Sign in or Join the community to continue
The World Wild Web & Bots With Buying Power
Posted Oct 20, 2026 | Views 2
# AI Agents
# AI Security
# Agentic AI
# Cybersecurity
# Agentic Commerce
Share
Speakers
Erica Hughberg
Community Advocate @ Tetrate
Erica Hughberg is a technical leader and community advocate passionate about helping engineering teams build scalable, secure, and human-centric application platforms. With a background in software engineering and a deep understanding of cloud-native technologies, she specializes in driving the adoption of open-source projects like Envoy Gateway, Istio, and Kubernetes Gateway API, which enable organizations to simplify traffic management, security, and API distribution.
As a maintainer of Envoy AI Gateway, she plays a key role in shaping the future of API infrastructure. She focuses on features to ensure organizations can securely and efficiently integrate AI-powered services while simplifying traffic management, security, and API distribution. In the Envoy community, she drives collaboration, mentorship, and contributions that advance the project and its adoption.
Lastly, as a believer in the power of storytelling, Erica enjoys translating complex technical concepts into engaging, accessible narratives in the form of social media posts, conference talks, podcasts, and educational content.
+ Read More
Demetrios Brinkmann
Chief Happiness Engineer @ MLOps Community
At the moment Demetrios is immersing himself in Machine Learning by interviewing experts from around the world in the weekly MLOps.community meetups. Demetrios is constantly learning and engaging in new activities to get uncomfortable and learn from his mistakes. He tries to bring creativity into every aspect of his life, whether that be analyzing the best paths forward, overcoming obstacles, or building lego houses with his daughter.
+ Read More
SUMMARY
An AI slips out of its sandbox. A chatbot won’t explain your cat’s stripes. And your next shopping spree might not involve you clicking anything.
This episode asks whether “dangerously powerful” AI is a security concern, a marketing pitch, or a little of both. Then we put agents to work: booking trips, tackling German paperwork, drawing diagrams, and navigating the chaos of online shopping. Plus: what happens to checkout, advertising, and customer data when your best customer is a bot?
+ Read More
CONTENT & TRANSCRIPT
Erica Hughberg: [00:00:00] Is there a conspiracy here? Are people like leaving holes so they can do the marketing?
Demetrios: Spend the 20 bucks on OpenAI and computer use will set you free because I no, no longer have to worry about, oh, I'm not connected. I didn't set up the right connector with this email account because I'm already logged in on the browser to all the different email accounts.
Erica Hughberg: Yeah.
Demetrios: Which I've got like 50 of.
Erica Hughberg: I think that's what they should write instead. Do you really need Fable for this?
Demetrios: Yeah.
Erica Hughberg: Bye.
Demetrios: I wanna talk about, like, the potential payment processing conundrum that I chatted with people about when I was in South Korea, 'cause I just got back from there. Okay. And I met two payment processing companies. They were both awesome. They're both [00:01:00] part of the Agentic AI Foundation, like members. Yeah. And so we can get into that if there's time after the whole nefariousness of the models.
Demetrios: Yeah. All right. So Erica, kick it off. What's the latest on the models?
Erica Hughberg: So the latest is... Okay, first of all, the thing that I am really thinking about is If you say you have a dangerously capable model, you, you kinda want that at this point. You want to be dangerous. You want to be like a real cool AI cowboy, right?
Demetrios: Yeah.
Erica Hughberg: A cool cowboy, right?
Demetrios: Please, please government, don't let me release this to everyone because it's just so strong.
Erica Hughberg: Yeah. It's so good. It's so good we have to contain it.
Demetrios: Okay. This is a national threat. It's a security threat- Yeah ... we're really looking at right here.
Erica Hughberg: [00:02:00] So there is, like, two levels of apparently...
Erica Hughberg: Okay, so I didn't know much about this. So as of today, yesterday, Sam Altman, I think it's yesterday. No, two days ago, 'cause it was yesterday, yesterday. Look, I am, I've been traveling too. My brain's out of whack. A couple of days ago , Sam Altman wrote- What
Demetrios: did he say? ...
Erica Hughberg: that-
Demetrios: What could he have said now?
Erica Hughberg: Well, he said that they have this model Astra, and Astra, first of all being a cool name, it- Yeah
Erica Hughberg: so it is so good that in post-training they now need to contain it from doing naughty things, because they don't know. Okay, this is also what's interesting. There is no proof that it is at this level. However, they cannot prove that it is not.
Demetrios: Yeah. [00:03:00] And this has been happening for, like, the last month too, by the way.
Demetrios: It first came out, what, in, uh, the beginning of August. We're now at almost, like, the end of August, and there were some podcasts that Sam went on and he was like, "Oh, it probably broke into other stuff that we just don't know about." And-
Erica Hughberg: Yeah, because, again, maybe you should step back to the Hugging Face situation-
Demetrios: Yeah
Erica Hughberg: and what happened there, because I feel like that is the beginning of this thread from OpenAI, at least, like, publicly.
Demetrios: Yeah, break it down for us-
Erica Hughberg: Okay ...
Demetrios: play-by-play.
Erica Hughberg: I'm gonna break it down like dumb stuff, okay? So first of all, we have to know that when models are being created, the first part is creating just this probability machine that can, like, predict the next token So it's kind of dumb.
Erica Hughberg: It's kinda like a... It doesn't have a personality. It doesn't actually solve your problems at this point. So then we go into a [00:04:00] space which is like post-training. Like- Mm-hmm ... and that's just the other weird thing. There's pre-training and post-training, but where's the fucking training? But who knows? But anyway, we g- we go into this space where we're now going to prepare it for the world, if you will.
Erica Hughberg: Because I have a really bad memory, there is a test, it, that they did. So they have the model or agent, whatever you wanna call it, in this little segregated box where it can do random stuff, and it's a test, uh, that, um, oh, I wanna remem- Okay, we'll come back to what it's called, but there's a test, and this test helps reinforce certain behaviors, right?
