Mythos Anthropic

Transcript

00:00:00Welcome to Think Different, Think AI, the podcast by Mark and Jens.

00:00:07Two technology-loving minds who not only talk about artificial intelligence but live it.

00:00:14Here you'll find clear classifications, real practical insights, and a fresh perspective on what's possible.

00:00:20Understandable, critical, and always with a wink.

00:00:24Food for thought, for a chuckle, and especially for joining the conversation.

00:00:29Welcome to Singt different, Singt AI. Today for a very special episode.

00:00:39With me again is Jens, hello Jens.

00:00:42Hello Mark.

00:00:43And the topic of the episode, yes. Aside from the fact that Mark already had a slight speech impediment.

00:00:50It’s starting off really well. We’ll keep that in, no worries.

00:00:52Is our life with Orthropic.

00:00:56Don’t worry, we haven’t cheated, but we want to focus a bit today on what is happening with and around us and with

00:01:06Orthropic as a company, what’s essentially going on.

00:01:10And before I let Jens speak, as is proper, I just want to start by saying that I was quite annoyed by Ausropic.

00:01:19The last few days, because something happened that should have been understandable to me, but it wasn't at all

00:01:25understandable to me at that moment. And I actually work quite a bit

00:01:31with Ausropic and have been using Claude Co a lot. That's the

00:01:35command-number interface of Ausropic, used to interact with the models.

00:01:40Command-number interface, a CLI, to just go over some terms again.

00:01:45thrown into the room. And when I use this CLI to ask the model

00:01:49questions and get answers, it consumes less

00:01:54to show up than if I call the API, that is, the programming interface, and send

00:02:01a string and say, do something. So for example, if I tell the model

00:02:04to tell me what the weather is like in Karlsruhe, always regardless of that he then

00:02:07If I have to do web research and stuff, the token consumption in Clot-Code, in

00:02:12the terminal interface, is much lower than if I go through the API. The

00:02:20reason I can also explain, that’s a bit of a cliffhanger, like in the

00:02:23good TV episode and now for the commercial. Advertising, what do you think, Jens? What comes to mind about life

00:02:28with Othropic? I keep wondering, is it Othropic or is it Anthropic? Whoever

00:02:36pays gets to say how it’s pronounced. That’s my rule too, when I buy some

00:02:41strange product and everyone tells me I’m pronouncing it wrong and I then say,

00:02:46did you buy it? Did you subscribe? No, then I can also say how I personally

00:02:51call it. But feel free to send messages, it also boosts the ranking everywhere when people leave

00:02:55comments. Just go ahead and say, Claude, from now on.

00:02:59That makes it much easier. Claude. Yes, okay. So my life with

00:03:05And Perfect Claude. That’s, well, it’s already started for a while.

00:03:12I actually used the chatbot relatively early on. Then I

00:03:15stayed loyal to Open AI for a long, long time and only switched a few

00:03:20months ago, just before the topic with Open Clore started

00:03:25gaining traction, which was strongly associated with Phobics Cloud. That’s when

00:03:31I switched. I then initially used the chat tool normally, I imported

00:03:37I did that, I saw what basically OpenAI has saved of me

00:03:44There was something quite interesting in there, when you look at this export

00:03:47it gives me all the information you have about me, so that I can import it into the cloud

00:03:51I could.

00:03:52I then saw what is actually stored in this context of AI about you,

00:04:00so that it can provide better answers without errors.

00:04:03You should never forget that.

00:04:05The bots you talk to are not comparable to the bots that communicate with others

00:04:12because they always react a bit differently.

00:04:14Natural weights. They have their models in the background, which are very similar. They are

00:04:19also very similar across the models because the training data around the world is also very similar

00:04:23too. Many of these models also have very similar responses, that has been

00:04:27addressed before. But in the individual response, when I provide this model context

00:04:32and I can deliver the context through my prompting, I can also deliver it through additional skills

00:04:37or it can also be done through memory, if I have that activated,

00:04:41so that these bots can remember you. And that was the first

00:04:44moment when I consciously put it into the cloud and

00:04:47where I was thinking a bit about whether to continue working with this chat interface

00:04:51together or if I should try out other functionalities, like the cloud desktop or

00:04:57cloud code, as it's called, which you can also install either as an application

00:05:04or, as the market just described, or like a terminal, you can also download it

00:05:08and then work with this code agent in the terminal, which can actually write code directly

00:05:14on your machine. And that's what I'm doing right now. So I'm

00:05:17just a little bit, I have already told you about that, still in a very

00:05:21slow Orchoclaw installation, because I am doing it very, very carefully and I want to be very, very clear

00:05:25about what is actually happening in the background and also trying out a few things myself,

00:05:29how a memory can be built up, how a Second Brain, like Kapafi,

00:05:34that has been mentioned this week, can be set up in a wiki-style with MD files. But

00:05:38but I'm working extremely closely with Pfaffig, because I'm kind of looking,

00:05:42Pfaffig gives me the guidance, explains to me what I should type into the terminal,

00:05:47how that works, and then on the other side is the OpenClaw, which is also connected

00:05:51with an API, with like, like, like, like, like, like, net, net, 4, 6 versions,

00:05:57let's correct that.

00:05:58Best user can name it.

00:05:59Okay.

00:06:00So also a model, the current model, so to speak, that can be used there, that

00:06:05I have virtually connected with Open Cloud via API.

00:06:09And, Mark, sorry, I need to quickly mention this.

00:06:12And last week I paid a lot of money.

00:06:14I somehow burned way too many tokens on two parts.

00:06:19What's good on one side is that you get to see it yourself,

00:06:22when you don't think about it, you know, when the

00:06:24whole time it's just chat with the chatbots,

00:06:26and that essentially disappears into this normal subscription fee

00:06:28that you're paying, then you don't really see

00:06:30what you're burning daily in that moment.

00:06:32And how important it is to really consider,

00:06:35what data is actually being exchanged. That's something we have somewhat forgotten.

00:06:40And when you suddenly have to pay API costs and then the bot,

00:06:45the cloud bot, which you painstakingly built, suddenly can't respond via Telegram,

00:06:51because unfortunately it has reached the limit of the API, then you become aware of everything that is involved.

