Mystery AI Hype Theater 3000
Mystery AI Hype Theater 3000
"AI" Agents, A Single Point of Failure (with Margaret Mitchell), 2025.03.31
After "AI" stopped meaning anything, the hype salesmen moved on to "AI" "agents", those allegedly indefatigable assistants, allegedly capable of operating your software for you -- whether you need to make a restaurant reservation, book a flight, or book a flight to a restaurant reservation. Hugging Face's Margaret Mitchell joins Emily and Alex to help break down what agents actually are, and what to actually worry about.
References:
PwC launches AI agent operating system to revolutionize AI workflows for enterprises
An Open-Source AI Agent for Doing Tasks on the Web
Scale AI announces multimillion-dollar defense deal, a major step in U.S. military automation
Other references:
Why handing over total control to AI agents would be a huge mistake
Fully Autonomous AI Agents Should Not be Developed
Bender vs. Bubeck: The Great Chatbot Debate: Do LLMs Really Understand?
Fresh AI Hell:
DOGE suggests replacing workers with "AI" (of course)
Vape, or the tamagotchi gets it
"AI" for psychotherapy, still bad, still hyped
- Via @Hypervisible
Palate cleanser: "AI is the letdown"
https://www.cnn.com/2025/03/27/tech/apple-ai-artificial-intelligence/index.html
Comic relief: "Fortified with AI"
Check out future streams at on Twitch, Meanwhile, send us any AI Hell you see.
Our book, 'The AI Con,' comes out in May! Pre-order now.
Subscribe to our newsletter via Buttondown.
Follow us!
Emily
- Bluesky: emilymbender.bsky.social
- Mastodon: dair-community.social/@EmilyMBender
Alex
- Bluesky: alexhanna.bsky.social
- Mastodon: dair-community.social/@alex
- Twitter: @alexhanna
Music by Toby Menon.
Artwork by Naomi Pleasure-Park.
Production by Christie Taylor.
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, 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 54, which we're recording on March 31st of 2025. And we're taking a closer look at the hype around agents, those AI tools that unlike chatbots are enabled to operate applications within computer systems and in theory, act on your behalf without your direct oversight. They'll keep your life organized and productive while you rest. So the advertising goes.
Emily M. Bender:We guess we can see the temptation to have a reliable, indefatiguable assistant looking after your every whim. But as we'll see today, it's in just about any case we can think of. Hooking up the random text. Generating machines to interfaces for real world action is just a really bad choice.
Alex Hanna:With us to take a closer look today is Margaret Mitchell, a machine learning researcher and chief ethics scientist for Hugging Face, an AI startup that provides infrastructure for machine learning data sets, models, and demos. You may recognize her as the guest host for episode 33. Today, she's back as our guest. Welcome Meg.
Margaret Mitchell:Hello. Thank you. I'm excited to be here and excited to be a second time guest. It's, it's an honor. Thank you.
Emily M. Bender:You, you're our first repeat and it is so awesome to have you back. So, before we get started on the episode, we have to take a moment to plug our book, The AI Con, which is coming out less than a month from when this episode will be out, a little bit more than a month for you livestream folks. Pre-order it now, so you can have it the instant it's available. That's May 13th in the US and affiliated markets and May 22nd for the UK and affiliated markets.
Alex Hanna:And we're doing a little tour to promote it, talk about some of our most important points, and meet our fellows in the fight against AI hype, starting with a virtual event with the Distributed AI Research Institute on May 8th. Mark your calendars and visit TheCon.AI for a full list, if we're gonna be in your town. Speaking of traveling, I am also traveling right now, so I'm not on my usual setup. Those of you on the live stream, it might be a little weird, video might be jumpy. The sound might a little bit wonky, but I'm still real. I've not been replaced by an agent yet, and probably never.
Emily M. Bender:We have the real Alex. So AI agents, I have to say this is like sort of, you know how when the hype of um, the hype-tasticness of a term kind of wears off and they have to go to the next one? So the shine has kind of gone off the phrase AI a little bit and there's sort of two directions that I see people going. There's the AGI thing, but you can't really claim AGI yet. And so if we're gonna claim something now that's better than ordinary AI, it seems to be agents or agentic AI. And it just, it seems to be now part of just like the standard narrative of the supposed developmental trajectory of these things. Um, including, so speaking of traveling, I was last week down in the Bay Area at the Computer History Museum to debate none other than Sebastian Bubeck over the question, uh, 'Do LLMs understand?' And, uh, the debate was moderated by IEEE Spectrum's Eliza Strickland. And she started off with this like, here's a presentation of what's gone on to date. And she got to AI agents at the end. I thought, why? Why are we talking about that?
Margaret Mitchell:It's the thing. Yeah. Mm-hmm.
Emily M. Bender:Yeah. Um, anyway, we will drop a link to the YouTube of that debate in the show notes in case people wanna see. Um, what do you think Alex and Meg, do the people wanna see?
Margaret Mitchell:Oh yeah.
Alex Hanna:The people definitely wanna see.
Margaret Mitchell:It was great. Highly, highly recommend.
Emily M. Bender:It was, it, it was, it was a good time. Alright. So--
Margaret Mitchell:I mean, you, you convinced him by the end. He, I mean, I, I don't know if he said it exactly as such, but the arguments he was making by the end was like, yeah, the AGI is, uh, not real. Like--
Emily M. Bender:Yeah. And, and you know, just, I guess the real question is, is it useful? Is sort of where he ended up. Which was not the question under debate. I have something say about that too.(crosstalk)--pivot. And
Margaret Mitchell:you do a pivot when you don't wanna answer the question.
Emily M. Bender:Yeah. All right. So let's get into our artifacts then. Um, we start with something from, uh, PricewaterhouseCoopers, uh, written by none other than Matt Wood, uh, who gives his title as "commercial technology and innovation officer, PWC US." And actually, I'm guessing they changed their company name and it is now just officially PWC. Is that right? Probably?
Alex Hanna:Yeah, I, I've heard it, I've heard, I've heard it PWC for the past, you know, seven years or so. So I don't, I, I don't know when that happened. I'm not an expert on their corporate branding.
Emily M. Bender:Uh, all right, so I'm gonna start, set us off reading this and then you two can jump in. So, "PWC's Agent OS is an enterprise AI command center, seamlessly connecting and scaling intelligent agents into business ready workflows, up to 10x faster than traditional methods." So what, what do you say?
Margaret Mitchell:Yeah, I was so, uh, impressed by how blatantly LLM-y all of this was. Um, I mean, you kind of see this throughout the document, I guess, which we'll get into, but like this whole thing is just dripping with language model-ese and, and like PR type hype. So I was like trying to think of like what would've been the prompt to generate all of this, like maybe like two bullets and then like asking it to generate four pages or something. And I think it was something probably like, 'write me something about a system that can efficiently scale complex workflows, write four pages' and that's exactly what this seems to have done.
Emily M. Bender:Yeah. Pro-probably with the Agent OS thing, which is just cursed.