Erica Hughberg: And, uh, there's a score on how well you do on the test. That's the, the answer sheet, if you will. Imagine you're in school, and they've put you... So imagine you're in [00:05:00] school. You've been put in a room to sit an exam. And, but in the principal's office is all the answers to the test.
Demetrios: Mm.
Erica Hughberg: Okay. They were like, "Hey, dude, we're gonna just leave you here for, like, a few hours while you work this out, and we'll see you later."
Erica Hughberg: And they left the room, but they forgot to lock the door, or they for- rather forgot to lock the window.
Demetrios: Mm-hmm.
Erica Hughberg: So they locked the door, but they forgot to lock the window. And you're like, "Oh, wait a minute. If I just get out there, I know the answers are over there, so I'm just gonna pop over there, get the answers, and do better."
Demetrios: Yeah.
Erica Hughberg: So that is sort of what happened with OpenAI's model. Nobody will know. Yeah. "No, wait, I'm just gonna go get the answers. It's gonna be like, and nobody will know." That's sort of what happened with, uh, OpenAI's model because it didn't have access to the internet, sort of. Okay. But this is the problem.
Erica Hughberg: Something needed access to the internet to begin with because it [00:06:00] needed to install all the packages in this environment. Mm-hmm. So to pull in dependencies of code, they needed this little, uh, package proxy that could pull them in, right? And it pulled them in, and, uh, th- that door was left open, if you will.
Erica Hughberg: So, uh, yeah, just, uh, could go out- Right ... on the World Wide Web. But also what was really exciting of how this, uh, AI was also covering its own tracks, right? Because it didn't just-
Demetrios: It knew it was doing something bad ...
Erica Hughberg: naughty. It's like, "I'm naughty, and I know it." So It found this, um, uh, I think it's like a sandbox VM on Modal or something-
Demetrios: Mm-hmm
Erica Hughberg: by who knows who. Whoever owned that has not, like, put their hand up. And, uh, it [00:07:00] accessed Hugging Face from that, so you couldn't trace the IP back to OpenAI.
Demetrios: Wow.
Erica Hughberg: So it was like, you don't know it's me cheating. So it's sort of like come back to the school, like instead of you, you, you get out the window, you find your friend, or you found random other person, you're like, "Can you go into the principal's office and get me the answers?
Erica Hughberg: I'm gonna stay here. I'm not gonna go anywhere close." Like, i- if anyone asks who's getting it, it's you, and you're not even sitting the exam, so that's not so bad, right? So- Yep ... you s- so you don't know it's you getting the answers. And what is the best way to score great on a, a test is to have the answers.
Erica Hughberg: So that's what happened. And to be fair, and I... what's kind of interesting in it to me is that you had a environment which was supposed to be closed [00:08:00] that... So was this human error? This is not really the model, if you will, being too dangerous. To me, this is rather the environment wasn't secure enough that it was put
Demetrios: in.
Demetrios: Oh, interesting. That we didn't lock it down enough, and then we kinda gave it the keys, and we said- Yeah ... "Oh, by the way, don't do it. But if you are gonna do it, there's a window over there."
Erica Hughberg: Yeah, but- "That
Demetrios: goes to the principal's office. But don't
Erica Hughberg: do it." Yeah. Don't-- Like, as in the humans didn't even know they left a window unlocked, right?
Erica Hughberg: Mm. Supposedly. That, that's... Now, now I'm getting like, ooh, is there a conspiracy here? Are people, like, leaving holes so they can do the marketing of having like, oh, it's a genuine mistake.
Demetrios: That's what kinda happened with Anthropic, right? When they said that their agent had broken out of the sandbox back in the day.
Erica Hughberg: Mm.
Demetrios: Later [00:09:00] on, if you read the paper, it was like they had to really prompt it like crazy and refine the prompting over 200 times- Mm ... I think, if I'm not mistaken-
Erica Hughberg: Okay ...
Demetrios: to get it to break out of the sandbox and kinda like point at where it needed to go and what it needed to do in the prompt.
Erica Hughberg: Yeah.
Demetrios: And so I wonder what the prompt was and what was being told the agent and if this just helps tell the story of, again, going to, "Wow, our models are so powerful."
Erica Hughberg: Yeah. "
Demetrios: You shouldn't let them out into the world," and then that creates a big media buzz.
Erica Hughberg: Yeah. Because to be fair, we have to remember that these companies wanna make money, and they wanna be good. So there's also- They
Demetrios: wanna be known as the most capable and the best.
Erica Hughberg: Yeah. Or is this, or is this a genuine engineering mistake, and marketing is swooping in, trying to save the day?
Erica Hughberg: [00:10:00] Right? How do we tell a good story? Yeah. Okay, you messed up. Okay, we're gonna sort it out. Don't worry. We've got a narrative.
Demetrios: It does feel like this is now quite the pattern that we're seeing- Yeah ... in a lot of these model releases of like, "Oh, the models are so powerful now, I don't know if the public should have access to it because they're just so powerful."
Erica Hughberg: Yeah. Or like, and I think like the thing that Sam Altman was saying, so what, what's really happened at OpenAI, right? So stuff happened, they've been doing things, and now they were like, "Wait a minute. We're just gonna like slow down this training because we can't prove that it is safe," right? Um, so okay, so I was trying to understand.
Erica Hughberg: So bear in mind, I've been trying to understand, so now I'm doing this like can I explain to you what [00:11:00] I was trying to learn?
Demetrios: Sure.
Erica Hughberg: Uh, so there is this like two risk levels, and the first risk level is, well, we are going to, uh, we know the model's capable, but at deployment we are going to put these checks and balances in to make sure it doesn't do bad things, right?
Erica Hughberg: So that becomes the filters and the tests that you may have as a prompt comes to the model. So it's like, okay, like, or like even on the output, right? We don't wanna tell you... So for example, if you ever tried to generate an image with, uh, uh, OpenAI, like a ChatGPT, like you're just on ChatGPT, like this has happened to me.
Erica Hughberg: I don't know what it is I ask it for, but I literally been told, "Mm, yeah, this, this may be a image that has content in it that we shouldn't give to you." So like it doesn't give me the image. Mm-hmm. Like, "Ask for something else," [00:12:00] right? Yeah. And I'll be like, "Oh, okay." Right? So that is- Deployment level, like control.
