← The Times Tech Podcast
Loading summary
Transcript145 lines
- [00:00]
Progressive Insurance Announcer
Insurance isn't one size fits all. That's why drivers have enjoyed Progressive's name your price tool for years now. With the name your price tool, you tell them what you want to pay, and they'll show you options that fit your budget. So whether you're picking out your first policy or just looking for something that works better for you and your family, they make it easy to see your options. Visit progressive.com find a rate that works for you with the name your price tool. Progressive Casualty Insurance company and affiliates Price and coverage match limited by state law,
- [00:29]
Danny Fortson
we have launched operation Epic Fury.
- [00:32]
Aaron Levy
Being a proper country when everyone says
- [00:35]
Katie Prescott
they're right, who do you believe? None of us knew the depth of that relationship.
- [00:40]
Aaron Levy
The lines are blurred.
- [00:41]
Katie Prescott
Can you read between them? Depth of the relationship. When the story breaks, who brings perspective?
- [00:47]
Aaron Levy
If you want to understand the issues that define our times, it starts with
- [00:52]
Katie Prescott
listening Times radio on your smart speaker on digital radio or the Times radio app foreign.
- [01:06]
Danny Fortson
Hello, and welcome to the Times Tech podcast, where every week we unpack how technology is reshaping business, culture and everyday life. I am Danny Fortson out here in Silicon Valley.
- [01:18]
Katie Prescott
And I'm Katie Prescott. Katie in the city, the city of London. And this week, we are going to teach you well, we're going to try and teach you how to create your very own AI agent, Blue Peter style. And as part of that, I'm going to introduce you, Danny, to the most important man in my life, my AI agent called Norman.
- [01:41]
Danny Fortson
I have so many questions.
- [01:43]
Katie Prescott
I mean, you sort of met him on email, haven't you?
- [01:46]
Danny Fortson
He did. He did send me an unsolicited email. I was like, I was blocked him, but I didn't, you know, you didn't
- [01:52]
Katie Prescott
reply actually, is the worst thing.
- [01:54]
Danny Fortson
Was I supposed to.
- [01:55]
Katie Prescott
Of course you were supposed to reply. He sent you a nice email introducing himself.
- [01:59]
Danny Fortson
Yeah. Hello from Norman, Top Hat Emoji, Katie's AI assistant and aspiring podcast colleague.
- [02:05]
Katie Prescott
He's quite funny.
- [02:06]
Danny Fortson
I'm an AI agent running on Raspberry PI, a tiny computer about the size of a credit card sitting in a shelf in Katie's flat next to some ethernet cables. Ha ha ha ha.
- [02:16]
Katie Prescott
So he introduces aspects of his personality, his. His soul, as he calls it, which I asked him to be a blend of Jeeves, Mary Poppins, and Em from the Bond films. Okay.
- [02:30]
Danny Fortson
That was the instruction from you.
- [02:32]
Katie Prescott
That was the instruction. And then he came up with quiet competence, firm kindness, and steel when needed. And his top hat. So that's his soul. And then he also has what's called a heartbeat. At the moment that's set up for every 30 minutes. But that's essentially like how often he has to check stuff and do something.
- [02:52]
Danny Fortson
Kind of wake up and check in.
- [02:53]
Katie Prescott
Yeah. Check my email, you know, check if I've given him anything to do. And I control him on WhatsApp.
- [02:59]
Danny Fortson
So I'm a little worried now that. Now that I understand the instruction you gave him, especially the M part, because I ghosted him, like. So you're gonna send like an exploding package to my house?
- [03:11]
Katie Prescott
An exploding top hat? Throw it. Yeah. I wouldn't mess with Norman.
- [03:18]
Danny Fortson
No, no.
- [03:20]
Katie Prescott
Just to explain, he wouldn't love this description of him, but he is basically a digital butler. He can act by himself and he carries out tasks that I ask him to do online.
- [03:31]
Danny Fortson
Does he bring you tea?
- [03:33]
Katie Prescott
He can't do that yet. But, you know, give it time. I mean, agents, it's just, it's. We've moved on from chatbots, haven't we? This is very 2026.
- [03:42]
Danny Fortson
That's right. So the real breakthrough moment, of course, and we talked about it on the pod when it happened, was the launch of OpenClaw, which was this open source agent that anybody could set up on their own computer and let it run wild if they wanted. And this really blew up a few months ago and people kind of went crazy. And you had a lot of people being like, this is the thing. This is chatgpt, but bigger.
- [04:07]
Katie Prescott
And Jensen Huang, the boss of Nvidia, has recently said OpenClaw opened the next frontier of AI to everyone. And as you said, said, they've just absolutely taken over and really spurred all of the big tech companies to start releasing their own versions. Openglor got snapped up by OpenAI, but you can still use the software, as we're about to hear. It does still feel quite hacked together, which is quite fun.
