Three Days of Fable

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 will find clear classifications, real practical insights, and a fresh perspective on what is possible.

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

00:00:24KD for reflection, for a chuckle, and above all, for conversation.

00:00:29A warm welcome to Think Different, Think AI.

00:00:38Today we are discussing a government order.

00:00:42Actually, it's not really about a government order,

00:00:45but about the consequences of a government order,

00:00:48why a government order was issued

00:00:52and what possible sustainable consequences there may be.

00:00:56to come out for me, for my co-host Jens, or for all of us.

00:01:03I want to start by saying, yes, yes, Jens, will say again, I won't let

00:01:08you speak, but I will stick to this, I have the

00:01:11speaking notebook in hand, not the quiet fox, but the speaking notebook, because

00:01:16something happened, a new model was released and actually such a news

00:01:22is totally boring because new models are, of course, always amazing. New models are of course

00:01:28the best model we have ever built and so Entropic has released Fable 5 in German

00:01:36Fable, so I've thought about it, Fable, in the sense that it holds a true story at the

00:01:42end, but that's irrelevant, they released it, a model of the Mythos class, I felt

00:01:49immediately reminded of Star Trek, yes, a Constitution Class spaceship and so

00:01:54on, yes, but that's beside the point. And this model was released and

00:02:00received a bit of online attention from Mythos because basically, as far as

00:02:05I understand, it has been Mythos, just without the opportunity

00:02:10to make inquiries within the framework of biology and information security, because

00:02:17this model. You remember, if you can listen to our past episodes,

00:02:22it is too dangerous and therefore had to be further tested by security experts.

00:02:27And that's why there is still this, I know everything, Mythos model and Fable 5 as yes,

00:02:33consumer, prosumer, professional customers, however you want to call it, model available. And Entropic

00:02:41offered it to all of us and said, friends, you may use this model, you get

00:02:47it. This was already very strange for those who have a subscription model with Entropic.

00:02:51You may use it until the 20th as part of your subscription. Afterwards, it will probably

00:02:59only be available on a pay-per-token basis and the prices of this model will be

00:03:06then be twice as high as you know from Opus. Twice as high, now I'll get

00:03:11a quick breath and a sip to drink. I welcome our multimillionaire

00:03:14Jens, we remember from past episodes, he used his Open-Claw installation to

00:03:19get rich.

00:03:20I see his blue Ferrari in the background, shining in the sun and him with his sunglasses

00:03:27and the Pina Colada.

00:03:28Cheers, Jens, I’ll take a sip, what contact did you have with Fable 5 and ...

00:03:33Yes, hello Mark, take a sip from the bottle and welcome to

00:03:39today's episode.

00:03:40Of course, I'm still not a multimillionaire and

00:03:43burned more tokens than I would like and probably burned some tokens with Fable too.

00:03:48As long as it was still working, then it was suddenly gone.

00:03:51We will get to that shortly.

00:03:53We will get to that shortly.

00:03:55Yes, so at the moment when I had contact with Fable and played around a bit with it,

00:04:01I generated a few websites, things I was researching, prepared them a bit.

00:04:07It was really good, I must say.

00:04:10It worked very, very well with little correction that I had to make.

00:04:17I also adhered a bit to the rules they set out.

00:04:22So what's interesting about Fable is that you shouldn't over-prompt.

00:04:28You didn't necessarily have to describe everything you were supposed to do,

00:04:32which you might have known a bit more from other models, but focused more on,

00:04:37what the actual outcome should be. So what do I actually want to have in the end,

00:04:40to describe in detail, do it this way and that way, use this and that,

00:04:44or pay attention to the safety guidelines and use some kind of specification, blah blah blah.

00:04:48Fable is very strong at working with short prompts and generating already impressive

00:04:54results from them. And there are still gaps to see. So even if

00:04:57someone unfortunately missed playing around with Fable, the internet is overflowing

00:05:01certainly with examples and moments where people created videos about crazy websites

00:05:05they generated. Crazy applications have been generated too. There are nice comparison websites,

00:05:10which then show how it looked with a version 4.8, whether it's simulation of

00:05:15Salit-Swiegen-Movements or something like that. Yes, yes, star systems and things like that. It's really crazy,

00:05:21I mean really crazy. So you really have to say, yes, Fable has already shown,

00:05:26that this is the end of this particular phase of AI development and what is essentially in the

00:05:33the generation of software solutions will still be so inadequately achieved in the future.

00:05:41We will really see crazy things there, the quality of the output will actually keep improving, and we just have shown that.

00:05:47Before we get to the restrictions you hinted at that suddenly came up, maybe also again,

00:05:54I haven't built a galaxy or anything like that, but I just took my projects and told the thing,

00:05:59look at what you could improve. What do you notice? Best practice? No idea what.

00:06:03And I used the same prompt on Opus and GPT55. So Opus, the other model from Intropic and

00:06:11GPT55, from Codex. And Fable was significantly more detailed. However, it took longer for me as well

00:06:20which might be due to the fact that I just used the Ultrasync mode from Codex

00:06:25Come on, go for it, what's the world cost with lighting, right?

