Episode notes
Anthropic's Mythos model chains zero-day vulnerabilities, Allbirds pivots to GPU-as-a-service, Perplexity does your taxes, and the EU AI Act countdown begins.
Chapters
- 1:23 — Explains Anthropic's Memphis model and its capabilities in finding and chaining vulnerabilities.
- 7:30 — Discusses Allbirds' pivot from footwear to AI-native cloud infrastructure.
- 12:20 — Details the EU AI Act and its implications for businesses.
- 25:30 — Highlights GitHub's opt-out option for training data and its importance.
Links
- https://www.anthropic.com/
- https://www.allbirds.com/
- https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32021R2477
- https://github.com/
- https://www.butter.dev/
Transcript
Note: this transcript is auto-generated and lightly edited for readability.
Tim: Welcome, welcome to episode number six of Digitalize Agency. Great weather today, we're back in the usual spot, not online anymore. Way better. Much better. The weather's great, I already said.
Sebastian: It's actually not that great. In the morning it was nice and sunny, but no. It ain't raining.
Tim: That's true. So it's good weather. We're not complaining. Guys, huge news, right? Which ones are you? No, but I mean, come on.
Finn: The biggest one is definitely, yeah.
Tim: We already talked about it two weeks ago. For sure. So make sure to always subscribe, because we talked about it two weeks ago, before the headline took over the news. But Anthropic finally released their rumored mall, Memphis.
Finn: Indeed. After it got leaked, again. But, well, that's a different story.
Sebastian: Yeah, and releasing is maybe a bit big. But they released the news.
Tim: They confirmed the news. So basically, maybe one of you want to catch the listeners up to speed of what methods is. Feel free to start. What we're dealing with here.
Sebastian: So basically, Anthropic released, or released, didn't actually release it, but trained a new model on, I think it was the biggest training run yet. That's why they also got so surprised that the model did such a big jump in accuracy and quality of output. But now it's basically one of the best models out there, and the model is so good. At least they say. They say, exactly. Good point. But it is so good that it can chain together multiple vulnerabilities in order to create a new zero-day or a new cybersecurity leak, I would say. And that is basically all the hype. So, but good news, the public doesn't have access to it. Yet. Yet. True.
Finn: I just wanted to say.
Sebastian: So they, right now, they only have it in preview mode, basically. It's for the companies. Exactly.
Finn: For the companies like JP Morgan and all the big AI companies. With the project class wing. Yes. Class wing, what's it called? Yeah. Yeah. But we have to mention, I guess, that they at least want to use it to kind of patch all of the vulnerabilities that the model can find before. Of course, they want to make bank of that as well. That's not a real secret. That's why OpenAI also does the same now. But it tries to not rush into FOMO.
Sebastian: But there are a lot of things happening right now because on the paper, it's all nice and good. And the MIFOS model is super nice in their benchmarks and whatnot. But if you read through the paper, they threw 20K on credits, basically, on that model. And security research... How many parallel agents runs? It was like 20 how many. I don't know. But the fun part is security researchers basically reproduce the same output. So basically, a normal model found the same security vulnerability like the normal one.
Finn: After it was found, though, no?
Sebastian: No. Yeah, but they basically reproduce...
Finn: Like they knew what they're searching for, at least.
Sebastian: No, exactly not. They did an experiment that they tried to see if they can actually also use normal models that are accessible to the public in order to get those vulnerabilities. And they threw also, I think, 20K on it for credits. So just 20K on API credits, which is crazy much. But they got it with normal... I think they used Codex and the new GPT 5.1 model. So not even Entropic, but they found the same. I think on four or five different bugs that only MIFOS found, they could reproduce the exact same output with a normal model. Crazy. So, yeah.
Tim: I saw some numbers of how much estimated, of how much the MIFOS training run consumed in a dollar value of electricity, of what the costs were. Guess. Only electricity or... Well, the whole training. The whole training.
Sebastian: Yeah, but do we also calculate how much of the GPUs that training run killed? Nah, I guess it's maybe...
Tim: Well, I didn't come up with a number, so I kind of tell you, but... Guess. Good question. Wow.
Sebastian: I could be completely wrong now, but I say 350 million.
Finn: 350 million. 350 million. 350 million, yeah. What'd you say? Yeah? I thought it's way less. Let's give it like a 300 million then, if you said like, yeah?