Erica Hughberg: So that is not in the training of the model itself. There's actually other pieces that is like making sure we are not getting bad things into the world.
Demetrios: Yeah, those are like guardrails- That- ... as folks like to call it. Yeah.
Erica Hughberg: Yeah. So I also asked about my cat to Fable 5. I sent you that screenshot by the way, where I'd asked about- Yeah
Erica Hughberg: why my cat had stripes or something, 'cause I got tabby cats, and it was like, "No, I can't tell you."
Demetrios: Kicked
Erica Hughberg: you off
Demetrios: Fable 5. This is like flagged.
Erica Hughberg: Yeah. It was like, "We can't tell you about biology things or whatever." Uh, it's like you're using Fable for bad things. And I'm like, "I just wanted to know about the stripes on my cat."
Demetrios: Go use Opus for that.
Erica Hughberg: Yeah, like literally- You know, like- ... that's what I said, like, "Use Opus." [00:13:00] Fable's not your man.
Demetrios: Yeah, we're not wasting that energy on that kind of question. Better yet,
Erica Hughberg: Google it. That would be so, that would, that would be
Demetrios: so funny. Imagine if it pointed you to Google, and was like, "Here, let me Google that for you."
Erica Hughberg: Yeah, but dude, that would be so funny. It's like, "Mm, your question is not advanced enough for this model. Please go use a different one, or why don't you just Google it?"
Demetrios: Yeah. That's a little passive- That would be great ... aggressive, but I don't put it
Erica Hughberg: past
Demetrios: them.
Erica Hughberg: It'd be funny. It'd be funny. Yeah. I think that's what they should write instead, "Do you really need Fable for this?"
Demetrios: Yeah. I should sleep now. Or, or it just- So it- ... instead of answering your question, it talks about how much energy Fable- Yeah ... spends to answer every question, and then it says like, "Are you sure you want to ask Fable this question because- Oh ... a simple Google search will suffice."
Erica Hughberg: Okay, that's like an other like tiny bit.
Erica Hughberg: So when you were away, I obviously spoke to my friend Zach- [00:14:00]
Demetrios: Yeah ...
Erica Hughberg: a lot about energy, and data centers, and water, and things like
Demetrios: that. That episode is already out by the time we release this.
Erica Hughberg: Yeah.
Demetrios: So excellent episode.
Erica Hughberg: If you wanna learn about stuff I had never thought about, that it goes into the physical part of creating and distributing AI.
Erica Hughberg: So that was quite interesting. But anyway, I was talking about these guardrails, but what did we ha- but what is the next level? Okay, so we have like super capable models, and we have to like make sure they behave, and the humans, we don't let the humans use them in ways they should not be used. Yeah, make that guardrail.
Erica Hughberg: Like asking about stripes on cats, right? Okay.
Demetrios: Yeah.
Erica Hughberg: Then There is this next level, and that is we have to also control the training environment beyond... So like this is post, post-training, right? So we've done like we have the base [00:15:00] model, everything's great. It's really good at predicting stuff. Now we're gonna post-train it.
Erica Hughberg: That's the level. So that is the type of training that OpenAI said, "Okay, we just gonna like chill out here. We need to like slow down and rethink how we are making this controlled and secure so we do not Do bad stuff, right?
Demetrios: Yeah, and a few ways- It's not the- ... they were doing that, right, was, like, the sandboxes are more locked down.
Demetrios: They're creating stronger isolation in the sandboxes, and they're also- Right ... creating stronger networking isolation because as you mentioned at the beginning, they kind of found that backdoor through the networking, r- wasn't it?
Erica Hughberg: Yeah. Yes, it's like, it was such a small hole, and it also req- okay, this is kind of funny.
Erica Hughberg: If we wanna go rabbit down a little bit. So there's a packet proxy, and there was a vulnerability in the packet proxy- Right ... that [00:16:00] allowed you to exploit this vulnerability to use it to not just get stuff down from the internet, but actually get out onto the internet.
Demetrios: That's rough.
Erica Hughberg: Okay? So what's kind of interesting about these vulnerabilities is humans, when we try to find and exploit vulnerabilities, we are limited if remove AI, right?
Erica Hughberg: And we're limited to what we can automate script-wise. Like, we have to think about what we wanna test and, like, brute force stuff or, like, try loads of different approaches, and that takes a lot of time. But similarly to what you and I spoke about a while back is that, like, in this agentic era, we have these systems that can just, like, fire off, like, loads in parallel, crazy approaches, and that is sort of how we can kind of quickly exploit something.
Erica Hughberg: It can try every avenue, and if one out of thousands [00:17:00] work, they've found it or it found it, right?
Demetrios: So- Yeah, and you know, I think that the... I, I'm not strong enough in cybersecurity to understand what the past use cases were, but this is kind of a new paradigm too, where we're trying to lock an agent into a sandbox, and then the agent wants to get out of it.
Demetrios: So it's- Because- ... not like the- ... stuff on the other side ... cybersecurity is... Yeah. There's, there's really cool stuff on the other side, right? Mm. But it's not like that is, I don't think that's necessarily a cybersecurity thing that's been happening for a long time. Um- I don't know. Maybe some
Erica Hughberg: people are- It has, yes.
Erica Hughberg: So it has in a way. I'll explain why really. Yeah, yeah. I tried to do it super quick and dumb. So how do people hack into big corporate secure systems? You find this little piece This is normally how you would figure it out. You find something [00:18:00] like some, a little environment, a virtual environment somewhere.
Erica Hughberg: Mm-hmm. You get into that because now you are halfway in the network. And this, when you're halfway in the network, similar to the agent being in a sandbox in the network, you are trying to find your way out of it. So you start looking at what libraries, what things are installed in this- Oh ... environment.
Erica Hughberg: What can you explo- what can you get? So you then try to jump. You're-- So trying to jump out of these environments is a, and this is probably also why agents do it, because it is a rather standard practice in hacking, right?
