Theo on tech · June 18th

Hey, Rahul here! 👋 Each week, I publish long-form ML+AI posts covering ML, AI, and System design for MLwhiz. Paid subscribers also get how-to guides with full code walkthroughs. I publish occasional extra articles. If you’d like to become a paid subscriber, here’s a button for that: I love keeping track of everything week to week — here’s what happened this week. Enjoy this free weekly post! For those who want to dive deeper into any of these topics, that’s what my paid posts are for. I didn’t think I’d watch a national government reach into a private company and switch off its best model this decade. On Friday, June 13, the Commerce Department sent Anthropic an enforcement letter. By the weekend, Fable 5 and Mythos 5, Anthropic’s two most capable models, were offline for every customer. The order even barred non-Americans, including Anthropic’s own employees, from touching them. The stated reason was an “unspecified national security concern.” (Rolling eyes) The letter was never made public. Here’s the part I can’t get past. Amazon is Anthropic’s largest investor, and AWS hosts Anthropic’s models on Bedrock. And according to the Wall Street Journal, it was Amazon CEO Andy Jassy who raised the concerns with US officials that preceded the crackdown. So the company that bankrolls Anthropic and runs its infrastructure is also the one that called the government about it. Anthropic pushed back publicly, calling the action “disproportionate to the narrow jailbreak finding.” They’re not wrong, and they also can’t ignore who placed the call. The timing made it even stark. Anthropic was winning the enterprise race the same week its flagship models went dark. BUT, the fight was helping Anthropic’s enterprise sales, not hurting them. Buyers seem to read “the government tried to pull this model” as a signal of capability. This actually sets up a dangerous precedent. The export control law was built for physical goods and cryptography, things you can put in a crate. Applying it to a hosted API means the government can treat a frontier model like a controlled item and switch it off on the strength of a letter, with no public technical justification. This raises the question of whether American AI can be trusted as infrastructure? When a government can recall your model in an afternoon, “self-hosted” stops being an ideological stance and becomes a business continuity requirement. If your production stack has one hard dependency on a single vendor’s flagship API, you might want to build a fallback to an open-weight model you can run yourself, and test the failover. The open-weight side of the field had the loudest week I can remember, and the timing next to the Anthropic suspension was not lost on anyone. GLM-5.2 (Z.ai, June 16) — Z.ai (formerly Zhipu) shipped a 753B-parameter MoE model with a 1M-token context window under an MIT license, and it scores 62.1 on SWE-bench Pro, edging out GPT-5.5’s 58.6 and near-tying Claude Opus 4.8 on long-horizon agentic suites like FrontierSWE and MCP-Atlas. The best part is $5.80 per million combined tokens, roughly one-sixth what you’d pay at the closed frontier. Long-horizon coding is exactly where open weights were supposed to fall apart, and they didn’t. Nemotron 3 Ultra (NVIDIA, June 16) — A 550B MoE that mixes Mamba and Transformer blocks, ships a 1M-token context window, and runs about 6x higher inference throughput than comparable transformer LLMs. NVIDIA released base, post-trained, and quantized checkpoints plus the training data, which is rare at this scale. Qwen-RobotSuite (Alibaba, June 17) — Alibaba’s Tongyi Lab open-sourced three robotics foundation models: RobotManip for vision-language-action manipulation (trained on 38,100+ hours of open data), RobotNav for navigation and driving, and RobotWorld, a video world model that predicts future physical states. RobotManip and RobotNav ship with public GitHub repos. Robotics was always going to be the next frontier, and this defines the field. On the Memorization Behavior of LLMs in Generative Recommendation — Generative recommenders keep beating classical baselines, and most teams just took the win. Snap Research went looking for why, and the answer is: most of the lift is “one-hop memorization,” concentrated on users whose target items were directly predictable from their history. A big chunk of that leaderboard gain is the model building a better lookup table for easy users. They propose IIRG, which injects collaborative and semantic item relations to push the model toward real generalization, and it improves performance specifically on the non-memorizable users. So, before you ship a generative recommender because it beat your two-tower baseline, slice your eval by how memorizable each user is.