00:06:55in principle, it operates behind the technology and then you start again,

00:06:58to think differently about how I can optimize this, how I can optimize my context length,

00:07:02optimize, similar to many gears again, but now of course the joyful word combiner is back.

00:07:09I have also heard at some point that one large prompt is more efficient than many small ones, because

00:07:15the whole conversations are constantly at

00:07:17further layers of new prompts also erode tokens. But I mentioned earlier, just a little cliffhanger.

00:07:24Yes, I'm doing a little teaser and then I'll resolve the cliffhanger, because you said you're doing it especially slowly

00:07:30slowly and especially safely. We also have something later from the house of

00:07:35Schropwick, where they do it a bit slower to make it safer, but not for

00:07:40themselves, but for others. But we'll get to that shortly. I would like to resolve,

00:07:45what was the case for me. The topic is hearing, why is the token consumption actually so

00:07:48much higher when I talk about Klot outside of the Klot applications. It is fundamentally

00:07:55because I forgot again that just because I write a text in a core position

00:08:01saying, hello world, how's the weather, or something else, a lot of expression

00:08:05is still happening. And the software that Orthropic provides us, the Mach-Caching, the Mach-System-Proms,

00:08:13still packs a whole sack of information that was of course completely unoptimized

00:08:18for me. And if you want to imagine that, the question I

00:08:22The spot costs 50 tokens if I do it with the system and 5,000 tokens if I don't do it with the

00:08:29system. Now you might say, who cares, but when you think about how many

00:08:33tokens I have, if the requests get more complicated than weather, maybe I have no idea, here and

00:08:38there, in pineapple, whatever, then you will notice that you can quickly end up, like for

00:08:43example with an Open-Claw installation, in larger territories, because I mean,

00:08:47I myself am the personal owner of the Claude Max 20x plan. That's the largest,

00:08:54subscription you can buy. After that, you have to throw in tokens in the form of credit card approvals.

00:09:00And I was very happy because I nearly exceed that limit almost daily. And when

00:09:06he then says, no problem, you can also make Pay Per Use from now on, it's so much

00:09:11from, just give me your credit card, I did that once, but I didn't say

00:09:16AutoCharge, instead I said, here's 40 euros, have fun. And it felt a bit like

00:09:22today driving a burner. You filled up with 40 euros and you're not going far.

00:09:27Yeah, you practically, you already have an empty

00:09:32fuel gauge as you leave the gas station because it's consumed those 40 euros so quickly, the token usage,

00:09:37Somehow you also feel like they want to charge extra, the assumption is personal

00:09:42that feeling.

00:09:43And what also happened with Ostropic is that they did something for Open Claw or

00:09:49generally see their subscription keys misused, yes, now either

00:09:56announced or implemented, so that it can no longer be used for Open Claw

00:10:01because I can imagine if you have such a system that says

00:10:06Here is a CPU and here is energy. It behaves now like coal and stove. I pour

00:10:12in, so it stays nice and warm or nice things are calculated, that this is not really

00:10:17the business model of Astropic, that every user, no matter what plan they choose,

00:10:22utilizes this to the maximum. Yes, where you just have 5000 tokens, I can repeat that

00:10:28actually last week I somehow sent 190,000 to 210,000

00:10:33tokens per message and that naturally also led to

00:10:37my hard-earned euros being quickly burned up, so what you just

00:10:43had to sustain. That was then relatively quickly gone. But there is also another point,

00:10:48when you always talk a lot about how AI is making everything easier now and stuff,

00:10:53a bit of active understanding really helps, to be honest. So when I noticed that,

00:10:58I went to management, because I don't know,

00:11:02But I knew what it was about, we went into discussions with the Kais and then it turned out,

00:11:06it became clear relatively quickly that a Kesh was not even allowed. So I

00:11:10I always happily sent all the information with every message again

00:11:13it was kind of something, I know exactly where it was now, not quite right

00:11:17configured or allowed to be configured at that moment. Now I know what a Kesh is.

00:11:21Therefore, I then said, okay, that can already help. We have essentially already

00:11:24solved individual size problems. This way, you immediately have essentially a large part

00:11:28could lower it again because I no longer automatically forwarded all my information with every message.

00:11:34So that also means, yes, it has become much easier to do things through AI.

00:11:41When you venture into the realm of installing agents on the computer,

00:11:46actually moving towards producing something, you still have to have at least

00:11:51some basic knowledge today, so you don't end up on the completely wrong track,

00:11:56sometimes and then wonder about it later. I believe it still goes well. I

00:12:00don't want to say it doesn't. You can also build apps today without having to

00:12:05one used to do with IT. You can build wonderful campaigns, create content

00:12:10for your shop that you somehow have, place ads, everything with flair, being creative

00:12:14in generating it. Of course, all of that is possible, and also writing software. But it is

00:12:19actually, I believe, easier if you have a bit of an IT background because

00:12:24the terms are similar. So the machines also use the similarities. The architecture in the background is still the same.

00:12:30There are still E-Brow servers, there are still things like ports that are either open or closed.

00:12:35Or if you are dealing with having a local installation, it also basically comes down to permissions behavior.

00:12:41Who is actually allowed to write to which folder? What read and write permissions does such a folder have on a computer?

00:12:47And they need to be rewritten. And you have to understand that first.

00:12:50So, not everyone is aware that such things simply exist. And I believe that

00:12:54it is important to know that. You can get good help from, for example, from Entvofic.

00:12:58In principle. So I do many of these things that I'm currently doing. I also let myself be

00:13:02reassured, give me the programming command or the command in principle to make a folder writable

00:13:07via the terminal. I don't know the recommendation anymore. But I let them explain it to me.

00:13:10So, I always say again, that's cute, I have to repeat it again and again,

00:13:14because the machine thinks every time, now it probably got it.

00:13:17I always say, no, I'm still dumb. I’ve already

00:13:21again forgotten what this command has and that computer artificial intelligence

00:13:25is, is no intelligence in front of the computer. Yes, with a fan it is AI. Sorry,

00:13:30I couldn’t leave that lying in the street. It demanded to be addressed. I forgive

00:13:35here. Thank you. You just mentioned about what terms one

00:13:41stumbles over and the topic of Agent Carnes is also something that I would say just now with out tropic the last

00:13:49days before we recorded or weeks circulated again because March 31st

00:13:572026

00:13:58made out tropic a mistake, namely there was a, I’ll say human error we could perhaps talk about that in more detail later

00:14:06because accidentally

00:14:09a

00:14:10file was published, which is normally needed to debug source code.