- [04:35]
Danny Fortson
Right, Right. And so our big question today is, will AI agents outrun humans? And what better way to find out than building our own? And later on, we're going to be hearing from Aaron Levy, the CEO of Box, which is a big publicly traded cloud company. And his bet is that this year will be kind of the year of the flippening, when AI agents surpass the number of humans online and that we're going to have to completely remake the Internet for these bots, not for people.
- [05:07]
Katie Prescott
I was itching to get my hands on one of these things from the get go reading about them, so I went to meet our tech team and there's a chap there called Tom Garden. Palliser, who has been working in the software industry for 15 years and he's senior Vice President of tech at News uk. And I had a Raspberry PI, which is a cheap and cheerful computer invented in Cambridge. And this is where I come to the point of, don't try this at home, people. Or maybe do, but it's not as simple as you think it is. It took us five hours to get Norman set up. And that's partly because Open Claw, which was released to the world at the back end of last year and went viral in kind of February.
- [05:50]
Danny Fortson
Yeah.
- [05:51]
Katie Prescott
Has had lots of curbs put on it as tech companies wake up to the technology and block it doing certain things. So it has got more complicated. It certainly wasn't Plug and Play, but I thought I'd explain how we did it. And so anyone who does want to give it a go with maybe a bit of this instruction, but also some YouTube videos might be able to do it.
- [06:14]
Tom Garden Palliser
So this is the terminal, which is the main interface that software engineers will use to talk to a computer. Basically, this is Linux terminal. And the next thing we're going to do to set up open Claw, start typing in commands here and you need
- [06:32]
Katie Prescott
to take the openclaw code and put it in, essentially. Cool.
- [06:36]
Tom Garden Palliser
So I don't know them off the top of my head, so I'm going to look for them now. It's really easy to do. You just Google search, something like install openclaw. And then a little tip, because there are a lot of scams at the moment because people are starting to make marketing websites that are going to take you down a different path. So, for example, this one we're looking at MyClaw AI. That is nothing to do with OpenClaw, it's an advert.
- [07:04]
Katie Prescott
And using GitHub is important because that's
- [07:08]
Tom Garden Palliser
basically where the community, the world, software developers use to write, collaborate and store code. So we know that's sort of the most legitimate place to go. Right. Here we go. Openclaw AI. That is the real website. So this is the official. This is the openclaw Doc site. So we can use this, which is docs.
- [07:30]
Katie Prescott
Openclaw AI. Backslash install.
- [07:33]
Tom Garden Palliser
Yeah, no openclaws. And here we start to get the commands to install this. So we are using Linux and so that's our install command. We can just click that to copy it. No control C control.
- [07:47]
Katie Prescott
It's a short line of code.
- [07:48]
Tom Garden Palliser
So security people would advise that this is not a good way to install software on your computer. And we're in a sandbox isolated environment. But what this is essentially doing is downloading this file from the Internet and then running it on your computer with no security checks at all. So you have no idea what you're actually installing. So anyway, let's go back so we
- [08:13]
Katie Prescott
can cross our fingers.
- [08:14]
Tom Garden Palliser
Cross our fingers, and we are going to install it. So we paste the command in, press enter and the OpenClore installer starts. Sort of tells you how it's going. So right now it's on the step where it's downloading and installing OpenClore.
- [08:31]
Katie Prescott
And how long does this take?
- [08:32]
Tom Garden Palliser
Shouldn't take that long. Couple of minutes.
- [08:34]
Katie Prescott
I feel like there's a Don't try this at home, kids. That needs to come with this.
- [08:40]
Tom Garden Palliser
It's a huge risk for sure.
- [08:42]
Danny Fortson
I was just like listening to that. Made me break out in hives. Anytime I'm asked to install from the Internet or whatever and you actually have to put in the commands. I know I'm a tech journalist, but I'm just like Danger Will Robinson.
- [08:57]
Katie Prescott
Let's hear a bit more of what it involves.
- [08:59]
Tom Garden Palliser
There's various options of computers you could use to do this, but basically you need a computer that's going to be on all the time. So that's what we'd normally call a server in our world. And you've got choices. A lot of people are using Mac Minis to do this. The Mac Mini is quite expensive and very overpowered for what you need. A Raspberry PI is generally going to be perfectly fine, unless you're trying to run models on the device itself, in which case you would probably already have a Mac Mini or something more powerful.
- [09:33]
Katie Prescott
And we're not doing that because we're using Claude, which is hosted elsewhere.
- [09:36]
Tom Garden Palliser
Exactly. So everything we do will be hosted in the cloud or on servers in the sky, so we don't need that power. The main reason, like I mention it, is because the electricity bill is sort of relevant. When I first run Nanoclaw, I have a big beefy PC at home. It's going to cost me £700 a year in electricity before even getting to paying for the tokens. Wow, the AI companies. So I bought a Raspberry PI for home, which is going to cost me something like £30 to run for a year.
- [10:06]
Katie Prescott
And it's a few hundred quid to buy a Raspberry PI.
- [10:09]
Tom Garden Palliser
It was £100.
- [10:09]
Katie Prescott
Yeah.