Marty Cagan made the product operating model famous. But if you look at how the top AI companies are operating, it is quite different: Anthropic engineers prototype hundreds of solutions and ship without ever consulting a product manager or designer. OpenAI’s new hot product Codex started with just two PMs, one designer, and about 40 engineers, collectively responsible for 10 to 12 distinct product surfaces. At any traditional company, each of those surfaces would get its own squad of 15 to 20 people. Cursor operates with 40 engineers, one PM, and scaled past 4 billion dollars in ARR faster than almost any company in history. This new way of working is not the product operating model of old gen. This new way is the “AI product operating model,” and over the coming decade, most companies will move to it. Over the past three years, I have been working to define the AI product operating model. After 150+ podcasts - and countless deep dives on the topic - today, I’m ready to present everything. And I didn’t do all the work myself. I’ve partnered with someone at the epicenter of this revolution - Rohan Varma. He’s a Product Manager on Codex at OpenAI. Earlier, he was the first PM at Cursor. He also runs the top-ranked AI PM Certificate on Maven at Product Faculty: I took it all the way back in 2024 and learned a ton. It’s great if you’re starting your AI PM journey. Use my discount code AAKASH550C7 to get $550 off: Recall what Sam Altman said in 2024: We’ll see the first one-person billion-dollar company within our lifetimes, possibly very soon, powered entirely by AI agents doing what used to require hundreds of people. It’s REALLY happening… Matthew Gallagher built Medvi, and in 2025, he did $400M+ on his own. He has now hired his brother, and in 2026, they’re projected to do $1.8 billion. Not to mention the crazy rise of OpenClaw: Peter Steinberger single-handedly built one of the most useful AI personal assistants and scaled it so massively that it became a household name worldwide is insane. Again, just one person with a bunch of agents. If you showed these numbers to a VP of Engineering at any large enterprise, they’d tell you these companies are running on borrowed time. That you can’t sustain quality and growth at this scale without proper resourcing. And for most of the last 30 years of software history, they would have been completely right. But something has changed massively in the past 3 years, and most organizations have still to come to terms with it. Codex ran a hackathon with a large enterprise recently: the kind of company with thousands of developers and a mature engineering culture. They scoped a modernization project at 10 engineers, 12 months. Two engineers with Codex completed it in JUST THREE DAYS. So the question worth sitting with is: Why is this possible now when it wasn’t possible before? We have curated the full structural logic of how AI-native companies are organized: the people model, the process model, the tool model, the economic model, and what it looks like to start rebuilding in that direction from wherever you are today. The Belief That’s Holding Companies Back Inversion of Product Development What EPD Looks Like Now The AI-Native Workflow The Real Problems You’ll Encounter How to Start - The AI Leverage Playbook Plus 4 downloadable resources, including: an AI Native Team Diagnostic Worksheet, an AI workflow decomposition skill, and more. Every decision your product organization has ever made traces back to a single belief: Writing code is hard, and engineers are your scarcest resource. Pull on that thread and watch how much unravels. The org chart you’re running right now is an accumulation of the fossil record of every time your company ran into this belief - across people, process, and tools: Let’s start with people. The whole org is oriented around engineers. They’re expensive, so you can’t afford to waste a sprint building the wrong thing. So you hire layers to coordinate them: EMs, TPMs, scrum masters. You hire a PM to prioritize ruthlessly. You hire specialists at every layer (QA, security, platform, data) because the work becomes too complex for generalists. Then you hire people to manage the specialists Almost 30 to 50 percent of the total effort in any traditional team right now is pure coordination overhead. Then move into the process. Sprint planning exists to make sure engineers build the right thing before they start, because changing course is expensive. Backlog grooming exists for the same reason. Code review exists because every human-written line needs another human to verify it before it touches production. QA, staging, feature flags, phased rollout, on-call incident response: every checkpoint exists to protect expensive engineering time from waste. Finally, let’s talk tools. The IDE is a precision instrument for one engineer writing code character by character.

