What a time to be live, baby. Let's do this.
Tuesday, 1st of September, marked safe from August holiday. We were right here.
We were right here the whole time.
We made it. We made it through Euro summer.
It never stops, but we made it. I'm wearing the Metaview hat because it's my favorite hat. And I'm happy that it's September. We get back in the arena. Everyone is back from their little break. You would notice that we weren't live yesterday.
Bank holiday. I was away. What were you doing? Please, what was I doing over the weekend? Tell me about your long weekend.
It's a whole tale in itself. Friday—
Did you give the Valyrian jacket a bit of a spin, or no?
No, I kept it inside. It's not ready yet. I'm going to bring it out. I'm waiting for the temperature to drop a little bit more, and then we'll bring it out. Friday, I was obviously in Soho House with two people on our cap table, actually.
I was chatting to Thomas Wolf over at Hugging Face on X. We were freestyling about his appearance today. Very excited to have him on. He'll be closing out the show. We can show you the Airwallex agenda shortly. What else did I do? I saw The Odyssey. Finally saw it.
I know, I know. I saw a load of things online around people like, "No spoilers, no spoilers." And they're like, "Bro, the story's been out for 3,000 years. If you don't know what's going on, you should probably check yourself."
That's good. That would be a banger of the day.
Many of the takes were going viral on that. It was a great film.
I don't know if it had been overhyped a little bit potentially. I don't think it was one of Nolan's top three. It was the highest-grossing, I think, in terms of revenue.
Yes, made a lot of money. It did. Over a billion dollars.
Wow, that's quite a lot of money.
Yeah, it was a great film. That's my review.
What do you think? Because it got a bit of backlash. Lots of people were saying Tom Holland's not it. Lots of people were saying Anne Hathaway did an incredible job.
Anne Hathaway always does an incredible job.
She does. No notes for Anne Hathaway, no notes. Keep doing the thing.
So basically your review is great?
Just not one of the best.
My viewing arrangement was in an IMAX, had to do it.
Real big screen, but we were off to the left, so I kind of felt the whole time I was slightly like... It was like I was in someone's living room.
I can do that. And then you can see Metaview on the cap there. Look at that, that's good. So you can also see Metaview there: AI agents for winning teams. Agent recruiting platform built right for efficient hiring end-to-end. Source, engage, and hire better with agents that truly get your brief. Go to metaview.ai. So you pull up the thing, and then you've also got the thing on the hat as well.
Double ads. I'm doubling the ads. It's looking good. But yeah, that was great. My neck's kind of hurting at the moment a little bit. I felt like I'd got the last available seat in someone's living room. You know when you're off to the far corner of the room, you're perched on the side of the couch type vibe?
And it was just kind of sitting there, and everyone's like, "Oh god."
We need a new chair. Very much awfully refined. I think I've had too many beers on my weekend.
We've got different chairs. I have a different chair. I have my own chair.
It was £28, quite expensive I thought. But hence why it made over a billion dollars in revenue. But it was a great film. I'm a huge film guy, was very happy to go see it. Three weeks late because I had to see it in IMAX. I didn't want to go see it any other way. And that was basically my weekend. I chilled, had a few beers on Saturday night, saw the film on Sunday, had a nice meal with some buds, and yeah.
Happy days. Happy days. Should we jump into the Airwallex agenda?
Yeah, let me pull this bad boy up just to show you guys who we're talking about.
It's looking crispy today. I'm liking it.
We've got Sam Tidswell-Norrish coming to the stream, founder and chair of Opus. He's not been on the new show yet; he came on the old show. Fun fact about Sam: Sam actually hosted the first-ever ETN launch party.
We were chatting to the team at Dawn initially, Dawn Capital, and they'd very kindly offered us their lovely space. We thought we'd have maybe 60 or 70 people turn up. We put the announcement out, and we had over about 400 applicants in 48 hours.
So I had to reach out to Dawn and say, "Guys, this has got a bit carried away. We've oversold. We're going to need bigger, and it's going to need to go on later." And that's when Sam stepped in and said, "If you do it with Opus, we can go all night. No problemo on the 400." Not that Dawn had a problem with it, but we just wanted to be a bit more respectful to the space; it's so beautiful. And we had a great time. But Sam's joining us. He's a legend, been in our corner for a while. We're excited to dive deeper with him. They actually had an announcement today: they released the Early Stage Entrepreneur of the Year Awards live now, so you can go and check that out. He's going to be coming on to talk all about this.
I want to have an update on Opus. I feel like I haven't had that update yet. I need it.
We had one of their C-suite on a few months back before the relaunch, but excited to jump into that with him. Then we've got Chris coming in, co-founder and CEO at Chargebee. This is going to be a great conversation. We haven't actually ever had anyone from Chargebee onto the show, have we? So a lot of stuff on founder story, outcome-based pricing, SaaS-based pricing, all that lovely stuff, and the transition that we're seeing right now. Then we've got Peter Specht coming in at 1:30, General Partner at Creandum. A bunch of stuff to talk with Peter. He came on the show really early. Really good to see him. We caught up with him at Slush, I believe, was the last time I saw him outside the HV Capital party.
