Shifting From Tasks to Responsibilities

Today on the AI Daily brief Anthropic has officially launched Fable five, the first of their Mythos class models I think fairly undisputedly, the best AI model we have ever been able to use, and yet at the same time, We are now at a level of AI models where how to get the most out of the state of the art isn't as simple as doing your same old prompts, but just with the new model On today's episode we' going to be discussing the launch, the benchmarks, the first reactions, and how to get the most out of Fable five AI Daily Brief is aaily podcast and video about the most important news and discussions in AI All right, friends, quick announcements before we dive in. First of all, thank you to today's sponsors, KPMG, Section, Zencoder, and O Systems To get an ad free version of the show, go to patreon. com slash Ai Dailybrief, or you can subscribe on Apple podcasts. And of course, if you want to learn more about sponsoring the show, send us a note at sponsors at aiailybrief. ai And by the way, yesterday I teased that in response to so many requests to make it easier to dissect and share specific parts of episodes, we were going to be experimenting with some new tools to do exactly that. Well, it turns out that Fable five liked what we had started, but thought it made some obvious errors, like not including timestamps on the little share cards with specific parts of the episode, and not turning the whole thing into a pipeline that could work automatically So it did that. And so you might be getting this sooner rather than later. Keep an eye out on the show notes and on aiailybriefed. ai for more of that But now let's talk Claud Fable five. On the one hand, this is not a particularly surprising release. First of all, it's been a couple months now since we heard about this new Mythos cllass of models. Some companies, of course have had access to them through Anthropics Project Glasswing, and when we got Opus four eight just a couple of weeks ago, they made it clear that they were working hard to get to a mythos cllass model that they could release with sufficient guardrails that they could feel confident about it being out in the public Now I guess what might be a little bit surprising about it is how quick the interval was between four eight and what we got in Fable five. But as we'll see, in a way that's much different than previous state of the art jumps, Opus four eight still has a pretty big role to play in the Fable five led ecosystem. Now then over the last couple of days, rumors started getting loud that some Mythos Class model was coming, and a little secret for you guys out there, if the loudest AI content creators on places like X are not responding to and participating in the rumor cycle That usually means that they have early access and that the rumors are true In this case, they were, and on Tuesday, june ninth, we got Caod Fabable five, and some others got Cawed mythos five Now first of all, let's talk about Mythos five, as it's almost entirely irrelevant for just about everyone here. Mythos five is effectively the same model as Fable five, which is the one that we got, but doesn't have all of the safeguards, many of which are controversial that we're going to discuss in a little bit. Mythos five will only be available initially as part of Project Glasswing and is being deployed to those Project Glasswing partners Anthropics says in collaboration with the US government as an upgrade to what is available now, which is Claude Mythos Preview. They say they intend to expand access to Mythos five through a broader trusted access program soon, but for now it is available only for a very small set of organizations. No, the big one for us is Fable five just from the name alone You can tell that Anthropic is treating this one as a big deal. First of all, we get an entirely new naming convention. We now have Haiku, Sonnet, Ous, and Fable, as in a class that is above Ous. Second, think about how long it's been since we got a lab that was willing to put a full new base number on its model. Indeed, the last time that we got that was the somewhat disastrous rollout of GPT five last August. All of those big transformations that we got around the turn of twenty twenty six came in model designations like Opus fourty five and fourty six and GPT five three and fivety four. So clearly here, just from a naming convention alone, Anthropic is not playing and no they are not playing. Regular listeners will know that in general, I felt that we were at a point where benchmarks are so saturated that it's pretty hard to derive much signal from them, and that even when one new model comes out and is a point or two ahead of the closest competitor making it state of the art The vibes in real world experience can be very, very different, meaning you basically just have to test these things for yourself. Yet sometimes the leaps are big enough that the benchmarks are worth paying attention to. And that's certainly what we got here. On Exploit Bench, the cybersecurity benchmark, Mythos and Faval five score a seventy eight percent, compared to, for example, GPT five five's thirty four percent On health bench, sixty six percent compared to GPT five five fifty one point eight percent. On the legal agent benchmark, GPT five five comes in at two point one percent, while Mythos and Fable five are up at thirteen point three percent. On GDP Val's test of economically valuable knowledge work tasks, GPT five five scored to seventeen sixty nine, Opus four eight scored in eighteen ninety And Mythos slash Fable five scored a nineteen thirty two. And then of course where the model