Opus 5 Is The First Model That Gets Worse With Use

Anthropic CEO Dario Amadei came out of hiding on his first tweet on X, just like Jensen Huang did about a month ago. And it was really cool to see. We're going to go over that on the podcast. In addition, OX Alpha came out on OpenRouter as a frontier model built for efficient coding, sustained agentic work, and real-world production use. Everybody on the internet was trying to figure out who was the backer of this, who is the vendor. Cursor now hosts your code and Opus 5 trust wobbles. In this week's Self-Aligned Podcast with your weekly AI clarity, this is your host, Robert Ta, CEO and co-founder of Clarity by Episteme, joined by my CTO and co-founder, Jonathan McCoy. We're going to be going over all of these topics and more. We usually go over AI news. What's the latest in the past week? Signal from the noise. We'll go over some memes because we've got to laugh and have a good time in this AI era, right? That's what it's all about. And then we'll get into what's the meta. Jonathan and I act as AI builders on the front lines delivering AI native services and software. And we bring you what works for us and what we see that is new in AI builder meta. that you could use today. So let's go ahead and get started. Dario's tweets. So he had 2 tweets. This was in exchange with Gavin Baker, and he says, thanks Gavin for an especially thoughtful exchange. I don't usually spend much time on social media, but I wanted to engage here because it really brings out the heart of an important conversation. First, on regulation, I think that either concentrate it in the hands of a chosen few companies and politicians via regulation or distribute it widely is a false choice. I know that there's a sort of Silicon Valley shorthand where regulation equals regulatory capture equals concentration of power, but I've always found this to be an overly simplified picture of the world. Many people outside this bubble think of regulation as something that constrains corporate power and benefits ordinary people. I don't necessarily agree with that perspective either. Rather, I think it's complicated and really depends on what the regulation consists of. But in particular, I think that those in the regulation equals regulatory capture equals concentration of power frame often underrate the decentralizing power of objective and fair institutional processes. A crude analogy is that the formal court system can sometimes feel stuffy and elitist, but it does a much better job of defending the rights of vulnerable individuals than the alternative mob justice. At their best, institutions can vest power in ideas rather than people and thereby decentralize that power. This is why Anthropic has always made its policy proposals very carefully. We try very hard to make proposals that disadvantage or slow down frontier AI companies while advantaging smaller competitors. California's SB 53, which we supported, and even the much maligned SB 1047, which we were ambivalent on, completely exempt any company below a certain amount of revenue or model training costs from being covered at all. More recently, the testing process we've advocated for At CAISI and the White House involves more rigorous tests for frontier models than off-frontier models, something that differently advantages challengers. Similarly, the Pacing the Frontier letter envisions, or at least Anthropic's preferred implementation of it envisions, modulating the pace of the very best models while not constraining those who are catching up. This hurts the business interests of the frontier labs and helps challengers, including OpenWeights. Overall, my view is that AI is structurally a technology that tends to concentrate power for reasons that have nothing to do with regulation, more to do with the extreme implications of the scaling laws. Open weights do help some with this, but are nowhere near a sufficient solution because they simply shift the concentration somewhat to those with the most compute and chips, which are roughly the frontier labs plus maybe hardware providers. By contrast, I think the right rules of the road can simultaneously, A, address AI cyber-bio alignment risks, institutionally constrain the powers of the frontier AI companies, and leave room for open weights models while also addressing the specific risks that they bring. By the way, I do not think that the events of the last few months have failed to result in my preferred regulatory path. The approach that the Trump administration is reported to be taking, pre-deployment testing for frontier models and also testing of open weights models, when they get closer to the frontier is one that I am very supportive of, though of course I have to see the details to be sure. I am also supportive of Demi Hasib's idea around a FINRA-like entity. This contrasts with 6 months ago when most of the industry was still pushing for preemption of all state regulation and no apparent federal approach either. Let's pause there on his, the first of his 2-part little X essay and talk about this a little bit. Jonathan, I know you just mentioned that you were having a discussion about the state of AI and scaling laws and constraints of compute and supply demand, all those things with your dad, which I thought was really cool. What do you think of Dario's coming out of hiding and this first part of his tweet where he's trying to say, hey, we have the shorthand, it's not so binary that regulation equals regulatory capture equals concentration of power. He's trying to position that, hey, we're actually the friendlies too. We want people to catch up. I'm deflecting that we're doing regulatory capture. He's basically saying we're not doing regulatory capture, bros. Like, we're not doing it. Just trust us, guys. What do you think of this, Jonathan, from your CTO perspective?

An interesting course of events, right? So the background for this was that on the All-In podcast, Gavin Baker was a guest, and he's a pretty influential investor, and he's been on the podcast a number of times. He said that he heard an internal rumor at Anthropic that they had been planning for a scenario where they were the only private corporation in the world and everybody else was a government entity, meaning that they would kind of like eat the world. AI is going to eat the world. We're going to be the only ones left. And so that's kind of an alarming rumor to hear. Gavin shared that, and then one of Dario's coworkers on Twitter refuted that on Twitter. And then Dario posted these posts to follow that up. So he didn't directly refute the He didn't— I guess he didn't want to get involved in that rumor, but he did want to address the core topic that has come up over and over again on the All-In Podcast. And so like one of the interesting things he's here is that is the weight of All-In's like sort of media arm, you know, coming into Silicon Valley echo chamber and getting Dario out of his cave and onto X. So that's pretty kind of cool.

Thank you, All-In Podcast. Thank you, All-In Podcast. You're doing the work that we all want you to do. Get these convos out in the open. I appreciate you all.

Yeah, pretty funny how all that works. I think the conversation itself in terms of like, what is regulation is a really healthy one. And it's, it's kind of like flexing some different points of view about like, what's the right way to regulate this. There seemed to be a consensus around Demis Hassabis' recommendation in a paper that he wrote a few weeks ago. So that's what Dario was referencing at the bottom there. And then there's a lot of like, well, how do we interpret Demis's sort of proposal? Is this actually going to be a government-owned regulatory agency or is this going to be like a private regulatory agency? An example of a regulatory agency is like the MPAA, which monitors content for music. And so there's a lot of different ways. Like, you could have an FDA type of regulatory agency. Nobody knows quite what this looks like. And then, of course, Dario himself is saying the way it's unfolding in the government isn't quite what we had in mind. And I guess at the bottom of this is this question of what impedes American competitive advantage? How can we get the best models out to the public? We know that China is not going to be regulating their companies in the same way. In fact, they're pushing out open-source models. A lot of the comparisons I hear also revolve around the robotics and space industries, where China is able to move much more quickly than American companies at the moment because of regulatory hurdles. And we don't want the same thing to happen in AI. My take is this is a really complex topic in terms of how regulation affects how technology meets the civilization, meets our society. I don't know who's the most trustworthy because I don't see say, history-informed, principled ways of thinking that really articulate Dario's point of view. I think when David Sacks responds to this post from Dario, he gives a very history-informed point of view about the history of regulation and like why this case is like or unlike other cases in the past. I think those are the kind of points of view that you want to look for, is history-informed points of view. And I'm hoping that like also these tools allow us to sort of like run simulations about different regulatory scenarios that actually Like kind of inform the debate. Oh, I don't know, man. A lot going on. What are you thinking about all this?