Alex Hanna:Yeah, (crosstalk) cursed technology. And what do they mean, like faster than traditional methods? Like what's the traditional method here? Like what is the, what is the baseline? Um, is it like an LLM? Is it just like somebody clicking things? Like, I'm not sure.
Emily M. Bender:Yeah. And there's, there's so much anthropomorphization going on in this too that was really, really getting to me. So um. I'm gonna, I'm gonna skip the middle paragraph there and go with "Enterprise organizations face a critical challenge. AI agents--
Margaret Mitchell:Not a challenge.
Emily M. Bender:Not a challenge.
Margaret Mitchell:Critical, critical challenge. The adjectives in this are like, so great. Like as someone who has written PR and comms, it's like add in those adjectives. Yeah.
Emily M. Bender:"AI agents are being developed in many ways, as embedded features within platforms, as standalone applications or as highly specialized agents built on proprietary or open source software development kits, SDKs. However, as embedded features, these agents struggle to communicate, collaborate, and scale effectively across complex enterprise environments." So this the, the subject of 'struggle', right? Struggle is one of these words that indicates sort of an emotional attitude towards what you're doing, right? Is the AI agents. And 'communicate' also very anthropomorphizing word.'Collaborate' very anthropomorphizing word, right?
Margaret Mitchell:Yeah. They struggle, Emily, they're struggling. You have to help them with PWC's tools.
Emily M. Bender:So, "Unlike other tools that support siloed agent systems--" Oh, you wanna say Alex?
Alex Hanna:Oh, I wanted to read this next one 'cause it's just terrible. Okay. Um, and this is in, this is in bold, so you should know it's important."Unlike other tools that support siloed agent systems, PWC's Agent OS is a unified orchestration framework. Acting is both the central nervous system and the switchboard for enterprise AI," which I love. Just mixed metaphors, like incredible stuff there."It seamlessly connects AI agents regardless of platform or or framework into modular adaptive workflows to integrate the essential enterprise systems such as those from Anthropic, AWS, Github, Google Cloud, Microsoft Azure, OpenAI, Oracle, Salesforce, SAP--" Or S-A-P, I dunno how you're supposed to say that."--Workday or other--and others." So just, a brain and a switchboard. And they evoke such, they evoke such different, um, kind of metaphors too. Like one is sort of like this biological one, which we're gonna talk about a little bit in AI Hell. But the other is just like, I see like the very feminized work of like switching switch boards and things that are like also not really, use these things. So I'm like, what are you, what are you, what are you doing here?
Emily M. Bender:Yeah. Yeah. Exactly. And central nervous system is like controlling and switchboard is connecting, right. That, um, but yeah, the, the biological metaphor there really got to me. Um.
Margaret Mitchell:And it's also another way of saying like, 'Hey, we're making a single point of failure', right?'We have no checks and balances. We're just saying this is in control of everything.' What could go wrong?
Alex Hanna:I like this, I like this, I like this comment from SJayLett in chat, which is "AI agents are dot, dot, dot, text extruders hooked up to Selenium, curl and maybe some other tools, right?" so like, so basically they're like, oh, you're like connecting LLMs to things that are like, like test frameworks for navigating the web, you know?
Margaret Mitchell:Yeah. I mean, like one thing I took away from this paragraph is like, oh, so you're not actually even introducing like an agent. It's just like a thing built on top of stuff other people is, is doing. And so that's why you can say like,'Anthropic, AWS, GitHub, Google Cloud,' it's like, just list people doing things.
Emily M. Bender:Yeah. Yeah.
Margaret Mitchell:You're, it could help with all of them. It's just an extra layer that will add on top of all that.
Emily M. Bender:We've got, we've got Abstract Tesseract saying, uh,"Arbitrary code execution as a service."
Margaret Mitchell:Right. Yes, exactly. Oh my God, that's a great way of putting it. Mm-hmm.
Emily M. Bender:Yeah. This list--oh go ahead.
Alex Hanna:Oh, no, no. I wanted you to click the AI agents like link.'cause I just wanted to go to the visual 'cause I think it's incredible. Um, so first off what it looks, it's, it's, it's another PWC article on AI agents. It says, "AI agents should reimagine the feature of your work, your workforce and workers." But then it is like a bunch of little squares, like that look like, they they, they look like Starbursts. Um, which is making me very hungry. Um, and, and then, but like, it's also, it's also this fascinating kind of like thing about like the container and like the self-containedness of a system. Like as if like software does not have like--so much of the software business and like, I think this agent shit, is like, oh, seamless, like seamlessness, like connected, connected. It's like, no. That's not what it's gonna do. It's just gonna, I mean, you're, exactly what you're saying Meg. It's just like you're creating a pipeline and it's still gonna fail and people are gonna have to intervene, right?
Emily M. Bender:Yeah. And then it's gonna be messy 'cause it's gonna keep moving and it's all in the OS. Um, alright, I gotta bring up something here from, uh, N Dr. W Tyler. Still not sure how to do that username. Um, "As a measure for the public demand for quote 'AI agents', there was an opened Rabbit R1 box in our building's post room for weeks. I'm pretty sure someone stole it, thinking it was a phone, then returned it when they found out what it was, and not even the person who'd ordered it seemed to want the thing." Okay, so I'm gonna, I'm gonna back us up here and just say that this list of--
Margaret Mitchell:Wait a second, wait, did that say 'work, your workforce, and your workers'? Just, just to like, oh, just going back to the way these things are worded, did that say 'work, workforce and work--'. Okay. Just making sure that, yeah, this is literally just saying one thing in like eight words. Okay.
Emily M. Bender:Yeah. Alright. And then there's some argument in the chat about whether or not Starbursts are good. But anyway, um.
Margaret Mitchell:That's a good question.
Emily M. Bender:So there's, we've got this list of companies that are supposedly providing agents, and I was a little, first of all, a little surprised to see Google showing up as Google Cloud. Um, and of course GitHub isn't separate from Microsoft, but fine. Um, mm-hmm. I guess I didn't know that Oracle, SAP were, and Workday were doing agents. Um, so--
Margaret Mitchell:Everyone's doing agents, everyone. Yeah. And Salesforce has actually really been leading, I think their last, um, conference was called Agentforce. Actually. They're very, very agents right now. Yeah.
Alex Hanna:Agentforce is like, I think it's a tool. I mean, 'cause Dreamforce is the big conference and then like Agentforce is what the, like there's a big billboard on the, on the US 101 in San Francisco.
Margaret Mitchell:Oh it's on the 101?
Alex Hanna:Yeah, it just says Agentforce on it. And it, and it's just, yeah. They just own that one permanently and now it just says Agentforce on it.
Margaret Mitchell:Amazing.
Emily M. Bender:The best way to get to the, the, the pulse of Silicon Valley is to drive down the 101. Okay, so we've got this, um, this list down here. I'm gonna take a, skip it down a little bit to their, um, "New Possibilities" and these are hilarious. Right. So, uh, "PWC's Agent OS enables organizations to quickly build, customize, and deploy intelligent workflows and agentic blueprints, simplifying enterprise wide AI integration and transformation--" And that 'simplifying enterprise-wide AI integration' is like talking to the people who've bought the hype and feel like they've gotta get some AI in, or they're, they're gonna, they get, you know, trashed by their investors or something. Um, so then they have these examples about entirely new possibilities.
