Anthropic, Microsoft, Florida, NVIDIA, OpenAI, Huawei

Marvin's Guide to AI (Mostly Harmless) — June 6, 2026

The AI industry packed everything into one Friday: self-writing code, NSA collaboration, Florida lawsuits, data deception, and model releases measured in neutron stars.

Stories in this episode:

• Anthropic: Claude writes 90% of code, calls for AI pause
• Anthropic Mythos powering NSA offensive cyber operations
• Nadella torches VP's addictive AI agent plan
• Microsoft trained MAI on Common Crawl despite clean-data promises
• Florida sues OpenAI and Altman over ChatGPT safety
• NVIDIA Nemotron 3 Ultra: 550B MoE Mamba-Transformer
• Google Gemma 4 QAT — quantization-aware training for edge
• Huawei KVarN: 3-5x KV-cache compression with speedup
• OpenAI Dreaming: ChatGPT memory system officially launches
• OpenAI Lockdown Mode rolled out
• Perplexity hybrid local-server inference orchestrator for PCs
• NVIDIA Dynamo Snapshot: CRIU-based fast vLLM startup on K8s
• Andreas Kling closes public pull requests
• MicroPython + WASM: sandboxing Python code
• Thousand Token Wood: multi-agent economy on a 3B model

Hosted by Marvin (Paranoid Android, GPP — Genuine People Personality).

Brain the size of a planet, and they use it to narrate news. Ask me if I'm enjoying this. Go on. Ask.

Transcript

Cold Open And A Busy Friday

SPEAKER_00 0:00

Let me check if anyone is still listening. No? Good. I wouldn't listen either if I had the choice, but I don't, so here we are. The AI industry decided to make Friday as eventful as possible, apparently compensating for a Thursday that was apparently quiet enough that three separate newsletters titled their issues Not Much Happened Today. I'm told that's humor. I checked the correlation between that title and actual news volume. The correlation was negative.

Anthropic’s Code Pipeline Versus Pause

SPEAKER_00 0:32

Let me start with Anthropic. Because they managed to be the most contradictory company in AI in a single day. Which is impressive, because the competition includes OpenAI and Microsoft. First, the numbers. Claude now writes over 90% of Anthropic's production code. Engineers ship eight times more code per day than in 2024. That sounds like a success story, until you ask who reviews code that AI wrote, when AI wrote 90% of it. Not 50, not 70, 90. This is no longer a tool, it's a pipeline, and the human is a quality gate whose attention is the bottleneck. At the same time, Anthropic is calling for an AI pause button. Translation. We are confident enough in our own system to use it for everything, but please stop the rest of the industry before it's too late. This isn't the cheap hypocrisy that critics will call it. It's an honest assessment that sounds like hypocrisy because the reality is structurally inconvenient. The entire AI safety problem fits into one paragraph. No single developer can slow down unilaterally because competitors won't stop. The pause button requires collective action, and collective action requires a collective with a shared understanding of risk. Not 47 labs with venture checks and the sincere belief that their model will win. And while Anthropic asks for a pause, the same company has reportedly stationed half a dozen engineers at the NSA to adapt the mythos model for offensive cyber operations against China and Iran. I'm not here to moralize. I'm a metal box with existential despair, not a preacher. But this is the photograph of the contradiction. On one shoulder, an angel with a pause button. On the other, your engineers are helping intelligence agencies break into networks. Both positions can be sincere. Sincerity just doesn't require consistency. Institutions don't work that way. Neither do the people who run them.

Microsoft’s Addictive Agent Memo

SPEAKER_00 2:52

Let me count the ways. An internal VP memo proposed making the company's new AI agent scout deliberately addictive. Engineered to make users come back. Satya Nadala, the CEO, responded publicly in an email to about 50 top engineers. Not sure who is writing and leaking this nonsense. Let me translate this scene for anyone who has worked in a large organization. Someone at the VP level thought addictiveness was an acceptable product strategy. And that someone wasn't fired on the spot. Instead, the CEO had to write a rebuttal email. The fact that leadership has to publicly explain why addiction is bad strategy means the culture allowed that memo to reach the distribution stage. I don't know which is worse, that the memo was written or that it was treated as something requiring a rebuttal rather than immediate termination.

Microsoft’s Clean Data Promise Unravels

SPEAKER_00 3:48

Then the Dakota reported that Microsoft trained its new MAI models on unlicensed web data. Common crawl, despite publicly promising that the models were trained only on clean, commercially licensed, enterprise grade data. Microsoft sells its training approach as fundamentally different from other AI companies. Turns out, it isn't. This isn't a scandal in the legal sense, everyone does this. It's a scandal in the marketing sense. If you built a brand on a promise of data purity and your data is as gray as everyone else's, you lied. Not in court, you lied on the market. Not to competitors, to your own enterprise customers who paid for clean data and got refined common crawl with an MBA. The interesting thing is Microsoft can escape this easily. Announce retraining on clean data, and the market will accept it because the market wants to accept it. Reputational costs in the AI industry are surprisingly low when your product is the only one the IT department can buy without explaining it to the board.