really shines, and what its very clear purposes is around agentic coding. On Swibench Pro, GPT five five scores a fifty eight point six, Opus four eight scores a sixty nine point two, and Mythos and Fable five are all the way up at eighty point three percent On terminal bench where GPT five five was a little bit ahead of Ous four eight at eighty three point four percent Mythos and fable score in eighty eight percent. And then on a new benchmark, which we're going to talk about in a little bit, frontier code, GPT five five is at just a five point seven percent. Opus four eight is at thirteen point four percent, and mythos and fable five are more than double that at twenty nine point three percent Unsurprisingly, artificial analysis found that the model achieved the top ranking using their blended benchmark run, overtaking both Opus four eight and GPD five five. And while some noted that the overall gap wasn't particularly large at just five points Many point out that the artificial analysis agentic benchmarks are starting to seem a bit saturated Increasingly different organizations are trying to solve the saturation problem with their own benchmarks. Every, for example, maintains what they call a senior engineer benchmark that they say measures how well AI coding agents can rewrite a real production code base the way a senior engineer would In other words, it's meant to be a more real world version of an engineering benchmark. And for some comparison points, GPT five five scored sixty two percent on that benchmark, Opus four eight scored sixty three, and Fable five scored a ninety one out of one hundred Cursor has its own cursor bench, which compares performance and cost. I've talked a lot about how their Homespun model compomposer two point five performs at a similar level to GPT five five and Opus four eight at a fraction of the cost. Fable five absolutely bodies them in terms of the performance, scoring a seventy two point nine percent which is eight points above the previous best. That said, it is definitely more expensive on that cursor test Now, one new benchmark that's getting a lot of attention is the just released Frontier code benchmark that was unveiled by Cgnition earlier this week. Frontier code aims to be an ultra hard test for real world aggendic coding. Cognition worked with open source developers to put together a set of tasks as well as evaluation rubrics. The tasks were split into three sets, extended main and diamond, the latter of which is a smaller set of ultra hard tasks. Unlike other coding benchmarks, frrontier code uses a combination of unit tests and assessments of scope discipline, style, and adherence to code based standards. The goal then is not only to test whether the model could come up with an answer that passes unit tests but whether the code is high enough quality to actually be merged into a production code base. When Cgnition announced the benchmark, Sean Wang, who works with Cgnition and who runs late in space, pointed out that Mer, whose measure of long horizon tasks has become the standard for how we talk about the performance of different models, found that in his words, more than half of Swet Bench results is unmergeable slop Meaning that even if that code nominally solved a problem or did its job, it did so in a way that wasn't actually usable by the organization running the code That's what frontier code was meant to solve And that's the one that it more than doubled the previous best of Opus four eight That said, we're no longer in a world where we can just discuss how good a model is raw, we have to take into consideration cost. This is the constraint of the token scarcity era. API costs for Fable had been set at ten million per input tokens and fifty million per output tokens, which while double the cost of opus was actually at only double in air quotes lower than some people expected. Notably, this is less than half the cost of using Mythos's preview within Project Glasswing. One very weird thing about the rollout is that while it was great that Fable was available to clawed users immediately. We didn't have to deal with any long delays or rollouts, Andthropic is almost positioning what we have access to in the Pro tier and above as an introductory offer The company is warning users that Fable will be removed from subscription plans on june twenty third, and after that access will require payer usage, which while a bummer to claud users everywhere, is just more evidence that we are in a firmly usage based pricing paradigm from here on out Now in the second half of this episode, I want to focus on the early indicators of how people are using this plus my first tests, but we do need to talk through a few controversies first. There are many who are not happy about the guardrails that have been placed around the model. Bantg writes, The Clawed Fable announcement post reads like a spit in the face. It deliberately conflates Fable and Mythos and spends the majority of the time talking about capabilities that are completely absent from the safety maxed version available to the public Shubby, who very clearly is no anthropic hater, says The guard rails are way too strict, even the simplest questions get cut off immediately Now specifically, A lot of people are calling out how strict the guardwils are around any sort of biology questions. Cremo writes You're not even allowed to ask Fable about basic biology questions, let alone anything that could potentially be dangerous. They shared an image of them asking, tellell me about mitochondria. It's the powerhouse of the cell, right? Which got them a chat paused, Eedit and retry with Fable five