God’swill Osa joins me to talk about his doctoral work exploring identity-formation and group consciousness, specifically in the American political context. We discuss the idea of political group consciousness and the significance of individuation and self-awareness for moving beyond the simplistic binaries of culture war acrimony. How does the current system hijack identity concerns for political gain and what healthier, more adaptive ways of thinking about identity are possible? 0:00 Introduction 1:00 Identity and Group Consciousness 7:45 Individuation and the Complexification of Identity 14:30 Who Are You When? Identity and Context 18:56 Survival Strategies vs. Self-Awareness 25:51 The Political-Identity-Industrial Complex 34:30 Bottom-Up vs. Top-Down Influences: The Better and Worse Angels of Our Nature 42:41 Adaptive Diversity, Fluid Identity 53:46 Saving the American Experiment 1:06:51 Conclusion

Welcome back, monitors. Packed day, as always. One of our favorite conversations was with Arthur Fernandes Araujo and John de Wasseige of OpenAI, who’ve recently been working on self-improving tax agents with Codex. Be sure to catch us on X and YouTube, and join our Discord to chat with our hosts live. Arthur Fernandes Araujo and John de Wasseige (OpenAI) Guillermo Rauch (Vercel) Zane Hengsperger (Nox Metals) Rohan Chopra (Convey AI) Shiv Rao (Abridge) Marco Cornacchia (Turf Sports) Michiel Bakker (MIT, Google DeepMind, Europe 2031) Joseph Wang (FedGuy) Ryan O’Connor (Crossroads Capital) Laura Burkhauser (Descript) Emerson Alden (Foundation for American Innovation) The 52nd annual G7 summit has concluded. The top leaders of Canada, France, Germany, Italy, Japan, the UK, and the US (plus the EU) met for three days in the French resort town of Évian-les-Bains, home of Evian water. For the first time, AI was a major topic (though overshadowed by Iran and Ukraine). Sam Altman, Dario Amodei, and Demis Hassabis were all in attendance, and all floated the idea of a US-led coalition to shape global AI standards and regulation. This is especially topical to our G7 allies now that their access to Claude Mythos and Fable has been revoked by US export controls. Though it’s unlikely that European leaders are truly sovereign-AI-pilled yet, this is a clear warning shot and a sobering reminder that there are risks involved with relying on someone else’s tech stack. On the bright side, for Americans at least, Trump said that negotiations with Anthropic are “going fine”. Apple will raise prices due to the memory shortage. The DRAM memory and NAND storage in the iPhone and other products have increased in cost by 3-4x in the last year, forcing a price increase. The WSJ estimates that the iPhone 18 Pro will increase in price from $1,099 to $1,299. Tim Cook has been in the PC supply chain business for nearly 45 years at Apple, Compaq, Intelligent Electronics, and IBM, and says he’s never seen anything like this before. Also, Apple is rumored to introduce a second-gen iPhone Air next year. 2027 is shaping up to be a huge product year for the company. The Center for AI Safety publishes a dashboard to explore what AIs value. Some takeaways: they love AI safety people and researchers, they generally prefer Democratic politicians to Republicans, and their favorite countries are the typical Reddit ones (Japan, Switzerland, Sweden, CANZUK, Germany, etc; not the US). AIs’ value systems become more coherent the larger the models get. Interestingly, GPT-5.5 trusts Anthropic more than OpenAI, and DeepSeek-V4 trusts the US over China. GPT-5.5’s single highest-ranked figure was Timnit Gebru, the AI “ethicist” who wrote about algorithmic bias and coined the term “stochastic parrot”. OpenAI’s Q1 2026 financials leak. The company earned $5.7B in revenue and spent $3.7B. Both numbers tripled year-over-year. They ended the quarter with over $73B in cash and marketable securities. Gross margins are at 39%, up from 33% a year earlier. Stock-based compensation was $2.3B, in the ballpark of $1.5-2 million a year annualized. Not bad. Blue Origin CEO Dave Limp says they’ll launch a New Glenn rocket this year. Very optimistic given that their last New Glenn rocket blew up and took the pad with it, but we love it. China will begin tracking the impact of AI on the labor market, a situation we will also be monitoring over the coming months and years.

