China Is Winning the AI Race

94 cents. That is what it costs to run a task on Kimi K3 versus $1.80 on Claude Opus 4.8. Claude 5 just dropped. Jensen just made his first post on X/Twitter. We have a lot of things to go over on the podcast today. Welcome back to the Self-Aligned Podcast. This is your weekly AI clarity. I'm Robert Ta, CEO and co-founder of Clarity by EpistemicMe, joined by My CTO and co-founder, 12-time serial entrepreneur Jonathan McCoy. Today we're going to report on the front lines of AI work as we normally do. There's been so much that has happened in the past week that it's almost overwhelming to put together an agenda. We have lots of stuff to cover today, and our usual structure: we're going to go through the AI news, the hot stuff. Kimmy K3 just came out. What does this mean for tech executives, AI builders, product managers? And then we're going to go into some American thought leadership coming out of Jensen Huang. Trying to unify different organizations around the need for open weight models and closed frontier intelligence. And it's Jensen's first tweet ever, so it's just kind of hilarious. Actually, we'll just read out some of the replies as part of AI Memes, which is our second section of the podcast. And the last section of the podcast is What's the Meta, where we go through different items that may help AI builders as we test things out ourselves from the front lines of acting as forward-deployed engineers. bootstrapping our startup Clarity. Let's go into Kimi K3. China's Kimi K3 beats Claude on frontend code and started a war inside the White House. Here's what happened and what matters for any founder or engineering leader deciding what models to build on. And there's open decisions from customers that we talked to, prospects we talked to around how do we optimize our token budget? And what do all these things mean? Fable 5 being taken off the market, then back on, and now, okay, it's just part of the plan. And hey, SOL came out and now Kimmy K3 came out and what do we do? And there was GLM 5.2, which is so like 3 weeks ago or 4 weeks ago. So there's so much going on. I'll give a quick recap. Moonshot AI from China released Kimmy K3 on July 16th, 2.8 trillion parameter open weights model. And the downloadable weights will be available in a couple days. It'll be available the 27th as promised under a modified MIT license. It debuted at number 3 on the artificial Analysis Intelligence Index score. It's the first open weight model ever that high, comparable to Opus 4.8 and GPT-5.5, still behind Fable 5 and GPT-5.6, but also number one on the front-end code arena ahead of Fable 5. So the first time ever, China has taken the lead over the US in front-end code arena. Should we be quaking in our boots, American companies? Some say yes, and we'll go into what Jensen says in a bit here. Demand was so intense actually that Moonshot paused new subscription On July 20th, they even said demand has pushed close to the limits of our current capacity. GPUs are currently feeling it. You could see that article here: Kimi K3 developer suspends new subscriptions amid compute constraints. Is the constraint and bottleneck in the AI warfare going on right now? Some industry takes. Rune says the era of the Chinese labs being far behind is over. Kimi is at least on par with the modern. public frontier models, people have to think differently without any competitive margin. Oh, I've switched to KIMI from Claude for a bunch of work. It's just so much more fun because it just does the thing instead of lecturing you. Woke lobotomized models are the enemy of American competitiveness. Some observations on KIMI. It's a very good model. Number one, I don't think its performance can be explained away by distillation or anything like that. In agentic coding sessions, it seems pretty much on par with the best public Models of Q1 2026. In my fairly limited use, it also seemed very token hungry. It's not obvious to me that this model is actually that cheap to run. Number two, I am personally surprised the Chinese state continues to allow the open sourcing of models this good given potential risks. To be clear, I myself might be fine with models presenting this level of marginal risk being open weight, but I am surprised that China is fine with it. I suspect the reason they are 75% explained by strategic blindness and lack of AGI pilledness. The CCP is very Yan Lecun in its views of AI. The other 25% or so is their lack of compute for customer inference, making China's open weight strategy an unintended byproduct of US export controls and the normal Chinese strategy of aggressive exports. Number 3, open weight models are inherently decelerationist, and I'm continually surprised to see the so-called accelerationists so excited about open weight models. I suspect the reason they are is that they know open weight models are effectively ungovernable, and they simply like the overall cloak of ungovernability open weight models create over the whole of AI. It's not a bad strategy. It reminds me of James Scott's recounting of the Hill People in The Art of Not Being Governed. Number 4, one probable outcome in an open weight model dominant world is full AI communism. What a take. Which is precisely what China proposes. Rather than a market product, AI is a public good which will ultimately be provided by the state as a kind of digital public infrastructure. He goes on to say that this strikes him as a dystopian hellscape. Move on to number 5. I would guess that the Trump administration will at some point realize that their best strategy here would be to create large amounts of regulatory risk around the use of open weight models. You don't need to ban open source, which is one of the dumber motifs of the AI policy discussion. You just need to direct every agency to issue soft law that creates FUD. And finally, number 6, it's probably true that open weight models of this capability make the world a bit more dangerous, but not so much more that you'll really notice. At some point, the models will be capable enough that you will notice. A non-living, invisible, dangerous, an infinitely self-replicating agent escaped from a Chinese lab, you say? Color me shocked. That's what Dean had to say. Let's check out what Emil had to say here. Emil is the Undersecretary of War for Research and Engineering. He says, every industry ecosystem has its supreme village idiot. Dean Ball, is that for AI? Congress passed a law in 2026 that restricted some uses of DeepSync High Flyer with waivers permitted. Only those models. It went through the democratic process, not some deep date scheme like he would prefer. Dean Ball has perhaps the biggest gap between an actual IQ and his own perceived IQ of anyone in the industry— about 40 points. Okay, real quick, this person Seth Bannon says, you're an undersecretary, that's a serious position with incredible responsibility. Posting like a school child doesn't inspire confidence you're treating the role that way. And then he goes on to reply, not interested in feedback from a terminal TDS sufferer like you who supported presidential candidates that called half of Americans deplorable. and ignorant for clinging to their guns and religion. You would rather a subtle and polite useless anti-Americanism, as is evident from your timeline. Any entrepreneur who takes financing from you should be embarrassed. Oh God, look at what Ethan Mollick had to say. Kimmy K3 is a very good model, but people are over-indexing on an arena score again. Yeah, I think that's very true. And a few other takes here, but I think that's kind of what we're hearing around industry. Those are some of the takes there. And for our own takes, I think I myself haven't I haven't tried Kimmy, but I could talk to it from more of a, I am an engineering leader at a company and I need to figure out the token budget for the board and product designers, product managers, researchers, engineers, they all want tokens. Like, what do I do? And one thing I'll say is I see that there's this AI gateway architecture that is emerging in industry where they want to have spend and controls management around token budget. Across all of their R&D resources, like engineers, product managers, designers, researchers. And that means let's route all calls to any models in one abstraction layer, like an API gateway, or in this case, an AI gateway. And then it's about routing what workloads or what tokens to use from what models. So one take that I have is even if Kimi K3 is only half as expensive, with nearly as much firepower as the latest Opus, then I think this is a really good option that engineering leaders should consider for different things. Like maybe you're not planning refactors, maybe you're like, some workloads moved to Kimi for writing tests and things of that nature. But yeah, I'm kind of curious, Jonathan, what do you think of all the Kimi news that is coming out and the takes around industry here?