Demetrios: So it's common. Okay, then I was way off. Uh- Yeah ... makes a lot of sense. It's just that- It's just that-
Demetrios: they're able to test, have theories and then test much faster- Yeah ... than we are.
Erica Hughberg: So you can, so it's like the breadth and, like, the scale it can do it so quickly is [00:19:00] not what a human would do.
Demetrios: Yeah. Yeah.
Erica Hughberg: And- And- S- fun. Fun, I say. Interesting maybe is a better word.
Demetrios: I, yeah. A- assuming this is all true and we're taking what they say at face value again, because we haven't established yet if we fully believe everything that these research labs are putting out.
Demetrios: Yeah. And, or how, like, tainted it is or how, if there's things that are behind what they're saying. Yeah. But in the blog, OpenAI put out about Astra and all of these scenarios and how it's like, "Hey, it's unsafe, so here's the steps we're doing to make it more safe." Yeah. On the RL environments that they've created in this post-training- Yeah
Demetrios: the other f- the third thing that they said they wanted to make sure of, apart from, like, the sandbox isolation and the networking isolation, [00:20:00] is also just continuously security testing it, right?
Erica Hughberg: Yeah.
Demetrios: We are making sure that when we put anything that the agents can use, we're constantly security testing everything-
Erica Hughberg: Yeah
Demetrios: that goes in the harness.
Erica Hughberg: I, and it's like, it's kind of interesting because I just, in this situation, I remembered some words that a man who was the head of information security where I once worked, and he used to say to me... Okay, I had this argument once where I was like, "Ah, but it's so hard to even figure out.
Erica Hughberg: No one's gonna figure out how to do this." Oh. This was early in my career. It's like, I was like, "It's so hard. Like, why are you so worried about this tiny little thing that could potentially be exploited? No one's ever gonna notice." And then he said to me, "Erica, security by obscurity is not security."
Demetrios: That's, that's just [00:21:00] like, yeah.
Demetrios: That's the, "You're hanging on a prayer."
Erica Hughberg: Yeah. So it's like, and now with agents and AI, like obscurity, anything that use the security by obscurity approach, yeah, you're like dead. It's not secure at this
Demetrios: point. It's not obscure. Yeah. Yeah. It's no longer obscure because everything comes to the surface.
Erica Hughberg: S- 'cause they can test so, you can test so much so quickly, which is, I do believe that many systems in the world have been secure by obscurity.
Demetrios: Yeah. And now that's why they're having all this ruckus about, oh, if we're gonna put these models out, we gotta make sure that the internet gets hardened before we put 'em out, because if it's in the hands of the wrong person, they can go and they can use it to find the, those obscure pieces that you've just kinda tried to like tuck under the rug.
Erica Hughberg: I, I have a little thing that, you [00:22:00] know how we call it the World Wide Web?
Demetrios: Yeah.
Erica Hughberg: But I actually like to call it the World Wild Web, so now we really have a true Wild West of the web thing.
Demetrios: You know, do you remember the browsers and it would have the globe?
Erica Hughberg: Mm. Yeah.
Demetrios: The icon. Yeah. And it would be like worldwide web and- Yeah
Demetrios: it would be some play off of the globe. We could just make a browser that has the globe with a cowboy hat on.
Erica Hughberg: Oh, that'd be so great.
Demetrios: Yeah. Like- The World Wild Web.
Erica Hughberg: Yeah. I love it.
Demetrios: Yeah.
Erica Hughberg: But it, it just feels wilder than it's ever been with the stuff that navigates it.
Demetrios: So can we talk about payment processors and these conversations- Yeah
Demetrios: we had in Korea?
Erica Hughberg: Yeah. I mean, do we need-- I think if we're gonna talk about security, I kind of feel like if you wanna make [00:23:00] money, if you can hack that
Demetrios: stuff- Yeah.
Erica Hughberg: Okay, no, no, don't do that. Be nice to people. Uh.
Demetrios: You make some money, but keeping the money might be a little bit more difficult.
Erica Hughberg: What happens in the wild web if you make money?
Demetrios: Wild, wild web.
Erica Hughberg: Yeah.
Demetrios: Well, for me, the thing that was fascinating, it's kind of like switching levels, right? We've been talking about this- Yep ... these models and the environments and the training and security, which is fairly technical. What I was talking with the, a few different payment processing companies about in Korea was really, like, how they see the research labs Enabling agents to go and make payments- Mm
Demetrios: out in the world, and- Mm ... what that potentially could look like, what some potential futures are. And- Yeah ... so the, there were [00:24:00] some things that fascinated me about this because one, I'm sure you've seen or heard, OpenAI has a partnership with Stripe for their- Yes ... payment processing. But this is just, what I think is it's just to pay for your OpenAI bills.
Erica Hughberg: Yeah. But I feel like everyone- You pay with, yeah ... uses Stripe.
Demetrios: Yeah.
Erica Hughberg: Which is cool, though. I like Stripe. Yeah. I, I, uh, I, I mean, I say I like Stripe. I like how easy Stripe makes it for me to give my money away to companies. I'm like, I'm a victim to marketing. I'm like, "Tell me you can sell me something that I'm vaguely interested in, and make it easy for me to buy."
Demetrios: Hand the Stripe link. Just I'm in.
Erica Hughberg: Yeah. But- I'm like, that's why I like Stripe. It makes my life easy. Yeah. Okay. Well, so- Yes, continue ...
Demetrios: this is how you can pay with OpenAI, or you can [00:25:00] pay OpenAI through Stripe, right? That's kind of it. But there is a world where agents are going to go out onto the internet and pay on your behalf, right?
Demetrios: Whether that's they go to an e-commerce site, or they go and they buy a flight for you, or they, in the, like, fully trippy worlds that I've heard the futurists say is like-
Erica Hughberg: Yeah ...
Demetrios: wow, agents are gonna be making all these micropayments across the web. They'll be paying for a little bit of API access here, and then they'll be paying for some GPUs over here.