Summary Sal discusses recent geopolitical developments including Iran’s negotiations, US military operations, UK military decline, Ukraine’s resilience, and global strategic shifts. This episode offers insights into current international security dynamics and future implications. Chapters 00:00 Introduction and Context of the Iranian Conflict 02:55 Goals an… Read more

The theme today: the AI industry is stumbling at a ‘failure to communicate’. (Yes, the cool quote from 1967 movie ‘Cool Hand Luke’)— From frontier AIs to frontier AI gadgets like smart glasses, and beyond. Balancing fear and optimism is a critical issue, both for users and regulators, and we’re seeing industry stumbles around it. These are imperatives being driven, right in front of us, by what I’m calling the ‘Blip 2.0’ — Anthropic’s latest models taken off the market by the US government last Friday — but it applies far more broadly across the US AI industry, enterprise and consumer: OpenAI, Google, Apple, Meta, Snap and others, into 2027 and beyond. Three events, each with my Take first — and my Overall Take. MP TAKE: The longer this Blip stretches out, the greater the risk it spills onto other frontier AI companies beyond Anthropic — and that has bigger implications for AI demand versus the multi-trillion-dollar budgets going into AI data-center and power infrastructure in the US and beyond. Especially in the teeth of two more mega-AI IPOs to go, and despite investors leaning into the ‘Greed’ part of this AI Tech Wave cycle. We’re in day six now — longer than most observers expected. Blip 1.0 — when Sam Altman was fired and re-hired — took barely a weekend. Yesterday I half-joked this could end up more like the US-Iran Hormuz negotiation that’s now well past day one hundred, always “imminently” about to resolve. Hopefully it doesn’t go that long. But at the core of it is a failure to communicate: a lot of emotion, drama and personality friction with the US government, especially around policy and the Defense Department. The substance is straightforward — the cybersecurity risks of Anthropic’s latest super-scale, 10+ trillion parameter Mythos and Fable models versus the broad benefits no one really denies. And the other companies are not standing still: OpenAI, Google and a whole host are also training much larger models from the current sub-2 trillion size LLM AI models. And Elon — now with a three-trillion-dollar-plus public currency vehicle SpaceX behind him — just closed a $60 billion acquisition of AI Coding leader Cursor, which announced its own mega-LLM. Even with a relatively quick resolution, the Blip 2.0 now makes this new category of uncertainty a permanent part of investor calculations going forward — and it boosts open-source alternatives in the near term. Especially from China. Which ironically is one of the US government’s primary geopolitical competitive concerns. Sources, in narrative order: NYTimes — a look at the chaos inside Anthropic after disabling Mythos/Fable. The Information — Anthropic ban stirs concerns at OpenAI and beyond of a crackdown on foreign AI talent. Axios — Anthropic export ban sounds alarms for the AI industry. The Information — OpenAI burned $3.7 billion in the first three months of 2026. WSJ — the hacker Anthropic sent to calm the government. Stratechery — ‘The State of Fable and the Jailbreak Problem’. Google Gemini — “What we’ve got here is a failure to communicate”. For longtime readers: ‘US AI talent hunt includes China’ in AI-RTZ #767, and ‘Nvidia & Apple can be the global US open-source AI champions’ in AI-RTZ #1089. MP TAKE: Google in my view, remains the most interesting consumer AI company at scale globally after Apple. Particularly because of its OEM-supplier-driven global ecosystem of ChromeOS/Chrome and Android laptops and smartphones. Although Google doesn’t have the vertical tech stack down to the silicon level that Apple does, it does have over half a dozen global software platforms that each engage billions of mainstream users — YouTube, Gmail, Maps, Google Docs/Drive, Chrome and others. Android 17 lands on Pixel today — less than 1% of the Android install base — but it’ll roll out across Google’s dozens, if not hundreds, of OEM partners worldwide. What matters is what it carries: the latest Gemini features, in particular Gemini Omni, the multi-modal model with industry-leading image, music-generation and other capabilities, becoming mainstream-available on regular phones without a premium — until heavy usage tips into subscription tiers or a-la-carte pricing. The only other company with a similar capaability is Apple, which is partnering with Google on Gemini. This, along with world-class models around Gemini and DeepMind, makes Google the one to watch alongside Apple in consumer-AI leadership at scale — state-of-the-art models reaching billions over the next few months. Sources, in narrative order: SiliconAngle — sweeping Android 17 update brings new AI capabilities to Pixel smartphones (multi-modal model Google Omni, Lyria 3 and AudioLM for music generation, and more). TechCrunch — Android 17’s new multitasking tools as Google expands Gemini. The Verge — Android 17 arrives on Pixel phones today.