Tim: Yeah, one more try. Yeah, is it way higher? Is it way lower? Yeah, come on.
Finn: If I ask like this, it's probably not lower, right? Let's say 600. Yeah.
Sebastian: More.
Finn: That's crazy.
Sebastian: Okay, then I will shoot for 800. I'm not going on the...
Tim: So, yeah, there is a reason why it is the most dangerous. model, so to speak, and why it's so intelligent, but the number that I saw from Reuters, and it was just an estimate, was 10 billion US dollars, just on the training run.
unknown: Ouch.
Finn: Well, okay.
Tim: The longest ever recorded training run.
Sebastian: Yeah, that is also the reason why... It explains a lot, though. The one trillion. Yeah. There was always that little, or not bottleneck, but always that star in the horizon. One trillion parameters, you mean. Exactly. Yeah. Labs wanted to train a model that big, because it costs that much money.
Tim: But that much? I mean, the second they had the results, worth every buck. Yeah, for sure. I mean, it found bugs in, we're not talking about, you know, small SME systems, but like gigantic multinational corporates. Yeah, forget about the corporation.
Sebastian: Think about operating system.
Tim: They found a bug in whole Linux. Yes. And the problem is, it not only finds one bug. But that's one thing about it. But the problem is, it is able to find so many, and then, you know, one doesn't get you far, or multiple doesn't get you far, if you cannot... Change it together. Yes. You cannot connect the dots. But what the model is so good at, is finding those, and then building the web between them, and see, okay, how can I use this exploit to get this, and this, and this, and then kind of like, consumes it all.
Sebastian: Super fun.
Tim: Super scary, but also super scary.
Finn: Super scary as well, yes. But to be honest, it's also a bit of, I don't know, Entropic seems to have done that quite a bit in the past, to kind of scare people into, and then, oh, we're selling the solution, by the way, only to those 40 companies that pay the most, but, I mean, they have to get those 10 billion back in their books somehow.
Sebastian: Do you know if the big corporations are paying to be in the class wing project? I would very much believe so. Because I read somewhere that Entropic wants to subsidize them, to give them basically, I think, a shit ton of credits, in order to lure them kind of in, in the long run, I guess. Yeah, I think not lure them in, but probably for more data.
Tim: But yeah, they are giving some compute of their resources, allocating it to the project. I'm guessing there is some value for them as well, 100%.
Finn: I mean, if it's not money, then it's data, it's the new money, so it's the same shit, I guess.
Tim: For sure. But very exciting news, I have to say. Yeah, and also, of course, also very interesting timing of Entropic, every time they release a new model, they have, like, these ground-breaking papers, which kind of have, like, a huge frenzy around them. Could also be a great marketing move. Yeah, I just wanted to say. That's what I'm saying.
Sebastian: They kind of scare people into a... No, no, I would say they're actually good in both. They're really good in training LLMs and building those systems, architectures, whatnot, and in marketing, because they know what they're doing nowadays.
Finn: And marketing is also really nice.
Sebastian: Yeah, super awesome. Also, with those ads against the ads of open AI and all of that stuff, so they're more going into that edgy direction and bothering other companies, I would almost say, and basically, you know, take that as their marketing, which is also super fun.
Tim: Fair enough. But we already see some of the, like, open AI, I think you shared this this morning, with open AI jumping on the train and also releasing... Cyber model, yeah. The GPT 5.4, was it?
Finn: I don't know if it's 5.4, like a 5.6 or something. I just read cyber behind, and they're definitely trying to tap into the same thing, like, but...
Sebastian: I think it's a 5.4 architecture.
Finn: Do you guys actually think that it's, like, a competitive model, or do they just... Did they, like, you know, quickly train, like, a specialized version that should be able to, you know...
Sebastian: I mean, that is the question, but...
Finn: They didn't really have the time to... Yeah, but I think the...
Sebastian: Or the thing is, with those LLMs, especially if they have now such a big training run, those training runs, they take time. Yeah, that's what I'm saying, they didn't have the time to... Most of the time, it's a month or longer, depending on how many parameters you want, and how many data, or how much data you want to put in, but they are always planning ahead, and I believe that there was already a project in OpenAI that was some kind of cyber-related thing, and they just now pushed it further, because you cannot not do any cyber...
Tim: Yeah, I mean, I also, I mean, I remember after the Code Red email from Sam, I remember they also pivoting towards, or not pivoting is maybe a bit too much, but I remember them...