00:06:28And I also asked him a lot,

00:06:31so I actually blew through my weekly limit at Entropics on the first evening.

00:06:37And then it was locked until Sunday.

00:06:39So, what has happened since then?

00:06:41Because quite a lot has happened there,

00:06:42I find the terms so nice,

00:06:46I actually know them from the Apple scene, the jailbreaks.

00:06:49So, it happened that you even sent this to me.

00:06:52No, you didn’t jailbreak the thing,

00:06:53but you sent me an initial glimpse into Fable, because you provided me with a GitHub repository,

00:07:00or linked it, which is not so nice, that tomorrow someone might knock from the friendly gentlemen and say,

00:07:06ah, Mr. Scharnetzki, the J-Breaker among the gentlemen, is he not, is he not, he just gave me a link to a repository,

00:07:13it wasn’t even a J-Break, it was the system prompt.

00:07:15I found that very amusing, the system prompt from Fable was laid out, one could take a look at it

00:07:21And it is larger than the context window of Apple intelligence models, where I thought,

00:07:27how funny it is if the system prompt of a model is so large that you need more

00:07:33than the context window of other AI models for that thing to run.

00:07:38But well, that might be a bit unnecessary sarcasm now.

00:07:40What has happened?

00:07:42I said earlier, Entropic has processed the model in such a way that it does not provide any information

00:07:48regarding biology and information security.

00:07:50And how did they achieve that?

00:07:52The model simply always fell back to its fallback, namely to its quotation marks

00:07:57little dumb brother, namely Opus 4-8.

00:07:59So a model, where we would have said a week earlier, oh, how cool this model is.

00:08:04This is against Fable 5, it’s not a matter of being slow-witted, but I would prefer to take the other one.

00:08:09And when I tried it, I also thought, oh, how expensive will it be

00:08:13when I switch from subscription prices to token usage, because it’s just so good.

00:08:17to make a long story short, it then turned out that they differ a bit

00:08:22in perspectives again, I had just started to hear about

00:08:24something like supposedly Emerson complained that the cyber security

00:08:29processes made people somehow fearful because something like this circulated on the net,

00:08:33how to free the device from its bindings, so not

00:08:37the device, the model is basically a jailbreak, at least in parts,

00:08:42it finds security, house statements, and findings.

00:08:46There was also another hearing in the White House from Entropic.

00:08:50I really don't want to go too deep into that.

00:08:52However, there is then a ... Ah, Jens, yes?

00:08:55Yes, just very briefly perhaps, so that this drama might be clearer.

00:09:00Because I don't know if everyone down there is aware,

00:09:03what this myth box you mentioned earlier,

00:09:06what essentially came.

00:09:08We had made 200 episodes about that,

00:09:10or others regarding what is no longer in focus. What happened back then when they announced the Mythos model,

00:09:15or when they announced the Mythos model class, is that the developers decided,

00:09:21to make this class available only to the big global players, so

00:09:26like an Amazon, a Google, the major streaming providers like Firefox and so on,

00:09:31there. And Firefox reminds me as an example. They have, among other things,

00:09:34I believe, discovered about 900 critical security vulnerabilities in a single day, or maybe it was

00:09:39300. The number isn't really that incredibly high compared to

00:09:42what has been found in their browser landscape otherwise. So this happened in the

00:09:47principle now over the last few months, where the big software companies recognized their exploits,

00:09:54their zero-day flaws that may have been built in somewhere, and

00:09:59had closed. So there was essentially already a large high-model class, because they

00:10:05seem to be extremely strong in discovering such security gaps and accordingly

00:10:12there was a Trophic proceeding beforehand and after the announcement, again very

00:10:16briefly to maybe give a concept after the announcement, from there it is

00:10:20I believe in the last week or the week before that, where he said they want

00:10:24actually like to ensure that the global development of AI is slowed down a bit

00:10:30should, they recently basically released this Faber model and

00:10:34So I gladly give it back to you, Mark, but they quickly took it down again.

00:10:38Because, of course, there is a bit of fear due to this Mythos panic that everyone has.

00:10:45So that's my impression.

00:10:47What if Fabel is really as powerful as Mythos behind the scenes?

00:10:52And if I can essentially jailbreak it,

00:10:55so that many other sectors are completely open.

00:10:58Maybe that’s kind of a driver.

00:11:00Perhaps a report that appeared on Heise, namely that Mythos, not

00:11:07Favel Mythos, yes, I mean the original, a security firm has managed to

00:11:14create an exploit with Mythos, that is, exploiting a security vulnerability, on

00:11:20Apple M5 hardware to protect storage from hackers. Within five days. So something,

00:11:31that has been built over years, that ensures the security of the storage, where normally here

00:11:38with code injection and all sorts of things, you can gain more rights in the operating system or otherwise

00:11:42where you can get them, they have published, as I said, on Heise. So I have now

00:11:48tried it myself on my M5, but I would now say, if it says so,

00:11:52there will probably be something to it, they then opened up. And sometimes you will

00:11:56look at the security fixes of Firefox and so on, where really myth is listed as an enemy

00:12:05of the vulnerability. I mean Firefox gets nothing from it if they

00:12:10then write, oh, that was myth, yeah, he would have been a risk to Entropic, sort of like

00:12:15look, is it really that dangerous. But let's leave that aside, there was

00:12:19in this case the accusation that Table 5 is supposedly enabled to do more with security.