I mean, there's a lot to cover because it's just such a large surface area that was impacted. by the news, like Ball and, you know, administration comment on this is a little bit weird, you know, just because it's like, they don't, you know, what's the government doing in this discussion about tools that I use for my work? So that's a little weird. I think there's, if you step back, was this a Sputnik moment or not? You know, it was one repeated theme I heard through the commentary in the week. And I think there's 2 sides to this. One side is, yeah, American innovation is being challenged and there are frontier model quality, you know, coming out of China. And this is like something, the DeepSeek moment. On the other hand, they pointed out, like, if you want to use GPT-3 right now, You can't because moonshots actually can't serve the capacity. And that was sort of the point of doing chip restrictions on Chinese companies. And that was sort of the point of like, you know, just the overall fact that they can't manage the infrastructure is really where the country to country relationship happens. And in one of the throttles that the U.S. is trying to play in this game between U.S. and Chinese companies. So I think there's a lot going on. You're looking at the comments from the Ball tweet right now. And I think like his point of view and like open models being decelerationist was not very convincing. I think people who are in the industry, uh, just are like, what are you talking about, dude? And this kind of point of view, it seems to protect like OpenAI's interest in their IPO, protecting their valuation. I think this is a really important juncture for them. You're talking about a company that's gearing up to go for a trillion-plus IPO. So that's one leg of like what's going on is that Anthropic and IPO are trying to protect their valuation. And on the other leg is security risks, right? Like all of a sudden you have a model, Kimmy K3, which is like as capable as Fable. 5 in terms of cybersecurity. So seeing how OpenAI and Anthropic on one hand had this point of view about how we really have to be regulating these open weights, then everybody else saying like, well, we don't have to protect your valuation. And we're not really sure that we believe that this security problem is real. And I think the biggest story of the week on the security side was the escape of OpenAI's model attacking Hugging Face. And Hugging Face didn't have an American foundation model to protect itself from it. So they had to use GLM 5.2. A Chinese open source model to protect themselves from an attack by an autonomous AI tool. So I think both legs of like this kind of support structure that OpenAI and Anthropic have been lying on to like protect their valuation and go public sort of got crippled this week. And I think it was a pretty devastating week actually for the foundation model companies, Anthropic and OpenAI in particular, because like they've been exposed to some extent to a large amount of risk that wasn't present before. I've heard some estimates that their valuation cut in half. As soon as next valuation come around. And I have more to say about like how it's affected me personally, but my overall take on the situation.

Yeah. How's the KIMI situation and this OpenAI situation affecting you personally as an AI builder?

Well, I think one thing is I'm seeing distillation as kind of a natural practice that happens once these models are released and that you step back and you say, these companies, OpenAI, Anthropic, took the corpus of human knowledge and communication language and they trained it, you know, sort of without compensation to the data of the countries, the people that created this. And is that right? And now it's kind of coming back around where it's like, okay, you know, open source is now going to distill and sort of liberate these open weights for everybody to use. That feels like how it should be. You know, it feels like that language is a property. And Chi said this on his speech about China and open models that like more of a public utility. I think of like our language and these open weights more like electricity than I do something like a nuclear weapon. Like it's something that touched every part of what we do. And it came from an ancestry and history of us. And language that we speak and like the knowledge that we have accumulated through our ancestry doesn't feel like it should be bottled up and sold in token by token basis. That seems like the wrong model for the long term. Like it's the right model for the short term, but the wrong model for the long term. So like you were talking about routers recently and routing for the right token, I think potentially moving to someplace where we can even distill our own use of these models, create our own tokens locally that are sort of like, you know, this is what I use in my workflow and route say 5 to 10%, maybe an increasing amount of traffic to these locally distilled models. Frankly, I started researching that, but then I had this like kind of moment of Opus 5 came out. I need this for my work. And then engaging in the sort of like changes that Anthropic was doing, seeing where they're really innovating in the harness and not the model, made me think like, oh, I've got to adopt these practices right away for my business because like this is going to make me more efficient and more productive. So it's kind of like some love hate going on internally about the foundation models, what they offer, what they represent in the world. It's kind of like a weird, weird feeling.

Yeah, there's so much change, which is like, oh yeah, heck yeah, change. And then it's like, well, we got to run a business. Fuck. Well, what's up with this change? The nerd technologist in me loves it, and the entrepreneur in me has to make money. So we have to be able to absorb and make sense of the world and understand like what advantages can we gain from the new alpha coming out and how to integrate that in our processes and our execution. Skip ahead real quick. The new rules of context engineering for Claude 5 generation models and the whole idea of dynamic harnesses. So a harness for every task, dynamic workflows in Claude Code. We're going to talk about that in a second. Just giving a quick sneak peek of this. Everybody in the media and industry the past 2 weeks going wild on Twitter to be like, oh my gosh, like open source, like what's gonna— China gonna win the war? Like what are we doing, US? Like, you know, there's a lot of that angst going around the Media hype cycle in AI. Now move into Jensen Huang, CEO of Nvidia, says, for my first post, I'm sharing a letter Nvidia signed on why open models matter. AI will transform every industry, power every company, and be built by every country. Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty. The world needs both frontier closed models and frontier open models. Go in here and read a few excerpts, but you could see here Open Weights and American AI Leadership. I'll just read this. In the 1980s, early open-source software pioneers challenged the prevailing belief that software would advance only if companies kept tight control over their code. This movement pushed for a transparent ecosystem where developers around the world could study, modify, and improve software. Software developed by the open-source community now supports most of the internet, and underlie systems used by the world's largest technology companies, as well as the US military and federal agencies conducting scientific research, cybersecurity, and other critical missions. Open source did more than lower the cost of software. It created a shared foundation of knowledge on which generations of American engineers and entrepreneurs built their institutional sovereignty. The United States now faces a similar choice with artificial intelligence. Our AI leadership will be judged not by one frontier AI model, but by whether the United States builds a strong open ecosystem that diffuses into every sector. This is essential for creating opportunities for innovation and prosperity across the country. It requires expanding access to AI, encouraging competition, robust application layers, and giving Americans greater control over the technology they rely on. Open weight models, AI models that anyone can download, inspect, modify, and run on their own infrastructure, are an important part of that foundation because they make advanced AI more accessible, adaptable, and widely available. Open weights expand access to the AI economy. Startups, established businesses, universities, and public institutions can build on advanced models without training one from scratch or paying frontier model prices for every task. Open weights let every organization match the right model to the right job at the right cost, reserving frontier-scale capability for genuine frontier problems and running efficient specialized models everywhere else. That discipline is what will make AI economically sustainable as its use scales into the billions of everyday tasks. America wins the AI era by diffusing it into the workflows of factories, hospitals, farms, classrooms, and Main Street businesses. Open weights also strengthen competition, and competition is what keeps the gains of AI broadly shared rather than concentrated in a few open hands. By allowing many organizations to build, adapt, and deploy advanced models, open weights create rivalry not only among model developers but across cloud chips, applications, and services. That competition spurs innovation, drives down costs, and distributes the benefits of AI broadly across our economy. Open weights also give customers greater control. As organizations invest in AI, they want to know that they will not become locked into a single provider or lose the knowledge and capabilities they build over time. Open weight models help provide that assurance by allowing organizations to control their own data, evaluate and adapt models to their own needs, and deploy them wherever their business requirements demand. And as organizations create value with AI, open weights allow them to own that value through self-improving models. Specialized capabilities, and accumulated knowledge that drives American sovereignty and prosperity. To be sure, open weights carry real and distinct risks. Once released, the weights are beyond the original developer's control, and modified versions are difficult to trace or reverse. But the right response to this risk is not to prohibit open weights. In a world where cybersecurity attackers use advanced AI, defenders need access to models with comparable capabilities so they can detect, simulate, and respond to emerging threats. Open models broaden defensive capability, increase transparency, and allow vulnerabilities to be discovered and remediated across many teams. In fact, openness may be one of the most important paths to AI safety and security. Relying solely on closed models is not inherently safe. They can be breached, misused, or fail in ways that outsiders cannot detect. And concentrating advanced AI capabilities behind a small number of closed models compounds that risk. It results in a small number of single points of failure, weakens competition, and leaves critical technology in the hands of a few providers. Open weight models, on the other hand, allow a broad community of researchers and developers to examine their behavior, identify vulnerabilities, develop safeguards, and improve them over time. Just as open source software demonstrated that transparency can be more secure than obscurity, AI safety may depend on giving more people the ability to test and strengthen the models on which society relies. It allows for rigorous benchmarking and evaluation, red teaming, and protections tied to real and demonstrated harms Rather than assuming that closed systems are safer by default. A strong AI ecosystem is not a foregone conclusion. Policymakers have an important opportunity to act. This includes expanding access to compute for startups and researchers, investing in shared training assets, datasets, tools, evaluation frameworks, and keeping the frontier plural by avoiding premature restrictions on open models that stifle competition or drive innovation overseas. These measures must also look at how strong application layers can expand sovereign use of AI across the economy. In shaping this ecosystem, policymakers should be careful not to conflate legitimate model development techniques with misappropriation. Distillation, or the practice of using one model's output to help train or improve another, is a widely used technique for model improvement, evaluation, and validation. It reflects a long tradition of learning from, building upon, and improving existing technologies, a tradition that has helped drive innovation since the rise of the open-source software movement. By contrast, unlawful efforts to extract value from closed models raise legitimate concerns. Those concerns should be addressed through targeted legal and commercial frameworks rather than sweeping restrictions on techniques that play an important role in AI innovation. The age of AI can be one of prosperity. With the right choices, open weight AI can expand opportunity, strengthen competition, extend American technological leadership, mitigate risk, and ensure that the benefits of this extraordinary technology are shared broadly across our economy. That future is worth building. And the United States should lead in building it. Some of the signed or co-signed entities: American Innovators Network, Andreessen Horowitz, Acquia, RCI, Arena, Black Forest Labs, Box, Cisco, CrowdStrike, Dell Technologies, DoorDash, Emergence Capital, GitHub, Hugging Face, who was just hacked, IBM, the Linux Foundation, Meta, Microsoft, Mozilla, OpenAI, Open Claw, Palantir, Palo Alto Networks, Perplexity, et cetera. So. Yeah, that was a lot to read, but also quite concise for the directionality and thought leadership and concepts sort of conveyed by Jensen there. Clearly, he is for open weights, and Elon Musk actually also supported Jensen's proposal. Elon's tweet here: This has my full support. Jensen is right. Retweeting and showing his support of Jensen's very first post on X. Corinne Oops Guess says: Even Elon Musk had to come out and back Jensen Huang. We see some other replies here to Jensen. Kind of funny. Nikita Beer. I love his design thinking. He says, hello, follow for follow. We got a troll. Yan Lecun, a post worth disseminating on all media, even including the present one. Lex Fridman, welcome to X. This will be fun. Brian Johnson, our friend, friend of the pod, friend of the company. Jensen, it's great to see you here. This man prioritizes sleep, eats nutritiously, is kind and measured, and has a wonderful partner, wonderful relationship with his partner of 41 years. Also, he runs in founder mode one of the all-time most successful companies. Peter Steinberg. There's so many people that came out. What is your CTO perspective on Jensen Huang's first tweet and letter from NVIDIA here on why open models matter?