Finn: I wouldn't even say scramble, like, what I see in OpenAI right now is that they're trying to find, like, a path, you know, like, how can we get out of this money kind of hole that they dug themselves in, like, they promised NVIDIA 500 billion or whatever, they promised this and that, and this and that, and they took so much...
Tim: And slowly, every now and then, you hear the news where it's like, oh, this and this data center is not being built anymore. Yeah, exactly. Yeah, exactly.
Finn: Project Stargate. Paused. Paused.
Sebastian: Paused. Crazy, right? Nothing. But they also didn't communicate, or at least I didn't see...
Tim: No, but they're trying to... Yeah, I mean, that's pretty much on purpose.
Finn: Yeah, of course, but not a big media or whatever. Yeah, no, it was Sam Altman's, like, kind of, only way of getting out of the... The video? Yeah, to kind of build that, what was it, giga... Gigawatt. I don't even know, gigawatt factory or something, not factory, data center. But no, it didn't happen. They tried to get SoftBank in and stuff and whatever, and...
Sebastian: I think they already had the deal with the SoftBank, but...
Finn: They did, but I don't know what happened, to be honest. I guess...
Sebastian: Yeah, but I think... It takes too much time with connecting into the grid, or not? I could be. Isn't that one of the biggest problems?
Tim: I think the whole thing, like, when they said, okay, a deal, then they also kind of stepped all back from that statement and said, okay, that...
Finn: Could be like a marketing thing.
Tim: The deal, yes, but we never said, okay, it's not like one bank transfer, you know? It's like a collaboration through the years.
Finn: I was also a pro, I would say. For sure.
Tim: But as such, Stargate, a pretty good initiative, in my opinion. I would love to see something like that in Europe. Yeah, please. And maybe more than two billion. I know two billion sounds a lot, and it is a lot, a lot, but looking at AI, it's time to get some more money involved, in my humble opinion.
Sebastian: Those AI companies destroyed the relation to money, for me, or not? I mean, if you... Yeah, I mean, not for me privately, but... Yeah, no, but if you talk about all of the bubble, I would call it, no? That's crazy. 100%. I got a new post from a really nice company, it was around about, I think, 20 million or something, and I thought, oh, only that little? Because normally, AI company, above 200 million, or at least 100 million.
Finn: Yeah, but think about what you can do with 20 million. Yeah, it's crazy! Of course, in AI, you need to buy stuff, but still, with 20 million, you can do a lot.
Sebastian: With 20 million, you can. You can do a lot. I would buy the biggest data container from TinyCorp. For example. And it would be... Five of them. Small, yeah.
Finn: No, it's interesting, but I think let's move on.
Tim: I have one more news that I saw this morning. Please go. Or even not this morning, but a couple hours ago. Oh. You guys know the shoe brand, Allbirds? I think... Do you know? I heard about it, but... I read it. It's like a very basic shoe. 10 minutes ago. Yeah, it's a very basic shoe, but they... So how do you connect now shoes with tech? Yeah, Wade. The pivot of the century, let's put it that way. The pivot of the century...
Finn: Don't tell me they're going to AI now.
Tim: They are. They are leaving everything behind. Allbirds says... It's moving away.
Finn: Do they make like high quality shoes? Huh? Do they make like high quality shoes?
Tim: No, it's more... I think like a basic... It's more popular in the US, but they're moving away from footwear and towards GPU as a service. What? AI native cloud infrastructure. Okay.
Finn: And...
Tim: That is the pivot of the century. And there's like a new brand called... Yeah. New Bird AI brand.
Finn: New Bird, they called it.
Tim: New Bird AI.
Sebastian: And I think they're stocked at a 12% jump or something? No, more. More? It's crazy. Just because AI is seed round GPU. Put AI in the name and better than shoes. But it's another hyperscaler and they all have the same problems with the GPUs. You cannot hold a GPU for AI usage for six years. It's done. After four. Especially if you run good training runs.
Tim: Yeah, 12% was a good estimate. What was it? Wait, let me... So today, guess! Guess!
Sebastian: Guess! Guess! Guess! First of all, I want to know their seed round. How much money did they raise again? Oh, I don't know. I don't know.
Tim: I don't think they raised anything. They didn't?
Finn: I mean, if you say it like that, let's give it like a 20, 25.
