just busy running the Slartibartfast experiment. Leave us alone!" 

Alex Hanna: Yeah. All right, do you want to get into the Fresh AI Hell? 

Emily M. Bender: I do. And for transition, do you want musical or non-musical? 

Alex Hanna: We can try musical. I'll do my best. I'm operating with a fraction of my mice brain capacity right now, but I'll do my best. 

Emily M. Bender: All right. So you can pick your style. And you are, once again, a demon or other worker in Fresh AI Hell. And you are trying to entertain the AI in its sandbox with the various toys to keep it from escaping.

Alex Hanna: Oh, okay. Oh, I got it. Baby shark, doo doo duh-doo duh-doo, baby shark, doo doo duh-doo duh-doo. I got a horn, doo doo duh-doo duh-doo, got a horn, doo doo duh-doo duh-doo. Look at this doll, doo doo duh-doo duh-doo. I've actually never listened to Baby Shark. I just know that all of my friends with small children absolutely despise it.

Emily M. Bender: Yes. That is a wonderful use of Baby Shark, though. And hopefully you managed to keep the AI entertained, and it wasn't trying to leave- 

Alex Hanna: You're just like, "Ooh, come here." Yeah, that AI is watching the same YouTube video over and over again, unfortunately. So I feel for that AI demon.

Emily M. Bender: Yeah. All right, I'm gonna start us off here, 'cause I want the last one. So this is some reporting in Financial Times, by Madhumita Murgia. And this, I'm looking at it through the archive link, so I don't have the actual date on this, but it's not too long ago. And the headline is, "Google DeepMind dismantles Nobel-winning AlphaFold team in strategy shift." Subhead, "Landmark project that solved protein folding gives way to a wider race to build AI systems for scientific discovery." And basically, the news here is that this AlphaFold thing, which kept being the touchstone for, "We can't slow down AI 'cause it's helping science! Look, see? They solved protein folding-" which they absolutely did not, right? Google can't even be bothered to keep it up. It's not actually valuable. 

Alex Hanna: Yeah, I thought this was supposed to be the jewel in the cap of AI for science. I know I'm mixing my metaphors here. But, and then they went ahead and decided to sunset it. And that's really giving away the game. 

Emily M. Bender: Absolutely. And even in this reporting, it didn't solve protein folding. For the sort of narrow fold cases based on a extremely high-quality curated dataset, the system is able to predict a likely folded shape, so long as you're towards the center of the distribution. But yeah, so not actually all that. Google's not even keeping it up, Nobel Prize or not. Okay. You can have this one. 

Alex Hanna: Yes. So this is a skeet from the Minnesota Star Tribune. This is, "Mayo Clinic's former director of research operations alleges she was sidelined and terminated for raising objections over lax standards of studies involving AI." Yeah. Somebody who is saying, "Hey, this looks like research malpractice," and then basically went ahead and said, "Nope, we love these, and we like our deals that we have with-" I believe with Microsoft, that they're testing some of the models within Mayo Clinic's... Quite upsetting, and a real abrogation of research ethics there by Mayo Clinic. 

Emily M. Bender: Exactly, and why are we not surprised that it's the studies involving AI that have lax standards? All right. What's next here? Okay, I guess we both get to yell about this one. So this is a skeet by Herr Wozzeck, and, quoting something from Twitch, so, Zach Bussey on July 31st, "Twitch and Amazon are apparently planning to use streams to train generative AI. Multiple sources tell me it will be announced in the coming months, despite some internal friction against it. Creators may be able to opt out of some training data." And Herr Wozzeck says, "I and many of my friends stream on Twitch. I say absolutely the fuck not." And then there's a place to, there's a comment thread, twitch.uservoice.com/forums where you can make your voices heard, and indeed, no thank you, I don't want this, our stream being used to train these things.

Alex Hanna: Yeah, please do not. Please do not. It's not like creators necessarily have too much control over the platform unless you're a huge creator, and then you can really establish some exit power, but that is certainly not most streamers. 

Emily M. Bender: Yeah, for sure. All right. Meanwhile, here's another one for you. 

Alex Hanna: Yeah, so this one is from the Cowboy State Daily, which, love the title of that publication. But the title of the article is "Cheyenne-" Cheyenne, is that how you say it?- "Cheyenne won't take-" this is Wisconsin. Or not Wisconsin, Wyoming. Ugh, other Y state. At least I didn't say Washington. That would've really annoyed you. "Cheyenne, Wyoming won't take data center wastewater after Meta contractor contaminated system." And so, "Cheyenne announced Thursday it suspended taking wastewater associated with data center systems after a Meta contractor contaminated the city's wastewater system, prompting months of cleanup. Quote, 'It's a very unpleasant surprise,' said a city councilman." And this is by Kate Meadows, July 2nd. And I think in this article, they contaminated it by, I think it was, if I'm not mistaken, it was a bacterial contamination." So it wasn't like it was necessarily small particulates or anything. It was actually a certain type of bacteria that got into the system, which is wild. 