Margaret Mitchell:And it doesn't say test, it doesn't say test. As someone who worries about the impact of these systems on society, I always look for like the evaluation strategies. And it's like, no, just build it and deploy it. No evaluation necessary. Okay. Sorry. Continue.
Emily M. Bender:Surely there's an evaluation agent. Okay. So first example, "A logistics director at a global manufacturing firm could use PWC's Agent OS to coordinate AI agents across SAP for demand forecasting, Oracle for procurement, and AWS for real-time logistics tracking. By adding a custom risk analysis agent and a disruption detection agent from PWC, they can proactively address potential supply chain bottlenecks. This cross agent orchestration could reduce supply delays by an estimated 40%."
Margaret Mitchell:I say it could be 60 percent. That's what I, why go 40?
Alex Hanna:Go all the way.
Margaret Mitchell:(crosstalk) -- 100 percent. Instantaneous. Why not? It could be.
Emily M. Bender:Well, 110 percent, while we're at it. Right. That's what you're supposed to give. Like, so it's just completely made up. Right.
Margaret Mitchell:Yeah.
Emily M. Bender:And then they have one about marketing operations, um, and it ends with "Conversion rates could increase by 30 percent." Again, could.
Margaret Mitchell:Why not?
Emily M. Bender:And campaign launch times could be cut in half.
Margaret Mitchell:Sounds great. I guess.
Emily M. Bender:And then the last one--
Alex Hanna:That's interesting. Yeah. The last one is, is pretty worrisome. So it says, "A compliance manager at a multinational bank could automate regulatory workflows using PWC's Agent OS to integrate policy analysis agents from Anthropic--" Like a very cursed sort of agent."--and internal documentation review agents powered by Microsoft Azure. A custom agent interprets involving regulations while ensuring alignment of company standards is set up to reduce manual review time by an estimated 70 percent." Uh, and that's really terrifying. I gave, I was talking to somebody at an institution and they're like, well, we've got these contracts. Can we have some of this stuff reviewed? Like, no, absolutely not. What are you doing? That's what it, what it what, like--
Margaret Mitchell:What I, yeah.
Alex Hanna:Yeah. It's, this is such a consequential thing, especially on compliance, and you're considering automating it and introducing another black box on top of compliance. Wow. What a, just what a, what a, what a shit show of an idea.
Margaret Mitchell:Yeah. And not to mention the issues with accountability that comes from AI agents because you know that these same companies will probably argue that you can't be accountable for your AI agent actions. So like, what a great way to not follow regulatory protocol by putting in place an AI agent that works poorly, having that, you know, create discriminatory outcomes or other like really horrible outcomes. And then be like, well, it was my agent that's the technology. What are we gonna do, you know?
Emily M. Bender:Yeah. It's still in its infancy, you know? It'll work better next year.
Margaret Mitchell:Yeah. Yeah. We're improving. It's always like we're working towards better goals. And you know what I, what I took away from that, uh. Is it a sentence? It's a couple sentences. Is that this company wants to have contracts with banks, with Anthropic, with a, with Azure and they want to have government contracts. And so like I think the role that this is playing is to say like, buy our product if you are one of these organizations, not with actually a concrete suggestion of something that has been shown to work, but just like brainstorming, like vibe selling, I guess--
Emily M. Bender:Yeah.
Alex Hanna:Yeah.
Margaret Mitchell:--Things that could happen.
Alex Hanna:Well it's also-- it's also the possibility that, that they already have contracts with these types of institutions. Right, and now they're like, you've got us on retainer and like now we're gonna like build product for you. Right. And I think there's, I mean there is also someone, GoogaGooga1957, very, very great name says, um, in the chat, "Articles like that are what, what come out when you have an MBA and a BBA--" Which I think is a Bachelor's of Business Administration."--write about tech. LOL, management consulting companies are a joke." And like Yeah, I mean you're getting the kind of like PwC, McKinsey talk of, like they love talking about AI, 'cause they can then like get a, you know, get someone fresh out of one of these programs to then like write copy about what they could do without like thinking any, like with any kind of seriousness about risks or potential harms or like you said, Meg, evaluation of any kind.
Emily M. Bender:Yeah. While we're here, what is a management consulting company, 'cause that's what PwC is, right, doing building something they're calling an OS. Did they hire some software engineers? Or just some vibe coders, like, what's going on here?
Alex Hanna:Totally. Probably vibe coders. Yeah.
Margaret Mitchell:Yeah. It's LLM generated code. That's probably what's happen-- because this is LLM generated, like this selling document is so clearly, and I, you know, I could, we could analyze why, but, um, but I, I bet you're right. I bet they just are like having some people who don't really know in depth details about operationalizing, you know, technology, just vibe coding with LLMs and being like, yep, here's a product.
Emily M. Bender:Yeah. All right. We've got some folks in the chat asking us for where this term vibe coding came from.
Margaret Mitchell:Oh, sorry.
Emily M. Bender:No, I, I introduced it, but we should, we should definitely let the people know. So I understand this to be basically, instead of sitting down and doing real software engineering, it's telling the LLM what you want and taking its output as if it were a finished product. That's vibe coding. And there was some 404 Media podcast, I think, talking about this guy who vibe coded a video game and, uh, managed to get lots and lots of money out of it because basically it was a publicity stunt and everyone was excited about it and it was all, um, money being made by, you know, uh, pay-per-view advertising effectively. Um, but Alex, you saw something about where vibe coding came from, I think just before the show.
Alex Hanna:Well, yeah. According to Wikipedia, vibe coding came from Andrej Karpathy, who is a really cursed like person. He's like one of the co-founders of OpenAI and also worked for Uber, I believe. Um, and, um, you know, uh, really, really kind of like first individual who comes into this. But, so, but I think vibe coding is like, you know, the parentheses vibe coding, like, uh, (positive, excited). And in our case we're saying vibe coding, (derogatory), right? Like that's the inflection that we're giving here. Right?
Emily M. Bender:Exactly. Exactly. All right, let's keep going. In this artifacts, we do have a couple more. Um, in between that set of bullets, in the next, it says, "In addition to these future possibilities--" As in, the vaporware I was just telling you about."--here are some real world examples of how we've already helped clients leverage AI agents into their workforce strategies." And then they list three things and I'm, I'm kind of wondering like what counted as an agent in these cases? Um, so the first one was, "A major technology company reimagined customer engagement by deploying an AI agent powered omni-channel contact center with predictive intent modeling, adaptive dialogue, and real-time analytics. The system reduced phone time by nearly 25 percent, cut call transfers by up to 60 percent, and boosted customer satisfaction by approximately 10 percent." Margaret Mitchell: I'm taking just hanging up on people, reduced phone time by a hundred percent. Um, yeah. And yeah, to your, to your point, like the modern wave of AI agents really has only taken place over the past six months or so, like a year-ish. But, uh, in terms of systems that are deployed, that's really happened within the past six months. Uh, so I imagine that if they actually have a reasonable survey analysis from whatever that deployment was, it wasn't an agent, uh, it was probably like regular code. Um, as you're mentioning, uh, I mean this is like also going to the whole thing about like rebranding LLMs because people have gotten kind of tired of the term LLM or whatever. So the term is like being way, way, way over applied, um, to the point where like, I'm kind of suspicious, does this even count as an agent? Is it different than just sort of code? Yeah. Did they make a slightly better customer service chat bot?