Tim: No, I still will stay with 15. So they did the pivot of the century and it's 25%? I don't even know. I mean... It's AI, bro. How bad was it before? Come on!
Finn: Number! Hey, let's give it like a... If you say it like that, let's give it like a 50%. Ah, 35. That's more.
Tim: Okay, I will start at the back, okay? Mm-hmm. Go on. Zero. Zero. Six. 600%. Nah, in one day. In one day. It's right here. And it's still going up. Yeah, let's short that shit.
Finn: To be honest, it was nearly a penny stock, so...
Tim: Let's short that shit! I mean, it was a penny stock for sure. It went up... Yeah, 607% until now.
Finn: So whoever's listening and didn't invest in... We all missed out!
Tim: Well, nobody could have known.
Finn: Life is life. Life is life. Well, that's actually interesting. Pivot of the century.
Sebastian: Good name for the episode, actually.
Tim: Yeah? Pivot of the century. Pivot of the century. Yes, we also decided that we want to try to always find the name within the episodes. So, this is going to be a tradition from now on.
Finn: So, now it's Pivot of the century.
Tim: Okay.
Sebastian: Pivot of the century from shoes to AI.
Finn: Damn. Phobosopher.
Sebastian: Sometimes my brain still works, huh?
Finn: I'm crazy. Yeah. So, my topic. Go ahead. It's not a big one, actually. It's a quick one. Yeah. It's just because you mentioned the EU, the AI Act is actually now... I mean, we all heard about it. I don't have to introduce the AI Act, right?
Tim: No.
Finn: The EU AI Act? Yeah. Yeah. It's basically now getting... Lower? Like, enforced in a way. Like, yeah. Basically, in six months, everything, all the AI systems that are connected to healthcare, to legal, to employment, everything that is privacy-wise relevant, is being enforced in six months. And it starts today, but in six months, it's the deadline, deadline. And the fines are actually quite... It can be painful. It's... Yeah. It's around 35 million, if you make a mistake. As a big company, yeah. No, no, no. 35 million, or 7% of your yearly global revenue. Not even profit. Revenue. So, 7% of a major company in Europe, or even a US company, if they have a problem, then...
Sebastian: Wait a year or two, then you're here. It's like the dot com again, with all of those data, what not, all of the big companies, Meta, Google, they all get a fine in a few years from the EU with the AI Act, and then they pay some money again, and that's it.
Tim: I mean, the thing is also, like, sure, the EU AI Act is as such, okay, morally, it's a good decision, but competitively, I think it's a death trap. I think what we're doing there, especially in Europe, I think... But it has a good side and a bad side. Sure. I mean, I see the bad good side.
Finn: All of my data is being sold for free to anyone, you know? I'm also happy that we have these privacy first kind of rules, but of course, business-wise, it's...
Tim: 100%, but what's going to keep us above the water in 20 years? The morals, or...?
Sebastian: I would say we will find out in 20 years.
Finn: We'll find out. I don't think AI is going to be 100% of the European economy, also not of the US economy.
Tim: Yeah, definitely not with what we're doing, and I think we're definitely not going to
Finn: be there. Yeah, but also not the US economy, it's going to be 100% AI.
Tim: Yeah, but I personally think we're not going to be able to compete in the big steps they're doing it. No. Yes, and that means we will not be free of their models and their compute and everything.
Finn: It depends if we finally pivot to local AI.
Sebastian: Okay. No, I also read something this morning really interesting. EU wants to fund Frontier AI labs. They are now actively looking for, I think, 10 or... Four years late, but... Yeah. 10 or 20 teams that get together, or basically not get together, but have one job to build another Frontier AI lab in Europe. And they are getting 24 months funding, so fully funded, and after those two years then...
Finn: What is fully funded? What does fully funded mean? It has to be like a...
Sebastian: They have, I think, 125 million dollars in the fund all together, and that is... All five? Mm-hmm. No, for all 10, I believe. Yeah, I mean... Yeah, but then...
Tim: Again, the idea is great. The idea is great. Yeah, I saw it today.
Sebastian: Okay, so let's give them the benefit of the doubt, let's say they are doing actually something... It is a good step in the right direction. Because then after those 24 months, they said that the best team, or the best two teams, three teams, whatnot, they get a billion dollars of funding. There is, of course, Arcturix in there, and you only get a billion dollars of funding over 25 years, probably, and in that rate, and under that condition, but at least they're waking up now and trying to push further. Yeah, no, I wouldn't call that waking up.