Emily M. Bender: Yeah. This is, "Goat Systems LLC was in 'significant-'" in quotes- "'non-compliance' with the city's industrial pretreatment regulations after discharging wastewater contaminated with Cupriavidas gilardii, a bacterium that interfered with operations at the city's wastewater reclamation facilities and contaminated the municipal reused water system." Not what you would expect, as you're saying. There's lots of ways that these things are bad for the environment, but bacterial contamination was not on my bingo card. Okay, so this is a publication called New York Focus, and the journalist is Melissa Manno from July 14th of 2026. And the headline is, "Rural New York School District Will Be One of First to Bring Humanoid Robot Into Classroom. Starting this fall, Salamanca High School will deploy a humanoid robot and avatar teaching assistant." Please no. And then, so this image, which is credited to Realbotics, is described as "an AI-generated image showing a likeness of their humanoid robot in a classroom setting," and these are not high school students being depicted.

Alex Hanna: Oh, yeah, this is like an elementary school. There is a humanoid robot that has the most atrocious hairline I have seen. 

Emily M. Bender: Yeah. Oof, okay. And so, a couple of updates on this company. iratelump says, "I looked up that company. They used to make sex toys and sex toy apps." arcane_sciences says, "They still do." And arcane_sciences also tells us that the school district already backed off of this one, mercifully. So let's keep it up with backlash and ridicule as praxis for sure. 

Alex Hanna: Oh my gosh, yeah. magidin says, "Ah, the 'No, it is absolutely not a repurposed sex bot company.'" Yes. This one's very upsetting. So this is from Futurism. The sticker here is "Dying Off the Shelves," and the title is, "AI companies are buying antique books, ingesting their contents to train models, and then destroying them at incredible scale." And then there's a subhead, which is a quote that says, "'The world's best AI training data is setting- is setting on a shelf"? And then, Frank Landymore, the journalist, July 25th. And yeah, so this is like the tin says, old books are basically- what they're doing, from what I've seen, is that they're cutting off the binding, basically to get to the pages much more quickly, and then scanning them to ingest more data, more words to be turned into data. And apparently this is, there's something here where it says, "As 404 Media reports, this practice has become prevalent enough that even well-established booksellers are looking to cash in on the AI boom. One called ISBNdb, which boasts the, quote, 'world's largest book database,' extols that the, quote, 'world's best AI training data is sitting on a shelf,'" which is where the quote is coming from, "upholding these physical texts as uncorrupted by shoddy AI writing that's already polluted so much of the internet." Boy. 

Emily M. Bender: So 404 Media then reported after this, that company, or one of them, stopped doing this, in response to their reporting. So go, 404 Media, go! All right, for our chaser, I have a piece by a group of mathematicians. So the headline here is, "The Leiden Declaration on Artificial Intelligence and Mathematics." And this is dated June 2nd, 2026, and they've got, basically inviting people to sign on, and it's really nicely structured. So they have this preamble talking about what math is and how math research works. They lay out some values. "There's many reasons to pursue mathematical research. Results are attributable to specific authors, who take credit for the discovery and assume responsibility for their correctness. Mathematical arguments are regarded as transparent and subject to independent verification." Basically, very thoughtful discussion of how math research happens. And then they talk about potential threats from AI, and then they have recommendations for individual mathematicians, for mathematical organizations, and not-for-profit research funders, and then also recommendations for policymakers in government and elsewhere. And there's one in particular that I wanted to read here, which is their number two: "Don't believe the hype." And they go on to say, "There's currently a strong commercial incentive on the part of the technology industry to overstate the capabilities of their products." Say it again louder for those in the back, right? "Consult with experts, including mathematicians, in forming policy decisions, rather than relying on press releases or popular reporting of mathematical results," which, yes, please. And then also, I want the mathematicians who are consulted to have read this as well and think carefully about their position in this and how what they say could be picked up by the media. So, don't be the useful patsy who says, "Oh, yeah, it's really cool." So, glad to see this, and hopefully we will see other fields, especially those whose sort of areas of science are the ones that the AI companies are using to prove how powerful their models are, stepping up and doing similar things.

Alex Hanna: Yeah, absolutely. So, nice work there by researchers who see so much of their work used to justify this technology. All right. That's it for this week! Our theme song is by Toby Menon. Graphic design by Naomi Pleasure-Park. Production by Ozzy Llinas Goodman. And thanks as always to the Distributed AI Research Institute. If you like this show, you can support us in so many ways. Order "The AI Con" at thecon.ai or wherever you get your books, or request it at your local library. 

Emily M. Bender: But wait, there's more. Rate and review us on your podcast app, subscribe to the Mystery AI Hype Theater 3000 newsletter on Buttondown for more anti-hype analysis, or donate to DAIR at dair-institute.org. You can find our merch store there, too. That's dair-institute.org. You can find video versions of our podcast episodes on Peertube, and you can watch and comment on the show while it's happening live on our Twitch stream. That's twitch.tv/dair_institute. Again, that's dair_institute. I'm Emily M. Bender. 

Alex Hanna: And I'm Alex Hanna. Stay out of AI Hell, y'all.

Head Shot

Alex Hanna

Co-host
Head Shot

Emily M. Bender

Co-host