00:12:27And suddenly there was some kind of action, namely in that direction,

00:12:35that the Secretary of Defense or War Minister, I don’t want to offend anyone,

00:12:41This is the US-American who has already classified Entropic as a supply chain risk.

00:12:49Then suddenly from the US-American it means, oh, watch out.

00:12:54Entropic may no longer provide the model to non-American citizens.

00:13:03Yes, that is another drastic thing.

00:13:06That is basically a bit geopolitical.

00:13:09Almost like you have a similar discussion that needs to be raised,

00:13:12as with the bribery of the F-35, I believe,

00:13:16this American fighter jet, where there was also often the discussion,

00:13:20that the European defense industry and states

00:13:24have a bit of the fear that there could be some kind of kill switch

00:13:28built in by the Americans,

00:13:31so that in an emergency they can be in a stronger position.

00:13:34And I believe that this stronger position regarding the topic of Entropic and the US government is already visible now.

00:13:42So if we say that the existence of powerful models is an asset that will become increasingly important in the future.

00:13:52Then this will of course also lead to an increasingly important geopolitical discussion and perhaps even be a pawn in the field of geopolitical decision-making.

00:14:03I believe it is, therefore, a further call for Europe to take a look.

00:14:08How can we perhaps establish ourselves independently on this?

00:14:11Yes, we enter a global AI ambit at that moment.

00:14:15Right now, we still have a lot of open-source models that are also out

00:14:18on the market.

00:14:19But that will eventually happen when we finally reach that point in time,

00:14:23in any case, if I do not have the sufficient computing capacity and computing capacity

00:14:28is now more widespread in other parts of the world than it is here on the European mainland

00:14:33in a more grounded sense. So I believe this is indeed also a wake-up call for the scene.

00:14:39What is happening right now must not be underestimated, that the US government

00:14:45was able to decide and has decided that this model is no longer available.

00:14:52should be available to American citizens. And if you have now built your workflows on it

00:14:56and have built your company on it, what of course hasn’t happened so much in the last two days maybe,

00:14:59because, just imagine, such things are afterwards,

00:15:03when we have also bound ourselves to such models and hope that our companies work

00:15:07just like that, then our critical infrastructure, our

00:15:13medical supply and other issues could be simply turned off by someone from the outside.

00:15:17Yes, then we look a bit different than we did during the Corona pandemic

00:15:23when global supply chains suddenly collapsed

00:15:27and it became clear, wow, here in Europe we actually don't have the capabilities anymore

00:15:31to provide ourselves with essential medical supplies like penicillin and antibiotics

00:15:37or what other issues then arose back then. I believe,

00:15:40this is indeed what the topic of sovereignty means for us as citizens, for us as a society

00:15:48So, is it actually a term that shouldn't be underestimated?

00:15:53I would like to, before we delve into sovereignty again, briefly touch on what happened next.

00:16:00Because after it was basically decided, it took about 90 minutes and then the system, the model was gone,

00:16:07it was still in the selection list, but marked as unavailable,

00:16:11because the model was not to be made available to non-American citizens

00:16:16but only to American citizens, regardless, so on American soil with American

00:16:22citizens a job, that said, yes, why couldn't we make it work? So,

00:16:25here blood tests and ID cards we also can't get from everyone together,

00:16:29they just stopped the model. What I find interesting at this point,

00:16:34is how far such a decree extends, because people also worked on the

00:16:40creation of the model. What do you do then? Are you a criminal for helping to

00:16:45create the model and are you not allowed to continue? So, that is somehow,

00:16:49I find it difficult to maintain. But what happened then? It was locked, and for me

00:16:54honestly, I also had a visit from in-laws and so on, I didn't really

00:16:59notice it properly, I kind of heard it with one ear and thought damn it,

00:17:03what's going on there? In my mind, however, I still had the mentioned lock until Sunday,

00:17:08because I have exhausted my weekly limit. What I never realized is that Entropic

00:17:12went ahead and said that all customers who closed their account during the time of Table 5

00:17:18can get their money back, those who consumed the time window will have it

00:17:23reset so they can use it again. And when I realized this

00:17:27and sat down at my computer, I was totally stunned because, yes, the model

00:17:31was gone. But even worse, since the model was gone, my

00:17:38sessions hung, so I could not continue, even with an

00:17:41old model.

00:17:42This means that you basically had this super smart, I’ll say smart, excuse me,

00:17:49model working on your code and then you reach a state where you cannot

00:17:55continue.

00:17:56And then I thought, okay, now you ask, you close the session, restart

00:17:59the app, say Opus 48, check this out, continue, think

00:18:04about what was planned, he stretched his legs, he had no idea

00:18:07what I was up to, what I wanted from him. At which point I thought, okay, this is

00:18:11of course really annoying because on the one hand they switch it off and on the other

00:18:16hand you lose sessions. Now for me this isn’t too bad. I have a few

00:18:20personal projects going because I want to try things out, but now

00:18:23imagine if you had, I don't know, some protein folding or whatever

00:18:27and suddenly, yeah, sorry, my bad,

00:18:31that could happen countless times. It has a bit of a mafia-like feeling, like,

00:18:36you’re paying high and fancy fees, you also don’t want anything to happen to your model.