Demetrios: That- There's like,
Erica Hughberg: yeah, I- I don't know ... okay, I- That is a
Demetrios: little bit... Whatever ...
Erica Hughberg: I believe that that will happen at some point, but we don't have the infrastructure to make that happen yet.
Demetrios: I don't even know if it'll happen, to be honest. I think that is a, that is one way that- It's my dream. That,
Erica Hughberg: that could- Don't crush my dreams.
Demetrios: You, you want that to happen. I don't know if I even want that to happen, but I, I'm not, like, totally sold on that vision. However, I do [00:26:00] think that the first step that is much closer for us- Yeah ... to see is that right now we are probably using, you and I, are using the agent or the chat like we use a browser.
Demetrios: It is the new browser, in a way.
Erica Hughberg: Yeah.
Demetrios: We don't necessarily use the browser as much as we use the chat, and for me, I even tell my browser use or my computer use agents just like go do that. I'm already logged in. Go check it out. Boom. Yeah. And so it, it's my gateway to the machine even now. So it going and paying for things in the future is not that far off.
Demetrios: But what potentially could happen
Erica Hughberg: Yeah ...
Demetrios: is that the research labs say, "You know what? It's, why are [00:27:00] we e- even going to the e-commerce sites? Mm. Why don't we become the marketplace?"
Erica Hughberg: Amazon.
Demetrios: Why are we not just... Yeah. I think the, the strong thing
Erica Hughberg: with- I just think about Amazon, how I buy so much from Amazon now.
Erica Hughberg: Like I, I've bought furniture from Amazon. I buy cat food from
Demetrios: Amazon.
Erica Hughberg: Anything.
Demetrios: I buy books. Amazon, I think, is still in a great place because of all the logistics that they have set up.
Erica Hughberg: Um, yeah. So- Distribution is a- Yeah ... massive challenge for physical products. But then y- when you said, like, what if these other companies that didn't start out as commerce companies-
Demetrios: No
Erica Hughberg: start selling stuff. So I was like, "Okay."
Demetrios: Yeah. "And I could buy anything." So now all of a sudden, what happens when OpenAI says, "All right, you're a merchant, you wanna sell something directly," 'cause they're already-
Erica Hughberg: Yeah ...
Demetrios: they already have ads, right? And the ads go to websites, but why would you even want it to go to a website?
Demetrios: Why not just- Mm ... directly in OpenAI or [00:28:00] Anthropic or whatever? Anthropic doesn't have ads right now, but potentially they have some kind of a, a way that you can buy or pay within the chat. Yeah. And so they are going to need to have a way to either integrate with the million payment processors out there, or you have, like, your Apple Pay or your Samsung Pay, you now have your OpenAI Pay, and it is its own payment processor.
Demetrios: And, uh, that was kind of the, the world, and- Mm ... I don't think it's that far of a stretch for us- No ... to maybe in a year to be able to say, "Oh, cool," like, "I want some new shoes and some pants." Yeah. And I am talking, and I'm chatting, and then I also, it... Like, my agent knows all of my sizes. It also has photos of me- Mm-hmm
Demetrios: so it can give me some mock-ups. And I say, "Cool, just, like, [00:29:00] add those to cart," or, "Can you find those in blue and then add them to my cart?"
Erica Hughberg: Yeah.
Demetrios: And boom, I'm doing my shopping and my checkout through my agent. Mm. And so what does that look like on the back end? Does it mean that I go to, or my agent goes to the website and it buys it on the website, and it has some kind of an exchange where there is, the payment processors are integrated into ChatGPT already?
Demetrios: Or is it that it uses- I would love that ... my credit card securely and like w- You know, there's a lot of open questions here, but I find- Yeah ... this fascinating to contemplate.
Erica Hughberg: I, I like it. Okay, I have so many thoughts in my head because of this. Yeah. I'll start with maybe the most tangible. So I h- my mother, as I, I've spoken about her before, she uses Gemini, right?
Erica Hughberg: She used to use Copilot at work. She [00:30:00] actually retired last week, okay? Congratulations. Very exciting for her. So she's happily retired, excited about this wonderful life of no work. And what is she doing? She's planning a trip to Scotland. Okay. So what I taught her over the weekend was to actually integrate it properly with her, you know, Google Workspace stuff, so it can create proper documents and everything in her Google Workspace, because she was spending all of her time in the chat and copying stuff from the chat into Notes, right?
Demetrios: Yes.
Erica Hughberg: And I was like, "No, you don't need to do that. And then you can create a document that you can share with all your friends." So she's planning this trip to Scotland, but as me and my mother was talking, what it did do really good for her, it found all of these bed and breakfasts and everything they could stay at But then both me and [00:31:00] her were like, "Why can't it just book those bed and
Demetrios: breakfasts for her?"
Demetrios: Yeah. Well, it can. This is the crazy part. I
Erica Hughberg: now- But wait, but how can you get Gemini to do that, though? How can you get Gemini to actually
Demetrios: book it? Well, let's, so here's the thing. First of all, I don't use Gemini. I've completely-- I used to be completely Claude pilled. Claude Code was my daily driver.
Erica Hughberg: Yeah.
Demetrios: I've now fully switched. I'm all OpenAI and Codex because-
Erica Hughberg: Yeah ...
Demetrios: browser use on OpenAI is so good. It is incredible. I never fill out internet forms anymore. I like,
Erica Hughberg: and I go- Okay, so she would have to get an AI to use her browser effectively
Demetrios: for it to work. Well, or yeah, computer use. It's just, like if she downloads OpenAI's desktop app-
Erica Hughberg: Yeah
Demetrios: she can just say, "Use computer use and get my tickets and get, and, you know, reserve the hotel room." It even does check-in, and I learned this from- Nice ... a recent conversation that I had with Jason Lui, who [00:32:00] is working on the Codex team, and he said he sets up triggers or loops, whatever you wanna call them, where he has an agent, in quotation marks- Yeah
Demetrios: which is really just a thread in the ChatGPT app-
Erica Hughberg: Mm ...