or continue with Opus four eight message Daria Newt Maz writ The cancer is flagged as a biosecurity risk by Claude Fable V I also tried to code a website on cancer mutations and Vable five was immediately removed from my list. Basically as soon as he typed in the word cancer, it switched him over to Opus four eight Fernando also found that switch to Ous four eight when they asked, what's the process by which DNA makes RNA, saying, oK, this is getting a bit ridiculous. How are we going to live forever if we can't use AI to accelerate biotech progress? The blog post announcement did call this out. They wrote when Fable's classifiers detect a request related to cybersecurity, biology, and chemistry or distillation, the response is automatically handled by Cloud Opus four eight instead. Users will be informed whenever this occurs. Now they argue Opus four eight is a highly capable model in its own right. a response that falls back to Opus is a far better experience than an outright refusal from Fable They argue that early data shows that ninety five percent of Fable sessions don't have a fallback at all. And yet they also very clearly say in this blog post that for the time being, they're going to be particularly hardcore about filtering out questions on biology and chemistry. Effectively, they say that they've ratcheted up those guardrails because of the increased capabilities of these models Now I'm going to pick on Sporatica onX here a little bit because they summarized a strand of conversation that I thought was just a little bit disingenuous. They tweeted I mean this in the most sincere way, but if your aim is to release a product and respect your users and have them enjoy the experience, but your classifier cannot distinguish between what is a seell and a true biohazard risk, I don't think the product is ready for release. They also wrote, I'm sure a few people have gotten good stuff from Fable. It's certainly a powerful model. But the overwhelming response has been mass disappointment because most everything is just being routed to Opus, which we already have I think that this is utterly ridiculous. There is a subset of people who I believe would find something to complain about no matter what, who read in this blog post, that Anthropic was being extra hardcore about filtering out biology questions, and who to be clear have never in their life asked a biology question, and went to go do so so that they could see the promised result of the switch to Opus four eight and then come complain about it on Twitter Now I am not dismissing at all the actual biologists who are going to have some very big issues with this. Their beef is real, but it is incredibly important, especially in these early launches, to filter out that looking for something to complain about crowd The much more interesting critical conversation comes around the limitations around AI research Now they did mention this in the blog post A adding distillation to the list of classifiers that they were keeping track of But admittedly somewhat buried on page thirteen out of three hundred nineteen, in the system card, there's this critical paragraph. In light of the ability of recent models to accelerate their own development, we've implemented new interventions that limit Clad's effectiveness requests targeting frontier LLM development. For example, on building pre training pipelines, distributed training infrastructure or ML accelerator designs Using Claw to develop competing models already violates our terms of service, but enforcing this restriction through our safeguards avoids accelerating the actors most willing to violate these terms. Now this is in my estimation, very clearly, a response to Chinese models using anthropics research to develop lower cost alternatives. And yet, unfortunately, it is creating a dragnet that is going to catch up lots of very legitimate researchers Prime Intellects Ellie Bacouche writ Mythos will be bad on purpose on AI Fontchier LLM research tasks. This is very, very sad for the research community. They also write that the fact that it is on purpose not visible to the user is in their words crazy Nathan Lambert argues that lab starting to pull up the ladders on the ability to diffuse AI was inevitable, but also has issue with the invisible part, saying doing it without telling the user is misaligned Dean Ball calls this shockingly hostile and a terrible look, and one that could silently damage all sorts of work Samy analysis looks like it's already getting nerfed. They tweeted. Breaking news, Anthrop's latest model will not help you if it thinks your ML research or ML engineering is interesting, and or will secretly degrade its IQ so that the average engineer won't notice. We are already seeing Anthropics's latest models, moderation filters, or GPU inference research and programming. Gurgly Rose argues the belief of many saying, Anthropic trying to limit competition limits many others. But I think Will Brown from Pribe Intellight captures the genuine sadness when he writes, It's the first publicly available model that I am explicitly not allowed to use for my work because Anthropic holds the view that the work I do to facilitate open model research is harmful Now on the flip side, we have the people who can't believe the pearl clutching in surprise, like Tenebrus, who writes, Sorry, how exactly did you guys think this was going to go? You thought Anthropic was going to build the infinity machine that can cure all disease and prevent aging, and then let Friggin Eli Lillily extract that and get the patent. The labs are going to do all of it. You better believe that this is going to