Yeah, this is really interesting. And by the way, one tweet, about 900,000 followers. Not a bad social media presence for the world's richest man.

I'm only kind of impressed.

Let's talk about where American thought leadership is coming from right now, because I think that's It's an interesting story on its own. You know, I thought the Chinese actually had clear thought leadership from their government, right? From Xi, from giving speeches, like they're all kind of aligned on this open weights being a good thing for humanity as a whole. We are fighting about it in the United States. And like, there's kind of a vacuum of leadership from our administration. Like they don't really get it. And they're still like trying to scramble and come and figure out what's what and what they should regulate. So like more of a defensive posture. I think the leadership has been in Sam Altman's hands and Dario Amodei's hands. for like what the future should look like with AI. And I feel like they dropped the ball, right? They went into like protect my company's valuation mode instead of like creating a narrative that was like that all of us could buy into. And I think you saw remnants of this when Fable 5 came out with the Alex Karp rant, with the Satya Nadella letter about the frontiers. I think Jensen has now synthesized this and said like, you know, okay, tech company leadership standpoint, we can have more of a stakeholder aligned view of like what leadership looks like, synthesized it into this document, you know, that protects consumer data and company data and it protects company interest. It makes a narrative about how America can be strong under this umbrella, right? So I think this was like a really positive example of American leadership coming through. And maybe it doesn't come from the place where you think, but we were talking about Alex Moisey has this view of influence of like, well, it's the richest people in the world who are the most influential, especially under American, you know, and you can sort of see that like it took, but de facto, like JP Morgan stuck Stepped in financial crisis of the late 1800s. You have people like Jensen Huang, the CEO of the most valuable company on earth, stepping up to like provide leadership. And that's sort of an interesting Americanism of like where, how we promote our leaders, right? But it's saying the right thing. And I think like it's a good direction that's being set that we could follow.

What I like most about what Jensen talks about in the Open Weights and American AI Leadership article here that was co-signed by over Fortune 500, Fortune 1000 companies in America, big ones. What I like most is that he explains it at multiple layers is like, why is this good for you as an individual?

Oh, cool.

Like, why is this good for jobs that you can have at companies or teams that would hire you as an individual? And then why was that good for the overall society, like your community and our nation? So I like that, you know, it was only 3 pages. We can even call it really just like 2 and a half, because this is just all logos. But in that concise 2.5 pages, it talks to all of those levels really well. So yeah, hats off to Jensen and the Nvidia team for stepping up in the chasm of thought leadership coming out. I will say, like, Jonathan, obviously, like, we're biased towards each other's thoughts because we talk all the time to make sense of the world. But yeah, I do think the thought leadership chasm, that was just in Dario and Sam's hands because they're running the frontier model companies, these frontier labs, OpenAI, Anthropic. And they've just been like on the All-In podcast, they just talk about how they've been just begging the US government to regulate and just like, not really thinking about the greater good of society, at least to this level of consideration, more so just thinking about— and my read is looking at tea leaves, just putting out perspectives or putting out marketing stunts and stuff like that to position better for valuation, their IPO and lining their investors' pockets and themselves and stuff, which is, you know, that is their jobs. But we're also at an era where we're changing the world. So It'd be kind of nice if the people change the world would step up and be like, yo, here's like a good vision for change that people would get behind. Not just saying like, yeah, you're all gonna lose your jobs, and then a year later be like, actually, psych, I don't know, yeah, maybe we were wrong on that one. Because that, that's basically the leadership coming out of them right now. It's like, you all should be fearful. Well, but what do we do? I don't know, fuck it, we're gonna try to make money. And then a year later it's like, ah, actually we were kind of wrong about the job thing. You know, that's just very different than what this is.

This is Highly considered.

They formed coalitions, right? And Jensen spearheaded the thing from his position as CEO of NVIDIA. That's to say, we have been missing this kind of thing. So it's nice that this is there. And I hope that Jensen inspires hopefully other leaders in America to step up in the chasm and the void that we have at this kind of level for a positive vision for change for all levels of the US, being a citizen all the way to like running a business or entities here.

I guess if you put yourself in Anthropic or OpenAI's shoes, You know, do you think that they, with this, this undercurrent, do they have incentive to invest in the research to have the best foundation models? Like, can they protect their business and their research and create, you know, really have a competitive advantage, you know, on the, at least on the model front?

On the model front, I think there's a few heuristics to think about, right? So this paper, thought leadership coming out from NVIDIA and Jensen Huang, talks about how there's billions of tasks. That we can give to AI now. And that's essentially, he's basically saying there's many businesses to be made to eat up that pie of the tasks in our economy that make up the nation. And if you're a multinational, you're operating in many countries. And so AI could literally affect GDP's worth of economies, right? That's what he's saying here. So opportunity is insanely massive. I'll back up into that to answer your question. We are seeing Anthropic and OpenAI Completely be like, yeah, fuck the consumer TAM. Like, we're going to go for enterprise TAM. And they're going at it so hard that they've spun up these forward deploy engineer plays where they've put billions and billions of dollars with investment partners to be able to land at a company with an enterprise deal, forward deployed engineers doing services work to implement their technologies, OpenAI's technologies, Anthropic's technologies. To get at those workflows, essentially, right? And we see these like managed agents coming out of Claude. Hey, like dynamic harnesses and stuff. And we're seeing things like Codex merge with ChatGPT and the Work product coming out from that as well. And it's very clear that they're starting with, if I'm a basic ass product manager, I'm like, where is the biggest ham that we could reasonably get? It was like, okay, well, like these companies could pay us lots of money. And so then it's like, all right, well then we have a lot of money. Let's try to make sure they're successful paying us for our tokens, get forward deployed engineers in there, and start dominating these workflows essentially. Right. On the flip side, we see other companies doing other things like OpenAI and Anthropic are concentrated in general enterprise, and they've sort of started with software, right? Because that's more deterministic. They themselves are dogfooding their own products. So easier to solve those problems, get to market. capture market share there. Back to your question, I think that's still— there's room for opportunity across many other companies beyond just those 2. And I think that's something that I'm starting to believe more, that the narrative that there is no moat is just completely wrong, actually. I think there are certainly areas where you'd probably be better off fighting in a different battlefield. There's plenty of money to go around is what Jensen's saying. And then look at all these companies that are signing this too and saying, We want open weight models. That's good for us. It helps us compete. So another way to put this, or said differently, if we have 100% of economically valuable workflows or tasks in the American economy, do I believe that OpenAI and Anthropic in 10, 20 years will own 100% of that economic value? I don't believe that. I don't think these companies believe that. And I also think this is a move to protect that future where people still can compete across domains, across verticals. I just think there's not enough infinite money And time and scope in the world for them to take over 100% of all workflows?