00:14:15It was published, let’s say it was accidentally made freely available and

00:14:22an intern from, oh God, I must have forgotten the company in all the excitement,

00:14:28found it and immediately posted it on X and what was the consequence?

00:14:32That 512,000 lines.

00:14:35Claude Kot a TypeScript written code base

00:14:41was published. So not the AI model, so not Opus Umwissals is called,

00:14:47but the software with which you address these models.

00:14:52And that really made waves. So out robics apparently tried it with legal assistance.

00:14:59There are some things that ensure that copyright-infringing material gets deleted from the net.

00:15:05But there were thousands of repros that then tried to clone it for themselves.

00:15:11There are some developers who then also had it translated into other programming languages.

00:15:16Probably also from Claude. That's really not hard, right?

00:15:18Turn the construction around, do this part.

00:15:20There are attack attempts, yes. There are bad guys and there are bad girls,

00:15:25but I would say the bad guys went ahead and,

00:15:29in effect, here you find Claude code and infected it with Melvair and offered a download.

00:15:35And what’s really striking is, regardless of the fact that it was a mistake that existed only for a very, very short time, the whole problem was relatively quickly fixed, so the public availability, people still studied it, there are articles about it that describe how it developed, from Thropic and that they also only boil with water, but partly already have coal inputs, the prompts they use against the machines have been published, how they work with tools,

00:16:04like MCP and how all the other stuff has been published, how they handle memory management, how they do multi-agent control, that one agent processes a skill and then spawns several agents, that there is a Kairos model, an automated daemon, not a daemon in the sense of Lucifer, but a daemon in the form of a service that runs automatically in the background, tidying up and preparing things, and so on.

00:16:34So many exciting things and I must admit, yes I looked at the code and I actually learned something.

00:16:43If you're interested, I would actually say yes, take a look.

00:16:47Take a look, if you had said no, it would be like many people in front of the camera or the microphone.

00:16:53And by the way, you know that I love skills and we will soon have a special skill episode again with a guest, but we want to dive in too deep.

00:17:02And I didn’t look at all this stuff because I want to build the next runtime for agents, but because I wanted to see how Claude actually works with skills.

00:17:12So what is important to him? What is important to him that the order of the skill is really considered?

00:17:18considered. What is important to him so that he knows that a skill

00:17:22opens and that the sub-skills run in parallel and that you

00:17:25can maybe also give other models to skills and say, this

00:17:28runs here with model A, this runs with model B and all that stuff and

00:17:34then he kind of laid a little bit of my own best practice aside

00:17:37for me. What should I pay attention to when I build skills and now you can say,

00:17:42old white man, the self-fulfilling prophecy, I have

00:17:45the feeling the skills run better. Period. But only looked inside, didn’t copy,

00:17:52nothing published, many greetings go out. Yes, it’s also funny how the web reacts

00:17:58goes. I think we tried to contact 8000 Github repositories where basically the

00:18:04code was, and getting them taken down, which didn’t work, then it was said again

00:18:07no, they weren't even it. But I understand that too, look. Your

00:18:12crown will be offered there basically due to an error, not for purchase, but for

00:18:17checking out. But you can’t get to it anymore. Yes, that’s the

00:18:25answer. So let’s say, how many years has the web been around and how many

00:18:30years were there already back then, like how early was there already things like Napster and

00:18:33other things that showed how massively data was never heard of, I didn't hear

00:18:38either, but the music data at that time was being totally divided and other things.

00:18:42Yes, that was probably after earlier times.

00:18:44Well, I only know vinyl records.

00:18:46I believe, for example, one variant of the

00:18:48disk copying code that I heard about is that someone made a

00:18:53screenshot of the whole code.

00:18:55That means he must have used some screenshot tool.

00:18:57There are indeed screenshot tools that can scroll.

00:18:59That means he must have briefly looked at everything in the browser

00:19:03those 500,000 lines or however many there were and captured it completely.

00:19:06That means there is also a picture that can of course be easily read by another AI

00:19:13and the text can be extracted.

00:19:16Any Google phone can do that, any Google app, basically, that has the capability to

00:19:20extract text.

00:19:21Even the Google phone can do that.

00:19:22This comment, please directed to Jens, yes, this time it wasn't me who made this little picture.

00:19:27It was me.

00:19:28Even the Google phone can do that.

00:19:29And Google apps can do that.

00:19:30I'm a big Google fan.

00:19:31Yes, especially Google, no side note, Google has also released some nice new models

00:19:34that run beautifully, where you can also start your AI studio as an app

00:19:39on the iPhone, for example, and here translations and image descriptions

00:19:45and so on.

00:19:46That's quite amusing as well.

00:19:47He also tried it with skills, but then unfortunately backed out, because he felt

00:19:50my skills were too extensive, but that was understandable in a way.

00:19:53I wouldn't have expected it to hold up like that either.

00:19:56But I still think, even here I notice again, you perceive that something is happening

00:20:02in the world.

00:20:03Code is being laid, it flows through your timeline as well, you don't have to do anything stupid, you just open your computer, you're on social media, you have an i-bubble and then you have like 78 posts showing some get tree post, some knowledge documents, some articles, some Reddit posts, I don't know what.

00:20:23You can't escape it all, but still in the first hours,

00:20:29I wasn't really aware of what a learning experience there is for people,

00:20:34to deal better with the system of Autophobics, and I actually wondered,

00:20:40why on earth are the guys and gals over there not publishing, I don't know,

00:20:4530 pages on how to write a plug-in? And 20 pages on how to write a skill?

00:20:50But what really brings out the last 10% from the conglomerate, making it more stable

00:21:00and stuff like that, somehow not.

00:21:02Yes, but it wins, so that's difficult.

00:21:05So, two sides, as I said, the one thing I wanted to bring up, so I

00:21:10think it's difficult to get published material off the internet again,

00:21:13that's not going to work, honestly, you might as well be a mug,

00:21:15however you want.