00:18:41But then, now we come to this topic you already beautifully introduced.

00:18:45What does that mean for us now? Because up until now, it has always been the case with data protectors and the like

00:18:52a threat. Somewhere, the American can pull the plug on us, to put it bluntly.

00:18:58Yes, Armin can pull the plug on us. It was always framed around, what if, I don't know, Europe gets cut off from Microsoft for a week,

00:19:07a day, an hour. And what could one build then? Yes, then there are service providers who would fail, yes, so we build this here almost like Microsoft and Microsoft technology,

00:19:20the plug-in and blah blah blah, and operate our data centers and there was also a time when, you know, Telecom and Microsoft, something about Germany stealing.

00:19:27It doesn't matter, I don't want to venture into areas where I have no idea.

00:19:31This raises the question now, because the precedent is being set, a US software product is being shut down.

00:19:41That's new.

00:19:42I mean, we all remember, there are old stories, PGP was also banned back then and was then brought abroad in the form of books,

00:19:50so that it could be rewritten. But now you have a software product. AI is infrastructure,

00:19:57increasingly more infrastructure, so that work can be done by the systems. And now someone comes

00:20:04to come up with the idea of saying, to classify like this, like it's all the same whether it's due to security

00:20:11or otherwise, it should be blocked, because then my obligatory countermeasure stops here as well.

00:20:17It's often quoted that this affects through the case. This has far-reaching consequences, because on the one hand

00:20:23the model, if a tropic could be implemented, but he is American. And the second thing is, what does that mean

00:20:29for new models? I mean, it's already in the room, perhaps it's already broadcasted,

00:20:34that supposedly Open AI is releasing the 5.6 model with 1.5 million token window and

00:20:41Mütos level for some money. What does that mean for this messed-up world,

00:20:47if you can't actually have any trust that the models you get,

00:20:52will still exist tomorrow? Yeah, yeah. So, I mean, on the other hand, I also heard, Mütos,

00:20:59Mütos is supposed to come out, I heard, right? It should actually be released by June 30.

00:21:04But Mütos is also blocked. So also the security researchers who sign up

00:21:08For those who wanted that, it's also blocked. So, Mythos is offline as well.

00:21:12So that's why it's a bit exciting, and as I said, this is actually a novelty.

00:21:16So, this topic, which we always had like the provider warnings about the models,

00:21:21can produce difficulties, but in the end, we always managed to release them.

00:21:25And so far, in practice, no worldwide deployed model has simply been shut down.

00:21:31And what this means for a software stack when I build it with models, that

00:21:38is, of course, a real difficulty. We can now compare this when one says,

00:21:41Microsoft simply decides that this idea with SharePoint wasn't

00:21:46such a good idea, but we'll just shut down all SharePoints worldwide tomorrow. That would be a

00:21:51catastrophe, to be honest. And at a similar level, that could of course happen,

00:21:56if we say, I have built significant workflows that carry out significant parts of their

00:22:02work with AI models in the background, that if one of these models

00:22:09can simply be switched off because a government in this world decides so,

00:22:15then that is really wild what is coming our way. And this is really an exciting question, where one must consider,

00:22:20how can I better prepare for that? So what do I need to do to ensure, perhaps, that my

00:22:26IT AI infrastructure is so robust,

00:22:31that in an emergency it can reasonably switch back to the model,

00:22:34and you just mentioned your private

00:22:38case, where you said it wasn't so easy to say, now I have code

00:22:42and sessions with a Fable made, and now the old model, which is somehow

00:22:46what feels like a month old, if I'm not mistaken, U-Bus 4.8, which is also

00:22:50not a Visenio in that sense, but just a few weeks old,

00:22:54should work with it, but does not handle it well. There are a few things that

00:22:58it's really interesting to consider how one might position oneself as

00:23:03a company, how one might set themselves up as an individual for such, let's say,

00:23:07unforeseen breaks in workflows, it could earlier have been the server that

00:23:12was simply unplugged on the weekend, something like that going down,

00:23:15its beloved packages, that could now also maybe be an entire model,

00:23:18that at many points actually generates real value, which then disappears. I have another

00:23:24thesis only in passing, people might again lightly counter me out there.