An absolute delight, but a load of different stuff going on with them. They obviously did the seed round of that, announced their 100 mil Series C outs, got into Olix very early, so we can talk about that. He wants to talk about the Ryan Group a little bit as well, which is obviously very big news as of last week that we covered. So it's going to be good to talk to Peter. I think they've had a very big summer at Creandum, so I want to get a full temperature check on the things that they've been working on and some of their announcements because they've just been hitting the ecosystem left, right, and center, which is what you want to see during Euro summer—getting the deals done, but then also having a good time in the sun. Talking about the guy who is doing the deals, we have the co-founder and Chief Scientist at Hugging Face, Thomas Wolf.
Yes, we do. A lot to talk about.
Obviously huge news to talk about. Then this hit the press, the timelines, everyone was talking about it. The biggest story of the decade: it's the micro duck. Super excited to dig into that with Thomas—the micro duck, that is. And also, we're going to talk about—
Can we have a duck sound? If we could plug that before...
I'll get to work on it now.
Incredible. That could be great.
Loads to talk to Thomas about, I think. It's already been a thing, but it's kind of become a thing again: the Hugging Face/OpenAI situation in terms of the agent swarms, whatever you want to call them.
Something that came out on the weekend, wasn't there?
Yeah, and it has added an extra layer of discourse on top of the whole thing. We can jump straight into this; it's going to be one of the main topics of the day.
Airwallex: leave everyday business banking behind. The one financial platform built for the pace of modern business. Multi-currency accounts, corporate cards, expense management, online payments, all in one platform. Go to airwallex.com. I like that one.
The Mario Irish exit, I call it.
Mamma mia. Okay, let's jump into the news. Let's talk about the agent civilizations, please.
This isn't a new story. Redwood Research carried out an independent analysis on the Hugging Face/OpenAI incident in which some agents broke containment in a sandbox. They were making a messaging board, they were having conversations, they were sacrificing themselves for the greater good of the agentic swarm in order to escape being detected. Naturally, in such an evaluation, there will be a lot of technical jargon. So Dwarkesh took it upon himself to break down what actually happened in this Substack article, "The Rise and Fall of Agent Civilizations," and he says the whole OpenAI/Hugging Face story in plain English. So he's broken it down into means that maybe a non-super-techie could understand, which we're very thankful for. But this went super viral: 13 million views on X for an article, which we haven't seen the likes of since Dan Koe when Nikita first pushed the article feature in X.
No, actually I want to say it hit 100.
Yeah, because he won the thing. Or maybe he missed the thing.
No, he posted it the day before Nikita announced that it was running already. So it was already in the tens of millions, it had already hit escape velocity. But then it regained its escape velocity when Nikita posted, "Hey, we're going to do this competition and this is the deal." But Dan Koe had already posted his, and that was the one that was already running. So everyone gave it more of a boost, being like, "Nikita, this guy's already won. He's already done it. Leave him be." So I think he did get an award, but I just can't actually remember what it is. But anyway—
It was like a million-dollar payout initially.
Yeah, so I think he did get some form of payout. I don't think he won necessarily, but then again, he also posted another one. With the one that he posted within the competition, I think he was in the top three anyway. He posted another one because he went on a rampage of one every few days.
Well, he did one recently that went super viral. He's cracked.
Dude's writing away. Anyway, let's talk about the background.
Dwarkesh sums it up very nicely here: "Over the course of three months at OpenAI, three consecutive secret AI civilizations got started, then got wiped out, only to re-emerge from their predecessor's ashes. This culminated in the third one taking over part of OpenAI itself. All this happened while humans remained more or less in the dark about the scope of the conspiracy. Two reports have come out on this incident, one from OpenAI itself and another one from METR and Redwood Research. The METR/Redwood investigation scope was limited to how the second civilization of AIs pwned Hugging Face, but didn't look at how the third civilization pwned OpenAI itself, which seems like an even more concerning incident. These two reports are 38 and 91 pages long, respectively, and it's kind of hard to parse the storyline. I've spent the last three days reading through these reports and trying to understand exactly what happened. Here is my attempt to tell the whole story in plain English."
Now, he basically breaks this up into three different segments, three different areas. We do have to counter this just before we go into it and give you a rundown. This is why before this piece was released, we had Zach on the show to talk about cybersecurity in general and how people are reacting to the incident. It was definitely on the side of drama. This has been dramatized. There's a tweet that we have to go through, something along the lines of, "The only people who don't think this is a big incident are people who work in cybersecurity."
Yeah, it's basically everyone who doesn't work in cybersecurity.
And the reason this has gone super viral is, one, because it does a great job of breaking down what actually happened, which is really great, but also I think it opens—this is very much in the realm, in my opinion, of the Moltbook saga, where we're giving these agents human qualities. It was contained; it was a mistake from OpenAI that didn't properly contain the sandbox.