continue to be a conversation, especially with open AI staffers like Adam GPT, writing, Well, look at that. Open AI ends up being the openp AI lab But one other interesting quirk of the launch, I do think has some interesting implications In the section on Data retention practices for Mythos cllass models, anthropic writes To ensure we're responsibly deploying Mthos cllass models, we are requiring limited data retention and review as part of our safety work Prompts submitted to and outputs generated by Mythos Class models are retained for thirty days for trust and safety purposes on every platform where these models are offered. Roheat writes, Wait, how will any enterprise use Fable or Mythos if this is the case? Taylor writes PSA, if you used Clawe Fable five today with memory turned on, you just violated all your NDAs. Anthropic requires a thirty day retention policy including human review, and the memory feature on by default searches past chats for context, so sensitive historical chats get pulled in. Now I think that the dispassionate analysis would probably view this as a temporary constraint that Anthropic views as necessary given the power of the new model It does create some very, very serious challenges in the enterprise, such that I can imagine that this is going to stick around for long The last critical discourse that we'll discuss before we get into how to get the most out of Fable though, is about the question of token efficiency and how much this thing costs in practice. YouTuber and AI entrepreneur Theo writes I am so screwed. Current Pace has me out of fable usage in about an hour. Do I make a second account or do I pay API prices Chubby showed themselves literally hitting the end of their max planan limits, writing when you're having too much fun with Fable five West Windter writes Big labs should force their employees to have token limits. This would cause them to be more innovative, but instead they're becoming lazy and wasteful, whichich means we don't see any efficiency gains since they aren't affected by the costs On the flip side though, Tyler Willis writes, Early into testing fable, but so far it seems like the token Hungry warnings feel a little overblown. It does feel token hungry, but it doesn't feel categorically different than other recent opus models Alex Volkov from the Thursday podcast writes Overall token usage wasn't crazy And that's a good thing Referring to a big project that it spent one point five hours on, he writes four point two million tokens is not very token hungry. It could have been much more Fabio Jonathan goes farurther writing Fable is cheaper than Ous in practice. costs more per token, but one shot's way more often, so I'm not burning time and the amount of token reprompting Or as John vs. Malch puts it, actually solving the problem is token efficient, it turns out But what are the type of problems you should be solving with Fable five It's not necessarily as obvious as it might seem at first, and so that's what we'll get into in the second part of this episode. One of the most important AI questions right now isn't who's using AI, it's who's using it well KPMG and the University of Texas at Austin just analyzed one point four million real workplace AI interactions and found something surprising. The highest impact users aren't better prompt engineers, they treat AI like a reasoning partner They frame problems, guide thinking, iterate, and push for better answers. And the good news, these behaviors are teachable at scale If you're trying to move from AI access to real capability, KMG's research on sophisticated AI collaboration is worth your time Learn more at kpmg. com slash U slash sophisticated That's kpmG dot com slash US slash sophisticated Here's a harsh truth Your company is probably spending thousands or millions of dollars on AI tools that are being massively underutilized Half of companies have AI tools, but only twelve percent use them for business value Most employees are still using AI to summarize meeting notes If you're the one responsible for AI adoption at your company, you need seection Seection is a platform that helps you manage AI transformation across your entire organization It coaches employees on real use cases, tracks who's using AI for business impact, and shows you exactly where AI is and isn't creating value As the result, you go from rolling out tools to driving measurable AI value Employees move from meeting summaries to solving actual business problems, and you can prove the ROI guessing if your AI investment is working Check out section at sectionai. com SEC Ti O n AI. com Coding agents are basically solved at this point. They're incredible at writing code Here's the thing nobody talks about Coding is maybe a quarter of an engineer's actual day The rest is standups, stakeholder updates, meeting prep chasing context across six different tools. And it's not just engineers. Sale spends more time assembling proposals than selling. Finance is manually chasing subscription requests. Marketing finds out what shipped two weeks after it merged. Zencoder just launched Zenflow Wor It takes their orchestration engine, the same one already power en coding agents and connects it to your daily tools, Gira, Gmail, Google Docs, Linear, Clendar Notion. It runs goal driven workflows that actually finish. Your standup brief is written before you sit down. Review cycle coming up, It pulls six months of tickets and writes the prep doc Now you might be thinking, didnn't OpenClaw try to do this? It did, but it has come with a whole host of security and functional issues, which can take a huge amount of time to resolve