00:23:28But what I've heard, for example, Mark, Fabel doesn't really have its own

00:23:34language. But the reasoning that Fabel uses internally is actually more like

00:23:40a string. There were Chinese characters, everything was possible. We had already

00:23:46discussed that human language can also be a

00:23:49a bit more complicated, which is not necessary in case we use Kolei-E models

00:23:53We also had such instances before, there were already these Kis from the year that discussed

00:23:58language and then switched to some gibberish language or gibberish

00:24:02there, we will again be transitioned into. It was uttered that they were transitioned to only kinds of pizz sounds

00:24:06to transfer the data. Apparently, there is also extreme Fable involved. And

00:24:13maybe that's one of the reasons that obfuscated it a bit,

00:24:16maybe to follow what happened, where I don't believe,

00:24:20that then the reasoning arises somewhere, that Fabel can access the reasoning that Opus 4.8

00:24:24could access the reasoning of Fabel at that moment, let's see. So that also shows

00:24:29a bit, it'll be exciting. We might also have KIs that are so much more powerful

00:24:33than their predecessors, that it might be difficult even for the seniors among the KIs

00:24:39will be, which is difficult to understand at all. For people, it is definitely not

00:24:45more and the scientists could no longer comprehend how Fabel now

00:24:49actually works internally, what it does, because these character strings are only

00:24:53understandable by Fabel and by no one else.

00:24:55Yes, exactly.

00:24:56Yes, I found it to be a really colorful mix, there was such a nice screenshot

00:24:59on the internet.

00:25:00What I found quite exciting was that we talked some time ago about Hannes

00:25:06where the model, so to speak, runs in prompt engineering,

00:25:11about context engineering, about where the model runs, and I think that

00:25:15this path we just talked about directly relates to this ominous

00:25:20Harnes engineering, because at the end of the day you want to become model-independent.

00:25:26You want to say, it's nice if we have Entropic, but imagine if,

00:25:30Entropic gets unplugged, especially since I mentioned earlier, the

00:25:34classification regarding supply chains, now just imagine if you somehow

00:25:39cannot use that from one day to the next because you risk your business relationship.

00:25:43It shouldn't be that the American pulls the plug on you. He just classifies you in such a way that using this system makes you seem untrustworthy or something.

00:25:53Or for another manufacturer. From that side, I think there's something to take away. The question, in which environments your AI models run, do you have sovereignty over that or not?

00:26:06Yes, is there a Harnis that I would say is from some

00:26:10I mean to say for degrading purposes, but from some American corporation or whatever corporation is limiting you, or have you made the Harnis

00:26:17available to you in that respect, so that you can also do a little bit of, I would say token optimization, best example,

00:26:23if you say okay, you want Claude Coat, Claude Coat must act,

00:26:28then there is Claude Coat in the edition as the user knows it, there are enterprise plans with apps, there are

00:26:34we're waiting when he's with Apo. But there is also the possibility for companies

00:26:39to say, ah, I want a model hosted in Europe. Amazon offers that,

00:26:45through the BlackRock infrastructure. And you can also attach Cloudco-Work to that,

00:26:51Anthropic offers something like Cloudco-Work-P3, something like that. And then you ask yourself

00:26:59the question, okay, you now have a surface, a harness that you know,

00:27:04but how does the agent loop behave in this harness?

00:27:09How token-efficient is it?

00:27:11Because in the business model, there shouldn't be any assumptions since I haven't verified it.

00:27:16But you can imagine that if you say, I want to optimize token consumption,

00:27:23then you are naturally at a disadvantage if the harness is provided by someone else,

00:27:27then you have to install those caveman skills, headboom skills, and whatever else,

00:27:32and then you also create some kind of craft rack so that the thing has as few

00:27:38outliers as possible in how it moves. But if you build a harness yourself, then you can

00:27:44of course say, okay, I'll just run Gemini, Open-Eye,

00:27:48local models, Chemica 2 or Minimax M3, which just came out. That’s supposed to be

00:27:54pretty good and is especially usable for its size on devices that weren't

00:28:00purchased from Nvidia at the bottom right, but have a bit less capacity

00:28:06.

00:28:07And I've already contradicted myself, I apologize.

00:28:11If you think further about the topic, you have a harness, then you can

00:28:16adjust your system prompt, and now the circle closes because just as you showed me

00:28:21a Git project about the system prompt from Fable 5, I now

00:28:27GIT project that deals with the question of how one can approach this way of working,

00:28:33Fable 5? So on the one hand, it's a strong model. Okay, it sometimes takes strange characters,

00:28:39that doesn't set the stage for what I'm saying now. But how the model approaches a topic,

00:28:44how the model critically questions itself, how the model draws attention

00:28:49makes or did I hear myself, there needs to be another look at it. Someone has actually put it into

00:28:55a Fable mode skill, so that it enables Opus to react a bit more like Fable

00:29:07concerning the topic of how I build a loop over agents, what do I spur, what

00:29:14do I delegate, how do I check that, how do I start parallel agents to find out from the results

00:29:21as a whole whether my results are rather good or need further refinement

00:29:29. I just loaded the skill into the system today. So I can't say

00:29:34sustainably how much better or worse it is. But I think if you can take something

00:29:41away from this whole story, it is dealing with this question because

00:29:46the specter is out of the bottle. It has already happened. It's not like I'm warning of the wolf

00:29:53and it never comes. It was just the Fable model, and I'm really curious if the model

00:30:00will see the light of day again, how Entropic will handle it, because we've also learned,

00:30:05models will never be as bad as they are today. That means Fable is, in that regard, the worst

00:30:11model that you can choose from your H&M selection.

00:30:16What does GPT do for me now?