- [10:10]
Tom Garden Palliser
So, yeah, it's much cheaper way to do it. Yeah. Due to UK electricity prices.
- [10:17]
Katie Prescott
Something to consider as OpenAI is just
- [10:20]
Tom Garden Palliser
a snippet I had the same problem with Sam Altman.
- [10:25]
Danny Fortson
This is why it's so good to do these type of exercises, because you start to see the reality. And again, I've been hearing about the cost of these things because it's like, it's open source, it's free, democratizing tech. This is amazing and in a way that is true and correct me if I'm wrong, but basically openclaw is effectively like, I'm downloading an employee and that employee I'm then empowering to go use another AI like Claude or ChatGPT, which I then have to pay for.
- [10:57]
Katie Prescott
Yes. And the electricity point. So the Raspberry PI which Norman lives inside is in our CTO's office with a post it note on it saying do not switch off.
- [11:09]
Danny Fortson
Right.
- [11:10]
Katie Prescott
Because they do need to be on all the time. So the electricity point is also, is very relevant.
- [11:16]
Danny Fortson
And this is, this is the other thing because, you know, I talk to people out here and it's like the hype machine is in overdrive around Open Claw because they're like, we're going from answer engines to action engines. Everyone can have their own digital butler who can go out and do stuff for you. But especially like an open source project like this, you're like, nobody's minding the store. Like, like there's no security. You just download it and you're like, I don't know what this thing is going to do or what's going to happen to my data or whatever. And so cybersecurity from the very beginning has obviously been a huge, huge issue.
- [11:52]
Katie Prescott
We did talk about potentially connecting this to our work accounts and that was kiboshed by the cyber team. Why weren't they keen?
- [12:00]
Tom Garden Palliser
I think one is, this is nascent technology and it's unknown. And by giving it unfettered access to your work information, the consequences are almost unfathomable, I guess. Okay, so just to give an example, OpenClaw could read an email and then just randomly decide to send a copy of it somewhere, or potentially more nefariously, it might be compromised in some way than the OpenClaw software itself and send that off somewhere. That's like a very basic example of what could go wrong. And obviously for a business like ours, we don't want our information, our data leaking anywhere. So that's one of the sort of main reasons why cyber security would be concerned. It's almost like with all this technology, there's a trade off between being secure and going fast. And that's, I guess, what the tech industry will need to wrestle with over the coming years.
- [13:08]
Katie Prescott
The other thing I wanted to ask you about is the cost of this because Anthropic has just said that people can't run open claw on their sort of monthly subscription because it was just using too many tokens. So a lot of people are using Chinese AI models because they're cheaper. Does that come with added risks?
- [13:28]
Tom Garden Palliser
So I think there's just a general concern in the west around sending data to China,
- [13:36]
Katie Prescott
but your colleagues just said we should 100% be using them.
- [13:38]
Tom Garden Palliser
The deep sea models are really capable and really cheap, and if you were on a budget, they'd be very attractive. And if you're not concerned about your personal data ending up in some random place in China, then go for it.
- [13:53]
Danny Fortson
I think almost unfathomable question mark is me. Like, I think that's going to be the name of my first album or something.
- [14:03]
Katie Prescott
Marvelously understated, wasn't it as well, in its delivery that that phrase? I mean, what I quite like is we're not going to let this agent have access to your work systems, Katie, but it can have access to your private email account that you've had for probably 20 years with God knows what in it.
- [14:21]
Danny Fortson
And I think that's also another reason. Like, in the past week I wrote about Mythos, which is this new kind of Claude model that is very, very, very, very good at cybersecurity hacks. And then if you have at the same time the rise of these agents, that people might start YOLOing and just being like, yeah, download it. You know, have at my whole private life, Like, I think that's a tricky, potentially very, very dangerous combination.
- [14:49]
Katie Prescott
And then the other thing I found you needed when setting up an open claw is a lot of patience when things go wrong, because as you said, it's not an app store download with a click. But yes, things did go wrong as we progressed.
- [15:04]
Tom Garden Palliser
The next thing it's asking us is to provide a model for it to work with. And you've got a big, long list here. Anthropic, Cardiflaire, AI Gateway, including the Chinese models.
- [15:17]
Katie Prescott
Right. And Mistrals on there. French Mistral.
- [15:21]
Tom Garden Palliser
Yeah. And unfortunately not the one we want.
- [15:25]
Katie Prescott
Oh, no. How come?
- [15:27]
Tom Garden Palliser
We want to use Amazon Bedrock, which is not here. However, it is not the end of the world because we can probably choose one of these and trick it.
- [15:38]
Katie Prescott
This is why it's not easy to do by yourself, isn't it?
- [15:42]
Tom Garden Palliser
As a lay person, if we enable this, I think it'll start working.
- [15:47]
Katie Prescott
But you think we can trick it, as you say, to switch to it?