I don't think it's really about owning 100% of workflows. It's more like, is there an incentive for OpenAI to make the large investments into the pioneering frontier models? You know, whether it's 1% of the market from that or 10% of the market, can they get an ROI now if open weight distillation can, you know, close the gap on that and commoditize it or seemingly commoditize it within a short period of time? One thing that the administration did say this week was that they were going to protect against intellectual property theft. If there's like sort of watermarks on the open weights that are distilled from Claude or OpenAI, then the government is putting in regulations and using sanction controls, actually. They're going to sanction the companies that do that. OpenAI is saying it's a little unfair if like nation states and large institutional, you know, corporations are going to distill our models and then give them to the world for free, which is a little bit different than like, say, you know, somebody making Linux, you know, and building that community around it and getting the whole world to use it. I can see OpenAI, Anthropic's perspective that like, yeah, okay, we have, we've created this innovation. We, we can still create moats and things like harnesses and user experience and hardware and stuff like this, but we're not sure we have the incentive anymore to like make huge investments into the best models because the market's not friendly to that. So I think that's the counterpoint of like the movements going on right now is that you might just see like OpenAI or Anthropic say, we just made GPT-6. We're not going to make GPT-7. We can't afford it because we're not getting enough ROI on these things because that could be the kind of game we see Being played.

I think there's— the reason why I don't see that happening is because every time anything comes out, like look at Fable 5 that came out and then immediately OpenAI clapped back with GPT-5.6, right? And that, that caused Fable to be like, oh, we're gonna move the weekly limit. We're gonna raise your limits. Oh, we're gonna move the deadline on Fable 5 being pulled from subscription plans. And then it's like, actually it's just part of subscription plans. And what I'm saying is I think I think that competition then breeds adoption, which should show up on their bottom line at some point where AI builders like you and me, the innovators, Moore's Crossing the Chasm sort of technology lifecycle adoption model, they're the ones talking about benchmarks. They're the ones sharing the news to other developer, builder, founder friends on what actually works. I think those conversations are where enterprise AI adoption ultimately lands on what is the value. I think ultimately what this comes down to is tech teams want to be using the best shit. I feel that. Tech nerds in general want to be latest iPhone, latest Android, like Google Pixel phone, latest MacBook. That's a proven model, which is release the latest shit and the innovators of that area will want to buy it. There's a certain job to be done there of like, oh, I have the cool latest thing. Social job to be done. Functional job to be done is actual benchmarks. Does this actually help my company to provide value? I think this is just what I'm trying to say is I just don't see a world where OpenAI is just like, I think we're good on just releasing GPT-6 and We're going to chill out while Anthropic like then releases Fable 6 or Fable 7 or whatever. We'll just like, fuck it.

They would focus on lower-end models, right? There's a better ROI investment on this mid-range model that a lot of our users are using. And so we're not, we're going to delay the ROI investment for the next like major innovation in the field and get the most produced profitable models and focus our business on that. And remember that stakes are pretty high here because OpenAI is not profitable, right? So they're relying on this kind of like train where they can keep competitive advantage through their models in order to justify all the investment. Their valuation gets cut in a third. If Anthropic gets cut in a third, like they're going to get acquired, right? They're not going to make it. They're going to need to like take this huge infrastructure investment they made and roll it into acquisition scheme for one of the tech giants or something like that. And I said, the MacBook Pro is actually a pretty good example where Apple can't release a product for the top 10,000 users on the planet. They've got to release a product that is widely used across. So they can't get too innovative and their innovation. It's what you see in companies like Apple. You know, it's like, they're not that innovative anymore. They're just incrementally improving like things for consumers, lower basis. And so I think like you might see OpenAI or Anthropic have to settle into something like that unless they can differentiate somewhere else in the pipeline, you know, like other than the model, right? It might not be the model they invest in. They might just have to put more investment into their hardware design, right? And I think that's something we're all anxiously waiting to see.

That's true. I guess one of the things I was thinking is, do we think there is a business to be made out of serving frontier intelligence? And I think in Jensen's paper here, he basically says yes. Going to be needs for frontier-level intelligence. He sort of talks about it like tasks. The way I think about it is that way too. There are lower-value tasks, higher-value tasks in terms of just capitalism. He's saying, hey, we want startups that have established businesses, universities, public institutions to be able to build on advanced models without training one from scratch or paying frontier model prices for every task. And then following up that statement with, Open weights let every organization match the right model to the right job at the right cost, reserving frontier-scale capability for genuine frontier problems and running efficient specialized models everywhere else. I suppose there's a question there, which is how valuable are these frontier-level intelligence tasks? To who? To what businesses? To what research institutions? And what is the ROI thesis of that? For example, what I've seen that's been happening, let's just talk about distillation and stuff for a second, But this is a pattern that just happens in, say, cars. The whole car industry is like this, where they have a leading edge, like grand touring, has all the trim, whatever model. And then the next few generations after, they innovate on that one. The innovators buy it. And the next few generations, they roll out the features of that into the lower-end models. And that's kind of what's happening, right? Without looking at all the how do you train, how do you label data, how are you doing the evals, Any of that. Anthropic and OpenAI have been competing at this thing could solve more complicated coding tasks or this thing could solve longer horizon, longer running tasks with fewer errors. They're trying to describe essentially, I think they're trying to, with that move, they're just competing with these higher value tasks essentially. There's a hypothesis there that we're seeing play out in real time, which is The enterprise companies have the most money and the highest value tasks. Okay, let's roll a bunch of money into FDEs and then try to get there to learn what are those and then invest. And so I could see a model where it's like OpenAI, Anthropic, Frontier Labs, they sell a SKU, which is, this is the leading edge shit, and then they charge some crazy rate for it. And then that just helps them learn what to bleed down to their lower end models, which Which might end up making a bigger business. Who knows? Just sort of segmenting into large enterprise, medium enterprise, and large enterprise including multinationals like Walmart and stuff like that. Then medium enterprise, you know, your 1,000 to 5,000 employees, and then below that, lower and small business. So I think it's playing out where it would make sense to sell the frontier intelligence, invest into that if it makes a lot of money and solves these really high-value tasks. For these high-value customers. Take that, make the features that come out of that learning loop and those product lines, embed those features over time, over generations or whatever, that meets market expectations for the lower market offerings. I could see that happening.

I think that split of tasks is exactly the right way to look at it. If you're an enterprise, what percentage of them can you serve through your own locally run open weight model? Then what percentage of them do you serve through the range of models that OpenAI, Anthropic, or Google are going to be offering? It's just like what percentage of of an enterprise task have ROI using foundation model tokens that are going to be the most expensive. And I think the calculus on that shifted with KIMI-3 this week, but it's unclear how big the pie is and what pieces of it— how big is the task pie and what pieces of it is everybody served by with the latest models?

Yeah, I think we're just seeing it play out in real time. We can conjecture more, but it's an interesting dynamic landscape, to say the least. To wrap up the Jensen Huang Thought leadership section. Let's just move into AI memes real quick because we're here and this is Jensen's first tweet, so might as well laugh at some of the tweets coming through. So I thought this was TLDR: GPU consumption should be increased, not decreased. Dash GPU seller. Makes sense. Everybody should need compute, and I want you all to consume my chips for for this. It doesn't really matter. Like I don't want just demand to come through. A couple Frontier Labs. I want all the— yeah, I could see that take too. SpaceX video. Here's a funny one. This is on Elon's tweet, but it's the Drake meme but replaced with Jensen where he's like, no to Facebook, Instagram, Threads, TikTok, YouTube, and yes to X. That's funny. This is also a funny one from Shree Stocks and Life. Never bet against somebody with this kind of track record. Dishwasher, busboy, waiter for 5 years, Denny's, and then boom, founder and CEO Most valuable company in the world. That's a pretty badass LinkedIn.