Tim: Yeah, okay, but they have to say somewhere, I guess.
Sebastian: I think that it's like... Still sleepy a bit, but...
Tim: Putting something out there... I'm very skeptical. Yeah, let's... I had a discussion a week ago about this. I really hope it works out for us, and I will try my best to be a part of it, but... Yeah, we'll see. We'll see. I don't know. I'm not sure if your regulations... Like, there needs to be regulation, but the way we handle it, and then comparing it to all the other... It's kind of like, you know, you need to win a rate, or, you know... We don't need to win it, but we kind of need to be staying...
Sebastian: Under the top three, at least.
Tim: We should stay at... Yeah, but if everyone, you know, gets in the car, and you're the only one running... Probably take... Hard to win the race. Yeah, probably need to take the car as well.
Finn: You know about it. No, let's see. But...
Sebastian: Let's see. Yeah, let's... Let's monitor the situation, let's see how the... We'll come back. AI act is enforced, and how the frontier AI labs in Europe will develop, and then...
Tim: Yeah.
Sebastian: I'm hopeful.
Tim: Yeah. Let's see. Me too. But I'm also realistic.
Finn: Yeah.
Tim: Let's see. I'm definitely hopeful.
Finn: You guys have something of a stack of skip? I do have one or two things. Go ahead. Let's start... Actually, I have three interesting ones, I would say. The first one is a quick one. It's Cursor 3.0. I just wanted to mention it because anyone that uses Cursor, they rebuilt the whole thing from the ground up, and now it's basically not based on VS Code anymore. I'm one of the only ones that it's not. And...
Sebastian: Cloud, rebuild VS Code. Yeah. Don't make mistakes. Thank you. Probably. Probably.
Finn: Do you guys think that's a good idea? Or... It's mixed opinions on it. My personal opinion is... It's okay. I don't know. It's a... What they're doing now is they're fully focusing on agents first. Like, they really push the agent window. They really don't really want you to work in the editor anymore, but to do everything through the agent, which of course there's also more tokens that gets used by them. But... No, I guess... I don't know. It depends if you're like a fully... Like a developer developer, then you're really... I don't think you're interested in it. But if you're more like a vibe coder, then... Of course. It's good. We're not...
Sebastian: Yeah, but I feel like they're spawning right now... For so many new IDEs that are no IDEs, but basically just a window manager for terminals. You are fair. Yeah. I don't know what I would... Superset. Yeah, yeah. Superset. Superconductor. What not. But... I tried it out and... I don't know. I actually like the feeling or... I like to see my code.
Finn: Yeah, no, same. Yeah, the techie vibe. Yeah. For everyone, I guess.
Tim: I think it's the same with the anti-gravity agent manager. I think anti-gravity made a lot of mistakes recently. Yeah, I mean, I'm not sure where my tokens go. A lot of stuff doesn't work anymore. When I do like one click, and it's like, hey man, you exceeded your...
Finn: After like two prompts. But at the same time, it's... I have a lot of errors in anti-gravity. Also commands that don't run anymore.
Tim: Yeah, true. Same.
Finn: I thought that was my PC, but... No. It's a universal problem. I don't know.
Tim: I don't know. Then they hopefully push an update soon.
Finn: I hope. Yeah, but... Composer 2, gotta give props to Cursor. Even though they stole Kimi-K and they kind of stole Opus, but whatever. Actually fun. It's a good model. It's a good model. It's kinda cheap. Kinda compared at least. We talked about it last week, right? Or the week before. I don't know if we mentioned it. We mentioned it. Yeah, we did. But still, Composer 2 is still a stack. It's really concise. It follows exactly what you want to do.
Sebastian: Yeah, I like it as a day-to-day model to scour the code base to find out what we're doing there. How we can plan out that. But at the end, I still use Sonnet.
Finn: Yeah, no. I actually use mainly Composer nowadays. It has a nice mix of reasoning. If it's a quick task, it doesn't reason at all. It just goes through done in a few seconds. It's really quick. If it's like a deep kind of complicated task, it really reasons for like a minute and then does the whole thing. But it doesn't like start doing something. It's like, oh wait. There's another thing and then everything is broken at the end. So, I like it. Try it out. Fun. But even more interesting, I didn't hear of it until now. Perplexity Computer, which of course we all heard about, they tapped into taxes now. I don't know if it's only in the US. No, I saw as well. Accounting and taxes.