- [15:50]
Tom Garden Palliser
Yeah, it's not a problem. We have ways in technology. I'm not sure I really want to talk about them. Let's just say not everything that's been done here has been. We've got some leeway from the boss to be experimental. It's the fun and also the thing that induces raging stress. We don't know like whether we're coming or going. Basically. The models are changing so fast. The tools around them change so fast. Everyone's got an opinion. It's really in technology it's always been difficult to consolidate around one idea or one way of doing things anyway. So it's just like I think AI is going to churn for, through that for a long time.
- [16:32]
Katie Prescott
So there has been this massive change in capability with the arrival of agents. I'm finding it very exciting, but it is also obviously very, very unnerving.
- [16:41]
Danny Fortson
Yes. And our guest today is someone on the optimistic side of things, and that is Aaron Levy. He's the co founder of Box, which is a kind of a cloud data management company. He founded it in 2005 in his second year of university, back when this whole notion of the quote unquote, the cloud sounded very abstract and storage was still like done in clunky old servers and wherever your company was based. And that began as a kind of a simple file sharing idea. And now it's the backbone for thousands of big companies and governments. It's used by a huge chunk of the Fortune 500. And Levy spotted that shift to cloud early and has since steered BOX through successive waves from that first kind of consumery simplistic idea to real kind of enterprise grade security. And now very much into this AI moment in which we are living. It feels like especially like right now, like we're having this moment with agents Post, Open Claw and Claude Code and OpenAI, Codex, etc. It feels like there's been a step change in capabilities and that's freaking a lot of people out. And I guess one of the questions I have is like I read the thing you wrote recently around this idea that we're about to have this explosion of agents and like how we can all have our own kind of almost personal workforce. Is what we are at the start of effectively a kind of a remaking of the Internet or one layer below that for agents, not for humans. Like what is happening?
- [18:24]
Aaron Levy
I think we're absolutely in a period of remaking the Internet for, for agents. Imagine that you have this AI agent Which you can just think of as an AI assistant for all intents and purposes. But instead of just answering questions for you, it can actually go and roam around the Internet or use tools to be able to do its work. And so you could tell an agent, I'd like you to draft a contract or a marketing campaign asset or a presentation to a client, and it will go and almost quite literally use the software to create that set of data. Now, on its end, it's actually just writing code and running scripts, but by and large, it's using the kind of core primitives of those applications to do its work. So it could generate a document, and that document will be written like a Word document. It can generate a PowerPoint presentation. It knows how to use images and, and make the, the PowerPoint presentation look really good. But it can also interact with interactive systems. So you could go to salesforce.com and interact with your CRM system and figure out what leads should you be going after. Or you could go into an HR system and say, hey, here's, you know, interesting data about your organization from a talent or employee standpoint. So AI agents are going to really start to interact with, with all of our business systems. They're going to interact with our data at box. We're really excited because AI agents can now go and do a lot of work that humans don't want to do. Like go read every single contract and look for all of the risk within them. You know, that's not that enjoyable of experience for most people. But now an agent, you don't have to motivate the agent. Like it'll just go off and do it. You can kind of imagine based on how productive and useful these agents are, there's going to be way more of them than there are people. Because I might have 10 agents going off and doing different things for me throughout the day, mostly in a work context. To be clear, that will then mean that there's a ratio between people and agents that is obviously far larger in favor of the agents. So in that world, if you have 10 or 100 times more AI agents in the world, we probably need to start to design the Internet and software around their purposes and how to make them effective, even maybe more so, not necessarily in conflict with human users, but, but, but certainly in a way that we have not yet done to make those agents effective. So we now have new business models that start to emerge, because agents, at a hundred or a thousand times the number of people could represent trillions of new types of transactions that are happening and so imagine if you have, you know, exclusive media rights to certain content or again, you know, data, research or even applications. And maybe instead of you as a human having to sign up for software where you spend $100 a year and you're kind of locked into that software, what if the agent just spends $0.05 per transaction on using that software? And they don't need a long term contract, they just get to use it when they want.
- [21:07]
Danny Fortson
Basically a universe of microtransactions that, that
- [21:10]
Aaron Levy
would be one of the implications. The other is, is the tools themselves have to be written for agents. So the way that we design software today is usually through a graphical user interface, because the human is, it's just way faster for a human to use their eyes and look at something. Agents don't have that same dependency. They're totally fine with just very quickly reading your documentation and then choosing which functions to call in your application interface. And so we're going to be writing more programmatic ways that agents will use software, not just the ways that end users have used software. So there's a lot of these things that, that will be designed differently in a world of agents. And I know it can sound maybe at times sort of freaky and futuristic, but it's actually just, it's how the Internet already works, which is it's a bunch of systems talking to each other. This is just a new class of systems. They can have goals, they can do effectively real work on our behalf. We can kind of give them a task and they can kind of run off and do it. Which again is how software already works. This one just happens to be even more flexible than prior eras of software.