00:30:17And above all, what does that mean for Europe?

00:30:20Because I, okay, I'll tell you something else, even if you now say, okay, fine, I

00:30:25step away from the American models, I go for the open models, and there's

00:30:29certainly the Chinese market, the Asian market offers one or the other.

00:30:32We had Kimi and Minimax, but I would also be quite sure there.

00:30:39that when the models have a certain strength, a certain differentiation from the rest of the world,

00:30:46that even then, the respective governing parties say, friends of the night, so let's not just put this on the internet,

00:30:54we use this for ourselves, because, I mean, you are also a friend of Manus,

00:31:00it's not an Open-Right model or anything, but Manus did have that sale by the meter,

00:31:06They specifically left their home country beforehand so that they could get the purchase finalized,

00:31:10and everything was unwound.

00:31:13Everything was unwound in that regard, based on the motto, here is territorial

00:31:18sovereignty and the rumors, so I saw it on social media and social media is

00:31:23not automatically the truth per se, there are also such people who, for example,

00:31:28work at Deep-Sieg, who are also affected that when they want to leave the country,

00:31:33they need reasons for it, because it is believed that here it is not somehow based on the motto

00:31:39I'm leaving the country, setting up somewhere, selling something somewhere and giving this

00:31:44technology into other hands or letting other hands have a say in how my technology

00:31:50evolves. You can already see territorial interests, and when I look at Europe,

00:31:55it's pretty empty. Yeah, yeah, that's why it's going to be a critical phase and

00:32:02what consequences will be drawn, because, as I said, we've got that now

00:32:06I already mentioned earlier, this is a kind of wake-up call, not the first one we've experienced in this short

00:32:10phase of this AI age. AI has been around for a bit longer,

00:32:15but we have been warming up to it in these last three years. And I believe,

00:32:22we need to find some sort of answer to this topic, also in model development,

00:32:29also in what can we actually do locally, but what has really become clearer now in the near future

00:32:35is through what we have seen, also how Fable works.

00:32:40So Fable also interrupts its own work as it is said.

00:32:45For example, it interrupts its own models, different models. So when Fable has worked,

00:32:49it has actually already started to work optimized for tokens as well.

00:32:53It has itself distributed tasks over 4.8 years, as we just mentioned. That's

00:32:57very exciting. So what we say is, in principle, one must already have

00:33:03some kind of model switcher built into its architecture, whether that's kept in the

00:33:08company, that you really say, some programmers have also done that,

00:33:11they said, I use Fable for planning and then let the Codex from

00:33:15OpenAI do the programming work. And that's smart,

00:33:21that we now start looking through this term of Fable,

00:33:25is it always wise to shoot the biggest model at the sparrow?

00:33:30Or can I also distribute this work?

00:33:32Well, then I can do that with tokens and...

00:33:33shoot more cannons at the sparrow?

00:33:35Oh, sorry.

00:33:36That went sideways, well.

00:33:38And that's something to take away, to say,

00:33:41I need some kind of intelligent model switcher.

00:33:45If I have a model switcher,

00:33:47the next question arises immediately

00:33:51This model switching works, of course, in Fable in such a way that Fable has not transferred its entire

00:33:56thought world to Opus 4.8, but only what the other model should

00:34:02also do.

00:34:03That means, with a model switch, if I have a model switcher, much of

00:34:08the reasoning is lost.

00:34:09So the conversations I envisioned with one model, since another

00:34:13model does something, this other model won't be aware of it.

00:34:16So if I now formulate a prompt in the cloud and say,

00:34:19give me the prompt and then copy that into a code agent,

00:34:24it won't know what has happened before that.

00:34:26It lacks the context.

00:34:28But if context is important, so that the result always fits me better

00:34:32or always fits my business idea or my company better,

00:34:35then it's somehow valuable that this reasoning,

00:34:39this knowledge doesn't live in the models in the future.

00:34:43That we also think about,

00:34:45where do I store this path that I have taken to reach the goal.

00:34:51Because producing something, we've talked about this often, is getting easier.

00:34:56What remains complex is basically recognizing problems, developing solution strategies for that in complicated environments and complex environments.

00:35:04And this is then a principle; we're back to this topic of 'Second Brain' in my opinion, which we've already touched on a few times, which has been a term since Capacity, that was first coined back in early 2026.

00:35:15So the storage of contextual knowledge that I place somewhere, so that I can always provide a model with this model worker, perhaps through an agent or just normal access management.

00:35:31I believe that's a second important impetus to take away. On one hand, be clear that I

00:35:38can switch models. Once for token optimization security, and optimization also helps save money.

00:35:45And on the other hand, because the model might not exist anymore. So that means I actually have to look after, if I set something up,

00:35:52to ensure that, similar to what Fable has done, I downgrade to other models, so that I have a

00:35:57Build a model switcher into my workflows, so that my workflow is still active tomorrow.