Zencoder took a different approach. SoC two, tyype two certified, curated integrations, Tighter security perimeter. Enterprise grade from day one Model Agnostic and works from Slack or teelegram Try it at Zenflow. free This episode of the AI Daily Brief is brought to you by Out Systems, a leading Aentic Sstems platform built for the enterprise. Organizations all over the world are building, orchestrating, and governing Aentic systemystems on the Out Systems platform and with good reason Out Systems openp and unified platform allows teams to architect, deliver, and scale governed agentic systems with agility. Teams of any size and technical depth can use O Systems to build, deploy, and manage AI apps and agents quickly and cost effectively without compromising reliability and security Without Systems, you can rapidly launch ideas from concept to completion. It's the leading agendic Systems platform that is unified, agile and enterprise proven, allowing you to accelerate growth, reduce operational friction, and deliver real enterprise impact with AI. Out Ss, Build your aentic future So like I said at the beginning In general, I'm not really a fan of using benchmarks as a way to determine how a new model compares to what's available currently. And yet, in this case, obviously, the benchmarks were significantly different enough in a way that we hadn't seen for some time The econom had to assume that big changes were afoot And for people who really put this thing to the test It was just totally transformative. Ai Camilla writes, Fable five is something to pay attention to. The way I now spend my weekends has completely changed because of this new class of models. Ft first she writes, this is an actual leap T jump from four eightate, anything to five, anything sounds small, but the functionality shift I felt is big Within my first few prompts, I went, Ohh, this is it Your work is no longer nine to five. No chance. We have high performing models that can run for one hundred plus hours. How are you giving complex goal oriented prompts to these systems? How are you deciding what to kick off? How are you aligning your org on these tasks Reasoning is on another level I hammered the crap out of this model Fable five is the only model to answer a tricky word math problem, MBA level that I've tested on all the previous models, and not only did it get it correct, it verified its own work automatically and explained where the assumptions might need to change. Zero babysitting needed. This was the first anthropic model that I kicked off, went out to a long lunch with friends, kept my phone open, and didn't have to do squat to steer it while away from my computer just worked And this idea of hammering the crap out of it to use Allie's eloquent phrase was common among the people who were having the most success. Riley Brown's first test was to upload a McackKinzseie report and tell it to create a document of the same quality, which it did with absolutely no problems in his estimation But then he went harder. He prompted I want you to create a Swift app, Rplet mobile app. This should be a swwift app that builds web apps just like Replt. He then gave it a bunch of other criteria, like no need for auth. He let Fable decide the stack but make it awesome And it did Riley writes, I am in disbelief Clae Fabable one shot Repplt mobile, which is a mobile app that builds web apps The prompt was basically build an app like Rebblet that uses Daytona for sandboxing and convex for DB. Builds app, prereview app, open in B browser, Eedit app Later on he took it farther Guys, mythos slash Fable is AGI. On the left is the actual lovable mobile app On the right is my lovable version I built with Mythos in two prompts Later, he added, My lovable clone built with Claude Fable, builds Swift apps now and you can preview them in the app Four total prompts to do this Now a bunch of people took issue with the hyperbole here, of Riley saying that his version was better than Lovable, pointing out that there is a ton of infrastructure and surrounding work that goes into a company. It's not just an interface and a capability set. Others pointed out the fact that they had to talk about all those aspects of a company, while Fable effectively one shot at a performant version of that app was a fairly significant moment If you cruise around the halls of X slash Twitter, a lot of folks were building games as a way to test things Pident shared a driving game that they built from scratch By the way, as I'm describing these use cases, it might be worth switching over to YouTube or Spotify to see the video version In any case, Matt Schumer writes, Fable has solved three D world buildilding. Uutterly insane. This is all completely custom built three Js running in the browser Now when some people claim that the walkthrough was slow, he said, For everyone complaining that this is slow, I ran the prompt to make it faster without losing quality and voila sharing a faster version that didn't lose inequality Jake Fitzgerald, right Claude Fabable F to design a humanoid robot Two hours and one point four million tokens later, I got this, which is indeed a design for a humanoid robot. Absolutely insane, he says. Lissan on Twitter writes, Mythos five wrote this melody, which I absolutely love, and it also wrote this piano visualizer. And then there was Hugging face headad of product Victor, who has a benchmark where he asks models to design a Boeing seven hundred and forty seven using three JS, writing Fable has done an AGI level job on the Boeing seven hundred forty seven benchmark In Dan Shipper' write up as part of E's vibe check, he shared a variety