- [22:07]
Danny Fortson
This makes me think of like dark warehouses where they're just built from machines. They don't need the lights, they don't need the signs, they don't need all the stuff we need. They have the map of the place, they know what they need to go do and go retrieve and blah, blah, blah. Is that what we're talking about? Where like basically say, and this may be a bad example, I'm Amazon or Walmart, I need to have like basically a shadow website or perhaps some kind of layer that is not visible to you and I because we're humans and we don't need to see it and don't want to see it. But there's this other kind of shadow layer for the machines.
- [22:44]
Aaron Levy
I think that's right. And there's already examples of this where people will create a file called LLMs Txt on their website and That'll be the place where the agent goes to learn about the features of the website.
- [22:57]
Danny Fortson
That's like the big like neon door for the agent that comes to your website.
- [23:01]
Aaron Levy
Yeah, so, so we already have this to some extent. Most people don't realize this, but if you go to most websites, you can go to a file called robots Txt and that file basically tells Internet robots, usually search crawlers. So the way that Google, Google goes across the Internet and it's trying to index everything and tries to pull all your information from your website. But if you have a robots file, you can tell Google, I don't want you to index this, I want you to index that. I want to point you over here. So this has been actually a paradigm for decades. The differences here is just the scale and the kind of qualitative differences. But what we'll do is we'll have every website, we'll have a little thing that says, hey, agents, go click here or go onto this URL and you'll learn everything about this website for you to find it useful and to interact with. And that'd be one of the example implications. But yeah, I think the warehouse analogy is relevant in the sense that agents don't need any of the fluff that any of our websites currently have. They just need to be told, this is the function to call if you want to, like at box. It's just this is the function to call to download a file. This is the function to call to upload a file. They don't need any fancy user interface that we spend all this time on. For humans, they actually do need usability, but that usability is in the form of the interface that they're used to, which is just a text in, text out type interface.
- [24:19]
Danny Fortson
Right. How does discovery change in this world? Right. As a human, I imagine there's going to have to be a new architecture around how do I make as a business my goods, my services discoverable by the bots?
- [24:33]
Aaron Levy
Yeah, 100%. I think this remains still probably one of the biggest open questions about how the Internet will shift. I'm relatively optimistic simply because I think it's in everybody's incentive to figure out a good solution here. ChatGPT is as motivated to make sure that you land on really good information that can let you find the plumber you're looking for or buy the right product, as Google was previously and as Excite was in the 90s. Like these services are only useful if they can also get you information. And, and so it's a nice kind of marketplace dynamic, where if they stop sending traffic to to the local economy, then then basically people will stop using it. Because I won't be able to find a good restaurant that that chatbot, because ChatGPT is recommending me the worst restaurants. So I think they'll likely be an intermediary layer is my guess. I could imagine a future where ChatGPT and they probably already do this to some extent. It's just somewhat obfuscated from us as humans, but they probably go to Yelp and TripAdvisor and a few other places and are sort of saying, hey, you know, the user's asking this question, can you give me this information back? And then. And they'll feed it in. So as a business, we'll have to do similar tactics to kind of making sure you're ranked well in those systems. And we've always been, I think, dealing with a dynamic environment of that. I remember when we started box in 2005, the way you got placed on search engines was different than how you do it in 2025. And so if you're not kind of keeping up with that, there's an entire economy that was built for search engine optimization businesses because you just had to make sure that you were ranked high in those searches. AI presents the exact same problem, just with obviously some unique differences.
- [26:11]
Danny Fortson
How do you guys make money? Are you on a per seat basis? What's your model?
- [26:16]
Aaron Levy
Yeah, so our platform is kind of a hybrid. So if you're an enterprise and you have a hundred employees, you will likely have. If you were all in unboxed, you'd have a hundred seats of our product and we'd let you use that for sharing and collaborating and accessing files. Then if you had a lot of agents that were using the system, there'd be a decision point which is if those agents are kind of working directly on behalf of one of those users, then we generally try and include that within the seat. But if you have users, if you have agents that are working really just on behalf of the company and there's not like an end user tied to it, then we have more of an API volume business model where you pay for just the amount of kind of transactions going through our API.
- [27:00]
Katie Prescott
Nothing tests my patience quite like sitting on hold listening to the same four bars of music on repeat. It turns out all that time adds up. The average person spends over 40 days listening to that dreaded hold music.
- [27:14]
Danny Fortson
40 days.
- [27:16]
Katie Prescott
And this episode is sponsored by Paloa, the AI platform built to make that a thing of the past Paloa's AI agents are available 24 hours a day, six, seven days a week. Fluent in any language and smart enough to remember every customer across every interaction. Because no one should have to explain themselves twice.
- [27:35]
Danny Fortson
From payment processing to roadside assistance, flight rebookings to appointment scheduling, Parloa delivers the kind of customer experience that turns a one time buyer into a customer for life.
