Latest / Beyond Vibe Coding / #S2.02 Fable Ban - Now What?
Transcript
- André Neubauer: then we only need to get an Autig, no, not an autig but an outtake. Sebastian: yeah, outages. ⁓ Outaches, many people over the weekend. ⁓ We just need outtake, to a special episode ⁓ of the Beyond Vibecoding This time it's only and me and The reason why we're talking alone, or the just the two of us, is ⁓ we wanted to cover the fairly recent Fable V incident. those who have been living under a rock or ⁓ or are not that ⁓ fascinated ⁓ like we ⁓ we'll briefly explain what happened André Neubauer: you Sebastian: how we ⁓ look at this André Neubauer: Yeah, for those of you who live under a rock, Sebastian, you want to summarize what happened last, I think, Friday, right? It was already Saturday in Germany, but like what happened on Friday evening? Sebastian: Yeah. Yes. But maybe before we get there, of all, what is Fable Five? Fable Five is the first Mythos model that was publicly available to yeah, the the global ⁓ base of Anthropic. And there was really a lot of hype around it. So I saw several people from my network on LinkedIn giving it a go and they were really amazed. So it seemed to be a very capable model, specifically also for coding use cases. And then on Saturday morning there was a shock because no Fable Five anymore. ⁓ we like got the message in in a WhatsApp group that that we are in where somebody was was posting it, right? ⁓ I I had actually ⁓ only Friday evening I had started ⁓ to a new app and test it out a little bit and see with the long spec how far can the model go. And then I wanted to iterate on it on Saturday morning and I was like, hey. André Neubauer: Yes, true, true, true. Sebastian: Well, what's going on? Why why is it not working? Right. And then yeah, the message, and suddenly I I knew why it wasn't working. So what happened on Saturday was apparently Jesse, or that's at least the rumor made the US government aware that was apparently a jailbreak in Mythos or sorry in Fable 5, which is ⁓ Essentially a mythos model with some restrictions, and the ⁓ government immediately issued an export control directive on the which means that no non-US citizen, this by way, anthropic employees. ⁓ who are living in the US actually, they they are not allowed to use this model anymore. And only US citizens were allowed to use this model. ⁓ And this really was big topic the and everyone was talking about how this ⁓ must be seen actually as a geopolitical move, as as a big issue for European companies, but not just European companies, globally. André Neubauer: Yeah. Yeah, I think the word you have, I at least have often read was a wake up call, right? That is now a wake up call for XY debt. So, but honestly, I also planned my week and differently. So I also started ⁓ playing around with Faber-Five ⁓ on Friday, started writing an app, like was very fascinated. I also shared you ⁓ like a screenshot. ⁓ very, very hungry, very aggressive. And then, so I was actually really looking forward for the weekend, like giving it another try and like get a better understanding on the impact. then yeah, weekend went, weekend went differently, right? And beside, think our, our plans for the weekend, I think the question and haven't actually haven't like, well, like a view, but I haven't came to my final conclusion yet as Like what is the takeaway, right? Because like you could look on that event from very different angles and you could be scared, right? Saying like, we all like shared that all the time, right? Like in the risk assessment likelihood of like they turned us off zero, right? And I think we like, we now learned like it's different, right? Then you could also say like, it's just Fabel five or the other stuff remains. like, why care? ⁓ and then, ⁓ well, like you always need to think, see that from the. context of your company. ⁓ So I think very different views and also very different reactions are currently, you can read it currently at LinkedIn ⁓ or like whatever your source of information is. Sebastian: Absolutely. Yeah. What I find fascinating to begin with is that my initial response to Fable Five was also like, wow, this is super impressive. And then I tried the app that I tried to build with Fable Five and it didn't work. So I had to iterate. I had to iterate then with Sonnet, which I normally use for coding, because I didn't have access to Fable anymore. And the the fixes then worked and I recognized that there Is just a new way how Cloud Code, I think, approaches ⁓ coding projects. I I didn't start such a fresh project in a while with Cloud Code, and I'm usually working with Pi, that's why I wasn't aware of this new mode. But my initial impression was was different. So I was as impressed as I was initially. ⁓ And I have to admit This might be due to my like lower level because I I read in every post where they said everyone on they have an own scale of AI adoption level 0 to 8 or something. And on theirs levels, everyone on level 7 and above was getting amazing results out of Fable 5. Everyone lower than that, not so much, when was not so impressed. And maybe I'm I'm lower than seven, I have to admit. Anyhow, I I think the the The gist of it stays that Fable V is an impressive model, no matter what. And happened now, or this door is open and it cannot be