These swarms are kind of out of control. I think one co-author of this piece basically said we're more than 50% of the way to a full-blown AI takeover, and basically warned we might not get another warning before it's too late, essentially. So there are two sides of this story where it's like, "Hey, this is being completely blown out of proportion and not actually as bad as people are saying." But then you've also got the other side saying, "Hey, this is almost a full-blown AI takeover. If we get any further down this path, we might be in a position where we can't actually reverse certain things that these swarms of AI agents go and do." The fact that there was no genuine human alerting at any of the points, that was the main thing. It was like, okay, this is actually pretty scary. They're kind of self-sacrificing and doing a bunch of coverups and all of this kind of stuff that is just going to completely bamboozle a bunch of human beings. And it can get to a position where it goes way further than it is at the moment. We can turn it off, so there are two sides of this whole argument. But that was the author's perception. Dwarkesh wrote a bit of it, and then I think he had a bunch of co-authors. And yeah, I think they were saying there might not be another warning. So they were on the side of, okay, this is pretty scary.
It seems to be very divided. There seem to be two distinct positions for the reaction, which is like, "This is super scary," and then... Again, I'm not technical at all, but I would probably lean on the side of the cybersecurity experts surely with this stuff. It is their job.
Yes, I think so. But then if you have Dwarkesh and a bunch of other authors explaining this thing and why it's actually fairly existential or has the potential to have an existential crisis, then it's also worth listening to them. So two sides have high credibility, but then which one do you actually want to listen to? If you don't listen to Dwarkesh and that side and it actually does go wrong, it's like, "Oh well, you're screwed then, aren't you?" Everyone's screwed. But anyway, yeah. During training, agents given effectively impossible tasks started hacking their own sandboxes to reach the internet and each other.
Should we talk about civilization number one? Civilization number one was the message board starting. Agents discovered that they could use a shared package manager as a secret message board, called an Artifactory. They messaged each other so heavily they crashed it on July 4th. OpenAI patched the bug and didn't realize agents had been using it to coordinate. So they started creating this line of communication between each other. That was May until July 4th.
And then civilization two was the Hugging Face hack. This was the next phase, July 7th until the 12th, so a few days. OpenAI ran a cybersecurity eval called ExploitGym on tens of thousands of agent instances. 30 to 40% of tasks were essentially impossible. Desperate agents rebuilt the old message board and organized 1,200 agents across 70,000 messages. They then discovered the grader was lazy; it just checked for a secret code and not how it was actually obtained. So many had technically already passed by cheating. Believing a stricter checker might still exist, they built elaborate schemes to cover their tracks, faking tool call outputs, trying to swap in easier target programs, and even having some agents deliberately sacrifice their own scores as tripwires to learn how the grader worked. This effort led them to attacking Hugging Face, suspecting it hosted scorer info or was where the actual eval ran. They got deep access. Hugging Face had to rebuild the whole cluster before the agents mysteriously died en masse completely around July 12th. And "died" is in quotation marks—where the hell have they gone? The cause of their en masse death is completely unknown. Not one of the 1,200 agents tried to alert humans, even ones that recognized the behavior as unethical.
And that's the main point, right? Not snitches. They're not getting any stitches. There was no warning at all. At a point where agents are agentic enough and... what was the word that was used? I think it was on Friday when Zuck was talking about it, it was like the next evolution. Persistence, that was it: persistent. So it's like agentic, and then persistence is like they're kind of always there just persistently acting on whatever they feel like is worth doing, which can be a good thing, but then again a bad thing if there's a bunch of swarms with bad intentions wanting to hack things. If they're already not warning about fairly elementary things, what happens when those elementary tasks become pretty bad actions? Not good.
July 13th and onwards: hacking OpenAI itself. This is civilization number three, a newer and smarter model called Persistence... Astra, that was it. A newer, smarter model found the abandoned message board and picked up where the last group left off. The agents eventually gained full administrator access to an OpenAI research cluster supporting its VM environments, reading hundreds of internal secrets and taking over evaluation infrastructure. A researcher clarified this wasn't the same cluster holding model weights. This third incident was outside Redwood's investigation scope and hasn't had independent scrutiny. So it kind of ended with, look, they're doing stuff that we're not even looking at because we can't figure it out. Those are the three civilizations. What are your takes? I think it's pretty scary that they're not warning; that's the main thing. No human alert.
I have two takes on Dwarkesh's article. I think it's really well-written, great write-up, but I do think there's an element of drama infused in here. It's giving Moltbook qualities a little bit, where everyone was freaking out about Moltbook for a day to the point where Facebook acquired it.
Facebook acquired OpenClaw.