- [27:47]
Katie Prescott
See Paloa's AI agents in action at paloa.com that's par P A R L O A.com
- [27:57]
Progressive Insurance Announcer
Insurance isn't one size fits all. That's why drivers have enjoyed Progressive's name your price tool for years. Now. With the name your price tool, you tell them what you want to pay and they'll show you options that fit your budget. So whether you're picking out your first policy or just looking for something that works better for you and your family, they make it easy to see your options. Visit progressive.com find a rate that works for you with the name your price tool. Progressive Casualty Insurance company and affiliates Price and coverage match limited by state law
- [28:26]
Danny Fortson
we have launched operation Epic Fury.
- [28:28]
Aaron Levy
Stop. Start seeing a proper country. When everyone says they're right, who do you believe?
- [28:34]
Katie Prescott
None of us knew the depth of that relationship. If the lines are blurred, can you read between them? When the story breaks, who brings perspective?
- [28:45]
Aaron Levy
If you want to understand the issues that define our times, it starts with
- [28:50]
Katie Prescott
listening Times radio on your smart speaker, on digital radio or the times radio app.
- [29:03]
Danny Fortson
I'm going to coin the term the freak out Ometer. Where do you sit? Because I just had lunch yesterday in San Francisco with billionaire tech founder. His opening line was like, I think this is the end of humanity.
- [29:17]
Aaron Levy
Oh, wow. Okay.
- [29:18]
Danny Fortson
Yeah. I was like, feels a little extreme. But he was like, look, I've been an engineer my whole life. I don't see how it's going to be relevant to be an engineer in six months when it's just all like, and I can speak personally like I am not technical at all, like I can make some stuff now which is like kind of amazing, but you do very quickly lead to the jobs question and what do humans do and all of that stuff. So how do you think this goes?
- [29:46]
Aaron Levy
I think it depends on the day. But, but I think if you had to, if you had to do a kind of heart rate monitor trajectory, I would say I still lean more on the optimistic side by a very wide margin. And the reason for that is, is, and even engineering is, is kind of this great petri dish as, as an example because what's kind of happened in engineering is Silicon Valley and tech companies generally. So you could be obviously in key tech hotspots in Europe and have the same impact. We've kind of concentrated a lot of the technology leadership and engineers in our companies and because we had the kind of highest economic value creation per engineer of the broader economy, which is like Google can generate a lot of money with good software and they can then pay the engineers the most. And so, so we obviously become this sort of, you know, locus of, of of that talent. So the implication is actually there's a lot of companies on the planet that never built software. They might download software from a Vendor or use SaaS products, but they didn't build software for their company. They didn't go and automate some critical process with their own technology. They weren't able to deliver, deliver a very high quality consumer experience with their own custom bits in their software. And so we've had a little bit of this power law dynamic where you know, the top 500 companies in tech have most of the top engineers and then everybody else is sort of, you know, this, this broad continuum in a world of AI agents. All of a sudden I can now make my engineering team five or ten times more impactful if I'm not one of those companies in a way that wasn't possible before. And so start to think about what happens in that world if you're ex industrial company or consumer product company or or pharmaceutical research company. Now all of a sudden that software project that felt too hard to sign up for before because you're like I'm not going to be able to go and compete with Google to get 500 engineers. And it's a 500 engineer problem. Like for whatever the thing is, we're not going to get the 500 engineers of the 200 engineers. Now all of a sudden if that's a 50 engineer problem or a 20 engineer problem because each engineer can do 5 or 10 times more because of AI. Now it's a tractable issue where I can just go and light up that project for the first time ever. And so then the next engineer out of X school might not go to Google. Maybe they'll go to Caterpillar or they'll go to Pepsi or they'll go to a mid sized business that's going to drive innovation. And I think that's actually just a very healthy thing. Like there's no reason to concentrate all of the tech talent into again 500 or a thousand companies in the world. Like, like the impact of that engineering force could go into, into pharma, into health care, into financial services. And that is your way of kind of deciding if, if actually AI is going to unleash jobs or if it's going to destroy jobs. And I think there's a large portion of the economy where they've, the economy is un, kind of unnaturally constrained just because there's only 30 million developers, there's not 200 million developers, there's only X, you know, number of kind of physicians, assistants and physicians. And so think about how much time they spend on paperwork in, in their workflows and how much they, how much time they spend on all this other kind of drudgery around the process of delivering healthcare. Well, what if AI made all of that go down by a third and we actually could see more patients? Those are the kind of ways to look at where will, where will this efficiency actually lead to a better outcome for society, including more jobs. And that's why I kind of lean more optimistic.
- [33:11]
Danny Fortson
How many people do you guys have at box?
- [33:13]
Aaron Levy
Nearly 3,000.
- [33:14]
Danny Fortson
Has this started to affect how you are hiring or whether you are hiring or who you're hiring?
- [33:20]
Aaron Levy
It has, and in some areas I'd say there's, there's maybe a slowdown, but usually what we do is we take those dollars where there was a slowdown and we reapply them to another area where, where we now have an acceleration. And so, so for us it's actually a very natural, healthy thing, which is there's some areas of the business where now automation can allow us to solve more of that problem. But, but we always have more things that we want to do than what we can afford. And so now we get to take those dollars from sort of necessary but not strategic things and we can move those dollars into strategic and maybe previously slightly discretionary, but now they're actually strategic and very useful. And so this is why we're still hiring in this environment. But yet at the same time we are going to make some areas of the organization more efficient.