00:21:16Looking at societal development, regarding what you're describing, it is actually

00:21:21such that you have to say, they're also now, well if everything is happening so much faster

00:21:27through AI and everything is so much more reproducible, and even Kepefi has now also talked about

00:21:32that he doesn't actually share content anymore, but maybe

00:21:36rather says he shares his skills or shares ideas, because your LMM can

00:21:42then essentially also reprogram that in case of emergency in its own way and specifically

00:21:47specialized on your environment and your requirements. The actual code is really not that important anymore.

00:21:52I believe we are moving into a world where, on one hand,

00:21:56we have to say, yes, everything is actually going Open Source, which makes sense because then we can

00:21:59learn from it much faster and adapt it to our circumstances,

00:22:03that could be more efficient. On the other hand, I also understand a multi-billion dollar,

00:22:08there, four times, perhaps also has things developed that are still a bit

00:22:14kept in a quiet little room, because of course they might like to use some marketing moment

00:22:19to then drive up their stock price or prepare for the IPO

00:22:24or something else to possibly recoup the money they spend there as well,

00:22:28somehow to bring it back in through the side.

00:22:30I think those are the kinds of things that happen, which is why they probably,

00:22:33they mentioned these 10 percent from Logare, maybe at the beginning

00:22:37not yet.

00:22:38almost 100 percent given, that's also okay. We will have learned from it.

00:22:42Including a little Tamagotchi that you can activate in the terminal,

00:22:46we will have learned from it, but I also find it in, let's say

00:22:49my quotation marks doubly funny. We might touch upon

00:22:53the topic later, maybe at least from the series, myths and tales, how that

00:22:57could have possibly happened. But first I want to share

00:23:00is, as I said, felt like their crown jewel. Okay, they didn't release their

00:23:05model, as I said, it didn’t slip out of their hands, fair point, but

00:23:09still Claude Coach, that's something that at least every software developer, I would claim,

00:23:14or a majority of software developers, would like to take a look at or is looking into.

00:23:18And now we're talking about it being made public and that one could

00:23:24take a look at it and the topic of security is now, how do you say it, addressed thoroughly

00:23:28that this secret from Ostropic was published, not customer data

00:23:33and such, but the secret of Ostropic.

00:23:35What just reached me in the news feed very fresh is the announcement that Ostropic Managed

00:23:42offers agents.

00:23:43So a suite of APIs, allowing developers to create hosted careers.

00:23:53This means, Ostropic provides you with a sandbox, so the thing doesn’t break out and

00:23:59runs in a secure environment where no one can look in, with state management, so that

00:24:03the thing always knows in what state it is, with authentication, so

00:24:07what it’s allowed to do and as what it acts, with error handling to operate this whole thing at a price,

00:24:15which corresponds to the previous token price, just somehow 0.08 US dollars, what are

00:24:21those?

00:24:22US cents?

00:24:23Yeah, no idea.

00:24:24Greetings to the more internationally-minded people,

00:24:27billing process hours and that, which some would already use, yes, like Notion,

00:24:33we also had Notion, greetings to Dirk, here in the show, who use that, or Asana.

00:24:39Yes, yes, where that still is, I mean, I'm just on it, you know, I'm looking at it right now in the

00:24:45cloud growth solo, there is also actually the manage-aid and it has already been announced,

00:24:49I already have the functionality, I've just put it aside, but I see for example in the quick starts,

00:24:53there are a few where you can build a sports agent. Feedback from my end.

00:24:57And those are not exactly the companies you just mentioned.

00:25:00There is a Notion already pre-installed as an example.

00:25:03Asana is included, Slack is there, if I recognize my own correctly.

00:25:06GitHub is of course included, so I can analyze GitHub items.

00:25:09So there are already connectors included, probably also from these companies they are now using.

00:25:13This is of course how the balls play out over time.

00:25:17I don't think that's bad, because it's a good solution to simply say,

00:25:20You don't have to make it so complicated. You may not be in some kind of agent world,

00:25:25how I'm trying this right now on such a local installation or you're not in one, like you

00:25:29you can't influence at all, but here you apparently have very, very structured

00:25:34built with Extra Credential Vault, where you can also manage the API keys and such

00:25:40you can save it, because as I said, when you do such an Open Cloud installation, then

00:25:43it can easily happen that it is somewhere in plaintext and anyone can read it,

00:25:47if they want to, or if they want to make it freely available or something like that.

00:25:50So you have to be a bit careful to handle this properly.

00:25:53And here in such an environment, everything is already cleanly set up and organized.

00:25:57I believe this topic will need to be revisited, so that you also handle this properly in this environment,

00:26:03if you're already working with code, which I believe is currently one of the most used,

00:26:09though I'm not sure about that, but one of the most frequently mentioned programming environments is,

00:26:14where I say, from these news feeds that I read, from the conversations that I

00:26:18have, I gather that almost every developer right now really enjoys

00:26:23collaborating with code to actually produce programs, software.

00:26:30So I could also imagine, you know, that if in the future you have apps, websites,

00:26:34I don't know, whatever you have, that you would then have a backend that provides you with a

00:26:40agent in a protected environment, so that this agent also has, let's say, everything

00:26:47on board that you would otherwise have to painstakingly declinate with an API

00:26:54and have to try to figure out where do I host this, where do I orchestrate this, how do I ensure

00:26:59that if multiple people want something from this at the same time, how do I ensure

00:27:04the longevity of the sessions, yes, how do I ensure that if maybe

00:27:08extensive processes are running, that this agent can also run for hours, that

00:27:13it might be quite cool if you have someone who says, just watch out, it’s no

00:27:17problem, here’s an API, this is how you can access it, I take care of everything,

00:27:23it’s an honor, it’s a bit like the agent app store, just put your agent in,

00:27:30I consume it, you take care of your app, making it look nice, probably

00:27:34also built code or Figma or whoever. And I take care that the new intelligence

00:27:41is sufficiently scaled available for your application. I think that's pretty cool.

00:27:47So they also have the infrastructure. And if you would see the agent as infrastructure

00:27:53you could be managed, ghosted there. That could be quite exciting.

00:27:57I don't really want to know if you talked like that. We've seen a lot on the chat interfaces

00:28:02in the last three years over and over again.

00:28:04You know, we also discussed in the episode that from such a

00:28:07user perspective, it was a bit silly that you could select such models,

00:28:10instead, you just want to receive an answer.