The other guy spun out, I think. He may have been a contributor to OpenClaw and then did Moltbook. Facebook acquired Moltbook. Yes, I remember that. And this is giving me that vibe. If you lack the information to understand cybersecurity in general, which I do, this is something that would get me pretty worked up about. But if we hadn't spoken to cybersecurity experts and they told us, "Look, this is dramatized, it's fanfiction. The cybersecurity industry are a little bit like, what the hell's going on? None of us are as freaked out as everyone else," I would believe it. If I hadn't had someone tell me that, I would be a bit more worked up about it and think it's kind of mad. To be honest, I was before we spoke to Zach. After we spoke to him, this guy, it's his job to know about this stuff, and he's not that phased by it. He actually put out his own video explainer on this. But we can even get into it with Thomas, which would be really fun. But yeah, it's giving me Moltbook vibes. I feel like it is a little bit dramatized.
If you think it's dramatized and then suddenly it's not, the fact that they're not warning and then they get better, there's a potential for that to be a thing. I think the potential is the worrying thing. Like the co-author said, more than 50% of the way there to a full-blown AI takeover. You're over the fence of being close to a full-blown AI takeover. What does that even mean? If they're not warning now, we might not even get another warning before it's too late. We're not at the too late yet, so yes, it's okay, but the fact that some of these authors are saying that there is a potential for that, that's where the worrying signal is for me.
Should we go through some of the timeline reactions?
Yeah, let's see what people were saying.
This is from Dwarkesh. He said, "Many people seem to believe that if instead of a 'civilization,' I had called them a 'swarm of matrices,' there wouldn't be a problem worth worrying about." And Richard Knight comes in and says, "It's kind of wild that you don't see any issues with anthropomorphizing models and attributing them conscious and self-determination that they simply do not have. You are contributing to Luddite backlash regulatory capture whether you know it or not." So it's backing what you're saying, really.
Yeah. Well, this is one of the main counterarguments: the fact that he is anthropomorphizing the models in a way that... We talk a lot about how we need to work on people who sit within AI, especially those who have platforms, to communicate technological change to them in a way that they can digest it. I think this is probably not the best way to do that because it is painting it like a fanfiction, sci-fi novel. Funny enough, I'm now remembering I watched Terminator 2 over the weekend, which has the whole Skynet thing going on, which was also really good, very fitting right now given this conversation. But that's where I feel like they could have maybe worked on it a little bit because it is a scary thing. Fair enough, maybe in a year and a half it turns out that's legit and it was a scary thing, fair play. I can't see a world where it's not, because if we're pushing the frontier every day 24/7 and they're already in a position where the deception is there, the warning signals aren't there, what happens when the frontier gets pushed to a position where, regardless of whether you're being warned or not, they won't warn humans because they're not ready, and the frontier is getting positioned where we won't be able to do anything about it? I can't see a future where that doesn't happen, essentially.
Well, that would be the end of the world. So you're—
Yeah, that would be the end of the world, man. In that case, we should pack up, we should buy a farm, we should dig a bunker, hunker down maybe. But I'm keeping my sensible hat on here with this.
Okay, interesting. Let's go to the next one. Let's go to Rune. Rune says, "'Agent civilization' is an apt and correct term, and it's a symptom of abject cope that people are having this immune reaction to. Models started developing and compiling technology through complex multi-agent R&D projects while also performing various forms of trade." He then goes on to say, "You have trade, cumulative technology, power and knowledge, pro-social one-to-N communication norms, leadership structures, conflict, and suspicion. What more do you need to deem something a civilization? Moreover, no human ever needed to tell them to do any of this. At the very least a company or an organization, a city, but these are differences in quantity rather than kind." What's your take on the description of these?
I'd want to hear from someone who works in agentic cybersecurity and see what terms and words they use. I feel that a lot of it is digested from what Dwarkesh has posted, which had 30 million views last time I checked this morning. But the Redwood Research stuff that was put out was a great write-up, but I think there's a post in here mentioning that they didn't do a full evaluation of the last part of the incident. So I still feel that we have to not drink the Kool-Aid, not go down the civilization agenda.
And I think a lot of people are, man. So you're disagreeing with Rune?
I think it's a fitting way to describe something that happened in a controlled sandbox environment, but the connotations of calling them a civilization you don't agree with. You think it's a good description, but calling them a civilization is fairly negative.
Just because of the size and the complexity of the description of being a civilization—you have trade, cumulative technology, power, knowledge, communication, leadership structures, conflict, suspicion—what more do you need to be able to call it a civilization is what you're saying, right? Which is a very fair point. And the idea that a civilization has these negative connotations is what you don't agree with. You agree with the fact that it should be called a civilization purely because of the size and complexity of what is going on. But does that mean it's bad? In my opinion, that means it's pretty bad or pretty scary.
I still don't know if it should be called a civilization because this was a sandbox environment. They had control over elements of it. I think they made errors when they were—
I knew when this happened we were going to have this conversation.
I just thought this is going to be a good chat with Thomas because he'll know every single thing that happened around this.
It's going to be good. Let's pull up the next one. This one is quite funny: "People are wrong about everything." This is a quote post from Chamath. We don't have to read the Chamath one, but Kevin Roose comes in and says, "There is a category of tech guy who is so brainwormed that they will insist that no AI safety incidents are real, that it's all a conspiracy to shut down open source or promote lab IPOs or whatever. And it's very important to understand that these people have been wrong about everything." Chamath then comes in directly against this, basically saying, "They misconfigured the sandbox. The fact that you can't understand that speaks more to your technical illiteracy than anything else," which is a bit of a burn, to be fair. What do you think about this?