closed, because it is clear that this can happen. And to some extent, I would say that this is an own goal of the Trump administration, because yes, ⁓ they are. might have been not like geopolitical reasons, because this this would be a strange move, because in the end, what is still possible is US citizens can still use the model. If it would be so dangerous, then why would you allow US citizens still to use the model? Because they can also be like working for other institutions, for for ⁓ even adversaries. Right, and and this will be the case partially. So I I'm not entirely sure if the if the geopolitical sorry, the geopolitical explanation is true. It could also be that it's just like these the the Trump administration doesn't like the woke AI. This could to some extent explain this a little bit, but it will also backfire on open AI because this what's true for anthropic right now could be true for OpenAI in the future. And hence what is clear right now is that this can happen, right? And the question that is open now is is this actually really a wake up call this time after the many wake up calls that we specifically in the U received already? Or are we going to hit the snooze button again? André Neubauer: Yeah, maybe before we like, where we can also share our, ⁓ our view, maybe also our professional view and how companies may react on that. ⁓ Yeah, but I just want to elaborate again on, one on one aspect. And I think it was, it was Sebastian: I think that's the purpose of this episode, right? André Neubauer: underrated. there was, you mentioned like Fable is Fable five is the first version of the Mythos ⁓ series. And there was a lot of noise around Mythos and the capabilities and Mythos itself, as you also mentioned is like, like even more powerful than Fable. And I think they underestimated from my point of view, ⁓ like even like, either it's a flex, right, they would just want to show the power. Or they really underestimated the power of that model because they, I don't know, for how many weeks and months they have been able to really assess the model. So I'm not 100 % sure if that was on purpose or not. You could argue, well, that was All planned, right? They wanted Entropic to release that, have a how you say that, a huge reaction from the market and then turning it off to really show who is at the end having the power on this model. Because again, I'm repeating myself, if I would have the chance to test the model, I would do this carefully. Like if I realized it's like, it's a real step up, right. In all, in all tests and all benchmarks. And then, right. Like you, you, you're surprised by the power, like, because someone else is informing you. It feels strange to me. It really feels strange to me. ⁓ but nevertheless, I think we now need to deal with the situation. Maybe one aspect I also want to mention is. I think it's not specifically on Entropic. So it's not specifically on Fable. It could be also like every other model. ⁓ So I think the state we are in ⁓ is at a level where ⁓ super high impact is not an exception anymore. ⁓ So might happen also to OpenAI. So their release, don't know, Codex 5. 6 Pro whatever ⁓ so that they may see a ban or yeah yeah so they may see a ban yeah Sebastian: Maybe subject to to these export restrictions, right? Export controls. Definitely. Yeah. And I have to agree that the story about the jailbreak, it doesn't seem to add up really. Specifically also since Andy Jesse from Amazon, right? Amazon is one of the big investors in Anthropic. They they would lose out if Anthropic would like have damages because of that. It's quite strange. André Neubauer: Yeah, yeah. Sebastian: Curious to at some point I'm sure it will come out. At some point we we will learn what happened and and how this happened. Also, I heard that Scott Besent, the the minister of the tre tre treasury, I think that's what it's called in in the US, right? He was the one like issuing this the w yeah. André Neubauer: ⁓ yeah. handing it directly over to the CEO of Entropic. Yeah. This is at least what I read. Maybe also a fun fact, sure if you're aware is like, do you know that Amazon is really invested into Entropic? So that is also like, this is, it's scary. Why would I harm a company I'm also invested in? Maybe the investment is so small that the risk or the downside is even larger, right? If you're not. Sebastian: Did he? No. Yeah, yeah. Yeah. André Neubauer: do not inform authorities, but I guess it's, it's a strange thing. Nevertheless, it is as it is. Sebastian: Yeah. Exactly. And at some point we'll learn what happened, really. Yeah. The thing is now what to do with that, right? And first and foremost, so I personally switched to Sonnet and iterated with Sonnet. That's it. And honestly, I I couldn't muster the energy to care a lot about this, to be very honest. I yeah. So number one, obviously I wasn't too impressed by Fable Five, even though I have to admit I didn't run Evelse, right? I just tested it once. It seemed to work well, but not significantly better than other models that I've tested lately. And number two, there the way that I approach the whole topic of AI or agentic engineering. André Neubauer: Why? Why? Sebastian: I'm using my Pi harness, right? Which is open source. I'm that the harness, it and you can also say that the harness is more