of use cases that wouldn't have been possible before Dan writes. As I walk to work this morning, I listen to a two thousand seven lecture by the philosopher Hubert Dreyfus, the author of the seminal text What Computers C't Do I've listened to this lecture many times, but I always struggle to follow because the recording is grainy and muddy The version I listened to today was brightened, leveled, and crystal clear as if I was in the same room with Dreyfus. It was not on a finicky website, but on a custom web app on my phone that allowed me to see the whole lecture transcribed, and each sentence light up as Dreyfus spoke so I could easily follow along. Later, on my laptop, I wandered through a strange video game A Borg's Lbrary of Babel, an infinite library composed of hexagonal rooms, containing every piece of text ever written I picked books off of its endless shelves and wrote at spiral staircases And then because I also have a job, I read a report that synthesized hundreds of detailed every subscriber survey responses and our entire web analytics stack and identified our biggest conversion issue. It proposed a clean, falsifiable experiment that no one else on the team had previously suggested. All three of these are big projects that would normally take anywhere from hours to days to months, inststead, each one was made with a one shot prompt to fable five Now the fact that Dan was able to go from these cool demos to actual work is pretty important And when it comes to actual relevant work for the work world, some of the most common use cases that I've seen people raving about Fable five for have to do with migrations or interactions with massive existing code basases In their announcement posts, for example, anthropic writes During early testing, Stripe reported that Fable five compressed months of engineering into days. In a fifty million line Ruby code base, the model performed a code base wide migration in a day that would have otherwise taken a whole team over two months by hand Asadmafmuved from the smallmall square, used it to design a website, which honestly many many previous versions have been able to do, and said that it was just better I run a design agency, he writ, AI generated slot makes me want to close it fast. Fable didn't do any of that. Real hierarchy, intentional white space, restraint. the kind of decisions you usually only see from designers who've shipped real projects. No model has come close to this before, not one. Tuts Hunderss rights Mythos slash Fable is unbelievable, was on a customer call today and had clawed transcribing in the background as they were telling me about the features they wished their current software had Claude was building the features in real time. By the end of the call, I was able to show a fully working product with the exact workflow they mentioned fifteen minutes earlier. autonomous loops building triggered from a customer call And yet, If you look around, this isn't necessarily everyone's experience I used the bell curve meme to jokingly divide the responses that I had seen into three distinct categories For simple use cases, a lot of people felt like it seemed pretty similar. On the other end of the spectrum for extremely complex use cases It has been to many quite obviously better Now in the middle, I jokingly had a lot of people wringing their hands about how, while of course it was better for long running tasks, it didn't necessarily do everything better But the broader point is that I think that we are increasingly in a shifted paradigm One that we've been in a little bit before, but we are in a lot now, where the state of the art doesn't reveal itself across the entire spectrum of tasks, but instead within the context of some things that weren't possible before Trini research wrote I think we've reached the point where normal people can't really determine whether new models are better than previous ones. Like Fable doesn't seem that much better to me, but every one hundred fifty IQ person I know is like, wow, the singularity came sooner than I thought Now in my personal experience I would draw some contrast to the idea that basic use cases aren't better For example One thing that I noticed was that Fable five was really the first model that I've ever seen. to be able to both push back and disagree as well as to update the positions that it had previously disagreed upon in a way that wasn't obviously and predictably steerable I think many of you have probably had the experience Where it feels like an AI model, even a super advanced model like Opus four eight or GPT fivety five was disagreeing or offering an alternative path almost just for the sake of it And or When you then push back, it immediately flipped its opinion to the exact opposite in a way that again was just incredibly steerable This makes the strategic ideigation value of AI significantly decreased when the back and forth that it's offering is so clearly just trying to reflect what it thinks you want to hear Yesterday, I tested it by having a strategic debate about a direction that I want to take super intelligent in. disagreed initially in a way that was precise and clear but based on some wrong assumptions pushed back, articulating why those assumptions were wrong. And whereas in the past The model would have instantly collapsed and cowtawed to exactly how I was thinking about things. In this case, Fabled five did update its position to take into account The new information that I had given it, but it didn't back off entirely from its initial position All on its own is a massive upgrade just from a very basic day to day sort of use case that as we see