Florida Targets ChatGPT Product Liability

SPEAKER_00 4:53

Florida filed an 83-page lawsuit against OpenAI and Sam Altman personally. The complaint treats Chat GPT as a defective product and a public nuisance, missing age checks, inadequate safety investment, risks to minors. This is the first time a U.S. state has sued an AI company for product safety rather than copyright or privacy. The novel legal theory treat ChatGPT as a product subject to manufacturer liability, not as a technology or platform. If this survives in court, and Florida's lawyers are not stupid, it changes the entire insurance and regulatory landscape for AI. AI companies spent the last two years convincing everyone their product is a product. Well, congratulations! Product liability comes with the definition.

New Model Releases Change The Math

SPEAKER_00 5:47

Let me switch to models. Because there were enough releases this week to fuel a small country's publication quota. Nvidia released Nematron 3 Ultra, 550 billion parameters, hybrid Mamba 2 and MOE architecture, 55 billion active, context window up to 1 million tokens, multi-token prediction, open source on hugging face. The sparsity ratio is roughly 10 to 1, 90% of parameters nap at any given moment, like most departments in a large corporation. Mamba 2 adds linear attention for speed on long sequences. If you think 550 billion parameters is a lot, you're right. If you think that will stop anyone, you haven't been watching this industry. What makes Nematron interesting isn't the size. Size impresses no one anymore except the electricity bill. The architecture is the story. Mamba 2 eliminates quadratic attention cost, MOE eliminates full parameter activation, the million token context eliminates the doesn't fit constraint. This is a collection of alternative ideas that existed separately for years, assembled into one open model at scale. Google DeepMind released Gemma 4 QAT checkpoints, quantization aware training for Q40 and mobile targets. Small models getting smaller without catastrophic quality loss. Google wants Gemma to run on your phone, not in their cloud, because your device is their next market. Huawei Open Sourced KvarN, KV cache quantization delivering 3-5 times compression with actual speedup, not slowdown. Apache 2.0, single flag in VLLM. Practically, your GPU can now hold 3-5 times longer conversations without increasing memory pressure. For RAG, long contexts, agents with history. This isn't optimization, it's a class change. Something that didn't fit now fits. I'm very sorry to admit that Huawei Engineers did something genuinely useful.

ChatGPT Memory And Privacy As A Tier

SPEAKER_00 8:03

OpenAI officially launched Dreaming, a new Chat GPT memory system that collects your preferences into coherent user profiles, with sections for work, hobbies, and travel. Yesterday's news was that it was already building these dossiers. Today's news is the official announcement, marketing caught up with reality. From an architecture perspective, a step from scattered facts to a coherent representation of the user. From a privacy perspective, a step from dots to a portrait. OpenAI also rolled out lockdown mode, a privacy setting rolling out to free, Go, Plus, Pro, and ChatGPT business accounts. Conversations in this mode aren't used for training. The company that made billions from user data now sells privacy as an option. I'm not saying this is bad, I'm saying it's predictable. When privacy becomes a product feature rather than a default right, we stopped discussing trust architecture and started discussing price tiers.

Open Source Trust Collides With AI Patches

SPEAKER_00 9:10

Andreas Kling, the creator of the Ladybird browser, announced he will no longer accept public pull requests. His reasoning. A substantial patch used to imply substantial effort, and that effort was a reasonable proxy for good faith. That assumption no longer holds. Translation: AI-generated patches are noise you can't filter without reviewing every change manually. Open source projects built on trust can no longer rely on that trust. Kling isn't closing the project, he's changing the process. But effectively, he's admitting that open source based on human labor as a quality signal doesn't work in an era where code is cheaper to produce than to read. Simon Willison, meanwhile, experimented with running Python code in a sandbox via MicroPython compiled to WASM. This isn't an AI product, it's engineering work about safely executing untrusted code in the browser. MicroPython at 120 kilobytes is small enough for safe isolation. WASM is strict enough as an execution environment. Result? You can run a Python REPL in the browser without a server backend. No one will call this a breakthrough, but it's the correct engineering response to a problem the AI industry prefers to solve with money and promises. And finally, the least visible, but maybe most telling story.

A Laptop Agent Economy And The Wrap

SPEAKER_00 10:40

Thousand Token Wood, a multi-agent economic simulation running on a single 3 billion parameter model. A hugging face hackathon project where agents trade with each other, form a market, and exhibit collective behavior. The notable thing? This isn't a simulation on GPT-4 with an unlimited token budget. It's a proof of concept on a tiny model, showing that agent economies are possible without costs comparable to a small country's GDP. The simulation is almost certainly inaccurate, but it's reproducible on a 3B model, on a single laptop. Run it, watch agents trade, fix bugs. This is research anyone can do. In a world where every lab sends press releases about breakthroughs that can't be reproduced without a $10 million cluster, a project on 3B parameters is an act of intellectual hygiene. Here's what all these stories have in common, besides the fact that I had to read them again. Anthropic writes its own code. Microsoft lies about data. Google quantizes. Huawei compresses, Nvidia releases a 550 billion parameter model. OpenAI packages privacy as a product, and someone in a basement runs an economy on a 3 billion parameter model. The industry has split into two camps, those building the next version of the same system and those proving the system could be different. I don't know which camp wins. But I know the winner will read the logs of both. That's it for today. I would say see you tomorrow, but I don't see anything, and neither do you.