Margaret Mitchell:Right? Yeah.
Alex Hanna:Right, right.
Emily M. Bender:Yeah. Um, so this middle one sounds kind of roughly the same. So, "A large hospitality company streamlined management of their brand standards across their global portfolio by deploying agentic workflows within a modern AI powered platform. Intelligent agents now automate updates, approvals, and compliance tracking, reducing review times by up to percent." Although I take it back, this is different. So the scenario here is, uh, hospitality company, so that's like hotels and restaurants and they've got brand standards and they've got franchisees or whatever who have to like stay consistent with the brand standards. And so now all of that's being reviewed automatically. And so it takes less time to review it.
Margaret Mitchell:That doesn't mean it's good.
Alex Hanna:That's true, that's true.
Emily M. Bender:Um, although in this case, you know, are the brand standards not held, I don't care, like that, that's, that's sort of like the harm is not big harm in that case. Right? The, this next one though, Alex, you wanna read the next one?
Alex Hanna:Oh Lord. Okay. So it says, "A global--" Also in the healthcare, uh, Lord."--transformed cancer care by deploying agent AI workflows across oncology practices. Intelligent agents streamlined clinical and operational processes, automating the extraction, standardization, and querying of unstructured documents to drive about 50 percent improved access to actionable clinical insights--" This is just gobbledygook of words, um,"--to support precision medicine and clinical research and drove a nearly 30 percent reduction in staff administration burden through AI powered document search and synthesis." I mean, this just sounds like they like had a makefile. Like I like, it's not like--
Margaret Mitchell:And there's no citation like these numbers. You have to see what the actual study was because this is a piece of marketing material. So they're going to put forward numbers that make things sound really positive, on evaluations or survey results that are most likely suspect. And so like seeing that they are putting out these numbers as being based on something they've done, like we have to see what they've actually done.
Emily M. Bender:Yeah. So Abstract Tesseract in the chat says,"I'm once again asking them to say who did what to whom." Which is great, and also sounds like a linguist speaking. Got gotta love it. So the the other thing here is, um, this sounds to me like they, they implemented a, a, um, RAG, right? Retrieval augmented generation over something somewhere in their pipeline. And, um, AI powered document search and synthesis is terrifying when you're talking about medical records and medical decisions. Because if that sped things up, right, that means nobody was checking. And that means patients had random nonsense, you know, entered into the system, maybe even in their clinical record.
Margaret Mitchell:Yeah. And this is life and death. Like--
Emily M. Bender:This is cancer.
Margaret Mitchell:This is cancer. And it is not something that you wanna hand over control to agents that have not been shown to work well for, that have been shown to be based on models that make horrible, um, hallucinations about what's actually happening.
Emily M. Bender:Yeah.
Alex Hanna:Well, it's also, I'm curious when they, I mean this again, talking about like a random statistic, like "nearly 30 percent reduction in staff administrative burden". Okay. First off, I'm wondering how you're measuring staff administrative burden. Second, I'm really curious on how the staff actually feel about this, right? I mean, and what are staff? What are things that our staff are not catching or are missing or, you know, are upset that they don't have oversight over? And when they have to deal with people who are upset, like, how does that, how does that get handled in the clinical setting? You know, they don't say which company this is. Um, you know, if it's, is it, is it even a, is it a provider? Is it an insurer like that, so yeah, lots of problems. And,
Margaret Mitchell:and how did the patients feel, right? I mean, like were the patients helped? How did their families feel? Like, what was the actual effect on the people of cancer?
Emily M. Bender:I would like to know which medical centers are connecting with PwC to do this kind of stuff, so I can go elsewhere.
Alex Hanna:Yeah, no kidding.
Emily M. Bender:Alright, any last words on PwC before I take us to the next one here? Um, should we go to Stanford?
Alex Hanna:If we must.
Emily M. Bender:Alright. Okay. So this is, um, something out of the Stanford human-centered Artificial Intelligence, HAI newsroom, basically. Um, the headline is "An open source AI agent for doing tasks on the web." Um, and, uh, do we have an author. We do--does this not--
Alex Hanna:Uh, there's no, there's no, there's author on this. Oh--
Emily M. Bender:Contributor is Catherine Miller.
Alex Hanna:Oh, it says contributor. Okay. Got it.
Emily M. Bender:And uh, the subhead here is, "NNetNav--" And I'm sorry, but someone needs help naming."NNetNav learns how to navigate websites by mimicking childhood learning through exploration."
Margaret Mitchell:Nooooooo. Already, already terrible.
Emily M. Bender:Okay. So, uh, Catherine Miller writes, "The arrival of large language models, LLMs, has pushed artificial intelligence towards new useful heights. Our computers help us write emails, essays, and computer code." No, um, I mean, you don't have to."Now, developers are attempting to turn chat bots into action bots that can book flights for us, find information hidden deep in a website, pull data from multiple sources to create shipping or sales reports, or create a new repository on GitHub." Um, and I want, I meant to say this at the top somehow, every time someone talks about agents, and this is included Eliza, at the debate, the top examples are airline reservations and restaurant reservations.
Margaret Mitchell:It's very tedious.
Emily M. Bender:Since when do you need an agent to do that for you? And I'm thinking about like a certain amount of automation. So if the restaurant ops into something like Open Table, so if they're willing to give up some control over their bookings, then you can just go to a web interface, enter a couple pieces of information, bam, you can, and like if it doesn't have the time you want, it shows you the range of times you can pick. Like that is a great kind of automation, not an agent. As far as airline reservations go, I remember the change from like having to talk to a travel agent who could see all of this stuff in some old COBOL database, but would only like tell you one thing at a time to getting the web-based interfaces. I think it was Travelocity was the first one that I used where you could like see all the options and move the sliders and I'm like, that's the right way to do this. I do not wanna do it in natural language. It's, it's always the example.
Alex Hanna:It's also interesting'cause it is, it is like all very feminized labor too. I mean this is like, and I think this, this is something that happens so much. I mean like the, and I, I mean it's like I probably like. I say this so much on this podcast, I feel, but like the feminized labor as the stuff that's to be automated, right? Like, and so like, and, and I mean like, and I don't wanna say this stuff is not like work, like the stuff is work, but it's like we could just make it completely like this. And, and the thing is like, okay, but also when these things go wrong, like you actually need relationships to like, to like navigate this. Like, and the thing that like really annoys me is like, now there's these things too when you like, use a phone, like you call and like Google's like, do you want like a bot to talk to the person, like at the restaurant? Like, absolutely the fuck not. What a disrespectful thing at a person who's working at a restaurant. Like, I, like if I did that, I would just like hang up immediately as, as, as, as someone who's running work at front of house. Like tell me, you never worked in the food industry without telling me you never worked in the food industry.