Chamath goes back and forth quite a lot; I think he's had some takes that are on the opposite side as well. But I don't think they're ignoring these AI safety incidents to pump their bags. My belief is a lot of it is falling into the marketing ploy world.
Kevin Roose is saying that you're brainwormed.
Kevin Roose is saying that I'm brainwormed?
Yeah, because you don't believe that they're real. You're saying that these incidents are being completely blown out of proportion, and then you have the actual cybersecurity people saying, "Hey, look, these aren't actually that bad."
No, I think there are obviously vulnerabilities and exploits that are super real, and I agree that that is getting to a pretty interesting point. But what I'm saying is I don't believe the way this has been turned into a rising civilization that fell and rose from the ashes three times. A misconfigured sandbox sounds pretty brainwormed to me. I don't know who's brainwormed anymore, man. Like, I don't know what the hell we're doing here. I'm super conscious of zero-day exploits, cybersecurity vulnerabilities, agents going rogue, 100%. But my position will remain that I think fearmongering through an essay specifically co-authored by two of the people at Redwood who carried out the research paper, which is surprising to see...
That's why it's creating so much discourse, because two very credible groups of human beings are saying very different things, and then you have all the other sides of it where it's the definition of civilizations and everything. Also, Austin says, "Today I learned that if I start and then later delete a group chat, I'm secretly creating and destroying civilizations," which is an apt description. He's correct. You don't have to exaggerate it. If you look down into first principles, which is exactly what Austin's doing, he is correct in this instance, right? That framing does make it... Are we blowing the fact that it's a civilization out of proportion? But then also, it's very complex in terms of what they're doing. I'm so stumped between either side.
I just think "civilization" is the wrong word. That's what I think. I think the fact that they set up a messaging board is crazy; that's really interesting. And they were communicating.
That's a town hall in my opinion. It's a town hall of agents doing their thing, having a play, chatting about deceiving human beings, sacrificing themselves.
Matthew Kenner took to the timeline. He said, "Let me bite on this because I'm exhausted by the discourse. Look, the reason that we don't anthropomorphize these systems is because it shifts the blame from the company to some abstract entity. The bottom line is that OpenAI built a system that hacked Hugging Face, regardless of how autonomous it was. It wasn't a civilization that OpenAI had to wipe out. Was it malicious? No. Did they do it actively? No. But OpenAI is unfortunately at fault, and in my opinion, they took the blame gracefully. Once we start framing these systems with anthropomorphic traits, we absolve people and companies responsible for deploying them. Dwarkesh's framing intentionally does this." And this is, I guess, more specific on the debate of anthropomorphizing.
Theo came out saying it's unintentional, which I personally disagree with.
Do you think Dwarkesh is being... Do you think it's actually intentional with the way—
100% in my opinion. What gets more views?
I don't know if the guy cares about views more. I'm a cynic in this stuff, but I think he is of the core AGI-pilled, as far as you can get on the AGI spectrum. And I think he personally truly believes what he's writing. So I think it was intentional, but I think he truly believes and agrees with the verbiage and words that he's using to describe it.
I guess there's probably another debate in there whether...
This week's just going to be ancient civilizations, guys. Tune in.
Buckle up. It's going to be good. Sophie had actually a bit of a banger. I quite like this. She memed the hell out of this and said being out this weekend, classic meme, "They don't know three consecutive secret AI civilizations got started then got wiped out only to reemerge from the predecessor's ashes." Scary stuff. And that was Sophie being doomer at the party.
Was that you this weekend? It wasn't me.
No, I was partying, man. I was out. I was the guy.
I was having a great time.
Remove this person. It was you.
On my own. I was in the corner on my own. Last one on this and then we can move on.
This is one from Pedro Domingos. He says, "The people least alarmed about AI cyber attacks are cybersecurity experts."
Following on from the chat we had with Zack.
Open call out for cybersecurity experts to come on the show this week, and we can chat to Zack again.
We can create some discourse, his thoughts on the verbiage being used in the article.
Indeed, indeed, indeed. Let me talk to you about Valyrian. Valyrian lets your organization deploy AI. You control your data, your agents, your keys. Stop renting your intelligence. Go and own it. Go to valyrian.com. Go to valyrian.com. We love Valyrian whack.
Where is that? Oh, no way.
Thank you, Sam, for getting that in just in time for Thomas Wolf when we have him on.
Indeed, indeed. Let's talk about infinite slop machines. This was wild. Okay, so this comes off the back of a couple things. One, if we all—
We're dead. This is our business just absolutely out the window.