than just the tool itself, right? It's what you do with it, right? The workflows that you use, the skills that you use, the extensions that you use. And this builds some resilience into my workflows. Because I really on one project that that I'm running, I'm using my Chat GPT plus subscription. And if this is fully consumed, I'm in my limit, then sometimes I'm working with Cloud Code where I have a subscription. And then sometimes even with my anti-gravity subscription. So that being said, obviously, this is not like this is not a posture for a company. How ⁓ could a company like position like that? There always needs to be a balance between Absolute maximum peak performance and resilience because if you are fully focusing ⁓ purely on performance, you always go with the best, right? And always go with the fastest, then you you are naturally not that resilient because you're so focused on on this, right? And I think to me, this is the the trade-off to be taken. So it would be wrong right now to say, okay, we don't use these kinds of models at all any longer. strictly go for local models or open weights models from providers. I don't know if open router would be safer or if you even have to go to European providers, be it Mistral, be it ⁓ like you you ⁓ book your own GPUs in OVH, Yonos or Stack It or whatever and then run your own models. ⁓ it would definitely reduce the The performance of the team if you would not use the the top models right now. However, building the possibility that you can switch into your setup so that you only lose, I don't know, 5% of performance and not fall back to 1% because you suddenly have to type code by hand. I think that is that is how I think about it. So looking at what are the factors that make André Neubauer: Hmm. Sebastian: us make the team, make me more resilient in the now while still being able to use the top notch models and and use the most performing that are available. André Neubauer: Yeah. Yeah. Makes, makes sense. I mentioned earlier that, I haven't quite made up my mind yet. ⁓ so like from a tactical point of view, I see a bandwidth of options or measures you could do, like maybe on the one side, like YOLO mode, right? Don't care. Like just move on, ⁓ except, ⁓ the dependency, except the risk. don't care. ⁓ On the other side, you could say, okay, that is now the wake up call. Like we need to, as you, as you just said, like we need to move everything to, to Europe, like, ⁓ get our own infrastructure. Like best case you do this in, in the seller, right? Like, so like you buy hardware. so, and I think, think you have a point where you say, I think the advantage of running, of using ⁓ like the frontier models is ⁓ too large. the downside, not just in regard to investment and effort to build that up, like building your own stack. think that is maybe for some business. this is what I meant earlier. Like it really depends on the context you are in, but if you are not really like dependent ⁓ on that, and well, like I hardly can think of a business which, which is, which would be maybe like. There might be, ⁓ like our listeners will tell us. ⁓ But other than that, I think the downside not using the latest models, they're in, they're out, given the fact that they change still so rapidly. Yeah. And then on the other side, I think going YOLO mode is also not an option, especially if you're working as like, if your company or in is already reached a certain state, I think as you also said, ⁓ And separating your harness, right? Like, so having some kind of abstraction so that you can in the worst case, right? Which actually like didn't happen. Like, let's be honest, right? They removed a model which was available for some hours. So I think maybe just a few companies probably have adopted it. ⁓ In case, Like really things go south that you could then like switch to something else, like go for like another on-premise as another model as a service or going for something on-prem. But like still being able to deliver the same quality because like your harness is ⁓ externalized, if you want to say so. Sebastian: Yes, absolutely. And fun fact, I actually had two meetings today about AI hardware and running models locally. But in our case, it's because they're on the horizon. There's the potential that we actually need this infrastructure for the use cases that we are supporting. And we also have one we have a DGX in our cellar, basically. We have a data center nearby where We have a DGX with four, no, sorry, eight A100 GPUs, where we I think in one of the episodes I described how we tested it was a Quen 3.6 with 35 billion parameters there, and it was like blazingly fast. So it was a lot of fun running it on just four GPUs. Couldn't saturate it with 16 users in LMPerth ⁓ benchmarking. And The model is quite capable for a lot of stuff. So I tried it even locally for coding use case. The speed was bad, but the quality was decent. So I I would say, again, right, I didn't run eval, so it would be hard for me to describe the percentage, but I wouldn't say it's even only 80% of the frontier models. It's probably even better than 80%. So being able to run this in a speedy way already adds a lot of Value, I would say. And then for specific use cases, like if you want to do knowledge retrieval, either you have like your data sources attached