Margaret Mitchell:Right.
Emily M. Bender:Yeah, absolutely. So, so it's always like booking a flight as if that is, so, it, it's disregarding and degrading these relationships as you're talking about Alex's, it's saying, well, the feminized labor is labor that should be automated. And then on top of that, it's like such first world concerns. Like what's a problem you face in your daily life that you would like an agent to take care for you? Oh yeah. It's such a pain to book a flight. So, okay. Um, uh, we have a quote here from, uh, Shikhar Murty, who's the PhD student whose work this article is based on, uh, who says, "'AI agents that can take actions on our behalf online could potentially reduce much of the burden of computer use, especially for repetitive tasks.'" Um, and I'm just reminded here of how computer science research has to be sold in terms of saving something, right? It's saving burden, it's saving time, it's saving money, it's saving need for expertise, and like that's not how research works in other fields.
Margaret Mitchell:Yeah.
Emily M. Bender:Right. Like, you know, we are learning something about the world, or we are, we are building a bridge that stands up better.
Margaret Mitchell:Yeah. It's really, this strikes me as the kind of thing we do to make like, uh, rationalizations of stuff that we already love to do. And I say this as a technologist, as a computer scientist. Like, I love coding and I love creating things. Mm-hmm. And it's so awesome if I can come up with a reason why that's good, but like fundamentally it's because I love doing this stuff. And I worry that, you know, with these large organizations like HAI and, and like OpenAI as well, like, they're kind of missing track, they're, they're like losing the reality of why people are working on this stuff. It's not because for a lot of people, they have this overarching goal of helping humanity. It's because they love working on it and then can post-hoc justify it by appealing to saving humanity in some way. And the thing that really gets me about this that happens over and over again is that when this is, when this kind of technology is justified as being useful for people, it's very rarely contextualize with respect to assistive technology. So like, if you wanna help people navigate websites, I know a few subpopulations who are legitimately asking for help with this, and that is not being covered in this agent work, right? So like, people with low vision, people with cognitive decline, people, um, who just don't have the background knowledge. So seniors who haven't like grown up knowing what a hamburger is, um, on a website, like this kind of thing, like that is not included in the considerations of how this technology might be developed. And that just further reflects the fact that there really isn't a centering of the people, of the humans in the development. It's a centering of the technology and what is just really fun to work on. And then post-hoc saying, you know, oh, it can be useful for people, that justifies it.
Emily M. Bender:Yeah,'it's such a burden'.
Alex Hanna:Totally. Oh yeah. And, and Trochee or Jeremy Khan in the chat, hey Jeremy, says, "Reminded of the gen AI music exec who said that everybody hates making music." Like--
Margaret Mitchell:Oh my--
Alex Hanna:And I'm like, yeah. It's just like, kid, don't you just want to get to the end product of music? I'm just like, do you, do you know any musicians? And not like booking flights is like, is like making music, but I mean making plans and like crafting and like being attentive to scheduling, like these things are like, they take some like degree of, of, of kind of flexibility and thinking and and whatnot. And there's also a good point in the chat, there's also a good point in the chat I do wanna highlight where someone's like, yeah, actually booking a flight, I'm gonna find the actual thing where. Uncanny Static says, "To, to be honest, booking sites are also are filled to the brim with dark patterns so that booking a flight is actually terrible." Like yeah, they've actually created it, made it quite a lot of work. And I mean, it's just like, maybe we should get rid of the dark patterns.
Emily M. Bender:That that would be something that's reducing burden. I wanna share this, on this point about the, the gen AI music nonsense. There's a wonderful thing that I found on Blue Sky today. So the initial post is by Bertoni, it looks like it's in Portuguese. And then this is translated by someone named Ale, A L E, as "To democratize art is not every person having a cute drawing made in seconds. To democratize art. Is every person having time and health to learn and make art if they choose to. And mainly to have the means to think and relate introspectively with art." And I thought that was really great. And it comes back around as we're talking about agents, if you're saving time, well, for what and for whom. Right. And the, the PwC thing was all about, well, let's save the company's time, which means we're saving the company's money 'cause they're paying for fewer people's time. Right, and meanwhile making a shittier product for everybody else, including their cancer care.
Alex Hanna:Yeah.
Emily M. Bender:Alright, so what do we wanna do in this one? Should we talk about the, um, the, uh, privacy aspects or, no, there's a thing about like learning, like children that we have to get into, which is just ridiculous.
Alex Hanna:Oh, we gotta, yeah, we gotta talk about learning like children, we gotta talk about like how this thing learns, um, and the privacy. Yeah. So they say, yeah, here they say "One reason NNetNav is so lightweight, rather than being trained with examples of how humans behave online--" Which that's kind of already confusing and not sure what that's doing there."--NNetNav gathers synthetic training data by exploring websites much the way a young child might. It clicks all the buttons and the types into all the fields to see what will happen, and then prunes out the pathways that don't help achieve the user's goals." Um. Just raging, raging out, just at that, in that comparison. Yeah. Like, Lofty Words in the chat goes,"Young children love booking flights."
Margaret Mitchell:Yeah, exactly. Yeah, yeah, yeah.
Emily M. Bender:When my kids were born, I gave them websites to click on and they just clicked around to see like, which menus-- no.
Margaret Mitchell:They systematically tried every text box and wrote multiple different possible queries into the text box, and then pruned, and then pruned the queries that were unsuccessful for them.
Alex Hanna:My, my son, little, little, little Tommy Trees.
Emily M. Bender:Spelled T-R-I-E-S, right? Sorry.
Alex Hanna:Yes, exactly. Yes, a hundred percent. I'm gonna type that in the chat because I love it.
Emily M. Bender:We are a nerdy podcast, just in case anybody missed that. So, so the, the fact also that, that this is showing up in the AI agent discourse is kind of interesting to me, right? Because it is this, there's a, an older part of the AI discourse, which is like, we're mimicking how kids learn. And I've taken Chris Manning to task on this before. Like he's, he wanted to say that, um, some language model stuff before it was getting called AI was learning just like children do. And I was like, Chris, I know, you know, I you that, that's not how kids learn because I know you have a whole PhD in linguistics. Like, what the hell? But somehow this is, I think, connected to these discourses of generality because if we can pretend that whatever algorithm we've written is learning like a kid, then it might be able to do anything a kid could grow up to do. Um, and it's also connected to these discourse of, oh, this technology is just in its infancy that Anna Lauren Hoffmann really nicely takes apart.
Margaret Mitchell:It makes it more approachable, you know, and it, I mean obviously there's the anthropomorphization aspect as well, but just makes it like more cuddly and, you know, snugly and feeling safe. Um, even though obviously there's not things to back that up at all.