Yeah.

Are you actually a hard worker, Jansen? What is this? The Last Supper? Is this supposed to be the Last Supper but replaced with—is this Zuckerberg on the left here? Yeah, Tim Cook. That's pretty clever. Satya. Yeah. This is Elon Musk. Oh, that's funny. That's a good one.

I like how they're all wearing like Greek robes and then Jensen's leather jacket in the leather.

Oh, this is hilarious. Denny's Twitter account. Denny's and Nvidia. What's the overlap? They both know the importance of staying open. Seeing like a completely non-AI company try to jump on. Like I'm thinking about the social media marketing team. They're probably using AI shit. So they're like, well, let's keep an eye out on all those handles and then they jump in. That's kind of hilarious. Firefox. Welcome to X, Jensen Huang. We support your first tweet. Being used to highlight the need to protect open weight AI models. Me showing up to the open source AI revolution with a budget GPU in a dream. X-Girl says Jensen talking to all his new friends on X. This looks like a good life. I would like this life, actually. Pliny the Liberator, one of us, one of us.

Pliny's the guy that like reverse engineers all the system prompts, uh, whenever a model is released. So he's like, oh, Mr. Open Weight.

Yes. The open weight. The Liberator. Pliny the Liberator. So moving on to What's the Meta? In this section of the pod, we talk about more AI news, mainly double, triple clicks, and sometimes demos for tools that we've encountered, topics that we're deeply considering to integrate into very upcoming sprints as experiments. So let's read through some of these. Today we're going to read through Cursor, their agent swarm model economics, give a quick high-level overview of that. We're going to go over this TASTE skill to inject Some people have asked me about how do you get cool looking websites that don't look like AI slop or like everybody else's AI vibe coded website. So we'll go over that and then we'll go over dynamic workflows and cloud code and the new rules of context engineering. So let's start with this. All right. Cursor rebuilt SQLite 6 ways and the costs range from $1,339 to $20,000. $57 on the same task. Here's what it means for engineering leaders staring at a Runway AI bill. Cursor had an agent swarm implement SQLite in Rust from scratch against an 835-page manual graded by SQL Logic Test. The cost to rebuild the SQLite by model mix, I'll just go over this. You could see Fable 5 here with the highest cost, $20,000. Lowest cost, Opus 4.8 plus Composer 2. 2.5, where they split up the workloads to rebuild SQLite. That came in at almost at a LEET score, actually $1,339. So only $2 off of LEET. That's kind of cool. Real quick nerd thing. That's fun. The architecture was planner agents being the frontier models. They decompose goals. Worker agents, fast, cheap models execute on them. And Wilson Lin says a planner doesn't implement, so its context never fills with low-level detail. Worker never plans, so it can spend all its context on one narrow piece of work and And that was the architecture they went for. They saw that GPT-5.5 alone burned 14.7 billion tokens and Fable 5, 12.2 billion tokens. Frontier as planner plus cheap worker mix delivered near identical quality at 7.9x lower cost than Frontier only. Once a Frontier planner has collapsed the ambiguity into a detailed explicit instruction, less expensive models simply have to follow it and the quality bar. The new Swarm hit 100% test pass on the Fable-5 mix using 9,908 lines of code versus 64,000 lines of code for the old Swarm, which is about 6.5x less code. So the thesis is really frontier for planning and orchestration, and then lower-level models that cost less for the actual implementation work could work really, really well. Jonathan, from your CTO perspective, what does this say from Cursor around the whole narrative around like token maxing and stuff like that? I know like you and I have talked about that a lot, but it's clear that they've proven out this meta of frontier planner, lower level worker. Thoughts on this?

Yeah, I thought it was a cool experiment, right? Like, you know, personally in my work, I've been using Babel 5 as a planner and then GPT-5.6 SOL as the implementation. I know it's not the cheapest way to get it done. But I think for unbound planning, especially creative planning, Fable is like a really great model. And then it can either orchestrate the other agents or run them. And so here they showed their own Composer 2.5 working with Opus 4.8. That was the cheapest one, right? Opus 4.8 on a bounded task, like recreate something with a fixed spec, is actually a pretty good model to do that. And Fable 5 is kind of like overkill, you know, there's not enough creativity needed, you know, where Fable can really show its value. And I think you could see that I would use Fable in something where it could produce a better result than Opus 4.8. And expect to see results there. It's curious they didn't try with just Cursor 2.5 alone. I wonder what's up with that. Like, it isn't well suited for sort of like the planning and orchestration tools. But, uh, yeah, these are exactly the kind of experiments you're kind of like running every day as a builder right now, right? I mean, they just did it at scale and put some numbers to it, and it's really helpful for the rest of us.

I like this section here, specs as prompts. Each jump in AI capability has raised the level of abstraction at which an engineer can work. Autocomplete, let Engineers work one line of code at a time, and early models raised that to a block of code, and agents raised it to a file or a feature. I love that breakdown, though. With swarms, the unit of work becomes the spec. We talk about this a lot, but spec-driven development is really the meta that I've seen stick for the past year for people to coordinate their actual coding agent workflows and hold them accountable to verifiable tests and exit criteria. So anyway, for that to work, the swarm has to actually follow a spec, which is what much of this post is about. We gave the swarm 835 pages of prose and it came back with a database. What was scarce in this experiment and what we expect to be scarce in software engineering going forward is the right description of intent. That's a really great way of putting it. It does just come down to, are you good at English and articulating your goals and making— and then also working with, say, like a coding agent from a planning perspective on unbound creative sort of planning tasks. Doing some iterations and then you come out with your intent at a higher fidelity through those cognitive loops with your coding agent. I do think that's what the meta is all about. And it's cool to see that Cursor here is not letting down on their research arm with the new acquisition into SpaceX. So looks like they're still putting out cool numbers for us to be like, okay, I had this kind of feeling that maybe Fable 5 was going too ham with tokens, but now it's like, okay, these are real numbers. Great. I like real numbers. This, they did this test. I don't have to do this. Now I know. What's better meta? So for all AI builders out there, if you want high value, low cost, think about this meta, which is frontier intelligence for the planning, and then anything else for the workloads where it's execution and implementation compute. Opus 4.8 and Composer 2.5 was able to get a similar quality of job done rebuilding SQLite 13.39. That's about 15x cheaper or so. Than Fable 5 alone. So this meta is clearly a good way to go. The next topic, let's talk about the Taste skill. So Taste skill, that's, that's what this skill is called. It's called Taste skill. A teenager's anti-slop skill hit 65,000 stars by banning the em dash and AI tells. Here's what this means for anyone whose AI-generated UI looks like slop. It looks like other vibe-coded slop. So if you use this skill, I'll include a link in the comments, so check it out. This is the Taste skill. It is the anti-slop front-end framework for AI agents. And you can actually see there's 67,000.6 stars now, 4,700 forks. And we could open up the skill itself, but just reading a bit more of the GitHub marketing on the README: portable agent skills that upgrade AI-built interfaces, stronger layout, typography, motion, And spacing instead of boilerplate-looking UIs. This repo also includes image generation skills for reference boards. Pair them with ChatGPT images or similar generators, then hand the frames to Codex, Cursor, or Claude code for implementation. I'll actually just fire this off. Actually, this is so interesting because our next topic is gonna be about how context engineering has changed, and this repo definitely hasn't changed with that in the last week. I'm not gonna go over everything, but I'm just gonna flip through here to show you what they're thinking about anti-slip. Optimization. They talk about things like liquid glass, magnetic microphysics, perpetual micro interactions, layout transitions. So there's a lot of literal taste encoded in this is what it looks like. It's like very domain-specific front-end taste here. They even have a forbidden patterns section. No neon outer glows. I've certainly seen a lot of that. No pure black, use off black. No oversaturated accents, no excessive gradient text, no custom mouse cursors. Are people doing a lot of custom mouse cursors? That's funny. If you are interested in leveling up your taste, perhaps try the Taste skill. This is the anti-slob front-end framework for AI agents. So you could make a webpage, make sure your friends, peers, colleagues, customers, prospects don't know it was vibe-coded. Next topic, the Claude stuff. I think this is some of the most interesting stuff that's come out. Anthropic just put out the new rules of context engineering for Claude 5 generation models. Every AI builder engineer, product manager needs to be reading this. They say we removed over 80% of Claude Code's system prompt for more advanced models. How to apply the lessons we learned to your own context engineering in Claude Code and with your own agents. There's a bunch of sections in here. I'll gloss over a few of them at a high level, then we'll deep dive into each one. One is called Unhobbling Claude. So with the new models, they basically say, hey, we had to trim a lot of the context. We used to have to say things like, like, don't do this, leave this documentation as appropriate. And they would assemble the context in this way where the system prompt would literally have things like leave documentation as appropriate. Skills would have certain statements in there like don't add comments or in this way. And that's stuff that Claude had to reconcile and remember all the time. And now they found, the Anthropic team has found that you don't have to do Instead of giving it rules, give it judgment. Instead of giving Claude examples, give it design interfaces to understand. It'll figure it out. Instead of putting everything up front in one prompt, use progressive disclosure. Instead of repeating yourself, use simple tool descriptions. Instead of memory in Claude.md files, it has auto memory. And lastly, instead of simple specs, use rich references, including the meta that we've introduced around HTML files as specifications or design systems, lightweight design systems. systems that are portable. I'll pause there at how you might apply this. Your prompt references system prompt, Claude.md, skills, memory. They even have, actually the very last thing I'll say is they have a Claude Doctor command. I will demo that in a second. But coming up for air for a little bit, Jonathan, I know you deep dived into this article. What are your thoughts from a CTO perspective on the new rules of context engineering for the new Claude 5 generation models?