to it, or you have actual like pre pre processed data in in a database that you can use and you want to do RAC with it or whatever, this model would be totally capable and more than enough, right? And there are tons of models in this range. You could for smaller use cases you could go with a Gemma for there's a I think what is a 12B model which would comfortably run in on a sixteen gigabytes of RAM for and for really a lot of use cases this will work perfectly fine, right? Also like text-to-speech models we all know Whisper and faster Whisper and that there's a ton of different models. They don't need that much hardware. This is all working fine, you can run it locally. So it always so my my my case is rather it always depends on the use case, right? And the the most advanced reasoning and whatever capabilities that's something you would actually probably need for coding or if you don't exactly know what you need it for. But then even for coding, there are models that are now specialized on these use cases that are really good at function André Neubauer: I am. Sebastian: ⁓ sorry, at tool calling, specifically also for tools that are relevant for coding, right? Just I did not test it yet, but had planned to test the latest Kimmy, what is it called, 2.7 coder model, which seemed huge. So you it's not easy to run it on on your own hardware easily. But you can obviously get it from non US companies and I think even international companies because it's ⁓ open weight. So There are several providers that that run this model, right? So there are options. And even if going into hardware, there's also different kinds of options that you would have. So you wouldn't need to go the full range and go immediately to the 200 plus K euro mark where you run the latest, I don't know, H100 or 200 ⁓ GPUs. You could also n just use like something like a make mini that we discussed, right? Or the What's it called? I think the GB10 DGX Spark chip by Nvidia, where there's also from ASU A ACES ⁓ from Dell, and different providers have different also price points for for these computers. I think this could also, even from a from a cost perspective, depending on what you're doing, if you are constantly going above your subscription, basically, right? You're running complex use cases and you need a model that's more in the eighty billi billion parameter ballpark, then one hundred twenty eight. gigs of RAM computer might actually be no bad option also from a cost perspective, even though also they are getting more and more costly ⁓ because the components are so pricey right now. André Neubauer: True. I think in regard to infrastructure costs, you're absolutely right. I'm wondering, well, maybe this is also, I can admit this is a blind spot ⁓ of me. So ⁓ I think the effort bringing that ⁓ or making it work at scale, like production grade, ⁓ I think that is, you should not underestimate it, but maybe I'm wrong there. Maybe I'm wrong. Hopefully I'm wrong. I can also disclose we had a meeting today on that Actually wanted to double check whether we are still aligned on our ⁓ strategy. me, maybe also sharing one approach or one ⁓ I think it's good to ⁓ separate two things, ⁓ separate between two things on two axes. So on the one side, I would separate between internal and external use cases. like internal stuff maybe has lower right? You can mitigate that better ⁓ compared maybe customer facing stuff. And then I would separate between what is impact? So ⁓ in regard to, is it really blocker or is it just slowing you down, for example? And I think if you do the metrics, right. And I think then you know which ⁓ topics you need to work on, right? So you might have some features in your company, which are customer facing and which might be a blocker in case they are not available, these models. So that is maybe a good activity to end care, like very specifically on that product and have a plan B or some kind of, I don't know. mitigation plan, I think it's not disaster recovery, like it might be an aspect of that, by the way. So instead, or compared to ⁓ making the entire company bulletproof, right? Or resilient, ⁓ think like really focusing on ⁓ might have a larger impact on the company rather than ⁓ securing everything just ⁓ quotes, right? Just because of that. that incident. Sebastian: Absolutely. Yeah. I also thought about this case. My I mean probably Fable Five was too new to be actually used customer facing, I would assume, right? So there are probably no companies aside from Anthropic that that would bring this in front of cum ⁓ customers yet. But still the same is true for any other model, right? But then thinking about what is an actually an an actual use case where you would André Neubauer: ⁓ Most likely. Sebastian: in a product use such a strong model, even the latest opus model, probably for the for the product use cases that most companies are running right now, and I'm not sure if if there's actually already so many aside from big tech obviously and and the typical companies that are that are running AI use case or LLM based use cases at scale. They would probably not necessarily need the full thinking or I mean multimodality probably yes, because there could be different kinds of inputs, right? But then it's usually, I would assume, more ⁓ Rug kind of use case where you receive