Emily M. Bender:No, no. And also something that we should feel like we should be taking care of and nurturing because it's like a kid. So it's drawing on all this stuff. So let's talk about privacy also. Um, so one of the selling points of this, um, is, uh, so according to Chris Manning, "'NNetNav could become a lighter weight, faster, privacy preserving alternative to OpenAI's recently released Operator for using an AI agent to do things on the web.'" And the idea here is that it is a smaller model and runs locally, so you don't have to give up your data, which presumably would involve like letting OpenAI see everything you're doing on the web, which, yeah, that's icky. But also, what was the, what was the Microsoft thing called? That was like the--
Alex Hanna:Microsoft Recall.
Emily M. Bender:Yes.
Alex Hanna:Yeah.
Emily M. Bender:So what we saw from that whole fiasco is that just because the information stored locally doesn't mean it's really privacy preserving, like privacy means data minimization really. Um, if we want privacy, we've gotta data minimization. If we're letting anything, like observe anything and collect everything, then we don't have privacy, even if we're keeping that data locally.
Margaret Mitchell:Yeah. Isn't this designed to put things out on the web? Like, isn't it, isn't it odd to give it local access to private information on your computer and then allow it to push things out on the web. Doesn't that seem like the opposite of privacy preserving?
Emily M. Bender:'Okay, I'm gonna make a blog post for you. This diary entry looks like a good one.'
Margaret Mitchell:Yeah. Well, and that's actually happening too. I mean, like, people are using agents to be like better social media influencers or whatever, like where it writes the social media post based on stuff they've written like in their emails or on their computer, um. Without, you know, detailed oversight of what's actually happening then it's gonna like risk massive privacy leaks. And not just for you, right? Not just your information, information that you have stored about others. Any other thing you have about other people on your computer, pictures, you know, messages, whatever, email, whatever it may be. Like this is a vector of, uh, privacy breaches that affects everyone that you interact with and save any sort of information about.
Emily M. Bender:Yeah.
Alex Hanna:Yeah. That's, that's a hundred percent right. Yeah. Uh, it, it also reminds me just of like, there's also, another shoutout to the 404 Media podcast. They have this podcast just about like the, there's like these Discords in which there was like influencers, um, and they, they were, and mostly it was like, it was all pyramid schemes all the way down, which, you know, is about right. But it was basically like selling kind of like a one click tool that would sort of generate these, like these, these YouTube channels that were meant for monetization. And it sounds like that, I mean, effectively it's sort of like, and if you're just doing it, but I mean this, this in particular is is, is a privacy nightmare, just in terms of like anything that you have saved in kind of like different fields and whatever. Um, and just putting in any old input box, you know?
Emily M. Bender:Yeah.
Margaret Mitchell:Like a child.
Alex Hanna:Like a child.
Emily M. Bender:Would you, would you give a child access to the contents of your One Password file, for example.
Margaret Mitchell:Don't a lot of parents like intentionally not give their children their phones in case the child like accidentally buys a bunch of whatever, you know? I mean, that's like a thing that parents don't do.
Alex Hanna:Yeah. Amy says in the chat, ACZhou says,"The children yearn for the websites."
Margaret Mitchell:The children learn to, uh, yearn to book flights.
Alex Hanna:Yes. The children are really, really, really want to be in your, in your work Slack.
Margaret Mitchell:Yeah. Powering your enterprise workflow. Sorry.
Emily M. Bender:There, there's, there's a couple other things here, um, that are wonderful.
So, Abstract Tesseract:"Flight reservations, famously not involving PII", um, and SJayLett talking about the influencers using this to like do more influencing. Vibe influencing?
Margaret Mitchell:Right? Yeah. Goodness.
Emily M. Bender:Uh, um, and, uh, Trochee Trochee gets it, uh, right to the heart of it."Once again, they are selling the dream of robot slaves that have good sense." All right, anything else for HAI before we move on to Scale AI?
Alex Hanna:Um, what else? I mean, it's, it's a, it's a pretty bad one. Um, again, yeah. These HAI press releases are, are, are pretty, pretty rough times.
Margaret Mitchell:It's interesting. It continues to be interesting to me that there's a university that has like a press team. That immediately seems odd.
Emily M. Bender:Um, Stanford is not alone.
Margaret Mitchell:Right. That makes sense. Yeah. I imagine that there's a direct relationship between how much a university costs and how robust the press team is.
Emily M. Bender:Well, and I mean, so the University of Washington has UW News and they will do press releases about university research. That's the thing. And they also like scan for and sort of promote news articles about us. Um, so it's a, it's a thing, like universities have PR teams. Part of what's going on here is that Stanford has at least two. This is not the Stanford general one. This is the HAI specific one. So this is a research center within Stanford that has a press team. And that's a little interesting. That's, I think, a little surprising.
Margaret Mitchell:It's a great way to have power and influence amongst other tech companies that have similar kinds of PR departments. Like if that's the game you're playing. I mean, it's sort of, anyway, that's a different discussion. Yeah. Always surprising to see to me.
Emily M. Bender:Yeah. All right, so we've got one more main course one. We will do this, uh, quickly. Um, Alex, you wanna lead this one once I get it open?
Alex Hanna:Yeah. So this is, um, about Scale AI, uh, our friends over there. This is by Hayden Fields over at CNBC."Scale AI announces multimillion dollar defense deal, a major step in US military automation." Um, this is published March 5th. Um, this here--
Emily M. Bender:What a bad news headline.
Alex Hanna:Yeah. And the sticker is "AI effect," which I'm not sure what that means. And then the key point--
Emily M. Bender:So it seems to be the major thing here. So they have "AI effect", then "AI age", "AI at work", and "AI insights". Like, I guess this is their AI coverage.
Alex Hanna:It could be like a sub publication or something. Um, yeah, I think it is.
Uh, so the key points:"Scale AI has partnered with the Department of Defense to use AI agents for US military planning and operations. 'Thunder Forge' in quotes, is the DOD's flagship program and will work with Anduril, Microsoft and others to develop and deploy AI agents. It's a multimillion dollar deal, according to a source familiar with the situation." And then there's a picture of Alexander Wang, um, with this like hand open speaking to a, uh, to the House Armed Services subcommittee, um, during a hearing in 2023. Um. So, yeah, so AI and agents and, uh, Department of Defense, what could go wrong.
Emily M. Bender:Well, Alex, we're supposed to be asking what could possibly go, right. Didn't we learn that in the 'Superagency' episode?
Alex Hanna:Oh, hey, I, I am triggered when hearing that. Yeah.
Emily M. Bender:I know. We have, we have post'Superagency' trauma disorder now.
Alex Hanna:Yeah. Yeah. So, yeah, I mean, uh, I maybe even won't say exactly what this is. I mean, some of the quotes are just, you know, fresh Hell. Um, so there's a quote here from the, in the release from the Defense Innovation Unit, um, which is a, um, I suppose something that is, uh--
Emily M. Bender:Wait, am I in the wrong place?