Yeah. I think it was kind of a surprise that the models could be this effective with less context. And I think all of us are thinking, what does it mean to compound and what's a good investment of what to sort of store in our prompts and our skills so that the next model that comes out, we can keep the learnings that we have and build on top of them. And I think one of the metas that we've seen is people compounding within the Claude MT Right? And turns out that's not that valuable. And so that's like a pile of learnings that you might have thought would be valuable the next time around that actually aren't. So pretty interesting, right? At the same time, it feels refreshing. It feels like pressing the reset button a little bit, kind of like, you know, cleaning up your hard drive or, you know, just kind of sorting things out. Yeah, spring cleaning. Yeah. And the chance to like work with Claude on a new basis. And when I ran Cloud Doctor initially, it said, hey, you have all this stuff in your Cloud MD that's just not relevant anymore. And that was true. So spring cleaning is welcome. Yeah, as to some of the specific modalities, I think when he talks about using trees to work alongside their dynamic workflows is a new meta, and the auto memory is a new meta. And so you're thinking about how do I meet those? And I think we'll get to this in a little bit, but I think the key is focus on the rubrics, you know, focus on the domain-specific rubrics that you can like let Claude code use on your behalf. Everything else, the orchestration level layer is just going to get like acid washed away, you know, by these models as they get better and better. So.

Yeah, I think, you know, the biggest takeaway you were just mentioning around the new rules of context engineering from Anthropic, I like that takeaway around rubrics are your kind kind of like moat amongst all the metas changing. If you accumulate tacit knowledge and compound your intelligence in your domain-specific workflows, economically valuable workflows that are niche, like other people aren't using these, like outside of your industry, for example, rubrics are where that goes. And so one way to think about this for everybody that doesn't understand harnesses or evals is remember when you were in English class and your teacher had to grade you and maybe 30 or 40 other students or whatever, on an essay. And an essay is quite subjective. We kind of know good writing when we see it versus bad writing. But to grade 30 students and 45 students, get their parents to be on board with fairness, what do you do? You have a grading rubric, you put it out there, and you're like, here's how I'm going to rate your essay, your paper, right? How well did you make the argument? You get a certain amount of points. How good was the grammar? Were there a certain amount of mistakes, like under a threshold or over a threshold? Did you have a good bibliography? What does that look like? The formatting, how many sentences? Like, those things are what we mean by a rubric. So a coding agent's rubric, what Claude Code and Codex compete on is they have people doing annotations, building evals, evolving their rubrics and their system prompts over time for these harnesses. And they're constantly evaluating what makes a good PR and how do we drive technical outcomes and business outcomes from that and improve the development experience in using these coding agents. So that's how they're competing. As everything changes, Jonathan just said, the meta is going to change where probably orchestration is going to get a lot easier. That's what this is telling us. Less is More than really where you should be concentrating your compounding effort on your team, if you're an AI builder or a founder, is the rubrics. Version them, keep them first-class citizen assets to your company, because I think that's where the judgment and taste from all your humans accumulate towards whatever product or experience or service you deliver in market. Yeah, another one that I really liked was this rich references. So in plan mode, Claude Code heavily relied on markdown files. with plans. Storing these files as plans helped Claude refer to them when needed. But we found that Claude can handle increasingly more complicated references instead of simple markdown files. Claude can reference HTML artifacts created by our new artifacts feature. And I'll actually show like one of the artifacts here is one of the ones I recently made for improving a referral system for a mobile app. And this had like visuals for design, what's needed from a different stakeholder to unblock dependencies like on the IT side, What's the verification look like? Any open decisions? And this was just like an easy way for me to like read a visual specification. I have more that are more visual. Like this one is more of a project plan actually. This is a project that Jonathan drafted for us just to discuss about dogfooding that we're doing. So an internal project that powers some of our customer development right now with us being the first customer. So we've been doing this and What the Claude team is basically saying is they're saying, keep doing this. The models are only going to get better at understanding these things. And there's something I really appreciate about that, which is, hey, these HTML artifacts are making it easier to align humans. And it's cool because now it's aligning the human to know how to verify against the agent. And the agents are— because the models are getting better, the agents are getting better at aligning against the HTML artifact. Yeah, we've come full circle to where We all just used to make slides in corporate America and now we're just making slides for agents. It's just HTML files. Congrats everybody, we've come full circle.

One clarification there, because you were showing like we've gotten in the habit of creating the HTML docs to share with one another. It's referencing this actual artifacts feature, which also creates HTML features, and you could do that in Claude Desktop or Claude on the web. I think we generate most of ours right from Claude code on the terminal and we just ask it to generate an HTML doc. And the added benefit you get from using artifacts in Claude Desktop is that they're shareable and it's sort of like maybe you can keep the artifacts within your threads. We found for our business, right, we want to like kind of make these HTML docs first foundational sort of way of communicating. We'll put them in the codebase. We've created tools now to like store these in the cloud so we can like comment about these artifacts. There's different ways to do it. So I think that's like one of the evolving metas is like, how do you best use these HTML docs and store them and sort of like synthesize the organizational knowledge that you're developing?

And that might actually be a good time to just show one of the tools that we've built to align internally.

The segue that one of the tools we use to align internally.