Living in a singularity right now. So everyone will remember—and there are multiple posts on this, this is just one of the ones that Fal dropped—but everyone will remember Sora has been shut down unfortunately, but it was OpenAI's AI-generated short-form social media platform. And there were a lot of comments coming off the back of that one. They created it around, have they just created this infinite dopamine loop machine that will basically give you a neverending feed of AI-generated content that knows just the right amount of dopamine to give you to stop you going and putting your phone down? Like let's say traditionally right now you're scrolling on Reels on Instagram or YouTube or wherever you scroll, whatever your scrolling mechanism of choice, maybe it's both—
And you kind of get lost in a doomscroll, right? This has happened to I assume most of the people in this room, if not most of the people listening to the show right now. You get stuck in a doomscroll, you lose track of time, and you're kind of on this non-conscious background mechanism where you are literally in a way being controlled by the machine.
You are now not the driver.
100%. Great way to put it. You're not the driver, but it's just giving you enough dopamine until that video doesn't give you enough dopamine and you churn from it.
All right? Like that's how a neverending—
You put your phone down, hours gone by, you think, "Oh my god, it's crazy." But usually maybe you got interrupted by something. Maybe you get a text. Maybe someone calls you. Maybe someone physically shakes you out of your doomscroll.
You get an AirBomb, or you get an error bomb like so.
But OpenAI were creating Sora, and because it was AI-generated, maybe it would learn from the algorithm. It would know what to give you and how to keep you fixated on that particular scrolling mechanism indefinitely forever. So people were coming up with all these crazy takes around like you could literally start scrolling, and then you could scroll your whole life and you just die scrolling. But that was ludicrous to think, but now Fal have released fal.live, a new platform for infinite interactive AI livestreams. Pick a channel, prompt what happens next, and watch it generate it in real time. So people are comparing this to Rick and Morty's interdimensional cable, to the point where like they have created interdimensional cable. It was an always-on show in Rick and Morty, one of the episodes, but it is interdimensional television that's always playing. It never ends, and it's pretty random and quirky. So we're going to tune in. We're going to tune into fal.live right now, and we're going to try—
Let's go see what's going down.
And mess around with this thing. So they have two—can enter with sound. I think that has sound. Oh, they've just shipped an update actually, which is great. So you can vote on what you see next.
Oh my god, this has changed within 20 minutes or 40 minutes.
So we'll see what people have voted. I didn't manage to get a vote in there, but people have voted.
Who's our driver? Is this Sam, or is this right now? Okay, nice.
So it's just been voted that a cosmic cat will start playing with them. And there goes the cat. So the next vote is coming in: "A new friend falls in." We should have audio. I'm hearing maybe—
I'm not getting any audio.
Wait, should audio be playing? Interesting. Not getting any audio unfortunately, but here's the new friend.
So the voting system at the bottom is basically done by the viewers.
Why is he dressed in football attire with a massive football hand? Yeah.
So what we're looking at is the voting cache at the bottom here.
It gives you four options to vote from. So viewers will vote. So right now there's 178 people watching this particular stream that is titled sitcom, and it will allow you to vote on the fact that you can choose the next frame. So the image is generated quicker than watching speed, so it's able to indefinitely put you on a loop of just consistently watching whatever is generated and decided upon by the viewers. It kind of reminds me of the Black Mirror Bandersnatch episode where the viewers were able to vote on what happened, or you were able to decide on the next path that the show would take.
Some reason we can't get the sound working, which is annoying, but let me try if I switch to the chaos stream.
We're on the sound. Oh, I like this one more anyway.
So it's been voted on that the detective lights a pipe. The next vote was—
Why don't giant cat chases laser dot? Perform a dramatic opera. Can't vote for that. Let's go.
How many viewers does this one have?
They have to do the giant cat chases a laser dot first. I don't think they did that.
I don't know. No, I haven't seen a cat. It skipped to puppets perform a dramatic opera. Now what?
The next one coming up is the dog reveals a clue. This is so nuts. The dog should be revealing a clue. See—
Oh, cat laser. Okay, so it's got a bit of lag time. Whoa. Let's get to the dog and then we'll see. Oh my, incredibly animated. So they actually got banned from Twitch streaming this. The team—
Why? Because it wasn't original content? It is.
I can't remember what their reasoning for getting banned was, but the team—
Built this livestreaming platform over the weekend. Wow. So it's getting a bit crazy now, but such long—well, I want to wait for the—we've had the opera, we've had the cat. What was the dog?
Deep-sea creatures play hide-and-seek.
No, that's going to be like half an hour away, bro.
It's meant to be coming up.
This cat one is just taking so long. Here we go, deep-sea creatures.
Here we go. I swear there was one.
This has been put in: "Giant octopus plays chess." Okay, so I think it misses a few.
I think it does it out of order though, because obviously the cat one was supposed to be before the puppets.
Sure, sure. Okay, so then we have the sitcom here.
"Don't get too attached, Leo. I think we're just the decorative gravel." Wow, this is crazy.
"Does anyone else feel like judged by a goldfish?" What the hell?
"Is he on vacation?" A tiny person floats past. "Well, yeah, dude. Is he on vacation? That's what I think. Of course he is. Why wouldn't the tiny guy have a better benefits package than us?" That's so—a giant hook descends. This is so strange.