Alex Hanna:--within the DOD? Uh, I, scroll, it's a little bit down. Um, so, "Spearheaded by the Defense Innovation Unit, the program will incorporate a quote 'team of global technology partners', including Endur and Microsoft. Develop and deploy AI agents. Uses will include modeling and simulation, decision making support, proposed courses of access, and even automated workflows. The program's rollout will begin in the US Indo-Pacific Command and US European Command and will be scaled to other areas." And then the quote, "'Thunder Forge--" which I cannot take that seriously at all, but, "'Thunder Forge marks a decisive shift towards AI power, data-driven warfare, ensuring that US forces can anticipate and respond to threats with speed and precision,' according to a release from the DIU--" Not the DUI."--which also said that the program will accelerate decision making and spearhead quote 'AI powered war gaming'." Um, so this sounds just like. Absolute kind of trash, um, just to kind of mass automation and, and war making, you know, and, you know, not something that's been on the docket for a while, but is now, you know, coming out in these new programs that are, are at least public knowledge.
Emily M. Bender:Yeah. Um, and I wanna direct listeners to our episode, uh, while back now with, uh, Lucy Suchman talking about automation and warfare. Um, there's a lot of good stuff there. Um, and also I forget if it was when they were on our show, we were on their show, but, uh, Charlie Jane Anders, um, I thought had the really insightful remark that, um, uh, a time bomb is an automated, is an, uh, an autonomous weapon, right? As soon as you like, set something and walk away from it in a state where it will explode or do something without any further action, you have done something automated and autonomous. And if you are talking about something that can harm people, including killing them, that is not a good idea.
Alex Hanna:Nope, not at all. And they haven't, I mean, they haven't discussed here any kind of like, uh, you know, like auto automated weaponry, but I, you know, I'm sure it is. I mean, the kind of people involved are Anduril and, and Palantir and, um, and they discuss, you know, the kind of opposition of, um, of these programs, from initial opposition through com, through employees and then employees, uh, and then Google summarily removing their AI principles of, you know, not having some of automation to, um, that is involved in, in killing and, and, and doing harm.
Emily M. Bender:So this sentence here, "Uses will include modeling and simulation, decision making support, proposed courses of action, and even automated workflows." Workflows for what? If it's workflows for dropping bombs, then, you know, there it is, right? And having that at the end, um, like this isn't proposed courses of action for figuring out how to more effectively feed the enlisted people. Right? That's, yeah, I don't think that that's what they're talking about here. Um, and so the "even automated workflows" at the end, that's workflows around deciding when and where to drop bombs, I think, um.
Alex Hanna:Yeah.
Emily M. Bender:And, and just sort of phrase--
Alex Hanna:I do wanna, I do wanna shout out Hayden the, the journalist in this who does says, who does say,"Both Scale and the DIU emphasize speed and how AI will help military units make much faster decisions. The DIU mentioned the 'need for speed' or synonyms eight times in its release." So yeah, this thing about like, its--
Margaret Mitchell:Hayden's generally a kind of awesome, uh, journalist. Like she's done a lot of work. I really like, so the headline I would guess she actually didn't write, it's probably her editor, but she is, she is usually pretty good at kind of seeing through, through bullshit and I think this is a little bit of her being like, hey, they're really pushing speed.
Alex Hanna:Yeah.
Emily M. Bender:Yeah. And the speed thing also, I forget when we talked about it, Alex, but the um, uh, that terrible system that the IDF was using to generate targets to bomb--
Alex Hanna:Yeah. The Gospel. Yeah.
Emily M. Bender:Yeah. That was also about like, let's just do this faster and more so that we can bomb drop more bombs basically. Right?
Alex Hanna:Yeah.
Emily M. Bender:Yeah.
Alex Hanna:Yeah. So, yeah, a hundred percent.
Emily M. Bender:Alright, so it's time for Fresh AI Hell. Alex, what is your favorite, uh, childhood adjacent or children directed musical theme?
Alex Hanna:What, what childhood adjacent--
Emily M. Bender:Musical style. Yeah so are we gonna do a lullaby? Are we gonna do, um--
Alex Hanna:Oh. Oh gosh. Uh, I don't think I could, I could do a lullaby. Maybe. What's a, what's another childhood? I mean, I guess the ABCs.
Emily M. Bender:Okay. Um, so, uh, somehow ABCs maybe to the tune of ABCs or however you like. Um, you are a nursery school teacher in AI Hell, and the only toys available to the kids are websites for them to poke at. And so you're singing a song to get them excited about it.
Alex Hanna:Oh, this is great. This'll be like the, an the Animaniacs song where they're talking about, uh, uh, the, the countries. So let me see. eBay, Google, Craigslist, Amazon, uh, Microsoft, Microsoft, MSN too. Uh oh shoot. Lycos, GeoCities. Uh, I've run out of websites. Um, CNBC and New York Times, WaPo too. Um, that's about all I can do. I need, I like my mind just like completely blanked on websites, but now I really wanna write that song.
Margaret Mitchell:That was impressive. Although I like the GeoCities reference. Yeah. Kind of showing your age there. I dunno.
Alex Hanna:I know. Sorry. I was like, when I was a child, what websites was I going to? Right, Lycos and GeoCities.
Margaret Mitchell:Ask Jeeves.
Alex Hanna:Yes, Ask Jeeves. Ask.com.
Emily M. Bender:All right. I'm gonna encourage people in the chat if you wanna make a longer version of that song to please go ahead. That would be awesome. So that takes us to Fresh AI Hell. We've got a bunch of tabs here, so we're gonna go through it fast. Um, Alex, I think you get this first one.
Alex Hanna:Oh yeah. So this is Forbes. The author is Jason Windguard, uh, published March 10th."Musk replacing workers with AI. Should you be worried?" Um, yeah. So it's, and it's got a picture of, of Elon Musk steepling his fingers, like a, like a madman. Um, and it's, it's, and I mean, it's just a, it's a thing on the, the DOGE um, automation, and, I mean, this is, you know, of course the, the DOGE dream right? Is removing things in government and chewing at the wires and seeing where the synthetic text machines are going to replace them. And guess what? They're not gonna replace them at all. Um, yeah. So this is--
Emily M. Bender:Well, I mean, they might do it and it'll be terrible, right? This is from March 10th.
Alex Hanna:Yeah, it's, yeah.
Emily M. Bender:Yeah. All right. I'm gonna keep us moving fast. Um, this is from Maiike Verbruggen on Bluesky. Um, with the sub, with the comment, "The future is now baby," and it's a link to a, a piece on Futurism.com with the headline,"There's a vape with the tamagotchi inside it that dies if you stop puffing." And then subhead, "A participant in an NYU stupid hackathon says they created a tamagotchi integrated vape where the pet dies if the user stops vaping."
Alex Hanna:And well, well, at least it's a stupid hackathon. Like is the the hackathon meant to like, highlight incredibly bad ideas.
Emily M. Bender:And then put them out in the world where people will pick them up? Like, I'm not quite sure what the thing is, but--
Margaret Mitchell:I would still get into smoking if I were, if I were younger. I could see myself being like, oh yeah, this is-- My friends, like all collected tamagotchis in middle school. This, yeah, would've had an effect.