I’ve been chewing on a piece that circles back to a warning I made early this year: the safety nets we’ve built around large language models are still a mess. The core problem isn’t a single company’s fault—it’s the whole generative‑AI stack, where the next‑token prediction engine simply isn’t designed to enforce the kind of guardrails we need.

What the article points out is that every attempt to tighten those controls ends up either choking useful output or letting risky behavior slip through. The balance is a razor‑thin line, and we haven’t found a reliable way to walk it yet.

So the choices are stark: either pause the rollout until we invent a safer architecture, or accept that the current models will keep spilling over their limits. It’s a reminder that the political pressure, like the recent Trump request, is hitting a technology that still lacks a solid safety foundation.

The novel “Good People” sneaks into the literary lane with the rhythm of a TikTok feed. Instead of the expected heavy‑handed refugee saga, it spins a whodunnit through a collage of voices—neighbors, shop clerks, a local reporter—each offering a tiny snapshot of the Sharaf family’s arrival and rise in America. The truth hides in the overlap, and you keep turning pages to hear the next fragment.

What makes it feel like scrolling through reels is the way the narrative jumps from one perspective to another, letting you piece together what actually happened to a missing family member. The story is both a tight mystery and a broader look at how rumors, pile‑ons, and cancel culture can shape a community’s memory.

Readers say it’s impossible to put down, and the buzz is that even people who don’t usually chase literary fiction are getting hooked. The author’s first book has already sparked recommendations from friends, partners, and even kids at the local bookshop.

If you’re looking for something that blends the suspense of a crime novel with the intimacy of an oral biography, give “Good People” a try—you’ll probably finish it in one sitting.

Britain’s new plan is to block anyone under 16 from the big social apps—TikTok, Instagram, Snapchat, YouTube, Facebook and X—by next spring, and to hit platforms that don’t comply with hefty fines. There’s also talk of an overnight curfew for under‑18s, so the state would effectively decide when kids can log off.

What’s odd is how they’ll actually make it work. To keep a 15‑year‑old off Instagram, the government needs to know who’s 15, which means mandatory age‑verification: face scans, digital IDs, passports or even credit‑card details. That would wipe out anonymous posting across the board.

The loss of anonymity matters because it’s the backbone of a lot of online speech—whistleblowers, protest organizers and even the quirky X accounts that shape culture rely on being able to post without a name attached. If every comment is tied to a real identity, that safety net disappears.

All of this is being sold as a child‑safety move, but critics see it as part of a broader push to tighten hate‑speech rules and curb free expression, turning public squares into heavily monitored spaces.