Oh, sick. Yeah. That was such a, I didn't even appreciate my own sick transition. Anyway, sick transition incoming. This is how we get aligned. We use Clarity Align. This is a product that we built internally. And you can see here, there's a little inbox, like people have left reviews on different HTML docs. That referral feature I was just talking about. We throw an HTML spec into here. I can open it on a whiteboard canvas similar to Miro, and then I could actually. This is really cool. We're experimenting with this and still dogfooding it, but we actually start a session, start transcribing, and the Clarity agent backing this Align app will look at the spec and see like okay referral as a platform not a path, and I'll say something like let's change the executive summary so it just says referral as a platform remove the not a path. And then make this more of an architectural decision to align with the product strategy overall of maximizing the amount of growth experiments we have for top of funnel referrals. In O1 engineering problem section, let's have a cleaner delineation what this is not. In the design section, let's go and remove all the iconography. Okay, so I just said a bunch of stuff and you'll see Clarity Agent It's processing stuff on the side here, but it's saying referral V2 framed as an architectural decision. And it actually records that decision here. So when we meet, we could look at these things. I could accept this. Okay. Remove iconography. Great. I can accept cleaner delineation. Cool. And then what I could do is generate a spec and then it actually will just modify this in place. And then that way we're all just kind of at a whiteboard. startup mode, founder mode, aligning on the whiteboard. What is the goal? What are we trying to do? What's the initiative? What's the project? How does this move our business outcomes or engineering outcomes, product outcomes? What's cool is then we could like confirm, okay, this is what I wanted. Now it says this is an architectural decision, not a point feature. It reframed the engineering problem to be more clear. Like I mentioned, it removed the iconography here. Actually shows me what it removed. This is a little dogfooding product demo, but this is how we think about, and I should say, this is how much emphasis we place on It is so important to get aligned. And I think the spec as HTML file to get aligned on anything, whether it be a project or a business strategy or a product strategy, or like literally like a new feature like this one, just doing this is really great. We vibe coded this. Honestly, it didn't take like too long. And I highly recommend people do something like this. We kind of made it to where it's like Google Docs-like. You could comment here and then you could see what's changed as well, version to version. Share this out to other people. Different modalities, synchronous, asynchronous alignment tool. And it all just sort of starts with the specification being visual. And yeah, you can see we exited the whiteboard view here and it took all my changes and I could look through and any stakeholder can go and look at what are the decisions. Like right here, there's D2 here on this artifact, the specification, and it says remove iconography from the design flow per D2. That's a decision we made. You can open the canvas back up and see who logged the decision, et cetera. Anyway, so we're dogfooding this. If anybody's interested in this, drop a comment or something like this. If you're interested in on the content that you're seeing this demo on, this is something we're starting to use more. The main idea being get aligned with visual specifications and your agents now are really good at understanding these visualized specifications in the HTML format. And you could do a lot with the HTML format. It's way cooler than making slides. Slide making sucks. I used to make a lot of slides in my previous career. And one of my favorite metas for getting aligned in this whole AI era is just easily making these very beautiful visual specs to align the team on. I could nerd out about alignment quite a fucking lot. So we're going to move on. By the way, I don't think you saw some of those things coming out. I added that since last time, which is the diff in place and look at decision that tracked the diff or whatever. Isn't that cool?

Yeah. So after the, you save it from the board, you come back, you have the version to like diff that you can see what changed.

Yeah. Isn't that sick? I think that's pretty sick.

Yeah.

All right, let's move on to the next topic from the Anthropic team. These are 6 new rules of context engineering for 2026. Instead of giving Claude rules, give it judgment. Instead of giving Claude examples, give it design interfaces. It's better at understanding that now. Instead of putting everything upfront in your prompt, use progressive disclosure techniques. Instead of repeating yourself, give simple tool descriptions. Instead of relying on memory and Claude.md files, Claude now has auto memory, which is amazing. And instead of simple specs, use rich references such as HTML artifact files that Claude has come out with, with Claude artifacts. So these are the new rules of context engineering, and this is the new meta for any AI builder out there, especially those building with Claude code. Take a look at this article, give these things a try. Next topic. This is going to be the final one. Claude Code now writes its own harness on the fly. Dynamic workflows in Claude Code. And here's what it means for anyone building on top of their own agents or orchestrating their agents for any kind of new task. You could stop hand rolling bespoke multi-agent pipelines and rubrics and evals and system prompts for every single job or workflow that you do. And you can actually just start having Claude Code write custom harnesses on the fly. So Anthropic published this article, A Harness for Every Task: Dynamic Workflows and Claude Code. Thesis verbatim is that Claude writes its harnesses on the fly, custom built for the task at hand. I kind of think of it like Tony Stark in Iron Man. Now, instead of just the Iron Man suit, it's like, hey, I'm just putting on new suits on the fly, on the go. Like, you want me to be accountant? All right, accountant suit. All right, want me to be a product designer? Product Designer Suite, and that's really interesting because they're going for obviously enterprise scale and enterprise productivity with all of their plays going around right now to compete against OpenAI to win the enterprise race. So how this works is Claude writes a JavaScript file with special functions that spawn and coordinate subagents, isolated Claude instances with their own context windows and focus goals. The workflow decides which model each subagent uses, whether it be Sonnet or Opus, whatever, and it could run. Subagents in isolated work trees. The real problem this solves for AI builders is when you ask the default Claude code harness to do a task, it needs to both plan and execute in the same context window. Workflows split those across isolated subagents and combat agentic laziness, is what they say. When Claude stops before finishing a particularly complex multi-part task and declares the job done after partial progress, that's sort of funny that they call it agentic laziness instead of just calling it a lie, which is what it is. Like it's like Claude, if you're telling me you're done. And then I look at your work and it's not fucking done, bro. It's not agentic laziness, although we might call it that as well. You're fucking lying to me. Like, come on guys. Like agentic laziness. Get outta here. Anyway, this is really interesting. How do you actually drive it? Say, use a workflow or the trigger word UltraCode to force one. I thought that was pretty cool. And that combines with the loop command and the goal command for a hard completion target with a keep going sort of command. And you could also set token budgets. In plain English, like only use 10,000 tokens, only use 500,000 tokens kind of thing. It saves workflows in the .clod/workflows directory, or you could share them via skills. Some real use cases that have used this, Bond recently did a rewrite to Rust, and that was a large migration refactor. And I think that cost them a few thousand dollars over a couple weeks on a long horizon task. That's kind of the high level. Claude writes its own harness.

Jonathan.

Jonathan, as a CTO and 12-time founder, you've seen a lot of businesses, you've worked on a lot of tech stacks. What are your thoughts on this moment of multi-agent dynamic harness on the fly? Like, I think this is kind of next level.

It should change the way that you're working with Claude code now. Then it's like, how do I best take advantage of this type of capability? By the way, you don't need to go over the exact patterns that they laid out, but I think those are really helpful.

So there's 6 new workflow patterns that the Anthropic team recommends with this this harness-on-the-fly feature. Number 1, classify and act. Task goes into a classifier that routes to different agents. Number 2, fan out and synthesize. Task goes out, it fans out, does stuff, comes back, and then collates it, synthesizes it. Number 3, adversarial verification. Many verifiers, and then running a worker against those verifiers. Generate and filter. Generate stuff, get to ideas, filter out with a rubric, deduplicate ideas, and then do a best of read amongst them and discard stuff against your rubric and filtering. Tournament style, where you have attempts to do workflows and then you pair them with judges, which are decision makers getting to— it looks like you can have different tiers of judges here and then a final judge to determine a winner from those attempts at workflows. And then finally, number 6, loop until done, where the agent does stuff and there's a checkpoint and it'll say, do you have new findings? No. If no new findings, Findings done. If yes, new findings, spawn another agent, keep going until done. So yeah, Jonathan, back to you.

So one of the things I did in our codebase is I just tried to talk with Claude Opus 5.0 and say like, hey, I want to inherit these dynamic workflows. How can I best use them? And we're kind of going inventorying the type of work that we do in the repository against these workflow patterns and trying to configure the work environment to like where the next time I do a task that I want it to sort of like trigger on these workflows and use them correctly. So that's the deal. And I think what you're mainly going to focus on is, is one, inventory the workflows that you're doing and what's important. 2, what skills have you provided, like as commands for the agent to use and make sure you're taking the new advice from Claude and simplifying your skills. Like don't leave these like really verbose skills anymore. Trim them up, use the doctor tool to trim them up. And then like leave those as breadcrumbs for your, your dynamic workflows to pick up on. And then number 3 is rubrics. And that is like, how do you take the domain-specific knowledge that you've accumulated, the good examples, like what is good in this example, and then synthesize that into a rubric that it can make so it knows it does a good job. When they actually use these dynamic workflows.

I was just going to say, so it sounds like the advice for AI builders there with these new patterns and the dynamic harness on the fly capabilities of Cloud Code is go look at your code bases, do some retrospective with your agent, and even copy paste in this article and be like, hey, like across all of our PRs and Git history, help me categorize our work and against these workflow patterns and help me start doing spring cleaning on my context and then how we're managing context in the repo. And skills and commands. And yeah, obviously if you're working on a team with other engineers, don't do that shit and just YOLO it to main. Like talk to everybody first, like make sure that they know that this commands and skills will change and stuff like that. But that sounds like the tactical advice you have for other AI builders, Jonathan.

Right. And like in terms of like sort of rehearsing for the next run that you do. Yeah. Talk it out, get a plan. It's almost like your coach football team or something like that. Like, you know, we're gonna practice these plays. plays beforehand so that when they come up in the game, we have them in habitual memory.