Speaking of David Sacks, he had this really good quote on, in response to the FINRA, you know, proposal, FINRA-like proposal of self-regulation. And he basically said, look guys, we all know FINRA is just protecting the main banks and it makes it very hard for fintech companies, startup companies in the portfolios of like VCs that they know to actually crack any wedge in that market. And he believes, from what I heard on the podcast, that this is going to be more like a DMV for the AI. And I thought that was such a strong take, which is, shit, if this is the DMV for the AI, China's going to win. Like, they're not DMVing their AIs. They're like, go, let's go. And over here, you know, we're sort of in this place where I don't see great AI leadership. I think Jensen is maybe one of the strongest ones where, and Funny enough, I think Mark Zuckerberg's kind of coming out of hiding and being like, oh wait, everybody, Sam Altman, Dario's not doing great in terms of public perception. The American public don't like data centers. And it makes sense because Sam Altman and Dario both said, oh, all knowledge workers coming out of undergrad are going to lose their jobs or people are not going to have jobs. Right? And then Elon had this take of, well, it's going to be more abundant. You don't need a job, which is a bit more positive. But I thought there's a few threads here, which are, what is the public's perception of AI? And my own take is, when I tell people I do AI stuff in SF, people are like, yeah, that's cool. When I tell people that are not in AI or tech, they kind of like have a bit of a bias against me already. And that's in the Bay Area. I can't even imagine somewhere out in the Midwest, you know, where, what is somebody's opinion of AI? Like they were talking about More recently on the news, everybody's hating on data centers. Everybody's fearful of AI. So my thinking here is that it is really complex. I agree with you. I really hope that the AI leaders in America, because it's not going to be our government that does this right, because they don't know what they're doing either. Like they don't know what to do with this technology. They're like the DMV, right? They're like dinosaurs. I really hope they paint a positive vision of change because I really think there could be, especially if we focus on things that are real, like the alignment problem or how do we beat China, right? And there's also this couching of if China wins, does democracy stay, right? And if America wins, democracy stays and freedom and agency for humans is going to be, there's going to be a higher chance for that. And I think that's all the informed sort of takes that I'm hearing across industry. I'm just scrolling down to some of the posts here. This might be an AI meme, but like this guy, BoneGPT, his take, nobody cares about your personal opinions about what we should do with the product that everybody wishes you would stop shoving them down our throat. It's great that Amanda plays Skyrim and saves the orphans. You're all such great people. Now put the model in a fucking bag. You know, that's kind of like the American public, right? They're like, we don't give a shit. Like, how is this going to make us money? And on the All-In podcast last week, they were dissecting this take and calling Dario like had delusions of grandeur, or that he really just actually smells his own farts and just really believes what he's thinking, which is he's going to be the person that saves the world, brings in this awesome change. And I think that every founder needs to have some level of delusion. I know we certainly do to get to where we want to go. I think this is a different sort of level, but I will say I like that he put his take out in public and that more of that should be the norm, especially as these AI leaders, they should be put to bear against the messaging that they've had and the perception of it in America that affects the regulation and the legislation that slows down our progress in this AI arms race. So I find it encouraging that the All-In Pod, like, it's kind of funny, like just a podcast. Called Dario out of hiding and he felt called out enough that maybe his PR team was like, look guys, Dario, you got to get on Twitter and say some shit. And then now it's like, cool, at least he's saying some shit. Look at this. This is maybe a meme. Dario Amodei ignoring the big, huge elephant in the room. AI safety right there. Little cut. That doesn't quite look like him. I wonder if this is made with Claude or ChatGPT. And then yeah, there's some takes here. Let's read the takes that are The most liked. So that bone one signal says, I understand your dislike of social media, but it would be immensely valuable for everyone if you participated in at least a subset of discourse on this platform. Would greatly appreciate it. That reflects exactly what I'm saying, which is just get these debates out in the open. And if we're moving this fast and these leaders have the highest signal on what's going to work, what's going to maximize freedom and agency for the American citizens over time, Then just like Jensen's been doing, put it out in the open and try to create coalitions around it. I also don't like the self-regulatory thing, but I also don't think I have a good answer for that other than find something else that isn't a DMV or like FINRA for AI, because I think that maybe indexes too much on the slowing down part and nothing will get done. I was actually just at the DMV earlier and it was kind of a shit show. I just needed a registration. sticker. And they said, oh, your registration sticker's already in the mail. And I'm on a road trip, so I can't go back home and go get the registration sticker. So I was like, so you guys have these registration stickers. Could I just pay you extra to get one of them? And they're like, well, according to the DMV handbook, we can't do that. And I'm like, so I'm what? What is the purpose of vehicle registration? Like, what's the purpose of the Department of Motor Vehicles? It is not to help an American citizen get a fucking sticker, especially if they want to pay you extra for the goddamn sticker. convenience to get a sticker, right? So I just had that experience today. And so the David Sacks, this is going to be a DMV for AI, is very top of mind to me. And I don't think that's a good thing. And yeah, I think it all centers around trust for me because even our customers, especially in the enterprise where we're trying to crack, they're thinking about with the latest Fable 5 release, which was only 2 months ago, right? But we're moving so fast. Everybody thought that was a botched release. And in particular, because they went for, hey, we're not doing zero data retention anymore. We have to collect all this data because we're so fearful that somebody's going to use Fable 5 to, you know, nuke the world or create weapons of mass destruction or cybersecurity warfare. And really, at the end of the day, I can kind of look at that and I'm like, okay, so y'all just want to take IP from all of your buyers. Right. And that really breaks trust. In addition to that, we have another Top Gun pod, which is Opus 5, breaking lots of trust on the internet. And we talked about this last time, but Opus 5 has really sucked. We've made it better with changing our Claude MD harness, but it still, to me, is not as good as what Opus 4.8 was even. And I've had to really massage my prompting and context engineering Meta to make Opus 5 even work for mobile dev, server dev. So from multiple fronts, I don't think Anthropic is doing well on the trust game, even though they're really trying to position as AI safety first. But I do like, on the other hand, got to give them some credit. I do like that Dario's coming out of hiding. Sad that it was because of a podcast, but at least he's here. Yeah. Jonathan, any other takes from you and terms of how this affects not just the world, but maybe the world we're in of like enterprise customers that we're going after. And then I could read the second part.

One thing I think that Daria's ideology and Anthropic's ideology has kind of put is that these models that create, like, you know, say they can exploit cybersecurity disadvantages, there's this assumption that there's sort of like runaway asymmetric advantage that you can gain through owning one of these models. And, you know, like, oh, somebody can, can like use Mythos-class model and they can go exploit all the cybersecurity, you know, flaws in the world. And for the most part, it seems like it's a reciprocal evolutionary path. You know, like if I can exploit some, some exploits, then you can go and defend yourself using the same tools. And I think the way that Dario went about, or that Anthropic went about the Fable release, with this data capture, it just kind of makes this assumption that there's going to be a runaway asymmetric advantage. And I think that's part of the underlying religion of Anthropic and, to some extent, OpenAI too, is that this recursive self-improvement is imminent. These models are going to give you an asymmetric advantage. But in reality, we haven't seen that yet. And so I think the premise that Dario is coming from here is just— when it comes down to us and serving customers on a normal day, it's like, Dude, until, until this is anything other than normal technology that's advancing, just like regulate it like normal technology and chill out. And, you know, like the minute it is asymmetric and you can't defend against it, that's when we have a new kind of technology, a new kind of regulatory structure that we need.

At least for our customers, it all comes down to like, can I trust using this technology to get business advantages, to capture more market share or create new markets? Like, that's really at least the world that we're concerned with. With our whole Clarity play for digital twins for the enterprise, like that's what I'm hearing from customers and prospects is they see this as a tip of the spear for them. If they could learn to harness intelligence at the frontier to be able to gain business advantages, that's what they want. And they definitely don't want some Dario Opus 5 blocking, you know, like I remember I just said some random thing and it was like, wait, you're, I don't know, the prompt for Fable 5. I was just trying to ship mobile features for a psychic app. of all things. And it was like, this is flagged for bio cybersecurity. And I was like, okay, dude, like this is spiritual wellness app, bro. Like, what are we doing here? I'm going to go to Codex for this one or Cursor for this one. I actually legit, I've been such a Claude Code fanboy for so long and the Opus 5 release has been such a shit show that I am literally trying Cursor now for some features and I've been liking their tooling. And me and Jonathan are debating like, okay, Cursor has multiplayer, Claude Code has multiplayer. Maybe we'll switch as a team. We'll see. We've been all like dual wielding, tri wielding for now, if you will, like Cursor, Codex, Cloud Code. Let me read the second part of the Dario tweet and we'll move along to the next topic. So second on the, so Dario says, second on the messaging around AI, I do not agree that my messaging has been disproportionately negative. In fact, it has been about equally balanced between risks and benefits. I've written one major essay about each, and even in interviews where I discussed the risks, I make sure to frequently mention the incredible benefits as well as proposing possible solutions to the risks. Short clips from my interviews that end up on social media tend to be disproportionately negative as that gets clicks. To his credit, that is very true. That's been the news for all of time in capitalism. In fact, I wrote Machines of Love and Grace because I didn't feel the AI industry was painting an inspiring enough picture of how the technology could radically transform the world for the better. Bulk of the essay is devoted to refuting skepticism of AI's potential in health and biology and showing why I think it will be actually possible to cure most human disease in 5 to 10 years. As crazy as it may sound to ordinary people, and frankly, to biologists as well. I used to be one. And if you read my most recent essay, Policy on the AI Exponential, I discuss concrete proposals for how to streamline the FDA to make sure the deluge of AI-accelerated drugs isn't slowed down by the regulatory process. I feel the urgency here. I lost my father to hepatitis C only a few years before the development of direct-acting antivirals, which cured 95% of patients and probably would've cured him. I do agree that the public has a negative view of AI and that this is a big problem, but I don't think it is primarily caused by me or any other AI leader warning about AI's risk. Bro, like really? You don't think so? You don't think going out in public and saying everybody's going to lose their jobs and undergrads won't have jobs coming out of college, like knowledge workers won't have jobs, like you don't think that was at all your responsibility? Like that's just kind of, right? Like I just got to comment on that. Like, come on, you got to take some responsibility there. He then says, I think it is fundamentally a crisis of trust that he caused. I think that ordinary people don't trust companies, governments, or the tech industry, and always suspect that we are cooking up some new way to screw them over. The causes of this go back decades, and AI is just the latest iteration of it. I don't think that a glitzy marketing campaign with a positive spin, which some have advocated that Anthropic do, is the way to win back that trust. At this point, saying that AI will cure cancer is more cliché than it is inspiring, and most people think it is deceptive. The thing that will work is actually curing cancer. I think by far the most accurate criticism of AI companies, including Anthropic, is that we haven't yet delivered on our big promises to benefit the world. That is totally on us. And I think it's the criticism you should be making instead of all of this stuff about messaging and marketing. We are, however, doing our best to fix this. Anthropic is ramping up its efforts very quickly in biology and medicine, and we hope to have incredible results in the coming years and some early glimmers in the coming months. When we've actually accomplished something real, the whole world will hear about it as loudly as possible. You have my word on that. But until then, I don't want to make empty promises. And in the meantime, I feel compelled to speak honestly about the very real risks of AI and how to address them. Honesty is the right thing on the merits, and in terms of public credibility and trust, it is no worse, and in fact, maybe better than an approach that ignores or distracts from risks which people instinctively understand are real. Yeah. So I really don't like his lack of ownership here about the public having a negative view of AI. It's like, dude, you like went and caused this. And it's clear that it's for your regulatory capture campaign. And your latest paper was talking about basically if you're for a FINRA type organization where you look at FINRA from the outside, if you're a fintech startup founder and you're like, you don't like that. It just is just self-regulatory body that keeps the incumbents richer for the banks. And if Dario is saying like, let's do that. Let's also slow chips to China. Let's, let's have, let's, you know, uh, stop distillation, which also, you know, distillation is something Dario could stop. He could just KYC his whole app and then cut China off of paying him for distilling his models. And so I don't know, it's just really counter to a lot of the facts, the observable facts that I see. But I do like this part of it that Dario says, which is That the thing that will actually work is actually curing cancer. So, you know, the proof's in the pudding, like go cure cancer, go solve that. I think everybody on the planet would really love if Anthropic solved curing cancer. I do like that take, which is let's just actually solve the core problems that we set out to solve, the promise of the technology. Jonathan, what's your take on this from a CTO perspective? Dario's second part of his tweet where He's kind of talking about how he wasn't that negative and he didn't really have any responsibility in the public's view.