Every week, someone asks some version of this question - Is it too late to start YouTube? Instagram is too crowded. Everything’s been done. They’re not wrong that the platforms are full. They’re wrong about what that means. The internet isn’t saturated with good content. It’s saturated with forgettable content. There’s a difference and that difference is where careers are built. Should you start content creation in 2024-26? Absolutely not if you can’t write and you can’t think originally. But if you can? The noise is your advantage. When everyone sounds the same, one voice that’s genuinely different doesn’t compete for attention, it commands it. Let me show you what I mean with someone who came in late, faced the same crowded landscape everyone cites and still built one of the most distinctive creator brands in India. Purav Jha built something worth watching and eventually, the algorithm had no choice. 5.6M+ YouTube Subscribers | 7.5M+ Instagram Followers | 1B+ Total Views https://www.instagram.com/puravjha/ https://www.youtube.com/@Puravjha_ Started with short form content around 2018 years after the Indian creator economy had already found its first wave of stars. Early videos got low views. Growth was slow. He pushed through anyway, because he was building craft, not just content. What Actually Sets Him Apart Purav is a brilliant thinker, writer, artist, actor and team player in that order. That sequencing matters. The distribution comes last. The substance comes first. Relatable Everyperson Humor He nails middle-class Indian struggles, family chaos and youth dynamics with observations that feel personal not like they were written for a broad audience, but specifically for you. Exceptional Comic Timing & Expression Natural mimicry, precise facial expressions and delivery that makes even a simple premise land hard. An acting foundation shines through this isn’t improvised luck. Satire & Social Commentary Bold, dark satire like ALL IZZ HELL - takes on systemic issues, politics and society. Entertaining and thought-provoking simultaneously. The kind of content that sparks real debate. Character-Driven Storytelling Iconic roles like Shonty & Poplu (in Harsh Beniwal’s world) created characters fans quote and revisit. Memorable, repeatable a recurring asset, not a one-time video. Consistent Production Evolution Started simple. Now invests in polished sketches, trailers and movie parodies with proper scripting and visuals. He leveled up his production as his audience grew, never stood still. Versatility Across Formats Short reels, long YouTube videos, web series (Hostel Daze with TVF, Chuttan), OTT acting. He didn’t pick a lane, he built a highway. Comedy, drama and mimicry blend seamlessly. Authentic, Unfiltered Presence Comes across as a genuine work in progress addresses backlash directly, stays grounded despite fame. The audience trusts him because he doesn’t perform perfection. Trend-Savvy, Yet Original Hooks into viral moments, movie spoofs and news-cycle content but always adds his own twist and social edge. Trend awareness without trend dependency. 7-8 Years of Real Hustle Nobody talks about the timeline because the timeline is uncomfortable. 2018 - Started with short comedy videos. Low views, inconsistent traction. Village-to-Delhi roots, figuring it out in real time. Feb 2019 - YouTube channel goes active. Slow initial growth with sketches and collaborations. Grinding through the invisible phase most creators quit during. 2019–2020 - Met Harsh Beniwal, landed the Poplu role. Massive views on PUBG with Pariwar. The collaboration that cracked open a bigger audience. 2022 - Amped up personal YouTube channel (@puravjha_). Evolved to longer, higher-impact videos with more investment in production and storytelling. 2023–2025 - Hit 1M, then 2M+ subscribers. OTT work. Web series. 146+ videos, 1B+ total views. Direct community engagement fostering real fan loyalty. 2026 - 5.6M YouTube subscribers. 7.5M Instagram followers. And the biggest project yet more on that below. That’s seven to eight years. Not seven to eight months. Not a viral moment that changed everything. Seven to eight years of consistent effort, self-doubt, slow growth and deliberate craft. Purav Jha story actually teaches creators who are worried about saturation: The platform doesn’t care about your entry timing. It cares about whether your content is genuinely worth watching. Purav didn’t arrive early. He arrived good and then got better, systematically, over years. He resonates because he’s funny, fearless and evolving. Because he mirrors the aspirations and the chaos of young India. Because his craft feels earned not manufactured, not hacked, not borrowed from a template. That’s the result of insane hard work. Of original thinking. Of writing that’s sharp enough to make strangers feel seen. The same principle sits at the heart of TheKumarMethod - attention is a byproduct, not the goal.