It's like, oh, this is a loop until done play. Loop until done. Loop until done. It's like, oh, we got it. Okay, fan out and synthesize, guys. Fan out and synthesize. The goal is ambiguous. The goal is ambiguous. We need to be able to generate and filter. Come on. No, I think that's right. This is the meta now, right? Like, you have to understand these patterns and then you recognize as the coach of all these agents. That's, that's how it feels, honestly. Like, like 4 monitors with lots of different Cloud Code sessions going. I just feel like I'm coaching and nudging them. So it's really nice to have a playbook that's official around workflow patterns for different use cases. So I really like that. And everybody should take a look at this. Just to summarize again, classify and act uses a classifier agent to decide on the type of task. Fan out and synthesize, split up a task into many smaller steps, run an agent on each of those steps, and then synthesize those results. Adversarial verification, for each spawned agent, run a separate spawned agent to adversarial verify its output against a rubric or criteria. Generate and filter: generate a number of ideas on a topic and then filter them by a rubric or by verification, dedupe duplicates, and return only the highest quality tested ideas. Tournament: instead of dividing the work, have agents compete on it. Loop until done: for tasks with an unknown amount of work, loop spawning agents until a stop condition is met. I've been using that one a lot. And use cases you could use this on: Bun was rewritten from Zig to Rust, like I mentioned, using workflows. Read more about that. But he says dynamic workflows and adversarial code review was part of what made it possible to rewrite Bun in Rust in 6 days. Oh, I thought it was 2 weeks. It was 6 days. That's crazy. And that was their— Anthropic uses Bun, right? Like, so to have a company as a customer, like Anthropic, using your thing, and then you're able to rewrite it and then ship it to production in 6 days, that's kind of an engineering feat in itself. You know, so that's super cool. The very last thing is let's go over Claude Doctor. So what the hell is Claude Doctor? Claude Doctor is a command. I saw this originally from Boris Cherny. He talks about running Claude Doctor and I just ran it. So I'll show people what to expect here. This seems like a really good thing to use anytime Anthropic releases a bunch of new stuff, which is kind of every 1 to 2 weeks at this point. So let me show you what happened when I ran it. So this is my Cloud Code setup. I ran doctor and it went ahead. Cloud Code health check. So that's what it does. Your install is healthy and up to date, but you have a duplicate copy of Cloud Code installed via Homebrew npm. Shit, I didn't do my spring cleaning. Worth removing so updates don't get confused. Your default permission mode isn't set to auto mode, which would stop prompting you for every routine action. Very cool. I've just been YOLOing dangerously skip permissions for a while, but I do know that this is the new mode. No unused extensions, denied commands, or slow hooks turned up. Everything proposed is reversible. So I ran this and it started analyzing, okay, where's Claude installed, where's installation, okay, I have this plugin, usage of the plugin, these MCP servers, what do I use, the Claude.md seems like it's not saying to do anything there, and it just gives me a bunch of checks. Gave me that summary and these are the more detailed So I went ahead and had it remove the duplicate install and I had it turn on auto mode.

Yeah.

And this isn't the meta around running, you know, a similar sort of process around your codebase. This is your own just individual user harness for my own Cloud Code. So yeah, if you haven't already and you're an AI builder and you haven't done your spring cleaning, run /doctor in Cloud Code and you'll just keep everything up to date, nice and tidy. So you can keep your agent sessions context maximized or optimized. Sweet. Jonathan, I'm curious, what did you find when you ran Claude Doctor?

I sort of did some pre-planning before I ran Doctor.

I'm gonna tell Claude Code about its own creator's articles. Are you ready for Claude 5? Here's the latest blogs on the newest Claudes to consider.

By the way, which model are you running right now? Is that Opus 5?

I, I was like, hey, are you on the latest Claude Code? I was like, update latest Claude Code. And it says I'm on Claude 5.

To you. It just lied. It lied to me.

Claude lied to me. I said update latest Claude code and it's like already on the latest 2.1.220. And then right here, 2.1.217. Claude, what are you doing? Claude, for the love of Claude, come on, Claude.

I always have to close out my session and then do it in the terminal and then come back.

I just updated my Claude code to v2.1.220 and I want Claude code to help me prepare by running the doctor command to prep for the new Claude 5 generation models. So I gave it this article, and if anybody wants a link to this article, drop a comment, I'll get it to you. So Claude Code is now running the read-only diagnostics, is reading through the article, and then we'll see what this doctor command comes up with now on the latest Claude Code installation running on Opus 5. So what I'm wanting to do is it helps me prep for the new rules of context engineering. Per this blog post to make sure that I'm leveraging the generation 5 models to their fullest. Jonathan, while we're waiting, what happened on your end? What did it suggest?

Yeah, it basically wanted to remove almost all of Claude MD, inventoried everything that was removing and why. I have a broader plan, you know, that like I created, but I was also trying to say, how can we meet our stacked harness architecture and what's the right way for the stacked harness architecture to co-evolve with Claude 5? So I went a little bit deeper. deeper in that direction.

Wait, say more about stacked harness architecture for people that are, haven't, you know, just tuned into the pod the first time.

We've been building our harness so that we can incubate agents and we can capture domain knowledge, like in a product manager agent. We have product manager sub-agents about like how to write stories, how to define requirements. And the stacked agent harness architecture keeps all of that learning inside of the harness itself so that you can incubate agents that are ready and have LLM judges for you to apply to your work. So you run it in Claude Code or you run it in Codex. And you initiate the harness that you want to use, and then it opens up a window that you can use to interact with the set of agents that are currently on your harness. So we were doing some orchestration tasks that I didn't want to be doing. Now Claude 5 can sort of take those over, and I'm hoping like Codex sort of offers something similar for one. But more generally, it's like, how do we get the stacked harness architecture to work with these dynamic workflows? How should we define the skills that the stacked harness architecture calls? So a lot of the stuff was already in really good shape and well prepared for this update. And other things like some of the PR lifecycle skills that we have need to be simplified and need to be kind of like like pulled more into like our harness and harness changes from, from Claude Code. So a little bit of a project to make the transition for us, but like, it's pretty much, I think we've anticipated this and I think we've kind of architected in this direction, sort of skating towards the puck.

As the models keep creating bigger and bigger waves, how do we ride these waves? Understanding that. I just ran Claude Doctor with the latest Opus 5 on the latest Claude Code in my Claude Code session here, and I'm asking it to help me prepare for the new Generation 5 Claude models, and it gave me a health report. It's telling me where I'm wasting context, and it's going through this list. It went through my past 50 coding sessions to make these suggestions. So I would highly recommend any AI builder out there to prep for the next Claude 5 generation models. Go just ship it, copy paste this blog Article that talks about the Generation Five context engineering rules, how to optimize what you get out of these models, the latest and greatest from the Anthropic team. Run it and then carefully read through this and upgrade your setup, and then you'll be prepared for all the new Generation Five models. Say, this is pretty cool. I'm gonna work through this, and I will clean a lot of my wasted. That's pretty. Sick. I love this Claude doctor command where it's like, yeah, you're right, I have gotten so lazy. I would love a skill for this. Thank you. Thanks for helping me do my spring cleaning, Boris, in just a slash command. I really appreciate it. Jonathan, anything else on the new roles of context engineering or dynamic harnesses or anything coming out from Anthropic here that we should discuss before closeout?

I think we covered a pretty good job of it. Docs that are 2 days old, And, you know, we're learning as it evolves too. I'm looking forward to getting my hands on Opus 5 and just with this new context situation, building some projects and seeing how it works, because I've heard a lot of good things.

I'm really excited. I'm going to try to— I've been building up a goal loop prompt. And so with the new Opus 5, I'm planning to hopefully one-shot a mobile app over the weekend here. So follow up in next week's pod on how that goes and hopefully be able to show you, show some of you that. Anyway, to the audience. If you've listened this far, please hit subscribe on whatever platform you've found us. It really helps us keep the pod going. If you are sick of all the noise and can't keep up with everything on Twitter, GitHub, new repos, new whatever, YouTube videos dropping, new stuff from Anthropic, OpenAI, whatever, and you just want a weekly dose of clarity, hit subscribe. This is your weekly AI clarity, where Jonathan and I, as builders and founders on the front lines of AI innovation, Building our own company and startup every day. We bring our takes on the latest news that AI builders, tech executives, and product people need to know about. So stay tuned for next time. This has been your weekly AI Clarity. Take care. Bye-bye.

Bye.