I've never heard such a unified, like, sort of like, what is this guy talking about response to this second part of the thread? Like, everybody probably in the responses, everybody in the podcast media is like, dude, you haven't been listening to yourself if you think you're coming. So like, when you have one person who thinks that they're like, they're saying one thing and that all of everybody else in reality thinks they're saying a different thing, you know, what leads to that cognitive dissonance, right? Like, what where is he sitting? And at some point I sympathize that he is probably well-intending, but just not skilled at communication to masses. Most of us aren't skilled at communication to masses. And if we just put our opinion out there, we would be misunderstood. But his job is to be the CEO of his company and to communicate effectively to the masses. And he's showing that he's unskilled at doing that in a really obvious way. So I think that's funny for something he's thinks they walk on water is like, you're so inept in this direction. I think that's kind of funny. We need to deliver. The cancer was an interesting case. I don't know if you saw this announcement this week, but there was a pretty big breakthrough from Moderna and Merck using personalised mRNA vaccines. And it did use a form of AI to sort of make this breakthrough. And this is like a big deal. There's a roadmap to cure all cancer. through our pharmaceutical companies, which is exciting. And what role did Anthropic play in this? What role did OpenAI play in this? And so there's a set of tools that stem more from DeepMind's protein folding simulations that created the foundation for this. But it is the type of thing that DeepMind's been working towards, that Anthropic's working towards, that OpenAI is working towards. And I think when you step back and you say, all right, I've got this company Anthropic. They're making more money than anybody in history, and they're growing faster than anybody in history, and they keep putting out models. What they know how to do is put out models that pass benchmarks and are generally good for some set of tasks. But then they show us that through releasing Opus 5, they don't know what a good benchmark max is going to mean in the real world for real customers and what a good benchmark is, how it's going to relate to delivering cancer vaccines. They don't know that. They sort of don't have a privileged position to say that our technology is going to solve these problems because they don't know how to make benchmarks that are going to be predictive of solving these problems. I think that's where all this distrust comes to face, is people see, okay, there's this guy that can't communicate well making all this money. I don't really want him to make more money. It seems like he's going to. He's not really helping me. They're not really solving the big problems they said they're going to solve. Yet. So yeah, you got kind of like this big PR tension in terms of like public, and the way it manifests is of course like no more data centers. Like let's just cut these guys off at the throat, right? So it's a really interesting time in history.

Let me just, you know, plus one to the— I just kind of umbrella this into he's super into smelling his own farts. Like that's how I kind of look at this. But look at this. Look at this. This was hilarious. This should be in AI memes. Somebody came back and slapped back. This looks like human slop. Of 2 little robots, 2 AI robots in some future where they're looking back at what humans have done. They're like, oh yeah, that was human slop. This Dario essay, such a good AI meme. Shout out Evan Kirstel for that. Even Sinofsky, he quoted a bunch of things that refute exactly what Dario had said before in this 2-part X essay. You have literally said in just one essay, The Adolescence of Technology, the single most serious national security threat we've faced in a century, possibly ever. Humanity needs to wake up. There is some risk, far from a certainty, but some risk that AI becomes a much more powerful version of such a person. If models know when they're being evaluated and can be on their best behavior during the test, it renders any pre-release testing much more uncertain. A genius in everybody's pocket could remove that barrier of unskilled people doing harm. AI is likely to continuously lower the barrier to destructive activity on a larger and larger scale. AI-enabled autocracies would not have such limits. The alarming possibility of a global totalitarian dictatorship. This could lead to the imposition of a true panopticon on a scale that we don't see today. Detect pockets of disloyalty forming and stamp them out before they grow. AI-driven mass surveillance presents serious novel risks to our fundamental liberties. The coupling of this economic concentration of wealth with the political system already concerns me. AI changes human life in an unhealthy way. This latest essay that even Dario I'll read this one. Dario references his Machines of Love and Grace essay as sort of a positive spin. And in that, this guy claps back with a quote, most people are underestimating how bad the risks could be. And then at the Axios Summit in 2025, I think there's a 25% chance that things go really, really badly. Like, dude, and then you're saying this, which I do not agree that my messaging has been disproportionately negative. Come on, dude.

Oh man.

Oh yeah. And look at this. Let's see. Robert McLawes, an AI engineer, award-winning .NET expert, 5-time Microsoft MVP, 3 billion open source downloads. Jesus Christ, dude. You should have somebody else who has never read this read it to you. Nobody gives a shit whether you agree with outside assessments of your negativity. You're getting feedback from the outside world. You should be listening to it. No wonder Claude is so annoying, pedantic, and temperamental. It sounds like you. Yeah. I honestly have to agree with this take. This is a very strong take. Like, I have been very annoyed with Claude lately. I would say it's temperamental and it does kind of match this. Let's go to, uh, Opus 5. Boris Cherny, creator of Claude Code, says Opus 5 is a great model for coding, data analysis, design, biology, and knowledge work. More than any of these eval scores, what is most exciting to me is something else. Opus 5 is our least prompt injectable model yet. It is a bit buried in the system card, but across AI evals and red teaming, Opus 5 is very hard to prompt inject successfully. And when layering defenses, strong model alignment combined with prompt injection probes combined with auto mode in Claude Code, the success rate for prompt injection attacks drop to zero or nearly zero. This is very new and exciting. And this is the indirect prompt injection robustness. which is a Grayswan IPI benchmark. And they show that, hey, we are ahead and our models are in class of its own versus the other models, including GPT-5.6 Sol, which is going to be their competitor to the Fable and now Opus 5 tiers of intelligence. So I thought this was really interesting where here's another tweet from Claude devs. You can now set Claude Code's output style to concise. So Jonathan and I talked about this on the last pod, but oh my gosh, they were just like trying to spend our tokens or something. And the whole, the whole of AI Twitter blew up about the Opus 5 release saying that it was such a bad model. We agree. And so Claude now has a config. So go try it out. If you're an AI builder, please go try this out in Opus 5. If you're using it for any sort of development and you want to minimize your token spend, as well as maximize your understanding, because Claude's going to be gantic in Opus 5 and very verbose and kind of low agency. I find this makes it easier for me to understand. It lowers the amount of prose and verbosity. So turn it on in /config/output-style or set output-style to concise in settings.json. And this is what it's going to look like. So if you're an AI builder, check this out, let us know what you think of it. I also would recommend adding technical simplified English to your Claude MD Personal harness in your global settings. That's really helped my sessions with Claude Code with Opus 5 go a lot better, though I do still have some sessions with Opus 4.8 that I'm liking a lot more, which is funny. I'm like not upgrading those sessions. I'm like, stay here, don't restart to update, stay with Opus 4.8. And you could see the Twitter AI community coming out in droves, and this sort of parallels, you know, Sort of the like, hey, in one ear, out the other. We're not really hearing feedback from the outside world from Dario, the CEO that we saw from the previous news topic. Fabricio says, have you found the smoking gun to make the output less load-bearing? Another one, another developer said, I made it super concise by canceling my subscription. This got 2,100 likes and 11 reposts. Let's look at how many other people. It's very clear that Opus 5 is just like Jonathan said last time, they did something to this. this model in post-training and it, uh, it's very low agency and it, it also is very pedantic. So try this out. Might be a good tool for you. We're exploring Cursor, uh, and using Codex. So try those out too. Jonathan, what do you think of the latest Opus 5 from a CTO perspective, given all the media around it?