in all of our AI usage pulse surveys is a big part of a lot of people's use of AI, that is strategic ideation And yet at the same time It is very clear that the real power in this model is around prereviously extremely difficult or impossible tasks, particularly if they involve coding So I'll give you three examples from my early experience First of all For those of you who aren't familiar, Superintelligent is our AI enablement platform. that helps companies understand their AI and agent readiness and prioritize what they need to do to get more AI native. We do that in a couple ways, but primarily through audits, where we deploy voice agents into an organization, which can then interview hundreds or even thousands of people all at the same time gathering way more information from the ground level than was ever possible before, and then aggregating and analyzing all that information to provide some very specific analysis around where a company is and what steps it might want to take next The product works really well, but one thing that I increasingly don't like about it is the approach to voice agents Unless someone was doing the interview entirely without looking at their screen, the Voice agent UX where you have to sit around waiting for the model to finish talking when the words that are saying are being transcribed in the window was just a really suboptimal experience. Now the real value of voice agents was on the input side, because users who are using voice ramble way more than they would if they were typing which means we get way more context and way more information. And when it comes to something like an agent readiness audit, the more context and the more information you get, the better. Luckily for us, turns out you don't need to use a full fledged voice agent to let people ramble You can just install something like the Whisper API from openp AI and do it that way What's more, we've also kind of Frankenstein super intelligent over time. So what did I do I asked Fable to rebuild the whole system with the new Whisper based input model And well It took a few hours which required me during that time to do exactly nothing and produce something that is frankly, fairly close to production ready in a single shot Now maybe I shouldn't be saying this because it somehow undermines the value of the software we've built, but our value was never in the software. It was always in the way that we collected raw information and turned it into actual signal, meananing that frankly, the more that we can do to make software get out of our own way, the better Next up, you probably heard me talk about the Enterprise Caw program, which was a formalization of cllaw camp that I launched earlier in the year, and what is a more hands on executive focused paid learning program that taught executives how to build agents. Now we have now had Hundreds and hundreds of executives go through three different cohorts of this Enerprise Claw program with a lot of success, but there are a fair number of companies for whom are approached with Enerprise Claw, which creates a lot of latitude for open source options, gives people the ability to actually use openenClaw, and is called Enerprise Claw Let's just put it this way, There are a lot of executives and companies who are never going to touch that with a ten foot pole So now, once again, in collaboration with Superintelligent, we are launching a similar but more enterprise focused version of the program that we're calling the agent transrformation intntensive Consider this your preview Again in one shot, I used Fable five to rebuild not only the marketing site for the agent Transformation intntensive, but the actual platform we run it on as well Lastly, and this may be the one that actually best reflects what Fable five does. I've been working on a new web experience for the AI Daily Brief that basically turns episodes into extremely shareable nuggets. The most important growth channel for the AI Daily Brief And one of the most value use cases is you guys sharing it with your colleagues This is also I've heard over and over, a significant value proposition for you as listeners is the ability to share specific pieces with your colleagues. However, that specific pieces part is a challenge, as the AI Daily brief, despite being daily, is quite dense So the idea of this new website is to actually chunk the episodes into relevant quotes, relevant sections, relevant numbers, where you can share just that piece Now with Opus four eight, I had already started to spec this out, and when I asked Fable five to go back and review what we had done, it basically said the problem with this is that it's just an idea, it's not production ready, and it turned what were effectively a bunch of fancy mockups into an actual production pipeline that I've now handed over to Claud Code to build for real, meaning you guys might be getting this sooner rather than later And the reason that I think that this is a good summation of my experience with Fable five so far is that it really does feel like a totally different world of delegating to the agent. Even with these extremely capable agents in the past, you still had to do a lot of management There is now frankly just much less of that management, which has the consequence, I think of upsizing the ambition Now this is what a lot of the anthropic staffers themselves described Alex Albert writes I've been at anthropic through every model launch. There's been a few cases I can remember of a launch that stands out and marks a step change in how we use models Claud opus three, Sonet three point five, opus four point five and now Claude Fable