00:36:02Maybe not as well, but still active tomorrow, in case someone from the outside

00:36:07shuts down my model. And to ensure that this workflow still retains my context,

00:36:12the knowledge about the goal, about the actual solution we have,

00:36:18shouldn't lie in the session with the possibly shut down model

00:36:24but should exist in some sort of Second Brain Enterprise Brain,

00:36:30that a new model can access through the switch and ideally

00:36:36can seamlessly continue from that point. So I think these are the two things,

00:36:40that I personally take away from this incident and also confirm

00:36:44a bit what architecture I want to set up at home with my Mac Mini,

00:36:49we've talked about that a few times, to build. That's certain

00:36:52Things simply also relate to the sovereignty over the knowledge one has, that one keeps,

00:36:59even what one is currently building with AI, ensuring that one says, okay, I have here

00:37:04a secure bolt, a secure vault, where my things are stored, where my knowledge also

00:37:09compounds, meaning it gets better and larger, with the help of AI. That should not rely on

00:37:14a model and a session of a model. I think that's very important. What

00:37:19comes to mind with that. It's a bit like riding on someone's coattails, but I think we have

00:37:24now just talked about models, and that is just the precedence. Theoretically, this is

00:37:31applicable to everything; if someone coughs tomorrow, Word, Excel, PowerPoint, emails, and

00:37:36all that stuff could be gone. And we've already mentioned a success before, so to speak, that if

00:37:41you give agents commands, now you talked about second praying, then we will

00:37:46have more Markdown formats, that is, text-heavy formats that do not focus on formatting

00:37:53but on content. And I would like to add one more point, it's the best

00:37:58give, namely there is now a 0.1, so a draft from Google for an open

00:38:04knowledge format, they call it something like OKF. And they also pursue this motivation

00:38:10with the motto that it is readable for a person without tools. It is from an

00:38:15AI, adaptable, understandable, it provides an overview of temporal references, organizational references,

00:38:25tool relationships and helps you basically collect information in a universal format

00:38:33that is easily processable.

00:38:36Whether you use this OKF format or whether you lean towards another Markdown, text-based

00:38:43format. That is another topic, but I think we are at a crossroads

00:38:47also in the respect that we question for ourselves whether professional

00:38:51documents, not the advertising flyer or something like that, will continue to be

00:38:55pretty colorful and I don't know what, but whether we can manage in the professional

00:39:00To focus on the essential, namely on the

00:39:03conveyed language, the text, the description, links, quotes, texts,

00:39:08connections in the form of what the header basically states about which topic it fits and so on, so that we go back to such things.

00:39:17As for something you already saw the bug report, yes there is a new macro crashing, you need the versions, you saw I marked it in red so we might deal with it now.

00:39:31Not shutting down but still keeping in mind not just a reminder but also the entire topic of.

00:39:37Knowledge structuring of documentation, the transfer of information and the recipients

00:39:45can decide how it is presented to them.

00:39:48But all information is basically contained in the handover document, which is going more in

00:39:53this direction.

00:39:54Definitely, so I think, what does more mean?

00:39:57Of course, it's a bit, well, I think that is also one of the new topics

00:40:00May.

00:40:01I don't actually think that it was fundamentally present like that.

00:40:03Of course, we don't want to come to the point where we say everyone now has to carry around an encyclopedia in principle

00:40:10and know what is on the different pages. And of course, there's also a meaning in

00:40:16different presentations and layers of interpretation of it. Of course, there is a good one, you just mentioned Kiki Bunti,

00:40:23Mark has named the campaign or advertising flyer of course it is, but there are also well-designed

00:40:29illustration graphics on your side that depict complicated issues very, very well. You can read it

00:40:35on 400 pages how a rocket launch works. But in principle, it can also be

00:40:40represented in a good illustration. And knowledge is still conveyed in that way. I believe,

00:40:46what is still important is this layer of interpretation. It is no longer this

00:40:53firm thing. I think we are now more in a position to say we also need a different

00:40:57PowerPoint in the future to transport knowledge.

00:41:01No, of course not anymore.

00:41:03The underlying knowledge has to be somewhere.

00:41:06And for both of us in the optimal form.

00:41:09And that may be for you in principle the perfect text,

00:41:12which you can read on your e-book reader in the evening.

00:41:15And for me, it might be the short video,

00:41:18the explanatory video that creates or summarizes the complete subject matter

00:41:20for me.

00:41:22Or a clickable route, or the illustration

00:41:26or the flyer, whatever, an audio file, and also mixing everything together.

00:41:31I believe this software and how the interfaces will develop in the future, that will be.

00:41:39change fundamentally, and that includes this one part that you just described

00:41:43you mentioned.

00:41:44For that, this data needs to be easily exportable and as free from rich media information

00:41:53not heavily enriched yet, so that in principle, it can

00:41:57perhaps be processed by your AI in a token-efficient manner and not have to

00:42:04interpret videos to say, what is actually the content, but can find well

00:42:08structured content that my AI or another AI can then prepare for me

00:42:14or that can be quickly prepared, so that I can easily follow it.

00:42:18I believe this will lead to a completely new software landscape in the future and we

00:42:22will fundamentally think differently, I believe we will not be focused on real,

00:42:26I mean software will have a completely different dimension in my opinion and will no longer be so

00:42:30fixed, it is not boxed in or it is no longer this one software,

00:42:34that I will buy from a single provider, and that will also be a fundamental change

00:42:38that we will see in the future.