Tim: Yeah, which is interesting. I think only being launched in the US first. Makes sense. But then being rolled out here as well. Makes sense. I think that's going to take a little bit more. But I think that is also an interesting topic. Also a pivot. Not of the century, but major pivot. I feel like they just changed the whole UI within one day.
Finn: Fully focused on computer.
Tim: Yeah. Fully focused on agents. So, just to get everyone catched up. So, Perplexity has their own agent suite, I would call it. Right? Like computer use, you mean? Yeah. Computer use.
Finn: Yeah, it's like a fully usable computer. Like an agent that can use a computer like a person.
Tim: They call it computer, but it's basically like an agent.
Sebastian: It is an agent with a few tools and one of those tools is basically computer use.
Tim: Yeah. But it sounds very interesting. Quite expensive though.
Finn: To be honest, I just read that at least the taxes part should be included in the pro subscription. I don't know. For now, I couldn't even test the computer use.
Tim: But the pro subscription for a computer, I believe. And that one I think is 200.
Finn: I don't know if that's the difference. Yeah, okay. I didn't.
Tim: I don't know. But I also saw that they now also have the ability or you have the option. There we go. Option to just load tokens or credits. You can just connect your card. Yeah.
Finn: You can do whatever you want. Yeah. I know. But it's still, I don't know. The usage is quite high. Exactly.
Sebastian: Yeah. But before we go on from that, did you guys hear about what the France government did? No. Because they are also right now the tax season in France, I believe, or a few weeks ago. And the actual government just released it. I don't, I cannot remember if it's an MCP or skills to do your taxes. Nice. Crazy. An official thing from the government. I mean, if I was the tax authority. Crazy, though. I think it's a skill. Crazy. And a lot of developers looked at it and said that their skill is better written than most of the AI companies' skills. Yeah. So, really nice.
Tim: I'm pretty sure it was in collaboration with Mistral. Oh, that could be it. That could be it. That could be it. Smart, smart, smart. Because they also, I think two weeks ago, announced a partnership. I love Mistral. Super nice. That Mistral is building same as the Maven defense operating system of the US military from Palantir. They're now on a, let's not call it a competitor because I think we're quite far from that. And there we then are at the European problem, in my opinion.
Sebastian: But something like that.
Tim: But something like that of integrating AI use in defensive, not offensive, but defensive purposes. Crazy.
Sebastian: But Mistral, super nice little lab. I know I work with it. No? Oh, I love it.
Tim: And I know them and so on and I follow them.
Sebastian: I use a lot of their OCR, so the object thingies. And their model for OCR is really good, really good. But nowadays there are a lot more local OCR models. But I just like the company because it's European, French, nice small. They have a nice design guideline. The models are fun. They are more concentrated on actually the good parts about AI rather than scaling to massive things. Yeah. And focusing more on little models, edge use cases and what not. So, shout out to Mistral. Super comfy.
Finn: But at the same time it gets easier to train your own models locally week by week. Yeah. That makes it also interesting for different normal users. Wait. It's also actually in my stacks. Hold up. Megatrain. I actually haven't had a look at it. It got released yesterday. But it's basically a framework that you can use locally on your whatever. You need one GPU. That's the interesting part. It makes it possible to actually train models on it that have more than 100 billion parameters on one GPU. Which then they export the memory usage and whatever. But it works. And I want to have a look at it. I don't say too much because I don't know too much now. But it's one of those massive barriers that local kind of smaller players still had. Like obviously you don't have a data center to build a model or train one. But now you, in theory, can. Quite cheap compared. You can take a lot of those gamma models, open source models and train them.
Sebastian: That was just my question. Is it focused on fine tuning or can you actually...
Finn: I suppose it's fine tuning but I don't know.
Sebastian: Yeah, it would make sense if it is. Otherwise I think on the 3070 you cannot...
Finn: I also think with one GPU they probably mean like a 5090 or something that has a bit more VRAM than... But it is fun.
Sebastian: I actually watched a YouTube video yesterday again about the RTX 6000. That shit was 96 gigabytes of VRAM.
Tim: Maybe catch everyone up what that is. Those weird numbers you just enter cryptic letters. Quite a nice graphics card. Yeah, it is.