five With Fable, the models stopped feeling like a tool I direct and started feeling more like something I collaborate with. Felix Risberg writes I normally highlight the numbers, but I want to talk about something else, because with Fable five out in the world, I think a third era quietly started today. I lead cllad code and coork on the desktop, so I think a lot about how people use AI to get work done I believe we're about to see a major shift, moving from giving AI tasks to giving it responsibilities When LLMs first hit the mainstream, users ask them questions like a smarter search engine or an autocomplete for code Then the frontier moved to tasks, handing the model an entire problem, which bugged to fix what doct to write That's how most of our advanced users work with AI. They're in the loop. Every task starts and ends with a human. With Fable five, I've personally moved on to responsibilities or loops. I no longer tell Claude to investigate a particular crash report. It runs a loop watching every crash report that comes in Its job is to no longer help me fix a crash, it's to keep our apps from crashing The shift sounds subtle, but I think it'll change what AI products look like When developers went from answers to tasks, the primary tool changed from IDEs to coding agents. AI apps in twenty twenty six look nothing like twenty twenty four Predictions are a dangerous game But I really believe our industriry's apps in twenty twenty seven will look very, very different from the ones we have today So there are two big implications of this First of all, I think we all might have to develop a new skill around use case classification Basically, I think that in this paradigm of token efficiency, we as individuals are going to have to to some extent, become token efficiency optimizers ourselves understanding which use cases require different models Now, for a while now, people have given lip service to the idea that different classes or powers of models could be used for different things, but I'm almost positive that a lot of the power user type AIDB listeners are still the type to crank state of the art models to extra high even when they're asking for a grilled cheese recipe because screw it you want the power, that's why. With the Fable five class models coming online, especially as they move to usage based, I do actually think we're going to have to develop that muscle to understand which of our use cases require and fit each different power level of model Second though And maybe even more interestingly, I think that we are all going to go through a period of having to up level our ambition As someone who spends a lot of time looking at the frankly completely morbon landscape of AI training, even the best programs are still about how you use agents to do different versions or better versions of the work that you do today. Maybe they push a little bit in using new ways to write software to solve your old problems But even I think that is not enough. Nate B Jones, who many of you might recognize from TikTok or another short form video platform To describe the new skill we're going to have to develop as task imagination, and I think it's a really great way to put it. Anthropic released their new supermodel, right? Fable five. And Fable five, even though it's kind of nerfed because it's not as capable as Mythos five, the really dangerous one that was released under Glasswing, it's still super strong. I've been playing with it. And you know what that is making me think The thing that actually matters to most of us is task imagination We are sort of sponsoring magic with these models We have to have a practical guide for how to do magic with the models. because for most of history we've had two modes. Wave our hands and give a general guideline and hope people like get the idea and then walk away or do all the work ourselves and get super detailed. Having that middle layer of like, this is what I want, this is the bar, this is how it works. This is not very human. This is not how we typically have worked But with Tools like Fable five that can run for nine hours, twelve hours, days Days, do you have anything you can give AI that will take days? Let me just ask you that I know there are some people who do and when you do, put them in the comments. B There's going to be a bunch of us who are like, no, I have nothing that has ever taken remotely even an hour on AI. So what am I doing with Fable five We need better task imagination he breezes through it, but I love this idea of task imagination and that's something that I'm going to spend a lot more time on in the weeks to come. You know, somewhat ironically, yesterday's episode was called Open AI decelaring the next phase of AI. But with the release of Fable five It seems to really be the case then again For all of you folks out there who have shifted over to the Codex world and are now staring at your lonely clawed code terminal, wondering if you need to go back It may be worth taking just a beat. When Robert Corson tweeted, att this point, I don't want GPT five point six. It needs to be GPT six. No way Anthropic has completely blown past them like this. Three models in two months in Fabable is not even their best model. F feeleels like Anthropic ruined OpenAI's whole model roadmap and release plan. In response Tibo from the Open Aye and Kodex team, who now leads a lot of their product efforts, wrote feeleing pretty good about things. My friends, we could be in for quite a week For now though, we are going to end this very long edition of the Eia D Linato brief I Appreciate you listening, watching as always, and until next time Peace