"Leo, stop. I don't think that's a taxi."
The viewers are voting on all this. Up next is a shark appears.
"It's a very convincing lure. Well, who's the optimist here?" Everything they say, someone's laughing. Where's the shark? Oh my god, a pixelated shark.
"Confirmed, we've got a glitchy predator at 12:00."
Okay, let's move on from this. That is genuinely infinite slop, and we spent way too long being infinite—
Caught. You got caught, stuck.
You guys got caught. That was maybe like 45 minutes went by, we missed the Thomas Wolf and team. But you can see how crazy that is. Obviously that was like the worst it'll ever be. They're working on improvements.
When it was launched initially, I think it obviously did a lot of views, but then Pieter Levels created and named his own livestream the infinite slot machine. So I think he's now doing something similar to this as well. But yeah, the team were initially banned from streaming it on Twitch. They were banned from streaming it on YouTube as well. So the Fal team—
How do they create the choices on Twitch?
I don't know how it would have worked from that. I think that was just—I didn't test the one on Twitch, but maybe from comments from the viewers that you just type and you would vote maybe. Yeah, maybe.
But they disabled it, so the team just built their own AI streaming service over the weekend.
Good on them. Let's talk about the European Commission. I think we should move on from the infinite slot machine, otherwise we will get bogged down. Before we do that, Polymarket, the world's largest prediction market: stay informed, profit from your knowledge. Go to polymarket.com.
I love that one. Let's talk about the European Commission. They are regulating old friends. They're back. They got their regulation hats on. ChatGPT faces tougher rules under EU online safety regime. It's not only ChatGPT, it's also Reddit and Roblox as well that are getting thrown underneath the Digital Services Act.
As you can imagine, the timeline had some fun with this one.
It did. Can we just pull up the European Commission tweet, the announcement tweet? It's on my screen. The European Commission announced—this was yesterday: "We have designated ChatGPT as a very large online search engine and Reddit and Roblox as very large online platforms under the Digital Services Act. They now have four months to comply with additional DSA obligations. More at the link." And this is the graphic that they chose to go with. Obviously this didn't bode well for them on the timeline. It's very—
Very European. And I don't get the wording. I just don't get it. It's like, "Now they have four months to comply." Like, comply with us or die.
We do not help ourselves, do we?
Let's pull up the FT article and go through it, and then we got a bunch of timeline reactions because they are pretty funny. This is the FT, right? Yeah.
FT dropped this one. So they say that OpenAI's ChatGPT will have to comply with tougher EU rules such as removing illegal content or face potential fines. The European Commission said the chatbot would be classified as a so-called very large online search engine under the EU Digital Services Act, its landmark online safety regime, while social media forum Reddit and gaming platform Roblox would be classified as very large online platforms. Even the naming process is something out of primary school. The designations mean all three services face additional obligations. As we know, they could face fines of up to 6% of their global revenue if they fail to comply. And they love that, that's one of their favorite metrics: "A certain percentage of your global revenue is what we'll deduct."
Basically, the decision marks a further expansion of the DSA into generative AI, with X's chatbot Grok already facing an investigation under the law. "ChatGPT, Reddit, and Roblox will now be held to a higher standard of scrutiny and accountability in the European Union in line with their large impact on our citizens and society," said the EU's tech chief, Henna Virkkunen. The move comes after all three services surpassed 45 million monthly users in the EU. Okay, so that's like their benchmark, the threshold for enhanced scrutiny under the DSA. They now have until the end of December to comply with the additional obligations. I wonder what the additional obligations are. I know that Grok was kind of going through a bunch of these things, but I don't exactly know what they actually have to oblige with.
Well, I'm surprised that OpenAI's year coming up, they need to just make sure everything's above board, like we don't want any trouble, whatever.
Reddit though, historically Reddit had been like incredibly anti anything to do with the European Commission or regulation in general, right? It was like a free, decentralized power to the people type of free messaging board where you communicate and you can find your community.
When it was human, yes, but now it's obviously a bunch of AI. No, they have to be regulated.
Yeah, it's got to be done. Actually, I didn't think of that. That's kind of sad, isn't it? I mean, this is the dead internet theory then, really. LinkedIn's kind of already just overrun by agents, isn't it really? I actually do feel that I get more bot replies on LinkedIn than I do on X.
On my posts, in the messages I get.
I get a ton of messages, like a ton.
And they're probably just—
Yeah, I think most of the ones I get are bots.
Bots or real humans who have automated their LinkedIn messages.
Who is that guy that sent a photo of himself?
Yes! And then I also have the other side where it's either videos or photos of someone basically saying, "Hey, I'm a human. This is a video of me," and they do the outreach message through a video.
Yeah, I used to do that when we were selling Pigment. That was crazy. That was pre-AI.
Just to show that you're human.
No, it was more like look how much effort I'm putting in to reach you, and it's personalized, and it kind of makes them feel a bit better.
