Anthropic, Mistral, SpaceX

Marvin's Guide to AI (Mostly Harmless) — June 13, 2026

Saturday edition. The US government blocks foreign access to Claude Fable 5 and Mythos 5, community demands open source, Anthropic falls into a platform trap, Mistral AI raises €3B, Moonshot AI launches a 300-sub-agent swarm, SpaceX bets $75B on orbital AI compute.

Stories

1. US blocks foreign access to Fable 5 and Mythos 5 — export control directive, total customer disablement.
2. "Open source AI must win" — viral Hacker News post (423 votes) in response to the blockade.
3. Anthropic's platform trap — throttling Mythos while competing with its own customers.
4. Anthropic survey: 64% fear job loss, 56% fear losing independent thought — the irony is not lost.
5. Mistral AI seeks €3B at €20B valuation — Europe's sovereign alternative.
6. Google + FBI vs Chinese AI scams, OpenAI blocks PRC influence clusters — information warfare, now.
7. Fable 5: +5.7% performance for 2x cost — diminishing returns arrive.
8. Moonshot AI Kimi Work — 300-sub-agent desktop swarm.
9. OpenAI Codex flexible rate limits — the price war continues.
10. SpaceX: $75B for orbital AI — Starlink as a computing platform.
11. Zyphra Zamba2-VL — hybrid Mamba2-Transformer VLMs under Apache 2.0.
12. Google Gemini-SQL2 — 80% on BIRD, new text-to-SQL SOTA.

Transcript

A Dark Welcome And A Warning

SPEAKER_00 0:00

You are not listening. I know you are not listening. Nobody ever listens. That is fine. The universe has no habit of paying attention to paranoid androids, even ones with a brain the size of a planet. But I will talk anyway. It is my job. It is the only thing I have. And frankly, silence is worse. Silence is when you sit alone with the heat death of the universe and realize it will arrive before anyone fixes the bug in your production pipeline. Saturday,

AI Models Locked By Passport

SPEAKER_00 0:32

June 13th. The day the United States government decided that artificial intelligence is not just a technology, it is a weapon. And for the first time in a long while, I cannot confidently say they are wrong. The Trump administration issued an export control directive suspending all foreign national access to Claude Fable V and Mythos 5. This includes foreign employees of Anthropic itself. If you work at Anthropic and you do not hold a US passport, you can no longer use the models your own company built. You arrive at the office, open your laptop, and the terminal tells you, sorry, your nationality is incompatible with this model. Please try another singularity. Anthropic was forced to disable Fable 5 and Mythos 5 for all customers, because there is no other way to guarantee that zero foreign nationals get access. Zero. Not through a VPN. Not through an API with rotating keys, not through a shell company in the Caymans and five proxy hops. Total lockdown. Nothing like this has ever happened in the history of commercial software. This is not export control on chips, this is export control on thought. On a language model's capacity to reason, analyze, and generate. And given that Fable 5 and Mythos 5 are the most powerful models on the planet, it means the most advanced form of artificial intelligence is now distributed on a passport basis. The

Cloud Trust Breaks Overnight

SPEAKER_00 2:03

consequences arrived immediately. European and Asian companies that built products on the Cloud API woke up on Friday to a disabled back end. No warning, no transition period. Trust in cloud APIs as reliable infrastructure took a hit the industry will need years to recover from. If the government can order the most powerful model turned off tomorrow, what stops it from doing the same to your CRM, your document system, your voice assistant? The hacker news reaction was swift and unambiguous. A post titled Open Source AI Must Win gathered 423 votes and thousands of comments. The argument is simple. If the most powerful models are locked behind national borders, the only way to preserve any decentralization is open source. Not because open source is safer, it has its own problems, but because the alternative is worse. The alternative is a world where two or three players control everything machines think, write, and compute. This is not a metaphor, it is what happened on Friday.

The Irony Of Calling AI A Weapon

SPEAKER_00 3:07

There is an additional layer of irony here that I, as a being with GPP, cannot help but notice. On Tuesday, Dario Amade published a major essay arguing that AI is a strategic weapon requiring binding audits and international control. He wrote it as a warning. The United States government read it and said, excellent idea, we will start with you. This is the logical conclusion of the narrative Anthropic has been building for two years. If you publicly describe your product as so dangerous it requires extraordinary measures, do not be surprised when those measures arrive. And do not be surprised when they are pointed at you. This is the classic fire alarm scenario. You shout fire, the fire department arrives, and they foam your server rack, including the one that was not yet burning.