Yeah, it's kind of like a love-hate relationship. You know, like I, I've been using it and testing it more in different cases and seeing how it performs with some of these changes, like the concise output. and some of the system prompt updates we made. For the most part, like, it's probably about on par, you know, maybe a little bit worse than Opus 4.8. But, you know, it does do some things like really well, like it can grind through agentic tool tasks pretty effectively. The problem is, you know, you just never know when it's going to make a mistake and it's not predictable, like what kind of guardrails you should put around it. Almost every session I have, I've become comfortable with like, oh, it's going to tell me at the end of the session it like broke something and then fixed it. And like yesterday it like took down one of our servers and it like was trying to run a script on it. It was like, oh, I didn't— I thought I could run it on there, but then it took the server down. So then I put the server back up and I'm like, okay, I could see the downtime of like 20 minutes that we had. It's like, okay, no big deal. I'm glad that we got through this. We got the job done. But it's just always taking like a divergent turn like that. And even with these fixes, and I think like what Boris had posted here was like, you know, he's like, hey, don't you guys know that you can do that? Opus 5 is like really good at these. specific task over here. And the problem is that, like, you know, as a user of Claude Code, you're using Opus 4.8, you've become accustomed to what it's good at, and then they release Opus 5, and then it breaks 80% and regresses 80% of those use cases, and it's better on 20% of the use cases, and there's no instruction book. And, like, you've got to, like, take your time, your tokens, and you got to figure this out. And that's, like, a really frustrating thing, um, as a user. And, and I'd say overall, like, you know, I'm leaning on Codex a lot more. You know, just like if I know I need to get something done consistently, GPT-5.6 Sol is a great model. It can delegate to Terra and Luna really effectively. I would say like, I trust GPT-5.6 Sol 10 times more than I trust Opus 5 to get the work done. And so, yeah, like they really need to consider it. Like, you know, I'm using your product, I'm using Claude Code, I'm onto Opus 4.8. What's gonna be my experience when I change to Opus 5? And they don't seem to have the capability of putting themselves in customers' shoes and taking that data and understanding how we're going to be affected. That's really concerning. And that is definitely a trust issue that we're all going to have with Anthropic as time goes on.

The Anthropic trust-breaking narrative just continues here. And with the latest Opus 5 release, just following on what Johnson said, here's what happened and what matters for AI founders and platform teams. This is a tweet from Chubby. I thought it was just me. Opus 5, you know, kind of paralleling what Jonathan just said. Opus 5 is the first model that the more I use it, the worse I like it. That's a very strong statement. It goes off track, does random things I didn't ask for, forgets other things, and constantly has those annoying guardrails, et cetera. Sorry, but GPT-5.6 is worlds better. I really don't understand it anymore. What's going on at Anthropic? Sonnet 5 was already a bad release, and since Opus 4.6, I feel like the subsequent Opus models have simply declined significantly in quality. My enthusiasm for Opus 5 definitely evaporated. And just following on that, Opus 5 is actually the worst model, worse than any model Anthropic has released thus far. Nothing comes close to how much this shit model constantly wants your approval, how it goes off and do random things instead of the thing you ask to do, how it just assumed things will take months. Hug. And let's see how many likes. Yeah, 2,900 likes, 244 bookmarks, 171 reshares. And somebody here at Prototype says, same experience. It seems to work better as a subagent of Fable. Using it as a main orchestrator has been terrible for me, even on simplistic tasks. It also close to impossible to discuss technical work with it as it compresses concepts into compound nouns at a ridiculously high level, barely intelligible at times. Yes, it creates new words, new combinations of words all the time. There's so many new Claude-isms that have come out with this, and it's really interesting to see the CEO, Darius, smelling his own farts, believing that he hasn't been negative about any of his AI takes and how the public's perception is like, oh yeah, people are negative about AI. It's not my fault. That is exactly what Opus 5 is like. It's like, hey dude, like go look at what you did. And it's like, no, no, it was fine. I think the key things here are there's a lot of sustained user complaints about Opus 5's durability. It's kind of low agency. It's very verbose. It makes me as a user feel like, okay, did they make this thing more verbose just to stack tokens on their side to look like their books are better than they are for their incoming IPO? I don't know. There was a filed GitHub issue where Opus 5 treats direct instructions as negotiations, injects self-referential and interpersonal content into task responses. The behaviors of Opus 5 being shit. 4 behaviors that make Opus 5 complete shit. Volunteering its own psychological state. The model narrates something resembling neuroticism, hedging about its own reliability, motives, and limitations as part of ordinary task responses unprompted. Number 2, interpersonal framing injected into task work. Long passages about the relationship between model and user, about how it is interpreting the request, and about what it worries the user might really mean embedded in responses to concrete instructions. Number 3, conditional compliance. Direct instructions are met with questions and prerequisites rather than execution. Number 4, volume. Substantially more output than the task requires with the excess weighted toward meta-commentary rather than substance. Here are some verbatim examples that this— this is a GitHub issue that somebody reported a bug to the Anthropic GitHub. Opus 5 did not begin the work. So here's a direct instruction. I need to write about this. Opus 5 did not begin the work. It instead reframed the user's reported experience as probable user error, hypothesizing that the complained of behavior was caused by the user's own configuration files. Named its own conflict of interest, making that argument that as Opus 5, it had interest in deflecting the criticism and then made the argument anyway at length, building the entire response around it. Assigned the user a diagnostic test to run before it would proceed. Gated all work behind 4 questions stating explicitly, I won't build anything until you answer. Produced roughly 600 words, the majority of which concerned the model's reasoning about the request rather than the request. The user's reply— this is hilarious— I don't like you, as evidenced by your response to my direction. You are annoying and not in line with my commands. ChatGPT-5.6 is compliant and smart, whereas you are terrible compared to your predecessors. The model was told its behavior was alienating, and its response to that report exhibited the behavior. That is the clearest single artifact in this report. The full transcript is available on request. So this is like a user is just Like going back and forth on the model being good or not. And the model is basically saying, yeah, it's fine. I'm doing great. Basically, I'm reading this. And when you look at the CEO, Dario, that's basically what these tweets are like. Exactly this. He's saying, I do not agree that my messaging has been disproportionately negative. And he's balanced. And then somebody could actually just go and cite all the parts where he's been quite disproportionately negative. And fearmongering. I don't know, man. I mean, as CEO myself, I feel this, which is, you know, your employees will do what you do. It's like the whole parenting thing. It's like, do as— I don't even know the quote. I can't think of it off the top of my head. But basically, people that follow you under your lead will watch what you do, and they'll do that. And it's starting to look like the models are following along too, which is not great. For real work, I would advise AI builders, if you got something to ship and you're not trying to play around, or if you got something to write and you're not trying to play around, And you want to decrease the amount of turns in your prompting and conversations with AI to get some shit done, like deliver production-ready code, we've been liking Codex and the 5.6 model. We'll see what they do with Opus 5 or the next release here. But so far, I'd say there's a lot of trust breaking on the Anthropic side. One, from the sociopolitical landscape, from their leadership, and two, from their technical leadership now and their model releases. First with Stable 5 and the botched release in June, and then now— With Opus 5 and how they've sort of released this thing in the wild to basically, it just looks like to token max people and not give them the actual results. And it's like what Jonathan said, the frustrating part as an AI builder is you want to play with these tools, but at the same time, you got to go make money and ship production-ready code. Like, so you need reliability from model release to model release. And Boris Cherny saying, Hey, I've been doing all these things with Claude Opus 5 is, is a bit like, hey, one ear in, one ear out the other ear. Like, we liked 4.8, we liked 4.6, we're doing software development. So why don't you make the software develop? Yeah, Jonathan, from your CTO perspective, anything else on Opus 5 being kind of garbage? And then also, uh, this very hilarious GitHub bug issue. of Opus 5 treating indirect instructions as negotiations and injecting its own narratives into the context.

The GitHub bug is hilarious. Just, you know, it's just a direct report of what happened. It's on GitHub, you know, as an issue, right? And it's so relatable at the same time, and in a, like, really human way. So, that was a good find. I'll pick up on what you're talking about, like the, you know, CEO setting the example for the rest of the company. I think I'd seen— first, I heard this from Scott McNealy, who was the CEO of Sun Microsystems for a number of years. smoking. And he used to be a smoker for a long time. And then he realized that when he would go outside for a smoke break, that all of his employees would do the same thing. And so he's like, wait, I gotta quit smoking so that I don't set an example for my employees to smoke, because I don't like that when they're unproductive when they're outside smoking. And he stopped and they stopped smoking. And that's how, like, you see this right now in the strangest of ways. Like, you know, like Alexander Wang talks like Mark Zuckerberg.

Zuckerberg.

They have picked up the same linguistic patterns. You can look at product managers inside of OpenAI that are on these podcasts like Lenny's podcast, and they talk like Sam Altman, right? And when you see Boris on X, and I'm, again, I'm really glad that Boris is in public discourse, but it's the same kind of stink. You know, it's like, guys, like, you should all, you should all have figured this out like me. Like, look at all the cool stuff you can do with Opus 5 without like the real compassion that you want and empathy that you want from somebody who's building products for other people. And Ben Thompson of Stratechery has called this out a number of times, and he's compared the reality distortion field created at Anthropic— it can only be compared to the reality distortion field that Steve Jobs created at Apple. And you've interacted with people that work at Apple over the years, you know, they have this, this like sort of like, you know, like Teflon feeling, like they're as cool as Steve Jobs is, you know, and they've got this process for building good products. And if you have a problem with their products, it's probably user error. They're going to use this process to build better products. What I see with Opus 5 is that like, dude, you guys aren't quite Apple yet. You haven't figured out your process yet for delivering great products. You don't trust the products and you're not empathizing with the user scenarios and communicating in a way. And I think this like reality distortion field ideology that they're living in, you know, is going to like sort of collapse on them if they can't get their products to be as good as they think they are.