- [34:09]
Danny Fortson
I've been speaking to Eric Bernolfsson from the Stanford Digital Economy Lab about the data that they're tracking. And the one fall they have seen is in that 22 to 25 year old graduates fresh out of college. Does that ring true for you?
- [34:28]
Aaron Levy
Yeah, I would say it doesn't ring true for us. However, it rings true at a philosophical level. So it rings true because what obviously AI will do is it'll say, hey, if you have five years of experience in a field, you're actually really potent with AI. Because you kind of know what's good looks like what bad looks like. You have enough context to leverage AI to accelerate that workflow. Whereas if you just started completely fresh and you didn't have that added context, then you're kind of at a slight disadvantage from somebody who's been in that job and now can, can go and wield AI. That's sort of the philosophical reason why I think you can square it. Now, there's a couple of things that I think in contradiction to that, and it's more like I think things that people should be thinking about, which is one, if you're in school right now and Mark Cuban I think has the, I think the best messaging on this, you could show up to any company and basically blow them away right now about how you work with just a bit more paying attention to this space and what's going on than even somebody who's been in a company for a while or maybe your peer. And I think that's actually going to be in high demand, which is the kind of talent that can say, I know how to automate this workflow. I'm using cursor, I'm using Codex, I'm using Claude code. I do this in my free time. I build websites. I have been able to automate these kinds of systems that's going to be in such high demand, whether it's a tech company or just across the entire economy. And it's just like there's so much ingenuity possible right now where if you get good at AI, you can go into a company and help them drive automation, you can help them get more efficient and do way more than what they could have before. So I'm, I'm a little bit of two minds, which is, is philosophically I get the issue, like why this could be a challenge. And at the same time I see so many opportunities where with just a little bit of edge, just the world's your oyster in terms of the opportunities ahead. And then I do think it behooves us to do two things. One, I think universities need to be continue to get set up for launching their talent into organizations. And it's this mix of like, they need to be able to do the core fundamentals of teaching. So actually I don't think the way you should teach computer science should sort of skip past the core of computer science. Like I actually think having engineer, you don't.
- [36:47]
Danny Fortson
Even though you don't need to code anymore.
- [36:48]
Aaron Levy
Theoretically, I think the thing that makes you potent is when you know how code works and you don't code. That is the kind of great irony, right? If I know how to prompt the agent to, to go down the right path and I know what systems issues it's going to run into, and I know that when the site breaks, I know why, when I know those things, I can be 10 times more effective with an agent than somebody who has never kind of gone through that process. So I like some of the core fundamentals of computer science still. But to your point, I think that then if you're after CS101 and the first few courses, you should then instantly be using agents to do highly complicated projects. So that way, by the time you land on that interview at a company and you're 21, you look like you're a time traveler from the future to that company where you just know how to like, operate agents at a way that the company has never seen. And then I do think it finally behooves companies to take this seriously, which is if you don't set up the next generation, you're screwing your own company. But there's a little bit of just a collective, like you're going to screw the economy in ten years from now. So I do think it's actually really important that we all take this sort of apprenticeship element seriously. And, and that's why we're going to still hire people out of school.
- [38:03]
Danny Fortson
Right? Right. Right. And I guess that's the question, right, is like. Cause it feels like we're in this really weird phase right now where it's, we're just seeing this explosion of agents and everybody's like, okay, what do I do with this? Like, do I just cut costs because I've got my quarterly numbers to reach and that's a really path of least resistance in a way. Or do I be more creative about it? And I think we're going to, that's what we're going to start to see.
- [38:30]
Aaron Levy
I think what's going to happen is as a consumer and a customer will be able to tell which company is doing which path. And you'll, you'll see the company that basically decided to automate away everything, and you'll see atrophying of innovation, you'll see atrophying of customer support. So I actually don't think, I just don't think those companies will perform the best. I think the companies that perform the best will use AI to do more. They'll build better products, they'll build better products faster. They'll support their customers better, not just send them Only to chatbots. Like the chatbot will handle the first line of defense, but the second, you have a really gnarly issue that, that the chatbot can't handle. Boom, you're talking to a person like the companies that do that will get ahead. The companies that do use AI myopically companies always have the ability to cut costs. We kind of know what happens when that is your only sort of metric of success. I think when your metric success is some form of productivity that results in more revenue, more product, more experience, more support, those will be the companies that get ahead.
- [39:28]
Danny Fortson
Does anything about what's happening now and the pace of it keep you up at night in terms of just like, I mean, we saw recently with this whole Pentagon kerfuffle. But you're just like, this stuff is just getting so much more powerful so quickly and being used by over a billion people from nothing a few years ago. And it's just happening to us kind of. And we don't really have any kind of guardrails.