Sebastian: I would say the highest end. It is the highest. Yeah, exactly. But it also costs 10k right now. Yeah, yeah, yeah.
Finn: I mean you're sitting on a gold mine with 100 gigabytes VRAM nearly.
Sebastian: That's crazy but I don't want to say anything but my framework desktop also has 105 gigabytes of VRAM. Okay, dedicated to the iGPU. That is something else but on paper I have more than the RTX 6K. But yeah. Let's... Let's... Let's circle it back to the normal topics again.
Tim: One thing off top of my mind what you should also not skip is GitHub is currently asking all its users to opt out of the training. They're not asking you to but the... They're giving you a little hint. They're giving you a little hint, yeah. A little banner. But if you're using GitHub and you have your code in there make sure to opt out otherwise all the models will be trained. In a few days now?
Finn: Yeah, I think it's...
Tim: We're getting I think 25th of April. Top of my mind. Yeah. But you should opt out if you like your code and you don't want anyone to train.
Finn: I mean if it's public code then it's definitely... It's already in there. It's used for training anyways but... True. Yeah, private codes. Yeah. Very true. Yeah, but those were my biggest stacks and skips and actually my biggest topics. You guys have anything else for this week or...? I have another...
Sebastian: I have a little tool for stack or skip but it's freaking niche. I think it's from a company called butter.dev, butter.something. Super small but super fun project. They're building basically containers or not containers. What is Cloudflare doing? Cloudflare. Well fair. Sandboxes. Thank you. Sandboxes. And they're basically building or focusing on building local sandboxes that spin up in under 500 milliseconds. Exactly. Exactly. So you can basically just create a bunch of sandboxes and execute code from the internet or whatever inside of those little sandboxes and then destroy them again and only take the actual data that you need.
Finn: I mean that's what Cursor etc does.
Sebastian: Exactly. The fun part is I haven't heard of one that is local, local because normally you always have to send it to an API. It's Cloudflare. Exactly. And not locally on your thing and I didn't know or I didn't want to spin up a Docker container because it always takes so long and they build everything in Rust if I'm not correct. So it's super nice and fast, lightweight and it's definitely something for our Odin.
Tim: I think this is a perfect stop to or point not stop but to stop the episode for this week. Or do we have anything else we need to bring?
Finn: No.
Sebastian: What's my biggest topics? I just have one thing. Let's go ahead. One more. Mintlify. Oh yeah. 45 million Series B fundraising. Mintlify is basically the underlaying knowledge management thingy. Basically…
Finn: They do documentation.
Sebastian: Exactly. Your favorite company has their documentation on Mintlify. And now they found out or I guess they already looked at it a while. But now the threshold crossed. So now it's 50-50 on what basically reads those documentation. It's 50% AI agents and only 50% humans now. I can imagine. It's still a lot. Yeah. A lot. I feel like 50% of pure agent traffic is already good. Sure. Yeah. But the fun part is they are now taking that money to focus on more the knowledge base. Because it cannot be that we're actually just sending MD files with skills and whatnot around to different agents for different use cases. No. But rather something unified that basically the agent can tap into the runtime to get their information, to get the context.
Finn: Totally perfect. Yeah. Yeah. Yeah. Yeah. Yeah. So many tokens being… Exactly. Wasted so much.
Sebastian: And I'm super excited for the future about how knowledge is actually getting stored, distributed and used in AI agents.
Finn: I don't think it's going to be the way that it's going to be a trillion parameters model that knows everything. No, no, no. I think it's actually going to be specialized models like a team of agents and then… Yeah. Yeah, companies like Mintlify that kind of manage the knowledge around it. Yeah.
Sebastian: The markdown files are basically infrastructure nowadays.
Finn: Yeah. That's definitely the more efficient ways than Mark does.
Sebastian: Yeah. So super excited about that announcement. I'm definitely going to… Yeah, that's cool. …look more into that. Great. Now I'm done.
Tim: Alright then. Yeah, I think we're done. Cool. Then… Thank you very much for listening. Make sure to subscribe, follow us on Instagram, on Spotify, on YouTube, everywhere where you can find us. Check us out. If you haven't grabbed your tickets for our AI Hackathon here in Harlem, June 6th, it's open for registration. Make sure to get your tickets and see you there. Until next week. Keep building. Keep building. Peace. Bye bye.