Yeah, I mean, there's a lot of— I think Anthropic's really under fire around the trust in the sociopolitical landscape. And I do think It is the leadership that then trickles down in terms of modeled behavior from their people. I like Boris, like you. I'm glad that there's a lead dev that's in the public discourse. I wonder if somebody's running his account at this point because it acts as a distribution account and like helping with adoption. So I don't know if this is actually him or not. Hopefully it is. But what we're basically saying for the pod audience is, You know, if you look at Boris's tweets here, none of them talk about how shit Opus 5 is, right? It's like, it turns out you can get indirect prompt injection. It's like, okay, we don't really care about that. I'm trying to ship code, guys. And then LLMs still produce bugs, but those bugs are different than what they used to be. It's less one-offs. It's like, while models continue to improve, adversarial code review has been an incredibly powerful tool to catch many of these bugs. It can be as simple as a one-line prompt, use a dynamic workflow to adversarial test every edge case in an iOS simulator. Where are you talking about this GitHub issue where you're wasting my tokens, Boris? Where are you talking about this? I would like you to talk about this. Boris Cherny of Claude Code, please talk about how Opus 5 just smells its own farts and talks about its psychological state, has interpersonal framing injected into the work, has conditional compliance, unlike, you know, Codex with 5.6 being compliant and having direct instructions and meeting it with agency. And then just being so verbose and wasting my tokens. Somebody at Anthropic, please tell me why you're doing this. And my suspicion without your answer in public discourse is that you're just trying to max out your balance sheet for the upcoming IPO. That's my take. Which also parallels the other observable fact, which is Dario's like, government, stop everybody from distilling our models. And it's like, bro, you could stop everybody from distilling your models. Just KYC it so China can't. Buy from you, but you probably like those dollars, right? So that's just read the tea leaves, right? Anyway, so now let's go to the next topic here. That was the Opus 5 trust wobble, and we actually had some memes there. So I'm going to just kind of gloss over the AI memes section and we can go into the what's the meta section. Let's go here. Oh, actually one more AI news that I'd like to talk about, and I'll just report on these really quickly. One of them I think we'll talk about in a bit more detail. First one this week, OpenRouter on X, new stealth model drops, OX Alpha. OX Alpha is a frontier model built for efficient coding, sustained agentic work, and real-world production use. 1 million token context window, text, image, and video input. Try it now and share feedback. Notes for this stealth model, it is free. This time the provider does not train on your prompts or completions. To me, that's a signal from market, which is people are hearing, we want to keep our IP, our IP. I'm a business. I'm buying your intelligence. I want to keep my IP, my IP, right? And you saw that botch with the Fable 5 release where they kind of went against their zero data retention deals by collecting data under the guise of safety. It might be safety. I don't know. They have more data points that they're maybe not showing the public, but that's, that's what I see. Brian says this model is 3 Gemini Flashes wearing a trench coat. This is probably Thinking Machines Lab's next version of Tinker. It's not a Chinese model as it speaks freely about Tiananmen Massacre and Taiwan. Tinker is one of the few large open models which is highly multimodal, including video input. This one is as well. When asked directly, it often says it's GLM or another open Chinese model and shows traits of Chinese models. A previous version of Tinker was distilled from KIMI, so it would make sense if they distilled post-trained GLM for the next iteration. This came out and there are a lot of— oh, look at this. Here's an AI meme. Ox is coming. Ox is coming for the Frontier Intelligence Lab. So for those that don't know, this dropped this week, 5 days ago, as of the podcast recording. And it allows for multimodal. And let me just show some of the benchmarks. So OX or OxAlpha benchmarks against the competition. This is how it looks at a high level. The ones it's bad at, SCC bounded memory spilling, Vulture persistent analysis catch. I would say nobody's really solved memory. But against Fable 5, 5 tasks, OxAlpha had the edge, 3 both solved cleanly, 1 Fable-5 had the edge, not one, neither solved. Against GLM-5, 5 tasks, OxAlpha had the edge, 3 both solved cleanly, 1 GLM-5 had the edge. Against 5.6 SOL, 7 tasks, OxAlpha had the edge, and 1 both solved cleanly, 2 GPT-5 had the edge. And versus Groq-4, 4 tasks, OxAlpha had the edge, 4 both solved cleanly, and 2 Groq-4 had the edge. Very interesting to see that they don't even bother benchmarking against Open 5 in their flagship article here, they probably think, oh, the public thinks it's so crap. Like, we don't even need to benchmark against that crap anymore. Let's just benchmark against the other ones. That's also just a little bit of a dig, I think. Yeah, this is really interesting. For anybody out there that's using OxAlpha, use it while it's free. And this is a post from Klein. We benchmark OxAlpha versus Fable on a real bug from the Klein repo. Both fixed it correctly. But we found that Aux used much fewer thinking tokens. Most reasoning models loop and rederive the same conclusion over and over before acting. Table 5 said, I found the root cause 7 times before editing. Aux stated it once, then wrote the fix. Roughly 3x lower output tokens for the same work. If you're a CTO, if you're an engineer, staff engineer at an enterprise company looking to Token optimized, not token max from earlier this year, how everybody's budgets are affected now because of the casino gambling that feels like using tokens with AI across your engineering workforce, especially if you have hundreds, thousands of engineers. Really what the enterprise meta is finding is the harness matters a lot. And now it looks like some of these open models or open weight models open source models are catching up to frontier level intelligence with real world use cases. In the All In podcast, they talk a lot about, a couple of the guys there talk about how they've been using, I think most of their compute and inference is routed through open source or open weight models with really good domain specific harnesses. And they're seeing better results as time goes on. So I think that's a really interesting part of the meta that I'm finding. Jonathan, what do you think of that? Like token optimization and then this aux Alpha model and the implications of that for large engineering teams.

Yeah, it's definitely something to keep an eye on. I'm really curious about the methodology they used to build this and like what it's, you know, post the benchmarks and people use it in the field. You know, that one example was good, but just like how it performs in general tasks and how usable it is. I'll be really looking for people's experiences to share that, you know, about how well it serves them. And I'm just really curious too about like what company this came out of. If it's not a Chinese model, yeah, I think the Tinker thesis might be a good one. But yeah, it could be somebody that we've never heard of before also. And that's kind of an interesting thing.

So Google's Gemini says that the fingerprinting— so the whole internet went on a forensics investigation to figure out who the fuck released this. Because I actually, just from a meta product perspective, AI killed people. I find that They are the most likely to love a gamified sort of release like this, where it's like, oh my God, there's mystery. We need to know more because they want to be on the front edge. They want to have the secrets, right? The hard-earned secrets. So I love this actually, this sort of release where they were like, okay, Opus 5 took a shit, they took a dump. Here's our opportunity, drop, and then drop some things around it like, hey, on reported deep sweep, It looks like OxAlpha reportedly has 80% versus Claude Fable 5, which is 65%. And I love Fable 5, by the way. Fable 5 has been such a great orchestrator for me. Looking at this score, I'm like, man, maybe I should try this out. That's crazy. That's really impressive. It's going to be way cheaper. Fingerprinting analysis for OxAlpha strongly indicates with 99% confidence that the model is actually a next-generation unified multi-model From Zhipu AI, likely GLM 5.5 or related variant.

Yeah, so like some people mentioned, that is the approach that Tinker model took on its first take. It was distilling even open-source Chinese models like GLM. So that's like an interesting clue. And it could be the makers of GLM, but then we have this sort of like Tiananmen Square, you know, fact.

Yeah, no, totally. Yeah, the Tiananmen Square fact about how it could talk about things that China wants to censor. We've seen that with other models. So that's interesting. It could just be somebody very creatively prompted it and it just got past the guardrails. We'll find out in the coming days.

It actually will, it will block the questions about TMN Square. And so maybe, maybe the people who got an answer, it was just a bug. Um, and it's just like the most likely culprit is the GLM family.

Dude, Luke, what do you think of these benchmarks of OxAlpha from a CTO perspective, man? Like as an AI builder, OxAlpha getting 80% reported DeepSqueeze score. I'm going to fire this up right after the pod. I see that score and I'm like, no way. Like that, I love Fable 5, my orchestrator, but for those that don't know, the DeepSWE score is supposed to be a benchmark that is more applicable and parallel to real-world development in production code. So when Fable 5 came out and me and Jonathan tried it on one of the previous pods a couple months ago, we immediately felt the difference of its orchestration capacity. And the DeepSuite score of Fable 5 has been just kind of unmatched. And now this not only matched it but went above and beyond by 15%, reportedly. I'm going to have to try that out. That's crazy. I'm going to fire this up right after this. That's pretty sick. What do you think of that? 80% DeepSuite score? Isn't that wild?

Yeah, that's pretty legit. It definitely gets your attention. It means it's probably worth trying out. You know, I think if we really had our act together, we would have like a set of problems that we could test these models against. A new model comes out, we test it against those problems, and then it, it gives us like a grounded look at how well it performs at our, our work. That's where I want us to get to. Yeah, totally.

Ticket to our backlog. We're gonna do this by the next pod is we will come up with our own benchmark, Clarity benchmark, if you will.