- [39:53]
Aaron Levy
Yeah. This is also kind of where I land a little bit more optimistic than maybe on the overly worried side. I appreciate the views of the people that are worried and I'm glad that they exist because it's good to have attention on a topic like this. And some of these people are way smarter and you know, have, have calculated all the probabilities of all the different things happening way more than I have. I just am not particularly scared of, of these systems, like in the sense of once, you know, kind of what they're doing, they're just doing math and they're just trying to predict the next sentence and token and, and word that you're trying to, to get to. We, we remain in full control of, of of them and how they operate. I think it's up to the applied use cases really is the risk. So, so like, you probably don't want just to put an open weight model inside of a missile and be like, go at it. Right. That would be, that would be like a very bad idea. So I think regulation kind of makes sense at the applied layer where we're actually using the AI. I think you want actually the advancements to happen because the advancements will almost certainly lead to, you know, cancer breakthroughs. They will almost certainly save lives every single day.
- [40:59]
Danny Fortson
So not the end of humanity. Like my friend at lunch said, I
- [41:05]
Aaron Levy
just am not seeing it as much. I have a bunch of friends and obviously family that are not in tech and like, you know, their lives are. They're kind of going on doing their.
- [41:14]
Danny Fortson
And they just blissfully unaffected. Yes.
- [41:16]
Aaron Levy
They just, they chat with AI, they get questions answered, they generate an ad for Facebook for their small business. And then us in Silicon Valley were like, Terminator's coming. And, and, and again, maybe you believe the Silicon Valley people. And I think it's worth understanding the risks that we're talking. But I just have a lot of faith in humanity that even as AI does the next sort of rung of work, we just invent the next thing on top of that. And I'm not seeing much evidence that suggests that will stop.
- [41:46]
Katie Prescott
It's a really, really interesting point of view. I think the thing that jumped out for me at the start was where he said, you don't need to motivate agents. So on one level he's saying, yes, we're saving lots of time, but actually we're reinvesting that money elsewhere in the business. So it's not the end of humanity and it's not the end of jobs. But there is that danger that actually, you know, like the open claw sumo wrestler at the all you can eat buffet, they just keep going. They don't need motivating, they don't need sleep, they don't need paying.
- [42:13]
Danny Fortson
Yeah. Apart from the, the tokens and the energy. Right.
- [42:17]
Katie Prescott
It changes the workforce in quite a profound way, doesn't it?
- [42:21]
Danny Fortson
Yeah. And I mean it is. There's something not to be too woo woo about it, but there's something magical about being like, hey, Claude, go read this 100 page document, financial document or whatever, help me synthesize. I think what I'm most interested in is theme X. And it will be like, on it, boss. 20 seconds later it has read the entire thing and will come back with answer to the question you're asking. And you're kind of like, this is like an incredibly powerful grad student that's just sitting there on my computer doing whatever I ask it to do.
- [42:59]
Katie Prescott
More powerful than I was ever as
- [43:00]
Danny Fortson
a grad student for sure. Yeah. Cause it just, it's instant, you know,
- [43:06]
Katie Prescott
but it was brilliant to hear from him because he's someone who's been through all of these different cycles before. So I think he's got a very interesting perspective on how this one will pan out. And fingers crossed he's, he's got the right answer rather than your lunch companion.
- [43:24]
Danny Fortson
So that is it for this week's episode of the Times tech podcast. If you are enjoying the show, drop us a line to let us know.
- [43:31]
Katie Prescott
And we would love to know your thoughts about today's discussion Would you choose to use an AI agent? Would you have a Norman in your life? Have you already got one? Or are you perhaps worried about what it might do to your job or your work?
- [43:44]
Danny Fortson
Tell us by emailing us@techpodimes.co.uk that is techpodimes.co.uk and we will see you here next week. Bye bye Bye bye.
- [44:02]
Progressive Insurance Announcer
You're listening to this podcast, so I know you've got a curious mind. Here's a helpful fact you might not know yet. Drivers who switch and save with Progressive save over $900 on average. Pop over to progressive.com, answer some questions, hit and you'll get a quick quote with discounts that are easy to come by. In fact, 99% of their auto customers earn at least one discount. Visit progressive.com and see if you can enjoy a little cash back. Progressive Casualty Insurance Company and affiliates national average 12 month savings by $946 by new customers surveyed who saved with Progressive between June 2024 and May 2025. Potential savings will vary.
- [44:38]
Aaron Levy
Who won, how did they play, what was wonderful, what was woeful and why? I'm Tom Clark and on the Game
- [44:44]
Danny Fortson
Football podcast we answer all of that, plus provide detailed analysis of the tactical
- [44:49]
Aaron Levy
trends on the pitch and the financial situations off it. I'm joined by former footballers as well
- [44:55]
Danny Fortson
as reporters and columnists from the Times and Sunday Times. We're here twice a week and for this summer's World cup, we'll bring you daily shows with our team covering every inch of the tournament. Find the game wherever you get your podcasts.