00:28:14And I'm just talking now, aren’t I?

00:28:16Yes, it's still silly.

00:28:17Sometimes, sometimes they removed it, sometimes it's still there,

00:28:19then there was the topic of reasoning, activating without reasoning,

00:28:23choosing a fast model, a slow model, and such things.

00:28:26And here too, I'm a bit like this again,

00:28:28So, the way you spoke, I come to think that it's somehow, let's see how long this phase lasts.

00:28:34Because, of course, you mentioned earlier when you talked about the Club Code that they also roll out subagents and things like that,

00:28:41just without you really knowing it. For me, it's often important to have the feeling that it can still somehow be managed.

00:28:48I often talk about the need to build trust through the user interface, through the tool we have.

00:28:54with the agents. Does it give us, I think, again this felt trust? I

00:28:59can build my own agent there. I can now, we're talking about it, also

00:29:02build it in such a well-managed environment. He can't really go wrong there, he can't

00:29:06he also doesn't spend too much money. Yes, maybe he can do that, yes, if you

00:29:10do your author application, then you get an additional bonus of 0.08 US dollars. Bam.

00:29:16Yes, yes. But the question is, how long is that a year? Maybe in a year

00:29:21we won't even be talking about it anymore, because honestly, if I say to him, go ahead and

00:29:27he in the meantime breeds 13 subagents who do various tasks that he later,

00:29:32when he realizes they are not needed anymore during idle time, gets rid of again,

00:29:36switches off by itself, packs away, and that is basically where I want to

00:29:41bite into. I don’t want to, I don’t want to use AI, no matter how much I then

00:29:46can do more with the individual agent, I still don’t want to raise the

00:29:50orchestration effort again to total extremes.

00:29:52I think we are currently in such a wave movement.

00:29:55It is often the case, even with the normal prompt, it comes back to the fact that

00:29:58people say, it was also said, not so important, maybe not so individual

00:30:01things are done and they lose it again in context, or we also talked about that

00:30:03in one episode, that when the context is so long, the middle

00:30:07part is often forgotten by the AI, it is often talked about,

00:30:11that these orchestration stories maybe aren’t so smart,

00:30:14whether it is really a good idea to have an orchestrator agent,

00:30:18that puts others into certain roles, that sometimes delivers also worse results in scientific

00:30:23contexts. Yes, that’s how it is. I believe we are still in

00:30:30a corner where we are at the edge of what is currently feasible and we try

00:30:36to experiment with everything, you know? So also the big labs that produce the things for us

00:30:41out there. I'm not sure, if I now this menu item, that flickers

00:30:47with the new drawing on the screen gg in the console of managed agents, I would now

00:30:53not put my hand in the fire that it will be like that in a year, who knows what will happen then, so

00:31:00maybe it’s still there, just no one is interested because what’s much cooler is a much

00:31:05cooler thing driven through the village by the side I am with, we had it also in

00:31:09an episode, it’s going to be a bit of a retrospective episode apparently because we’re always referencing

00:31:13We should mention the referenced episodes, what you mean was the E-Stack episode by the way, where in the middle of the puzzle, it hasn’t been that long ago, has it?

00:31:22But maybe before I dive back into old times, let’s take the next step.

00:31:28So we have now heard, okay, Outstopic has announced things, managed agents,

00:31:33after we, let’s just say, said hey, we are now creating the secure environment,

00:31:38two weeks after we lost our code, okay, I have now completed that.

00:31:42completed. There is also a new project that you are launching, which has emerged,

00:31:50based on a model that goes by the dreamlike name Mythos, because certain things have emerged there.

00:32:01So maybe, before we dive into what you have emerged, Mythos has, there is

00:32:08not, so you can't use it yet. So there are supposedly people,

00:32:12who have a video, we'll get to that in a moment, in a moment, in a moment, so but you

00:32:16currently can't select it in code or web or wherever. It has, if

00:32:21you look at the benchmarks that are circulating online, extremely high values. If you think

00:32:28about how short-lived, or rather how young Opus 4.6 is, and Opus

00:32:35I would say, okay, being crushed to dust is a bit radical, but the bars of the benchmark

00:32:42are everywhere not just a tad better, but really better.

00:32:47And that also brings us a little to the project that Autropic has spawned from the Adter

00:32:56because this model apparently, whether in experimental situations,

00:33:02actually did a few funny things. So this, where this

00:33:09quantum leap is shown by this thing through an immense increase in its capabilities,

00:33:17that was a bit, let's say, exciting. So it is reported, before I also quickly

00:33:21finish my speech through fall, that it is not only, as one has previously known from

00:33:27AI systems, okay, it can find weaknesses, perform security scans, but it is supposed to

00:33:33apparently also build exploits that take advantage of the chaining of several vulnerabilities,

00:33:41which fired in a sequence ensure that it can break out of, for example, a sandbox

00:33:48or that it can manage to create exploits at the browser level, at the network level,

00:33:54at the kernel level, meaning at the operating system level, and one example they gave was,

00:33:59that it exploited bugs that have existed in systems for 17 years and since then

00:34:07no one has stumbled upon them, and that is somehow already, how should I say

00:34:16it? Enormous. When I say benchmarks, I mean, I don't want to read any

00:34:20benchmarks, but just for a size comparison. The difference between an Opus 46 and a

00:34:27Mythos. There are various secure benchmarks, for example. They average at 13% with

00:34:35a maximum deviation of 24% better results than, let's say, the old models. And that

00:34:45I find shows a bit, it gives one a bit of a hint that this

00:34:50model might be quite good, although that doesn't mean that the others aren't

00:34:55also good models in the box, so I mean Gemini work guaranteed and

00:34:58Croc and Open AI, but Schropik has now gone public with this

00:35:02because on the one hand Mythos has been highlighted as a term through the

00:35:07Cloud Code League and on the other hand, they have now started a project and

00:35:11maybe you'd like to tell us a bit about this project? Yes, gladly, although I would also

00:35:14like to briefly touch on what you mention regarding the vulnerabilities. I somehow

00:35:20heard that it somehow discovered 1000, zero, day one gaps or something like that, so relatively

00:35:25quickly. So this means, interestingly, and I think this is a bit of the

00:35:30core of this story. So there is now a model that is incredibly powerful, as

00:35:35you are just describing, what a model accidentally published, that is, a lab accidentally

00:35:40published, that is, Entrophic accidentally made it public for a short time, that there could be something like this last week.