some input, right? Then you need to retrieve some some content from ⁓ some sorry, some some contextual information from somewhere. André Neubauer: Samples, yeah Sebastian: ⁓ and then create an answer for a very specific type of questions, right? In a specific context. And ⁓ you yeah, it should be relatively easy to use different kinds of models and even be fairly agnostic when it comes to providers. And at some point I would even argue it might be there might be, I mean, if you start with it, it's always cheaper, no CapEx and no operational complexity to use it out of the André Neubauer: Yeah. Sebastian: out of the box basically, right, from the provider. But at some point it might even become to see if you can run it cheaper if you host it yourself, right? I mean, for that you need a certain type of scale, specifically also because the older models are generally very cheap also from from providers, right? But ⁓ yeah, so that's why I'm I'm I I didn't even think so much about this case because it ⁓ doesn't to be so critical in a sense that it's probably fairly easy to find a different provider that provides a similar feature set basically, right? André Neubauer: Yeah. So I think you're absolutely right in that. So in that very special incident, think it's so only if a few companies have at all been affected ⁓ the other side, ⁓ don't know, right? Maybe next time they ⁓ down all services, ⁓ an order ⁓ Entropic to shut down all their services ⁓ of we don't know yet, ⁓ this ⁓ don't know yet ⁓ quite during the months and ⁓ So ⁓ I think ⁓ should be for the... ⁓ again, ⁓ not... I I would go for Frontier Models because of the quality, the productivity gains and so on so on. Still, I would not go for YOLO mode. I think you need to know the critical systems in your landscape. Sebastian: Yes. Yes. André Neubauer: and you need to be prepared for that. Does it need to be everywhere? Probably not. There might be some experiments which, where it's fine that they are affected and you can deal with that. You can mitigate still maybe fast. for the stuff which is deciding between your company is going bankrupt or not, I think you need to have a plan B. And if that is not, if this does not exist yet, then I would say it was a good wake up call. Sebastian: Absolutely. Yeah. Amen. Yeah. I fully agree. If you think about cloud code being suddenly taken away from engineers worldwide or even the cloud desktop app, right? That would be a big bummer. I mean, even there, you could probably just proxy it and route the requests to ⁓ something that you host yourself or maybe even open router or something. But still like just thinking about it and maybe testing it out and doing a failover test is something that should be on the radar. Yeah. That being said, I think yeah, that that's pretty much it, right? It's the the balance. So you should not just say now, wow, we freak out and we don't use anything anymore that comes from the US or maybe even China, if you wish. It's probably not not a good idea. On the other hand, also saying I I don't care, I I just go on like nothing happened. It's probably also not great. It's always the balance, right? And it it also again, like what we used to say in the past, it always depends, right? It always depends on what you're using it for. Are you using it internally? Are you using it ⁓ for products that you serve to to your customers, right? Which like there might be different criticalities and also the different use cases. I mean, a lot of engineering organizations right now without clawed code would be out of work, I would say, right? Or they they would not be able to work as long as Cloud Code is not working. And just like thinking about this scenario, thinking about what you could do in order to overcome it, either by proxying and routing to some other provider, or by saying no, we like invest a lot in the harness, the skills, workflows, extensions, whatever, and we are able to switch to a new harness very quickly so that we can use different kind of models in the new harness. All of this is relevant, I would say. It it should there should at least be a plan for that and ideally some tests already. So some capacity of some principal engineer or senior engineer or someone who likes to take care of that should go into These kinds of tests. André Neubauer: Thanks for the summary. I think we speak when there's the next incident. Sebastian: Yes. L so next week basically. Let's see. André Neubauer: Latest. Sebastian: Bye bye, see you next time. André Neubauer: Bye. Sebastian: to another episode of our new season of Beyond Vibecoding, partnering with Impala Search. ⁓ André Neubauer: the go-to tech and executive search agency in Germany. Sebastian: In this podcast, we are exploring the transformational change in software engineering and knowledge work in general. I'm Sebastian Heidemeyer zu Apen, CTO at NorthIo. André Neubauer: And I'm Andre Nobauer, CTPO at Trusted Jobs. Great have you back. This time, it's an emergency episode. Sebastian: Correct, our weekends went totally different than we had planned. Just kidding. obviously, ⁓ you may assume, we want to cover the latest event on Fable Five, the new frontier model of Anthropic. And in this episode, we present our view as two CTOs with a couple of years of experience. So without further ado, let's jump right