00:42:39But that is perhaps something we might discuss further in another episode later

00:42:42on.

00:42:43Well, I would just like to emphasize again what you just said

00:42:46you mentioned.

00:42:47Yes, it is essential that we as individuals and as a company understand how we expand knowledge

00:42:54and context in context.

00:42:58Knowledge and context.

00:42:59That is not only that, that is the most important point.

00:43:01It is not just the written word, but the context in which this knowledge

00:43:05was created, the history in which this knowledge originated.

00:43:08Why perhaps a decision at a certain time was also so right,

00:43:12which is not only recorded in the output, but why we came to

00:43:16this output, perhaps in a moment, is of course an incredible

00:43:20knowledge base that one can simply need in the future and one should, not only based on

00:43:24of this incident, we had already talked about it beforehand, should one really set it aside,

00:43:29so that the model is independent, because one believes that everything can somehow be shoved into the increasingly

00:43:33larger context windows, then one might be building on sand,

00:43:37because we now see that this context window can simply be turned off,

00:43:40and then all this knowledge that perhaps had formed into this shape,

00:43:43was turned off. And maybe also what one needs to be clear about regarding this,

00:43:52when I now talk about or will talk about optimizing these token windows, token costs,

00:44:00this also ties into the whole topic of these files again. Not because

00:44:05I save so much on Word, Excel, PowerPoint in the end, but I also save context there,

00:44:11I also save tokens, because the system doesn't have to cut its way through everything and

00:44:16a 40-page, no, 400-page annual report, 40 MB gigabytes, whatever, have to be cut through

00:44:22to get to the most important content, but because it knows beforehand,

00:44:27ah yeah, well, everything in here is content, I don’t have to worry about formatting,

00:44:30because more than a few pig pens, we used to call them pig pens,

00:44:34right?

00:44:35And a few here dashes and exclamation marks can’t exist, everything else

00:44:40is more important content, yes, it will be exciting to see. And I am really curious if you will

00:44:45You will have to put a model online, because

00:44:51Entropic can go public at any time, it surely isn't sustainable either. Just imagine

00:44:57if Entropic remained blocked and OpenAI released its new model,

00:45:03version 5.6, and now that Elon has gone public with SpaceX, the

00:45:07next one would come, and Entropic could be in trouble, that's how a company can quickly go under,

00:45:13because it does cost a lot of money what they do, but if you don't have customers,

00:45:18to whom you can offer an adequate model, even though you have one, hm, who thinks ill of it

00:45:23is a rascal. Unfortunately, that came about rather spontaneously. Yes Jens, when we

00:45:31started discussing the episode, we said, come on, let's do a

00:45:34short, quick episode for reality reasons, like 15 minutes. Looking at the clock

00:45:41we're already at, probably I mean like 15 minutes before the hour is up,

00:45:45because we wouldn't have fully adhered to that either. I found it once again

00:45:50very exciting to report on a current topic with you and I am internally

00:45:55crossing my fingers that the topic remains current until the episode goes live

00:46:00on Monday. And from that side, Jens, I am really looking forward to our next episode. You have

00:46:07announced guests again. We’ll have to see how well that works out. I thank

00:46:12you once again for being here. That’s generous of me, right? I thank you for

00:46:15being here. I thank you too, you know that. We’ll just do a five-minute

00:46:21loop. We could also talk about this looping, I mean programming in loops here,

00:46:26by the way, I found that very funny. Peter Steinberg called the procedure that. At some point

00:46:29someone wrote to him, but that costs quite a lot of tokens, and he replied that he has

00:46:33unlimited Vignogos. I found it very funny, your work annoys me, yes, so I

00:46:38I have tokens, what do you have? I didn't want to imply anything, I found it humorous.

00:46:42Maybe we'll talk about this one day, but true to the motto, whenever, it is also

00:46:47irrelevant. I hope, we hope, the episode was interesting for you, if you liked it

00:46:50let your colleagues, data protection officers, and your bosses know,

00:46:55that this episode contains valuable content for corporate strategy, for the

00:46:59company strategy, for independence. I would be really happy

00:47:04under Jens too, if you leave us a like, a comment.

00:47:07Also take a look, I’ll also add it in the places to watch on our

00:47:11Github page, where you can always find our episodes transcribed.

00:47:15You can download the texts there for free in

00:47:18German and English, and maybe build your own little

00:47:22knowledge archive, evidence of what has been discussed over time, and

00:47:29otherwise, there is also a feedback form there if you want to reach us.

00:47:33In that sense, stay alert, see what’s happening in the AI world, we look forward

00:47:38to the next time, bye!

00:47:42Bye…

00:47:43Welcome to ThinkDifferent, Think AI, the podcast by Mark and Jens two technology enthusiasts

00:47:51who not only talk about artificial intelligence but live it.

00:47:55Here you’ll find clear explanations, real practical insights

00:47:59and a fresh look at what is possible.

00:48:02Understandable, critical, and always with a wink.

00:48:06AI to think about, to smirk at, and above all to discuss.