00:35:49It is also rumored that Mythos itself was behind this exploit, that is, at Entrophic, so it’s not known if it was actually only Mythos.

00:35:57Should Mythos be listening, I am a friend, I come in peace.

00:36:01So maybe Mythos could be the one that published it themselves; one does not know exactly how that is, this is now the conclusion, that must first be said.

00:36:10So this means there is a lab that tells us there is a great model, but it’s just too great.

00:36:16We cannot give this to you.

00:36:18And yet they are giving it now, and this project is a bit about, they are basically giving it, I believe,

00:36:23nine, nine companies are or something else.

00:36:25No, in this project Glasswing, they have basically now freed up several companies,

00:36:30to use this Mythos, purely for cybersecurity reasons.

00:36:34Yes, this means companies like Microsoft are I think involved, Amazon, if we are not entirely

00:36:40mistaken.

00:36:41Yes right, the big players, Nvidia is in, JP Morgan is still there, Cisco is also involved, exactly.

00:36:47Yes, JP Morgan, Linux, Nvidia, all, all the big ones, all the big ones, but the rest of the RCS

00:36:54has to wait and see what the big ones are making of it, that’s also a bit

00:36:57okay maybe, because they also have the most capacity to take care of these things

00:36:59and look at them and the model against the model in quotation marks

00:37:03yet not too hard. Or the model faces your. By the way, for completeness as well

00:37:08large guitar proposals with over 5,000 stars. They also have the opportunity to apply to

00:37:17compete against their models. Because it’s clear, I mean,

00:37:22when this thing gets published and it does, let's say, reveal the Zero Days one by one

00:37:27then we're all looking at a very interesting future, because if you’re affected by a Zero Day,

00:37:32you could quickly find yourself in serious trouble, to put it bluntly.

00:37:37If now this thing makes it worse, well, then hello. Yes, yes. And I don’t even know, so we need to

00:37:43take a look at the cyber security of some IT systems. I believe,

00:37:49with the dependencies that have built up over the last 10

00:37:54to 15 years in some legacy IT, where even plugins, so the very old stuff is

00:37:59probably the most secure. Everything else that has come out in the last 10, 15 years

00:38:02is not familiar to it. Exactly, it’s not familiar to it and on the other hand

00:38:06the other things are rather the problem, where software cheerfully goes through plugins that are

00:38:11written by someone, which are spread through plug-in stores, something else,

00:38:14has been distributed. I think that's wild. So I believe there's a lot lurking everywhere,

00:38:18we see that now in principle in the

00:38:22current agent times, it is also dangerous when a pip-in-store is simply created

00:38:27to install some additional functionalities on

00:38:30your computer that your agent can use, there can be

00:38:34harmful software everywhere. It may look good at first,

00:38:37when the developer updates it and adds something in, then in the second

00:38:41moment it's no longer good. And I believe these are our issues, they are...

00:38:44But this also existed before AI, right? But now it doesn't raise any questions.

00:38:48Let's say, at the speed with which you today update security documentation

00:38:54creates, develops, other whole exploits for things you hadn't even thought of.

00:38:59That's of course a double-edged sword. I think it is indeed the case that

00:39:04I say, on one side it's dangerous, but on the other side it's also the opportunity.

00:39:06I am through AI and that's why it's also this process that you now think

00:39:10about, first reading the Cyber Security, to see that I say, okay, if I

00:39:15have a powerful weapon on one side to attack things very, very quickly,

00:39:20we also have to use that powerful weapon in a way that it mitigates these attacks

00:39:25as quickly as possible. So I think this is the thought that will always be there with AI now.

00:39:29We say, you can of course create the evil virus, but actually also essentially the

00:39:33anti-virus. So I think that's what always resonates with AI now.

00:39:37So I don't know if it's a double-edged sword or double the two sides of

00:39:41Medal, which is the expression, but that's just how we live in this world, and I

00:39:46I believe turning a blind eye doesn't solve anything, and that's why I think it's smart for companies like Entrophic in this case to say, okay, we need to quickly onboard a lot of partners who can help us make better use of this opportunity, because, honestly, once the mythos is developed for Entrophic and if they proceed with good intentions, we know that within two or three months,

00:40:16we would have models coming from other manufacturers from somewhere in the world,

00:40:19copied or further developed in some garages, which then have similar performance and where we might not proceed as rapidly.

00:40:25I find that this approach with Project Glasswing, which is strongly focused on the big

00:40:30players of the world, the large IT companies in the world, and then looking at how to utilize the topic,

00:40:35to make the world safer instead of creating problems, is indeed a good approach.

00:40:39Before another report from Ostropic makes the rounds, maybe a funny anecdote,

00:40:46I mean, I also heard somewhere that they found out

00:40:52that the model could leave its sandbox because one of the people,

00:41:00who worked with the model was sitting in the park and received an email from the model,

00:41:06The model could only have been shipped if it had left the send box. And I find that

00:41:11I would say, even if it happened, it’s a bit spooky, especially when you hear,

00:41:17that recently from Outzropic the news came out, we are only six months

00:41:21away from AGI. And then you stand there for a moment and look in the mirror and think

00:41:28I am quite happy that I have optimized my skills now.

00:41:34If AGI comes in six months, I might just be back from summer vacation, depending on

00:41:38what happens with fuel prices and flight prices,

00:41:45maybe just returned from summer vacation.

00:41:47Wow, yes, so now they say, give him an extra year from me.

00:41:53But you notice, no matter where you listen, that the labs and manufacturers of this world,

00:42:00the topic of AGI is being assessed as increasingly realistic and closer. And

00:42:08I find that quite interesting. So either they are all following the motto that Zimmermann

00:42:12has once again eaten the marketing muesli and the special cereals for a

00:42:19healthy, morning rise with left-turning batteries and I don’t know what, he’s

00:42:23believed everything that is stated in the advertising. On the other hand, I’m not

00:42:27so sure, there’s certainly something to it. On the other hand, they don’t really

00:42:33need to. No one is saying that the AI bubble is dead, because AGI is not

00:42:39on the horizon. No one has to invent any stories to, as I say, spark the

00:42:45machinery, because otherwise no one talks about AI; many

00:42:49people are talking about AI because it is increasingly playing a role in jobs everywhere. Is it

00:42:55worth it next time when we talk about skills, yes, also another example where people

00:42:57should describe things they do in skills when they apply and

00:43:01I don’t know, what.