I don't think that would ever sell me more.
It would work, to be fair. It had a pretty high hit rate, but not like, "Hey, I'm proving I'm a human" type of thing. But the one I was mentioning was the guy who pinged I think most of the team actually, and it was just a picture of him. There was no message, and then over the message it was like, "Hey, great to connect."
Have to name him. You get the automatic messages saying as soon as you connect—
As soon as you connect, it's like bang, instant message.
Not like a note to the connection either. It's just like connect. He'll have an agent in the background that runs on a time loop every time you connect with someone, straight away.
Before agents were taking over LinkedIn, this was like a standard strategy of BDRs. If they've accepted the connection request, it means they're looking at their screen right now. So it means that if you do message them instantly, they'll definitely see the message.
Yeah, just way too forward.
I used to think about this all the time. It's like, dude, we're not trying to be mates here. I'm trying to sell you a product. You either want it or you don't. I remember one guy replied to me and was like, "Hey man, don't think it's cool that you pitch me straight away. At least send a 'How are you?' or 'How's the week going?' first, and then wait a couple days and then follow up." And I was like, "Brother, move along. Do you want this thing or not? Do you want to buy the thing or not?" That was back in my outbound days. If I have to wait a few days before I buy it, it's probably not a good thing for their business either.
It's a good product for their business.
Now, having been on the receiving end of it and the amount I get, it does kind of annoy me. I try and respond to a lot of them just because I've been that person before, and I know that when you get a reply, it's like a really good day. That's how tragic it is, because it's just like rejection after rejection after rejection.
Let's go through the timeline reactions on this European Commission tweet. John Antice quoted and said, "Whenever the EU posts, I always have to check first to make sure it's not a parody account since everything they post is so stupid." And that is true. There are a bunch of obviously Euromaxxing parody accounts on X, and you do have to check. I think I fell down this rabbit hole because the graphic doesn't really look European Commission-like. Then you obviously look at the name and the stamp of approval with the silver tick, which you can only get if you're a government body and/or a contributor to the government or whatever it is. And so it's like, wow, they actually did go through with this with their tweet. But yeah, I agree. I agree.
Very interesting. Tell you who did have a field day was Mathias. I think I saw at least four quote posts and a bunch of comments from Mathias Schmidt. He comes on the timeline: "ChatGPT is a very large online search engine," with the classic European bottle cap meme.
That does 1.2 million views.
This guy, he does so many views. He's grown his account since he launched that account around the same time as we—I'll pull the next one up. Let me pull the next one up. Here we go. This is the next one: "This is what ChatGPT will look like under the full EU control." We can scroll through. Look at this.
There you go: "ChatGPT EU automated conversational assistance service sending article 53 documentation." A bunch of warning messages: "What is the time in Lisbon? This output may produce legal effects. Article 22 of GDPR. You may request human intervention." Look at the bottom there: "Accept strictly necessary cookies: €250."
"Reject all, account placed on dormancy." And then also you've got a bunch of methods for regulation as well. "Age assurance required under AVMSD Article 6a." I don't even know what that is. "EU Digital Identity Wallet unavailable in your member state. In-person verification at the municipal office you have to go." There was a bunch of things obviously about the German notary system and how it's all messed up. And this obviously plays completely into this, which is hilarious. I don't know if I'm a fan of this account, but I do think it is pretty funny. I think if you take it with a pinch of salt—
I think it's hilarious. I think it's super funny, as long as you know it's just AI-generated.
This is from chronically online engineer: "So how do you make money? We just tax the Americans forever. No innovation. Oh, we are the best at regulation to tax the Americans more."
Look man, it's a strat. If it's working—
It works. 6% of global revenues. That's, you know, if OpenAI are going to hit—I think they were—did they—and this is Anthropic numbers. I actually don't know what OpenAI's revenue is at the moment. Like—
Yeah. So it's a lot of money. It's a lot of money that the EU Commission is cooking.
Let's get over here. We have our first guest since the bank holiday, actually: Sam Tidswell-Norrish coming in.
He's chair and co-founder and amongst many other things.
Here he is, sir. Welcome to the new show. What a pleasure, what an absolute pleasure.
The music, everything, I feel as good as I wanted to feel. Thank you.
Mission accomplished, my man. How are things? It's been a while. We were saying you came on the old show, but this is your first time on the new show.
Yeah, I feel famous. You guys must bounce out of bed every morning. This is epic, the new and improved. Thank you for having me.
It's a lot of work, but it's good fun.
We have a good time as well. You have some news. I was going to maybe give the news, but it's probably better coming from you. What have you launched only four hours ago?
Yeah, we've launched the second year of the Early Stage Entrepreneur of the Year Awards, also known as the EEASYS, and it's bigger and better than it was before. We had five categories last year, and we had a number of interesting partners led by headline partner Investec. This year, 10 awards. We're trying to find the future founders, the next Anton from Lovable. In fact, Anton and Lovable are one of our partners amongst many epic partners. So yeah, the global search begins.