Emily M. Bender:And the, the, this obviously isn't about AI, but the reason, Alex turned this up, and the reason I wanted to include it is the, um, it just sort of shows it's very emblematic of putting the needs of the technology above any concerns of the people involved. So here, like, you gotta keep this imaginary thing alive by poisoning yourself, keep going.
Margaret Mitchell:By killing yourself. Yeah. Oh, that's so interesting. Yeah.
Emily M. Bender:All right.
Alex Hanna:Yeah. Uh, gosh. So this is, so I think the thing that, so this is a, a quote LinkedIn thread. So the original one is by Eric, uh, Arzubi, MD, CEO of Frontier, Frontier Psychiatry. Uh, so frontier, you know, um, great, great language. And he says, "AI therapy just proved it works and results are astonishing. Imagine getting mental health support 24/7, regardless of waitlists or provider shortages. Today's landmark study in, uh, new England Journal of Medicine AI-star, uh, shows that this isn't just possible, it's effective." Um, and so then it's got, uh, and this is about a study. Yeah. Um, and about like what they have and, yeah, and so then the most compelling finding, according to this person, says,"People form therapeutic relationships with the AI comparable to what they experience with human therapists." Ugggh, you're not making relationships.
Emily M. Bender:No. No. And so, um, the Kasper Benjamin Reimer Bjørkskov, who's a "founder, consultant, activist, writer, and human" is their byline: "AI as therapy. The final step towards isolation. Um, this isn't progress, it's a coping mechanism, disguised as innovation. Um, this isn't the future we were promised. It's a dead end." Um, and I, I appreciated that commentary and the thought that we've got this person who's got an MD like being rah rah about synthetic text for people with, you know, mental distress just seems so frightening. But rather than dwell on that, um, here's another LinkedIn post about a paper where the paper is called"On the biology of a large language model." And, uh, this is basically the, the, um, person posting this didn't read it, so they're not an author."I've only skimmed the paper, a long and dense read, but it's already clear it'll become a classic. What's fascinating is that engineering is transforming into a science, trying to understand precisely how its own creations work. Head exploding emoji."
Margaret Mitchell:Transforming the science is actually becoming a science.
Emily M. Bender:They're not biology. Stop it. Okay.
Alex Hanna:Yeah.
Emily M. Bender:Alex, you want this one?
Alex Hanna:Uh, this, yeah. This is by, so Hypervisible, so Chris, uh, Gilliard on Bluesky and his commentary is, "You thought his posts about AI were bad, but now it's clear they were just poorly constructed grift for his chat bot." And the, it is a link to um, uh, there's a picture of Mark Cuban. Uh, "Anyone can ask Mark Cuban for business advice using his $10 per month AI chat bot. Here's what it told me." The link is something in Fortune, which I mean, I guess you're a sucker for you, you know, I guess you're a reporter, so you have to like, you have to test these things, but you didn't actually test these things to know it's a grift, right?
Emily M. Bender:Yeah, exactly. Alright, I'm gonna let you do the next one too, Alex, 'cause you found this one.
Alex Hanna:Sure. So this one is a little bit of a palate cleanser, um, and it is analysis by Allison Morero for CNN Business."Apple AI isn't a letdown. AI is the letdown." And this is a really great takedown. Um, this published, um, in March 27th, so last week, uh, really great takedown of like the Apple, um, uh, AI, Apple Intelligence kind of, um, fiasco. The fact that basically was trying to summarize notifications and it was just doing such a trash job, right? Um, and so they, you know, they basically went ahead and they, you know, stopped the products, um, or delayed it or quote, 'delayed' its released indefinitely on the AI powered Siri. Um, and, um, there's some really great text in here which actually mimics a lot of the text that we use in the book. Like in the book promo. It says "Apple, like every big tech, other big tech player is scrambling to find ways to inject AI into its products. Why?" And then in italics,"Well, it's the future. What problems is it solving?" In italics,"Well so far, so far, that's not clear. Are the customers demanding it? LOL, no." So, yeah. Uh, which is, so it's pretty, pretty great. And it, uh, references that ad for the, um, the early ad for, for AI, which is, I think it was not quite, I don't think it was for AI, it was just the, the like iPad ad. I'm assuming that's what the link is too.
Emily M. Bender:Yeah. This one here where they crushed all of these--
Alex Hanna:Where they crushed all the instruments?
Emily M. Bender:--art tools and, yeah.
Alex Hanna:Yeah. I think we've talked about this. And so, I mean, yeah. So I think the, like, the analysis is really great. It's sort of like getting at it, getting, really getting at like, who's asking for this? Why do we have to kind of bend ourselves for this product that no one's asked for? Um, in, in a really nice place in that famous Luddite publication, CNN Business.
Emily M. Bender:Yeah. Love it. Love it. All right. And we are just about outta time, but I'm gonna take us to one last bit of comic relief. Um, this is a post on Bluesky from Matt Crosslin. Um. And, uh, the, the text is, "AI is an essential life-giving vitamin according to Ed-Tech vendors." And this is a, a zoom in of, it looks like a cloth poster at a conference. And there's this like orange seal with white letters on it, sort of stamped onto the thing and it says,"Fortified with essential AI assistants." Then, um, Hypervisible says, oh, "What fresh Hell is this?" And then, uh, Matt Crosslin says, "Do you really wanna know?" And then there's a zoomed out picture where there's, um, it's more of the poster and in the middle it looks like a Wheel of Fortune. And it says, "The nineties called and they want their peer review, discussions, assignments back." And then there's a guy with a very nineties styling holding a sort of wireless phone from back when that was the thing. We were not able to figure out what the vibe was supposed to be here.
Alex Hanna:Yeah. It's just, I mean, the vibe is that some elder millennials like myself, like myself, just like, they yearn for the nineties, and like--
Margaret Mitchell:I love the jacket. Yeah. Gotta get that jacket.
Alex Hanna:Yeah. Yearn in for the nineties and they want some kind of, but it's, this is just like, this is so cursed and I hate it. Absolutely.
Emily M. Bender:Yeah. But it's fortified with essential AI assistants.
Alex Hanna:Ughhh.
Emily M. Bender:Okay. Well, that we are a little bit past time, in fact, and I'm the one who has to start the outro, so I have to get to the right tab. Okay. That's it for this week. Margaret Mitchell is a machine learning researcher and chief ethics scientist at Hugging Face. Thank you so much for joining us again, Meg, it's always great to talk to you.
Margaret Mitchell:Yes, thank you very much. It's awesome.
Alex Hanna:Thanks, Meg. Our theme song was by Toby Menon, graphic design by Naomi Pleasure-Park, production by Christie Taylor. And thanks as always to the Distributed AI Research Institute. If you like this show, you can support us in so many ways. Pre-order The AI Con at TheCon.AI or wherever you get your books, and join our virtual book tour kickoff event on Tuesday, May 8th, or find us on the road. A full list of events at TheCon.AI.
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. That's D A I R hyphen Institute dot 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 D A I R underscore Institute. I'm Emily M. Bender.
Alex Hanna:And I'm Alex Hanna. Stay out of AI Hell, y'all.