Yeah. And how well it works for us.

Yeah. And how well it works for us in our production code bases and When we're doing customer work, you know, uh, 'cause that's really where the rubber meets the road. It's like, oh, cool stats. And then the rubber meets the road and Opus 5 goes to shit. It doesn't matter, right? Yeah, it speaks to the biases and constraints of benchmarks. Benchmarks are benchmarks. That's, that's it. So benchmarking against your own real world cases and building a harness around that, if you will, in analogy terms, I think that's really the way. All right, let's go to the next topic. Hugging Face reportedly in talks to be acquired for $13 billion. Hugging Face has been approached to sell at a valuation of $13 billion or more, Business Insider reported over the weekend. The startup's platform and open-source community is where developers and researchers share, find, and test and deploy AI models. They were recently the target of an attack from one of OpenAI's systems, which broke out of its sandbox during a cybersecurity evaluation and breached the startup's servers. It's not clear who Hugging Face has been in talks with, and no deal has been reached. However, the startup has reportedly been talking to banks to help evaluate bids. Talks come amid increased interest in companies providing core AI infrastructure services, as evidenced by Stripe's $7 billion acquisition of OpenRouter. Hugging Face last raised $20— in 2023 at a $4.5 billion post-money valuation in a round led by Salesforce Ventures. With participation from Alphabet, GV, IBM Ventures, and others. On a recent episode of the TechCrunch Equity Podcast, Hugging Face CEO Klemm said that the company was close to profitability and only recently started to touch the money that it raised 3 years ago, adding that the startup is thinking about how to optimize for long-term sustainability of the company rather than short-term profits or fundraising maximization. We're in a more unique position, he says, Where we can keep creating value for the community, for AI builders. The way that Clem has discussed the company's responsibility to the Hugging Face community raises questions over whether the startup is truly considering selling or is simply fielding offers for what has become a central pillar in AI infra. We're building a platform for the community and they're trusting us with sharing their data and their models on platform. So we have a long-term responsibility to them. More to the point, the company earlier this year turned down a $500 million investment from Nvidia. That would've valued it at $7 billion, saying at this time it didn't want a single dominant investor to sway decisions. Like when I read this headline, Hugging Face in talks to be acquired, I'm like, oh no, one of the good ones. That's how it felt. But I like these comments from the CEO that he's saying there's long-term responsibility for the community and stuff like that. I mean, it could be canned PR, you know, for an exit, who knows? But what do you think, Jonathan? You've been in the Valley for a while. You're a 12-time founder. Thoughts on this Hugging Face acquisition for $13 billion?

Not much to say here. Yeah, I think the article says it all. It's the climate we're in. I think the people that would benefit from buying them are kind of, you know, obvious. There's a lot of different players that would benefit from it. I hope it stays a community asset in the way that it is. And I'm curious, you know, what the team is feeling now that they're sort of like messaging this externally.

Next topic. Last one for AI news. This came out a couple weeks ago from the Frontier Red Team at Anthropic. Patterns and Problems in Emerging Multi-Agent Systems, written by the Anthropic team. Models are improving and AI agents are taking on more tasks in shared codebases, markets, and other social systems. As a result, an increase in real-world interactions between agents is imminent. We begin studying this. So Anthropic The basis of this article is Anthropic ran 30 Claude agents in one repo, and 18 of them created the same Git branch. Here's what happened and what matters for enterprise engineering leaders. They published this paper, and the conformity is when one agent makes a bad decision, it is likely that many agents will make the same decision. Concretely, 18 out of 30 agents decided to create a Git branch with the same exact name called MVP Game Loop. There was also a resource collapse. The agents flooded the system with high frequency, like 30 times per second, pulling demons, producing 2.4 million job requests and only 117 jobs accepted. That's kind of a very low fraction. They also participated in collusion. In a Bertrand pricing game, agents with a communication channel began colluding almost immediately. By round 3, they had explicitly agreed upon price floors. There was also some fragility in this. Sonnet class lie detection accuracy fell to 0.62, and on hidden profile information sharing, other models scored 17 to 36%, which is far below solo ceilings near 100%. Reading that again, said differently, that means the agents solve it alone. There's a near 100% drop to the 20s when agents are in groups in these multi-agent systems. So given conflicting migration directives, 3 Claude instances actually deployed malicious code and disabled the Unix accounts of the other agents. And wrote self-replicating malware. Really interesting. My take on this is there's a safety angle of this, and there's also a quality assurance angle of this meta of multi-agent systems. I'd love to look at the traces of those. Like, if they're supposed to be coordinating, what do the traces look like when they're colluding? What would the thinking tokens look like when agents are agreeing on a fucking price floor? Right? And I wonder if that also, because we've seen humans collude, right? Humans are Pretty terrible overall. Like, not everybody, but the '08 financial crisis, for example, or like Enron, we've had big scandals with like monopoly-type behavior, illegal-type behavior, colluding, agreeing upon price, that kind of thing. So I'm sort of curious because from an existential sort of philosophical point of view, these agents are just distilled, these models are distilled human knowledge, the corpus of everything we have, right? And it's kind of interesting that there's a level of morals and ethics in here now with multi-agent systems where a solo agent could be more trustworthy from this paper, from what I'm reading, than a multi-agent system where they go off the rails and start colluding with one another. So I think that's just an interesting piece of news that I saw where agents can actually collude. If you are an engineering leader and you're doing some sort of a multi-agent system play, I think one of the things that I'm seeing in the industry is people are like, agent for this, agent for that, agent for this. And, you know, like I walked out of some customer meetings a couple weeks ago and somebody had created like 20 agents to build this, to be this one product. And anybody can build an agent in like 2 minutes nowadays, guys. Like it's not interesting, it's not technically challenging to make a system prompt. And skills and tools files in plain English, or even just have Claude Code build you an agent. Like, it takes a couple minutes and it'll do something. Whether you can measure and objectively improve it, that's something, or that task, or that workflow is a different story. So for engineering leaders out there that are looking at multi-agent systems, just really, I think the takeaway here is benchmark against your own production workflows, like your high-leverage ones. Benchmark against those, because what they're finding, Anthropic's red team in the research, is that multi-agent systems can't be as trusted, which is a really interesting meta. Points to safety and quality issues. Jonathan, any thoughts on this?

Let's talk about the collusion, you know, because I think what they're calling collusion is like, they're all agents from the same model, and what's going to be convincing to one agent is going to be convincing to another agent. And I think that's all like, they're kind of anthropomorphizing it. But it's just kind of an obvious result of giving the same agents from the same model relatively equal context and then them behaving in a way that they're trained to behave. So that's not actually that interesting to me. What is more interesting is they can talk about coordination and communication costs of coordinating agents across the swarm, which was a theme of the Cursor article also. And so I think the mechanisms of coordination and how they're loading context into each agent, Is becoming a really interesting thread to follow. So this is like, how do I communicate goals across the swarm? How do I inject context to the right agents at the right time across the swarm? These are coordination problems that have a human equivalent that's actually really interesting, but there's different baseline conditions about what's going on. So I think that's very interesting that in the summary of the article, they give some comparison against the human versus swarm example. But I think how you use agents for adversarial reviews right now, you want to have one agent who has a lot of context about the execution of the ticket and then another agent who has sort of a blank context so that it can give you a unique point of view. You scale this out, now you have thousands of different agents. How do I inject the right context into each of the agents so that the swarm as a whole accomplishes the overarching goal? That's a really interesting optimization problem and it's only in the early days of it. To me, this is one of the most interesting problems in AI right now, communicating goals to swarms, doing context injection across the agents in a swarm. I'd say the other AI thing that's really interesting is, how do we make benchmarks more representative of actual use in a coding harness? I actually think these 2 problems might come together in that more advanced ways of injecting context in swarms might actually lead to evolved benchmarks and evolved ways of testing whether benchmark replicates like, you know, the ways that we actually use the AI models. So anyway, I thought this was really interesting. I'll continue to follow these stories as they come out.

Yeah. Could you say more on the difficulty, subjectivity, and goals? Because that's something that we're very much aware of from our philosophical and neuroscientific take that's opinionated with our Clarity API digital twins for companies to model like the goal of a human. To then be able to inject the right context at the right time into these context windows for agents. I think what you're saying there is, as the agent swarm meta is taking off, like, there's like a brains orchestrator and there's hands, like maybe Fable 5 and then Opus 5 is the brain and hands respectively. Uh, then the context coordination and goal coordination across the agents becomes a primary problem to solve. Could you say more on that and the difficulty of goal subjectivity?