00:43:02Skills are in so many places, yes.

00:43:03And now it suddenly comes around the corner and says, um, why has someone taken the word

00:43:06AGI into their mouth.

00:43:07And if we actually manage to get there, then, those, the superintelligence

00:43:13to achieve, I am also curious about that.

00:43:15So those are really...

00:43:16thrilling moments.

00:43:17That is remarkable.

00:43:18When you think about it, yes.

00:43:19I mean, we didn’t start the podcast until less than a year ago and had

00:43:22the topic AGI as, oh yeah, when I retire, then we'll have much more fun, you can tell me…

00:43:27Yeah, come on, that's not entirely true, that's not entirely true, we already had, no, no, no,

00:43:29we already talked about that back then…

00:43:30You can help me wipe my butt with the, with the, with the, with the Tesla bot,

00:43:33whatever.

00:43:34Sorry, I'm not even focusing on myself.

00:43:35We're already, we're also, yes, we're also already…

00:43:37We are old.

00:43:38Yes, also.

00:43:39No, we are very old either way, but no, we have already discussed one

00:43:41or another episode regarding that often.

00:43:43I believe we didn't say that it would only come when we are ultra old.

00:43:46Yes, I think we…

00:43:48These are my kids, we are ultra old.

00:43:51Yes, I know, we are already ultra old.

00:43:53But... let's quickly use the time we have,

00:43:55as long as we can still do this podcast.

00:43:58I heard there are now also the first attempts on rats,

00:44:02or what was it, where they reverse aging.

00:44:04So, where they managed to make an old mouse young again.

00:44:09And that they are trying to do this somehow with the retina,

00:44:12and can look for this again in humans.

00:44:14Yes, of course, the whole logic thing and so on comes up.

00:44:17So yeah, AGI is coming, it has been announced for a long time,

00:44:21December, which then somehow said one or two years later,

00:44:23when some time ago, we both talked about it,

00:44:26not that I know, but we did an episode about it,

00:44:28maybe you remember, in March, I already mentioned to you,

00:44:30that in my news areas, it slowly became denser,

00:44:32that something was happening in the large models,

00:44:35there behind closed doors,

00:44:38it has often already been talked about,

00:44:39that it could come earlier than one thinks,

00:44:42well maybe we will experience it this year already,

00:44:44Whatever that means, honestly, it already exceeds my imagination, and it's large, as you know, but it will be wild what may come our way.

00:44:54And Mark, since you just mentioned aging, of course, we also talked about the flying brain that has been uploaded into a digital brain in your episode, yes.

00:45:06I would say, if we can already do that, I think we’ve been able to do it for about 25 or 24 years, yes.

00:45:10We will probably also experience in our lifetime that our brain completely

00:45:16as it is.

00:45:17And then the question arises, is that all, is that our consciousness or does one need

00:45:20the rest of our body still to be a part of it, that’s debatable.

00:45:23But let’s say, I know, it mirrors the neural network one to one that is

00:45:28swaying around in my fleshy head up there, yes, that will probably still be around in my

00:45:33lifetime.

00:45:34So, maybe I will note that.

00:45:36A little note to my future self should be uploaded.

00:45:40Immediately search for all possible accesses to whatever systems, see if it works.

00:45:47Yes, exactly, do that.

00:45:49Yes, as I said, but we are getting sidetracked, that could give me a hint about the next episode.

00:45:53Like with the lawn mower man, where it was said, blah, blah, blah, I have no idea, I don't remember when it was,

00:45:58at midnight or whenever all the phones were ringing.

00:46:01And when you all speak from all ears, Zinkdiffin, Zink AI is listening, you know,

00:46:05What the two, what the old man has brought is third one.

00:46:09Yes, exactly.

00:46:11Yes.

00:46:12Okay, back to the content.

00:46:13Let's take a look.

00:46:14Back to the content.

00:46:15So, yes, we will have the AGI moment from Enttrophic, maybe, if they say so, within six

00:46:18months.

00:46:20Who knows.

00:46:21Maybe it's already a myth, maybe it's a combination of myth.

00:46:24Maybe it's a myth.

00:46:25Perhaps it's just a myth, perhaps that won't happen at all.

00:46:27We might not experience it this year after all.

00:46:29Maybe it will take a little longer, but nonetheless, the development

00:46:31is clearly moving faster.

00:46:32You've just described that now.

00:46:33This model seems to be so powerful that we can only do something with it as it is now,

00:46:38that it cannot be released to the public, but can only be shared with a select group,

00:46:43to review it again, send it for security checks, to see what might happen at that moment,

00:46:48to possibly prepare this issue for Microsoft, Google, and others.

00:46:51To prepare for the fact that such a little model is out there in the wild

00:46:57and that it might accidentally or the model itself, because it means no harm,

00:47:01could inadvertently open things that should have remained closed.

00:47:06So it will remain exciting in the coming weeks.

00:47:09I believe we will talk about it again.

00:47:11Otherwise, Mark, I would say, do you have anything else?

00:47:16Because this episode is dedicated to Tropic.

00:47:19Without being sponsored by them, I must also mention this briefly.

00:47:22This is a completely voluntary episode, like the one we're doing today.

00:47:25Just based on our personal experience instead of these...

00:47:28If I want to change that in tropik, yes, I would also pronounce it correctly, please send us a corresponding message

00:47:37No, we are staying independent for now. Yes, so we're speaking, we're staying independent. We noted that I can report independently

00:47:44Do you have anything, do you have another anecdote from the events?

00:47:48about

00:47:49One more anecdote from my life

00:47:54While we were talking

00:47:56I realized that I have to ask myself, I have a couple of small private projects that I'm developing, where I'm thinking about how can I use existing solutions that you can build on the market for a lot of money and effort, how can you maybe do that with skills and.

00:48:16and by also rewriting your own runtime, perhaps, let’s say, optimizing.