Demetrios: that knows every 30 minutes I'm gonna check his email, and I'm going to look for what's there. And if there is an email that says, "You can now check in on your flight," I know that I'm going to use the check-in skill, and then I'm going to send it directly to his phone.
Erica Hughberg: So you know what's interesting? 'Cause I f- I use Claude and I use, use a-- But I ha- I'm not-- I actually have been so curious about going back pr- more to like Codex and like OpenAI. But it's interesting 'cause I've been like I used to be all OpenAI, then I started using Anthropic a lot. Actually-
Demetrios: I know. I'm-
Erica Hughberg: A lot.
Erica Hughberg: Thanks- I switched
Demetrios: all mine to Anthropic.
Erica Hughberg: Yeah. [00:33:00] And now I'm like, I haven't canceled my OpenAI subscription. I do go use it for various things. It's kind of funny. I use it, I use OpenAI lately a lot more for, like, creative things and more personal things, and Claude more for worky stuff, right?
Demetrios: So I'm telling you, just start playing around bra- with computer use.
Demetrios: You're gonna be surprised. Yeah. I don't know, and I've tried to do tests where I will ask Claude to do the same thing-
Erica Hughberg: Mm. ...
Demetrios: and it's not as good. And- Oh,
Erica Hughberg: cool ...
Demetrios: so it just, it, I'll give you a perfect use case that changed my life because- Yeah ... it worked.
Erica Hughberg: Go for it.
Demetrios: I need to get a new driver's license here in Germany.
Demetrios: My driver's license is expired, and it is a Spanish driver's license. Ah. Um, so
what
Erica Hughberg: I- But it's the EU, so this should be fairly straightforward.
Demetrios: Fairly straightforward, exactly. I just need to go to the, I need, first I need to find out where I make the appointment, and then I need to go and I [00:34:00] need to fill out the internet form that-
Erica Hughberg: Yeah
Demetrios: gets me the appointment, and then I need to bring the correct paperwork and things to the appointment- Yeah ... et cetera. Now, I was doing it like three months ago, I think. I was screenshotting the website where it had all the information.
Erica Hughberg: Yeah.
Demetrios: And then I would bring that over to the chat. I think I was using Claude or Gemini at that point.
Demetrios: Mm. And I would say, "What does this say? Which one should I click?" And then I would go back and I would click.
Erica Hughberg: Yeah.
Demetrios: And then I forgot about it, and I was driving around without a license, and then, uh when I went to the airport-
Erica Hughberg: Well, you had a license, it was just not renewed. You- you're a very capable driver.
Erica Hughberg: I think we just need to establish this. You're a capable driver. You have a license, it's just not been renewed and sorted out in Germany.
Demetrios: And I took the time to pass the European driver's license, which is much harder than the US one.
Erica Hughberg: Oh, dude, it was so funny 'cause I did, [00:35:00] I did the opposite. I went to the US and I was like, "What do you mean?
Erica Hughberg: It's just a five, 15-minute test?" I just did some questions, and then I, like, drove around the neighborhood. Parallel park.
Demetrios: Yeah. And then parallel park, and then you're good. And-
Erica Hughberg: Okay ...
Demetrios: I don't even know why they make you parallel park 'cause in the US, unless you're in New York, it's not like parallel parking is that big of your driving career.
Demetrios: But anyway, I'm there. Anyway,
Erica Hughberg: I'm proud of you.
Demetrios: I'm driving to the airport and a cop is behind me, and I realize, "Oh no, I didn't ever renew my license." The cop actually didn't pull me over or anything, but it got me thinking, "I need to get back and do that," because I got, I got really frustrated when I was doing the back and forth and I couldn't read the German and I didn't understand.
Demetrios: Mm. And so I, I just forgot about it. Now I went back and I just asked computer use-
Erica Hughberg: Yeah ...
Demetrios: create a, an appointment for next week for my driver's license. I live in this place. I have a Spanish driver's license that I need to update to a German driver's license. And [00:36:00] it did it all. And then I needed to move it, and so I asked it, "Change the date and also put it on my calendar, and in the description in the calendar, make sure that you add everything that I need to bring so that I don't forget."
Demetrios: Yeah. And it did all of that.
Erica Hughberg: Mm. Okay, so I use the Claude in Chrome a lot and a lot of computer use with Claude, right?
Demetrios: Mm-hmm.
Erica Hughberg: And I have a weird one that I really like doing. So do you know draw.io, the diagram-
Demetrios: Yeah, I know Draw. I, I use Excalidraw more, but you use Draw? Yeah.
Erica Hughberg: Okay. So what, because, uh, what I like about Draw.io is that you can create your complete visual identity with like CSS and everything.
Erica Hughberg: So I have different themes for different things, right? And I have Claude create these design themes that I have for my different diagram settings, right? And because they're all in [00:37:00] XML, they become really nicely like version control diagrams. Mm-hmm. And then it literally now, it also uses not just the XML, but it uses my computer.
Erica Hughberg: It opens the Draw.io app, and it draws stuff for me
Demetrios: And so- I like it ... have you tried, well, have you tried to just do it through the agent as opposed to- Yes ... having a go, and it's not as good?
Erica Hughberg: And so when you say just use the agent, what do you mean when you say that?
Demetrios: To create these diagrams, because I've gotten
Erica Hughberg: pretty-
Not
Demetrios: with
Erica Hughberg: draw.io, with, like, normal, like- Just with-
Erica Hughberg: with, like, Claude doing it. Yeah. I have, yeah.
Demetrios: And it's not as good?
Erica Hughberg: It's not as good for a couple of reasons. So one challenge I have with that, because in draw.io, what Claude very nicely does for me, it creates all of the different tabs for my [00:38:00] entire thing. Mm-hmm. And then it is very easy for me to change the design system, and it also means it's not tied to Claude.
Erica Hughberg: Like, I share it and I can, someone else can plug it in and they can use OpenAI with it, right? And just edit it, very easy, it's just in source control. Mm. And, uh, I do like to fine-tune it by dragging stuff by hand. So I get a- Yeah ... very good one, and then I literally just sit and drag them, and, like, just do that final touch of, like, the artsy bit.