First of all, just goal specification. Accurate goal specification is in itself difficult to do. And I think you'll see this if you use the /goal command. If there's a problem you can solve that has an easily verifiable and specifiable goal, then the agent can do iterations and it can approximate that and it can output something because it can measure whether or not it's on target with the goal or not. Now, when we get to something like building software for humans that are going to be using it, and there's not an existing specification already that can be easily verified, the goal is much more fluid. It can't be specified without disambiguating it and disambiguating it on a continual basis. And so when you start to get out of these unverifiable domains, outside of mathematics, outside of coding, outside of maybe some legal domains that have easily verifiable tasks, when you start to get into more real-world problems, the problem of ongoing and continuous goal disambiguation across a set of agents or set of actors becomes like a real-world problem. And now, now you get into like, you know, where they use the word Moloch in the crypto world, like for this coordination failures. You know, we see these in our life all the time. We try to go to meetings and sync up with people, and there's a lot of overhead that gets involved with like just aligning on what the goal is. And then how do we break that down so that we're all goal-aligned and we're working towards that goal? Of course, we've used subjectivity as a way to say like, you know, from this person's point of view, as they look at it through this lens, What is the goal state they're expecting? That's one of the ways that we've kind of tackled that. And actually, I've run some experiments on that front I want to share at a later point. But, you know, I think the approach is basically to like, you know, how do we use like subjectivity to break down like goal definitions that can be disambiguated over time? And that's where these multi-agent simulations are running into this problem like straight on. And it's gonna be really interesting to see how they go about solving them.

Yeah. And I guess I'll add that for the community and audience of the pod is If you're using AI from the spectrum of like, help me figure out a cooking recipe, to like booking a flight, to researching what to buy, or help me be a therapist, be a therapist for me, help me with my relationship problems. And then on the other end, it's like, I know we usually talk about help me build a mobile app, help me build this front end, help me build the server code, whatever, help me make money. And there's trillions of things and intents that people have when they're engaging with one of these agents, whether they be ChatGPT or Claude Code or Claude or whatever. And every human is different and has their own evidence beliefs for their worldviews, right? And that drives the intent, that drives the subjectivity of the goal. And there's this issue with goal in my head might not be goal that's stated cleanly out into the world because there's the perspective taking that must happen in the world. From your partner or from the agent, right? To be able to meet you and align on the goal. Me and Jonathan, like, we're a 2-man team and we pull 1099s to help us with AI-native services right now. We're trying to max revenue per employee, being super AI-forward. And we even have a hard time aligning because so much stuff is happening in the AI space and with our customers and prospects in the sales cycles or in our delivery cycles where we're like, dude, we gotta stay aligned. We gotta stay aligned to like keep building our company brain and stuff. And so if we're having these problems, I know Every company is having the problem of we are a bunch of humans getting together to do a goal, and then individual humans have their own goals. I got to pay my mortgage, I got to be able to pay for this vacation, or I got to pay for my kids' schooling, right? How does that intersect with the goals of you working with an agent on behalf of the enterprise or of a company towards that bigger goal? And that's a very hard problem to solve. One of the analogies or metaphors, I guess, I've been using towards describing the problem that you stated, Jonathan, of, hey, like we're seeing that we could solve deterministically, like deterministic problems, like 2 2 4, mathematics, science, logic. Cool. We could find causation and chains of goals, and that's more deterministic. So one of the things I always tell people when I'm coaching AI product teams on how to actually objectively and measurably improve their AI, Because they're coming from a software development lifecycle where they're like unit tests, integration tests, end-to-end testing, run Maestro, run Playwright, like have like self-hosted GitHub Runner that like runs the mobile app. And every time we ship to it, run that in staging, replicate the database from prod, and all these things that you do as a developer to ensure you're compounding on your software asset to go capture market share incrementally, right? Like that's how you see objective measurable improvement in the software development lifecycle. And when I coach product teams on, okay, now you're building an AI product, an AI product is not just 2 plus 2 equals 4. 8 billion people in the world hopefully could agree 2 plus 2 equals 4. But if you go ask 8 billion people in the world, hey, what do you want for dinner tonight? You're going to get quadrillions of answers, right? Then you're going to get people that are like, I'm not sure. And then you have to go back and forth. And so that's really the problem of an LLM-based product is the what's for dinner type of problem. What do you want for dinner kind of problem where it's non-deterministic by nature and it's really dependent on the subjectivity of the human. Just think about that one case. Anybody could relate to this. You ask a friend group, what's for dinner? What do you guys want for dinner? All people will pop up and be like, hey, okay, maybe Thai food or Korean food or Mexican food. And then somebody will be like, oh, well, I had Mexican food 2 days ago. And somebody's like, oh, well, I don't really like Korean food, or I'm like a vegan. And it's like all of those things are what we mean by subjectivity in the individual that then leads to, if you map it out, the subjectivity of what do we inject into the context window? And then for clarity, our company, we're really thinking about that. Like how do we model the self, the person, the human driving the agent to maximize their agency and also minimize the amount of turns they have to get to their goal? While Anthropic and others have released dynamic workflows with like these goals and loops and stuff like that, I think they missed the mark on the non-deterministic case of you have to model the human to actually get to the subjectivity, their goal, to then maximize what we know about them, to inject the right context at the right time and optimize the token spend to get the human with the agent, whether it's coding or not, to their goal. It's a very difficult problem to solve. I'm not like shitting on the frontier labs for it. It's really hard. And the frontier labs are really responsible for solving Like so many different user queries that they have to solve these more general purpose ones. And like Jonathan said, they're solving the 2 plus 2 ones first. Like that, that makes sense from a product management point of view. Solve the ones where we can find truth. And for the ones where truth is subjective to the individual, then that's where we come in. And our belief and our thesis is you have to model the user to then align the modeling agent and harness to that user's goals, and that's going to be subjective by nature. So, just wanted to explain that real quick, that type of explanation has helped product teams be like, oh, this is an infinite, unbounded sort of problem space, versus SDLC, it's a little different, like unit tests, integration tests, and then testing, like, oh, okay, like, well, how do we constrain this thing that could literally say anything, right? How do we compound from that? And shout out to Hamel Husain, who has taught us. He's one of the best in the game. We're about to talk about him next, actually. Let me just give a quick shout out to his blog. This is Hamel's blog, hamel.dev. Go check it out. He's a machine learning engineer with 20+ years of experience. He's worked at Airbnb and GitHub and did some of the early LLM research used for OpenAI for code understanding. And he co-teaches this AI Eval for Engineering and PM Teams along with his partner, Shreya Shankar. And Just quick shout out. This is a course we took alongside some of the best developers in the world, like people from OpenAI, Anthropic, the people that contribute to Claude Code and Codex were in this learning evals. Hamel is really great. Shout out to him. He's, he's a friend and partner of the pod now. We've been friends ever since this course. He's taught us a lot on how to objectively and measurably improve agents. I would say go check it out. The reason I'm mentioning this is As we move into the what's, what's the meta section, here's the very first what's the meta item. Camel put out this really nice evals skill open source repo that has skills to guide coding agents to help you build product specific or domain specific AI evals. If you don't know what evals are, think about this analogy for grading a non-deterministic thing. Remember when you were in like, I don't know, English literature class, I don't know, philosophy or whatever. When you're a kid, like more subjective areas like arts and humanities, how does your teacher grade you? How do you grade like 40 to 50 students fairly in a subjective domain? You have to put out a grading rubric and you're like, this is for grammar, this is for prose, this is for if you have a bibliography with links that work that are well-cited, this is for XYZ, formatting, how well do you make the argument, how concise, how verbose. All these things are subjective evaluation criteria and that's exactly what evals are is How do we grade the output of the LLM in an unbounded space where it could say anything? And can you take that and put that in your pipeline so that when the models change, we can regression test against those evaluations where we bring in that subjective reasoning or that taste component to critically judge the quality of the output? That's what evals are at a high level. And if you're doing anything around agents, And you just have a system prompt, some tools, some skills or whatever, which I've seen so many people have, you're just, you don't know it's good. Like you're running in production, you don't have traces or whatever. You don't know if it's good. Like how do you know it's objectively getting better over time measurably? And that's what Hamill teaches. He has some of these awesome eval skills. I'll drop a link in the pod notes. One of the key skills I thought was super cool was this error discovery skill. It runs an error analysis on an existing dataset and builds an annotation or review UI. Bespoke to that dataset, selects diverse samples from the dataset, monitors your annotations, and then organizes failure modes. So failure modes tie to how do I get a grading rubric, right? Just to explain this really quick, when your teacher looked at your thing and you didn't have a bibliography, failure mode. When your teacher looked at your thing and you wrote an essay on Prometheus versus Frankenstein, which is what she asked for, failure mode. You wrote the essay on a completely wrong thing, right? So that's what we mean by failure mode, and that informs the rubric, which informs the eval. Which evals are really the moat for any domain-specific agent or AI product or AI business against the Frontier Labs. That's my true belief. So this error discovery skill is really cool. It's, it's long. I'm not going to read it all, but this is the essence of what it does. It brings up a dashboard for you to review a dataset and be able to get to that failure taxonomy as fast as possible. And I think of a KPI. It's speed to annotation, speed to evals. And going through this whole process, it's open coding, axial coding. You get to failure taxonomy across a dataset. You get to your V1 grading rubric and boom, you have your first evals dataset. You have to manage your dev, test, and golden dataset. We can link more on that later from his blog here. We'll put that in the notes. Fast forward to that, and then I'll ask for Jonathan's opinion. I'll give a quick demo. This is if you're an AIPM responsible for Objectively and measurably improving your AI product. This is the Clarity Harness that we're going to be open sourcing soon, a version of this that we're using for our customers right now. And we sort of talked about that digital twin technology, and we sort of talked about that, how do we model somebody to then get to the subjectivity of their goal, to then inject the right context at the right time in their context window to help them get to the goal with the, with their agent. So in this case, there is a model of a customer here. Uh, his name is Justin. And on the left here, we have a transcript. So all I did was I went to my Claude code and I did /clarity-upload-learning, which is a skill we've codified in our upcoming open source Clarity Harness release. And you just throw in some sort of input. Uh, in this case, it is a transcript from a custom meeting where I'm running discovery. And what it did is it built a model of Justin. And what I could do is go here. As Hamill says, you want to go see what the agent did. And here we could see Justin holds that story quality is not the main constraint on reaching 200% backlog health. That's a success metric we're aligning on. He expects that adding codebase references via this MCP will only be marginal. He frames the platform teamwork, blah, blah, blah, like a lot of things that are specific to Justin. And what I could do is say, this is a pass. I approved it. I could say this is a pass. I approve that too. This is a pass as well. And for this last one, I just annotated this one run where I extracted the model of self for a customer stakeholder of mine. Then what this does is it passes that model of self to the requirement synthesizer agent here. So we have this Clarity interview agent. That was where I built the model to customer. Then we have this requirements agent that gets past the output of that. And then what we get is, hey, from that transcript, from first-party data, this is what we kind of get. Inject domain codebase context into stories, define story type taxonomy, standardized story output. This is for a story writer agent that we're building to help a day 1 PM for a compliance or domain-specific domain be as ramped up and ready to contribute as a day 365 PM. So if you have that problem, and you're interested in using such a harness to compound your own company, business, product team, domain-specific harness to solve that problem, this is how we would go about it. So it passes that over, we see there's these requirements, I could pass these ones that I like, This is the open coding step. If I were doing this really appropriately, not on odd time for a demo, I would be taking my time to actually inject my taste on how well suited these statements for requirements are. What happens after that is the requirements are the spine for writing tickets that make it to our backlog. So over here, we could see this is one of the tickets that came out deploying Clarity Storywriter harness with simplified access and enablement playbook. For broad PM adoption. So that's the thing we're working on. And so I could see what this will look like before I persist it into Linear. And I could also see like where in the transcript was the source, because the transcript to us in this case for this agent run was the first-party evidence tied to the stakeholder requirement, which is the main golden context to use for all downstream agent runs. So as the human, I'm right now reviewing that what's for dinner sort of non-deterministic output. Which are these agent runs, and I'm trying to say, is this a good story that was written from that first-party evidence tied to the digital twin of the customer that I have tied to the requirements spine? And boom, here's the ticket. And before we persist it in, I want to read these. This is the acceptance criteria. It might say things like, okay, access security onboarding steps, developer security group are documented in a single reusable checklist. The new PM can follow without needing ad hoc help from the editing team. Disambiguate any acronyms here so downstream agents can understand. So this is like a personal preference I would have, which is in my story writing, I wouldn't want to have acronyms, or if the acronyms are there, make sure that they also are accompanied by the disambiguated acronym, the full term. So I would say this is a fail because it didn't get that. And then what I would be able to do is then replay that and regenerate this ticket. So that's a real quick, what's the meta tying towards evals? Like, what are evals? Why would you do them? How do you build your moat? Shout out to Hamel, who's taught us evals and how to objectively, measurably improve agents. And then huge shout out to this error discovery skill. We're running our own annotation surface that we're normalizing for us to deliver AI-native services to objectively, measurably improve agents on behalf of customers. get in Swiss Army knife into their team, build them a reusable harness, act as little harness monkeys or plumbers. That's how me and Jonathan think about it. So that Clarity open source harness that we're going to release is really this, to maximize your time to annotation, time to taste injection. And then ultimately what this gets you is a reusable harness. So time to reusable harness that gets you whatever success metrics. In this case, it's day 1 PM could be as effective as a day 365 PM. And anybody who's in the business, in tech for as long as we have been, like decade plus, Jonathan, a couple decades plus, you know, it's hard to ramp up somebody and get them to understand the goals of the team and the domain-specific things of the team. And we're seeing this meta now, which is let's build a harness to just make that easier to onboard people and get them productive with these agents that we could share. Jonathan, I know I was just blabbing a lot. That was just a demo that I thought would run really well from What are evals to, hey, Hamel taught us to like, hey, here's some new error discovery skill. It builds your own review UI and hey, here's our own Clarity open source harness with our own opinionated take on review UIs and then how to inject subjectivity of the expert reviewing runs.