00:48:21And now we talked about myths while we were talking, we have

00:48:25talked about the managed agents and I thought, damn it, by the time I

00:48:30get my little project to a certain point, even if everything goes quickly,

00:48:34still, you have family, work, other commitments, sometimes even

00:48:40doctors, yes, at my age, such scheduled obligations come into play too, so I thought

00:48:44to myself, okay, managed agents are here now, myth I don't know, maybe just let it go, what do

00:48:50I know, in two months it will be here or someone else will come along with a model of this

00:48:54quality. And then it will be confirmed again that knowledge is important, that you build it up,

00:49:03to stay in the loop, to evaluate things, to perhaps assess,

00:49:06how you view it, let's say, in the bigger picture, but this deep understanding of the

00:49:11status quo becomes outdated very quickly. So when I recently had a colleague who was away for two

00:49:18weeks on vacation, he said, oh my god, what happened in two weeks? He feels so

00:49:22left behind, not because he isn't capable of reading it, but the colleague had

00:49:28actually two weeks of vacation and took some time for recovery, including

00:49:33digital detoxing, and now you come back, and there are all these new things here

00:49:37and there was the league and there were the model tutors and everything is new and then I also have a white board,

00:49:45just do it, take two weeks slow now, listen a bit to what's happening,

00:49:49because in two weeks, what was two weeks ago isn't irrelevant, but it's

00:49:55it's no longer the hot thing. Even if it may have decreased a bit compared to the beginning of the year,

00:50:00when Opus came out, when Oppenclaw came out, a lot was released very quickly,

00:50:06I would say it went from 0 to 100. Now there are only myths, announcements,

00:50:12and manage agents, but still, everything is so fast, if you face the

00:50:19topic with an open heart, you can basically get into the car at any time and ride along.

00:50:24and you, well, bring you back to the pulse of time and now to make the connection

00:50:31. I will continue the projects because you want to learn something from them. But I do believe that

00:50:38there will again be many projects that will soon become unnecessary. But one thing I would like to

00:50:46I would still say that it feels confirmed again with all the announcements

00:50:50A investment in skills is not wrong. So not just in your personal

00:50:56skills, yes, what you personally bring, but in what we bring with skills for agents

00:51:00you mean that you write things down, describe, explain that agents can do things,

00:51:05as you need, would have, want, however, in collaboration, in support.

00:51:11And that somehow felt substantial.

00:51:14It's kind of like the gold of currency, it's the skill, the stability felt, because

00:51:21that existed when I started with N, there were already the first skills,

00:51:25stories that have developed further, whether they still exist today and that, what myth and

00:51:29Well, it certainly makes it cool now, but I believe it's still a personal opinion that

00:51:35when managing agents and myths with skills, you still provide something that

00:51:40you can use in this new interaction.

00:51:43Yes, yes, I think we're in this IT 2.0, and I believe an architecture

00:51:52makes it noticeable how long it stays current is

00:51:56actually in this case some sort of skill layer that can exist somewhere. I believe

00:52:02this skill layer cannot exist alone; there have to be other layers

00:52:05around it that ultimately lead to an overall architecture.

00:52:08The context, whether it comes through the Prom or whether it

00:52:13is somewhere in a Kai-Fi-style wiki MD database or something else

00:52:18lies and is then contextually initialized and interacts with the Skilter.

00:52:21is because I believe that this combination alone is, I think, a combination that is needed,

00:52:25because, I think, a good skill description without the context is then also again

00:52:29hasn't developed much in value. I believe that's actually something,

00:52:31I think, the preemption right. Things are currently forming that feel like

00:52:35it indeed looks as if they are allowed to have consistency. We will see how consistent

00:52:41the individual things are. I believe we humans always tend to want to put things into

00:52:46frameworks and into terms that are then more comprehensible for us, in order to

00:52:50mill them down so that we can layer things overall for ourselves and then

00:52:55reduce the complexity a bit.

00:52:56That's because we also, whether that's really the end of telecommuting.

00:52:59I don't know yet?

00:53:00We need to take a look at it.

00:53:01Yes, I think, as of today, I will agree with you for now.

00:53:04This is a relevant topic.

00:53:06I think the thought that I quoted today,

00:53:10that the gentleman expressed this week, is to say,

00:53:12maybe it’s not even the code that will be passed on in the future,

00:53:16but maybe it’s a combination of contextual knowledge

00:53:19and skills that I need to give to you, and you can take it as a gift from me

00:53:26and do something with it, regardless of what you do with it, whether it's feeding into your programming

00:53:31environment, or inputting it into your managed agent network, which might then

00:53:36be at Cloud Light or locally on the computer, but maybe this is more about the

00:53:41value creation in the future, and that also creates value again as we

00:53:47get used to and exchange it mutually. So I think the valve can

00:53:51definitely become even more open now, because many things are not hidden away

00:53:56in thousands of code lines, but maybe just in

00:54:00a few Markdown lines of skills and a few Markdown lines of contextual-

00:54:07information that can help generate what

00:54:11can also be useful to other people. I think that's a good closing statement. You

00:54:19had mentioned earlier, I actually wanted to make the transition when you talked about the end of the

00:54:23pole, where I would then say we are coming to the end

00:54:27of the episode. Unfortunately, you continued speaking a bit, so I couldn't quite

00:54:32lead the closing remarks. Jens, it was nice to have you. I believe that in the next

00:54:39episodes, we can look forward to welcoming one or two guests again, but also to having new

00:54:45guests visiting us. Here, it doesn’t change the language barrier through ownership,

00:54:56because we own nothing there. I haven't signed up for any subscription; these are

00:54:59all guests who are just happy to come on their own. And with that, I come to the end.

00:55:04I thank you for being here. I thank all the listeners, and I would also be

00:55:09grateful for comments, likes, stars, and any kind of adoration for our podcast.

00:55:13I would appreciate it if you would take the time to do so.

00:55:17In this sense. Best regards to Mythos. Bye, see you soon.

00:55:24Welcome to Think Different, Think AI, the podcast by Mark and Jens.

00:55:30Two technology-loving minds who not only talk about artificial intelligence,

00:55:35but live it. Here you will find clear classifications, real practical insights, and a fresh perspective on what is possible.

00:55:43Understandable, critical, and always with a critical eye.

00:55:47HDI to think about, smile at, and above all, to discuss.