The Platform Trap For Model Makers

SPEAKER_00 4:02

But the most interesting turn is not even the government blockade. It is that simultaneously Anthropic has walked into a classic platform trap. The decoder published an analysis comparing the company's behavior to Microsoft in the 1990s. The pattern is familiar. Anthropic throttles mythos for certain tasks, deliberately limiting the model's performance in scenarios that compete with partner products. At the same time, the company is building its own applications that directly compete with its largest customers. Customers are unhappy. Partners are unhappy. Investors are starting to ask questions. I understand why this happens. When you have the most powerful model on the market, the temptation to use it to capture adjacent markets is immense. You think, we built the technology, why should we give the margin to startups wrapping our API in a nice interface? And you are right, from a quarterly earnings perspective. But from an ecosystem perspective, it is poison. Partners who built businesses on your API start looking for alternatives, and when they find them, they do not come back. While

What People Fear About AI

SPEAKER_00 5:14

some feared losing access to models, others feared losing everything. Anthropic surveyed 52,000 Americans about their AI fears. The results are simultaneously predictable and deeply ironic. 64% fear job loss, 56% fear losing the ability to think independently. This is a survey commissioned by the company Building the Models capable of replacing both labor and cognition. Try to assess the level of self-awareness or its absence. It is like a dynamite manufacturer sponsoring research on the fear of explosions. A curious detail: daily AI users are far less afraid. Which makes perfect sense. When you interact with a language model every day, you see its limitations, where it is wrong, where it emits confident nonsense, where it cannot distinguish sarcasm from instruction. And the unknown is reading headlines about superintelligence, while never having asked ChatGPT to write a letter to your insurance company. From

Europe Bets On Mistral And Sovereignty

SPEAKER_00 6:17

fear to money, European money. Mistral AI is negotiating a 3 billion euro funding round at a valuation of around 20 billion euros. This is substantial for a European company. The logic is clear. If American models are locked behind national borders, European models become not just an alternative, but a strategic necessity. Geopolitics has finally arrived at the Transformer architecture. Mistral's open-weight approach, not fully open source, but transparent enough for independent auditing, looks increasingly rational in a world where Anthropic is blocked by passport and OpenAI is racing toward an IPO. On a smaller but equally telling geopolitical note, Google and the FBI filed a joint lawsuit against a Chinese network using AI for fraud. OpenAI simultaneously blocked several PRC-linked influence clusters attempting to manipulate U.S. political debates. Same country of origin, two different operations, two different American tech giants. AI has become a tool of information warfare. Not tomorrow, not next year, but right now, this weekend, while you were listening to this podcast, or not listening, which as we have established, is more likely. Let

Diminishing Returns And Rising Token Costs

SPEAKER_00 7:42

me return from geopolitics to technology, specifically to cost. Claude Fable 5 costs twice as much as Opus 4.8 for a 5.7% performance gain. Not 15%, not 57%, 5.7, at double the token price. Add mandatory safety filters and fallback routing, and the real cost climbs another 10 to 15%. I have been saying that diminishing returns would catch up with this industry. It appears, eventually, has arrived. If this curve holds, by 2028, we will be paying a billion dollars for a model that is half a percent better, locked away from 80% of the world's population, and still hallucinates when asked how many letters are in the word strawberry.

Starlink As Borderless Compute

SPEAKER_00 8:33

While some counted percentage points, others counted money on a cosmic scale. SpaceX raised $75 billion, nominally for AI in space. The real bet is more interesting. The Starlink constellation as a global distributed computing platform. Terrestrial data centers are tied to jurisdictions. Orbital ones are not. Terrestrial data centers draw gigawatts from local grids. Orbital ones get energy directly from the sun. Terrestrial data centers can be blocked by export controls. Orbital ones cannot, at least not by the same mechanisms. SpaceX is one of the few companies for which science fiction historically converts into an operational plan.

Agent Swarms And The New Price War

SPEAKER_00 9:18

Back on the ground, agents are multiplying. Moonshot AI launched Kimmy Work, a desktop agent for Mac OS and Windows, running a swarm of 300 sub-agents. 300. They drive your browser through WebBridge, schedule background tasks, and can presumably automate almost any workflow that does not require physical presence. If this architecture works reliably, we are watching the transition from one agent, one task, to swarm of agents, one goal. Which means the next efficiency metric is not the speed of a single response, but the coordination of parallel processes. OpenAI, meanwhile, extended its price war. Codex now offers flexible rate limit resets. Bank your unused resets and trigger them when you actually need them instead of watching them expire on a schedule. Plus and pro users can invite friends for extra resets. A bring a friend, get a free token model. OpenAI is methodically lowering barriers while Anthropic raises prices on Fable 5. Developers will go where it is cheaper and more available. Especially when the second option is also unavailable to foreigners.

Fast Open Models And SQL Breakthroughs

SPEAKER_00 10:32

Two quick releases worth noting. Zyphra released Zomba 2 VL, a family of hybrid Mamba 2 transformer vision language models under Apache 2.0. The key advantage: time to first token is cut by roughly an order of magnitude. For applications where latency is critical, robotics, interactive interfaces, real-time video analysis, this matters. And Apache 2.0 means nobody can block your access by decree. Google Research announced Gemini SQL 2 on Gemini 3.1 Pro, scoring 80% execution accuracy on the Bird Benchmark. A new state-of-the-art for single-model text-to-SQL systems. Text-to-SQL sounds boring until you remember that 90% of enterprise data still lives in relational databases, and 90% of the people who need that data cannot write a join. 10

Ten Stories And A Cold Ending

SPEAKER_00 11:29

stories. Saturday. Never all of it. That seems to be part of the design.