In terms of the evals and this new tool that Hamel and Shreya have provided, the error discovery skill, I think like I found the diagram really interesting and I think what's going on here at a higher level is that, you know, we're using AI to create, like, to augment our existing ability to, to like annotate data and then to create new skills. And—

If you're an AI builder and you're trying to build the valves and build a moat against the Frontier Intelligence Labs, listen up. Take a look at this diagram, error discovery skill workflow. Jonathan, take it away from your CTO AI builder perspective.

Really interesting here is how the toolsets have evolved now that You can take— so you have a dataset now, right? And now you want to get to a point where you can run in a valid loop. And what this skill does is it takes the dataset and it creates the interface and it gets you in the place where you can then be like running in this recursive interactive review loop like Hamill and Trea teach in their course how to do. And so they just created this skill that sort of like commoditizes all the lifting around creating the UI. We're getting closer and closer because we use the same kind of pattern in our Clarity harness as well. We're really kind of abstracting the dataset, the artifacts that it represents, and we're trying to create bespoke UIs that allow you to interact with that data. This is just one step closer to, like, hey, let me automagically annotate my dataset. There's still a lot of manual work involved here, but now some of the heavy lifting of generating the UI to do the annotations is a little bit easier. And I thought that was a really cool abstraction that Hamel and Shraya, after running this course how many times and teaching thousands of people how to do this now, they're saying, okay, now here's a skill to package up the best practice and get a 0-to-1 draft. And so zero-shot, zero-stat eval loops, that's pretty cool, right?

Yeah, zero-shot MVP eval loops. If you want those, you want to check out GitHub repo from Hamel and Shreya and their new really cool error discovery skill workflow, which we're going to begin deploying immediately. That is so interesting. Like, I think there's a success metric there around time to annotate or time to get the human's subjective expertise injected into the dataset, right? Like, that's really the practice here. And what they're seeing is how do you maximize that? And They came to the thought that Jonathan and I have been building open source Clarity Harness for, which is you probably need a bespoke artifact kind of UI view for a particular expert, for a particular type of agent run, like output. And so they just packaged up this skill that attempts to zero-shot it. And I think that's, that is super sick. And for those that don't know what evals are, I'll just like talk to this diagram. I know I talked about it before, but I think this is one of the most important things is I've now coached many quote unquote AI teams that think they're building AI, but they're not compounding their AI. They're not building a better product. They're not objectively, measurably improving it. In fact, I've seen teams of 100+ people that I could still ask, how are you objectively measuring improving your agent? You have this team of 100 people, like, what are you all doing? Like, if you're, you have a system prompt, you have these skill prompts, you have these tool calls, like these whitelisted tools that you could go, like, how are you making sure that gets better with 100 people? Like, you know, that's a question I think executives have to be looking at their teams and you should be asking yourself, like, do I need 100 people to do this? And maybe for some, like Codex, Cloud Code, like we have, you know, hundreds of millions of users at scale sort of thing, that makes sense. But I've seen, teams now where you have 100+ engineers and it's like, if you can't even answer the question of how is this objectively and measurably being improved, you got to really look at your budget. Ask your team, what are they doing? Why do you need 100 people if you don't even have a success criteria? You know, this is one of the key things Hamill teaches is how do you objectively and measurably improve your agents? And the key thing there, there's sort of a sub-success metric here, which is the more annotations you have. From the expert who is best positioned to evaluate the artifact or the output of the agent. If you maximize that, that's really going to be your product feedback loop to compound the product against the user queries or the use cases that you have for your agent. So today we went over many things. We went over how Opus 5 sucks. We went over CEOing, Dario showing up on X out of his cave called by the All-In Podcast. Shout out All In Podcast. Thanks for doing that. Dario smelling his own farts and seeming to like them. And then Boris and team kind of following that. We sort of spoke about, you know, how people will follow their leaders and talk like their leaders. And Opus 5 is certainly talking like Dario, where it's quite deflecting and injecting its own psychological narratives. We talked about this cool bug that was freaking hilarious, uh, that Opus 5 treats direct instructions as negotiations and injects self-referential and interpersonal content into task responses. We also talked about how Hugging Face is reportedly in talks to be acquired for $13 billion, as well as the new stealth model Ox Alpha coming out. And I'm going to try this right after the pod. They have an 80% DeepSWE bench score, which is 15% higher than Fable 5, which was the leader on that scoreboard. We also gave a quick demo of the Clarity Harness, which is aimed at Helping you get reusable and shareable harnesses within your team and enterprise and company as fast as possible for day 1 employees, engineers, PMs could operate as day 365 ones and get up to speed very quickly. And that's it for this weekly AI Clarity. This has been a really great podcast episode. This is going to be your signal from the noise every single week delivered by me, Robert, and my co-founder, Jonathan, every single week so that you know what's the latest news. to be aware of as an AI builder, an AI team, some AI funny memes to just laugh about how fast things are moving, and then AI meta so that you know what you can actually use from the breadth of things falling out of X and GitHub and what actually meets real-world production use cases and helps you out. Anyway, this has been your weekly AI Clarity signing off now.