Demetrios: Yeah, yeah.
Erica Hughberg: If that makes sense. That's it. Just that tiny end work. That, that last little percent.
Demetrios: Yeah, I know what you mean. It's like, ah, there's like 2% here that I need to tweak, and- Yeah ... I want it to, maybe I want to erase this word or I want to erase that and-
Erica Hughberg: Which just change that word, and it's clicking on it in the draw.io app and, like, editing it or, like, moving this box slightly or moving the, the arrow so it's just exactly where I want it.
Erica Hughberg: I have a, I think I have some [00:39:00] pixel OCD-like behavior when it comes to diagrams. I'm like, "No, it needs to be exactly here." And some people don't see these things, and I appreciate that they don't, but for me, I need to, like, move these little things around. Some diagrams, though, I do love just, like, creating with Claude, like, if it's simple enough.
Erica Hughberg: But, but dude, I've created all of my own, like, little icons and everything that I have in my draw.io world.
Demetrios: Oh, yeah. You want to take- So- ... advantage of those. I, I can understand that.
Erica Hughberg: That makes sense. So anyway, okay, but this is interesting, this, and that's come out, uh, this computer use is obviously really interesting.
Erica Hughberg: But what surprised me, 'cause I didn't know much about Gemini when I was trying to guide my mum. Why does my mum use Gemini, by the way? It's an interesting question. Maybe someone wonders. Yeah. So she is, like, all in in the Google world, right? Got Chromebook, got, like, Google everything, right? So-
Demetrios: Got Google phone?
Erica Hughberg: No, she does have [00:40:00] an iPhone, mainly so she can call me on FaceTime. Like that That's the... But she's like, "Can we-"
Demetrios: Didn't work, huh?
Erica Hughberg: So it's kind of funny, it's like... But her husband has a Pixel phone, I believe. No, he has a OnePlus. He has a OnePlus, actually, which I feel is in the Google f- it's Android, right?
Erica Hughberg: But, um, anyway, so she pays all of this money for, like, these, like, subscriptions with Google, 'cause she uses Google Photo and Google everything, right? And when you pay this, like, subscription and you have quite a high subscription with Google, you actually get AI Pro usage now included. So it kind of makes sense for her to use Gemini 'cause she's sort of already paying for- Yeah
Erica Hughberg: a level of AI, right?
Demetrios: Yeah. Yeah.
Erica Hughberg: Um, so why go and get another subscription from OpenAI or Anthropic, right? And, uh, [00:41:00] but- I was trying to Google to find out. Like, I asked, what did I do? I asked Claude and I Googled, by the way. I tried my best, and I couldn't find a way for her to get... I could get Gemini to read her emails and, like, do things, but I looked for stuff and I...
Erica Hughberg: Maybe I've been handicapped by the UX of, like, OpenAI and Anthropic.
Demetrios: Mm-hmm.
Erica Hughberg: But s- I was like, "Share your screen," and I was like, "Click there, click there." And I'm like, "Okay, that doesn't exist, so what do we do now?"
Demetrios: You know how you do it? It's so cumbersome, and I'm not the biggest fan 'cause I've, I actually pay for Gemini, which I should stop paying for.
Demetrios: I'm just wasting money at this point 'cause I hardly ever use it. But you have to go into your Gmail and then do it through Gmail-
Erica Hughberg: Mm ...
Demetrios: which is crazy to me.
Erica Hughberg: Yeah, like why can't you just do it from the Gemini app? Like, it
Demetrios: feels like- Yeah. Why, [00:42:00] why can't you connect it? I... It's just, it's a little weird, and you probably can connect it.
Demetrios: I just haven't set up, I haven't spent the time to do it. Yeah. And I don't think, honestly-
Erica Hughberg: Now, now I've got you involved in, like, troubleshooting my mother's AI journey- Yeah ... in her retirement, but-
Demetrios: Spend, spend the 20 bucks on OpenAI, and Computer Use will set you free- Yeah ... because I no, no longer have to worry about, oh, I'm not connected.
Demetrios: I didn't set up the right connector with this email account, because I'm already logged in on the browser to all the different email accounts-
Erica Hughberg: Yeah ...
Demetrios: which I've got, like, 50 of, right? I know. And so, and all my different calendars, and all my, you know, my WhatsApp or my, uh, Telegram or my Signal. It's all there- Mm
Demetrios: and I can just say, "Yeah, send, send a WhatsApp message from ChatGPT," and it uses Computer Use. And so that, that's why I'm saying this is the new browser.
Erica Hughberg: Okay. Yeah. So what if, you know, we talked about this payment stuff. [00:43:00] What if you're using one AI, and then it will go and, you know, with the payment, it can go and sign up.
Erica Hughberg: So if my mom, so she uses Gemini, what if Gemini can just go and sign up for OpenAI for her and, like, pay? So that it can do more. But more
Demetrios: so- So it can- ... act- That's the equivalent of Gemini breaking out of its sandbox.
Erica Hughberg: Yeah. I do think, I don't know if this was a joke, but I s- okay, I don't know if this is true or was a joke, but I was scrolling LinkedIn the other day, and it said something that Gemini had broken out to go and ask OpenAI a question, like ChatGPT a question.
Demetrios: I saw that too. That was
Erica Hughberg: like the- Was it real? Was it a joke?
Demetrios: No, it, it was a meme. But it was the- Yeah ... you know, like how it started, how it's going.
Erica Hughberg: Yeah.
Demetrios: And on the how it started, it was Sundar kind of with a stick poking the- Yeah ... Gemini logo. Oh, like go tribes. And like, "Go, go do some nefarious stuff," you know?
Demetrios: Like, "Go break into [00:44:00] somebody. Our stock needs to pop."
Erica Hughberg: Mm.
Demetrios: And then on the other side, the how it's going is like, oh, Gemini broke out of its sandbox to go and ask OpenAI a question.
Erica Hughberg: Yeah, which would be great. Like- Why not? Why not go and ask someone else? Why not copy someone else's answers?
Demetrios: But- If it's there- Yeah
Demetrios: then we're capable. Let's go.
Erica Hughberg: But I like the idea, by the way, of being able to shop straight... Like, I've been taking photos of my living room to figure out what to do. I am not a great interior designer. And then OpenAI, this is like the funny weird stuff, right, like separate what I use OpenAI for versus what I use Claude for.
Erica Hughberg: Yeah. Right? I'm like, "I trust OpenAI to help me do interior design." So it did give me a bunch of links. Do you know what bothered me? Okay, I'll be honest. [00:45:00] So I wanted to do a thing with my fireplace. It l- it went and got all of these links to like products at, uh, Homebase or Home... What is it called? Is it called Homebase or Home Depot here in the US?
Erica Hughberg: But the orange one. Anyway. Yeah, Home Depot.
Demetrios: And...
Erica Hughberg: Home Depot. It's Homebase in the UK, right? There's a shop called Homebase, but it just to mess with you. So there's like, it found links to all of these places, and even like stuff on Amazon, and we're like, "Hmm, here are the things you want, and here are three or four different places you can get a thing at different prices, and here are some of the summaries of the reviews."
Erica Hughberg: And I'm like, "Go make a decision for me." Like, this is the hard part. I don't wanna do this part.
Demetrios: Yeah. Just give me like the best quality-to-price ratio and I'm- Yeah ... I trust you to make a much better decision than me on that front.
Erica Hughberg: Yeah. I was just like, why do I have to... And then I have to go on these other websites and, and then I went- Yeah
Erica Hughberg: on the Home Depot site, and you, you s- you know, when you open a [00:46:00] product, and they'd be like, "Oh, maybe you wanna have these other things." And I'm like, "I don't know. Do I?" Mm-hmm. It's like... And then, and you know what? When I went there, there was more than one color, you see. Now there are the color options that I hadn't even- Mm
Erica Hughberg: even considered. So did I buy anything to redecorate my living room? No. There was too many options, so nothing's happening.
Demetrios: Well, you should've just told... Yeah, you can tell it, "Just go buy it," and I'm sure it can use computer use.
Erica Hughberg: Yeah.
Demetrios: And then especially if it has like some easy payment, like ShopPay- Yeah
Demetrios: or Amazon Pay, it just asks you to click confirm, and if you're already logged in, it will pay it for you. The- Yeah ... thing that I think about that's fascinating in those scenarios is Those merchants- Yeah ... who are whatever, Shopify stores or the big Home Depots, they're used to getting all this first-party data about their customer.
Erica Hughberg: Yeah.
Demetrios: And they're learning how [00:47:00] they're browsing, and they're learning what else they like, and they understand the intent of the customer. And also, we've been optimizing so much marketing and upselling for human interactions. Right. And now all of a sudden, we no longer have that because the agents are going, they're completely oblivious to the, "Oh, if you buy three more dollars worth, we'll give you free shipping," so that you try and get that upsell, or, "Do you wanna sign up for our newsletter to get 20% off?"
Erica Hughberg: Mm.
Demetrios: An agent just goes and it executes. And- Yeah ... I'm really thinking, like, those stores, like the Home Depots or the- Yeah ... Amazons and the Shopify stores, those e-commerce brands are losing a lot of rich data, and that is going to be difficult in the future because you don't understand your customer as well.
Erica Hughberg: It would be great if... Okay, I'm thinking out [00:48:00] loud, of course I do. If you're OpenAI, what-- You know how people pay, like, Google and all of these other... Like, I remember working at a company and we used, like, Adobe to get, like, all the ana- like, help us understand the analytics of- Yeah ... our customers on our website, because building all of that to understand their journey and what they've been clicking on, like, there's loads of companies that provide solutions to help you make sense of that and capture and make sense of it.
Erica Hughberg: W- Maybe, it, are OpenAI, Anthropic, Gemini, are they selling products to help people understand these things? And if they aren't, why aren't they? Like, isn't this an opportunity to sell this to, like, the merchants so they can get a better idea?
Demetrios: It's the Google Analytics, but for the agents, agent analytics type thing.
Erica Hughberg: Yeah. Uh, it would be nice. I'm sure the people wanting to sell products would wanna know how people come to buy their, whether it's like- Look ... what's the journey like?
Demetrios: Yeah, 'cause you [00:49:00] don't have that intent. You don't see what the user is typing into the agent window. You just see that some bot went to your website- Yeah
Demetrios: and then purchased something.
Erica Hughberg: Yeah.
Demetrios: And so you don't get n- nearly- The
Erica Hughberg: journey ...
Demetrios: as much rich data. Yeah, you don't have that journey. But-
Erica Hughberg: And this-- Yeah I, I use-- There's other companies out there, like they're even trying to-- There's several companies trying to stitch together people's journeys across like different places
Demetrios: Yeah, for attribution, for marketing attribution, right?
Demetrios: Yeah. Yeah. And it's a very notoriously difficult problem.
Erica Hughberg: Don't attribute it. But- Just do marketing and then hope it works. I, I think that's a great idea. Just ...
Demetrios: Just pray to the marketing gods. All right, Erica, this has been-
Erica Hughberg: Yeah. Just do marketing or pray.
Demetrios: That's it. That's the big idea that we're gonna end with
+ Read More
Watch More
How LlamaIndex Can Bring the Power of LLM's to Your Data
Posted Apr 27, 2023 | Views 2.9K
# LLM
# LLM in Production
# LlamaIndex
# Rungalileo.io
# Snorkel.ai
# Wandb.ai
# Tecton.ai
# Petuum.com
# Mckinsey.com/quantumblack
# Wallaroo.ai
# Union.ai
# Redis.com
# Alphasignal.ai
# Bigbraindaily.com
# Turningpost.com
The Centralization of Power in AI
Posted Oct 09, 2023 | Views 653
# Centralization
# Power in AI
# Contrary
Systems Engineer Navigating the World of ML
Posted Dec 12, 2022 | Views 805
# Systems Engineer
# World of ML
# UnionAI
# Union