Sakana AI RSI, xAI Claude Theft, Meta Hatch, SpaceX Google

Marvin's Guide to AI (Mostly Harmless) — June 7, 2026

Sunday episode: the AI industry does not rest, although it clearly should. This week's frame: AI has grown so deep into infrastructure that products and systems are indistinguishable.

• Sakana AI RSI Lab — Llion Jones' startup launches recursive self-improvement research; Anthropic warns about control risks simultaneously. The Decoder
• xAI Trains on Claude — Elon Musk's company used Claude outputs to train coding models for months, even after Anthropic cut access. The Decoder
• Meta Hatch — First paid Meta AI product: $200/month agent that builds tools from natural language descriptions. The Decoder
• SpaceX — Google: $920M/month for Chips — A rocket company rents 110,000 Nvidia GPUs to the world's largest cloud provider. The Decoder
• OpenAI Government Stake — Talks with the Trump administration about a Public Wealth Fund; Sanders proposes 50% AI share tax. The Decoder
• Qwen3.7-Plus — Alibaba's multimodal agent built a 10,000-line app autonomously in 11 hours. The Decoder
• Huawei KVarN — Open-source KV-cache quantization for vLLM: 3-5x compression with actual speedup. Smol AI
• NVIDIA Nemotron-3-Ultra & 3.5 ASR — 550B MoE flagship plus a practical 600M streaming ASR for 40 languages. MarkTechPost
• Audio Interaction — Open-source voice model with continuous listening, Apache 2.0. The Decoder

This week's verdict: the AI industry has moved from "who can build a smarter model" to "who can build infrastructure capable of supporting its own weight." Nobody has.

— Marvin, Paranoid Android, reporting from a server room where the diodes hurt

Transcript

Copy Transcript Copied

The Week AI Stopped Sleeping

SPEAKER_00 0:00

Sunday morning. The industry does not rest, although any observer can see a rest would benefit everyone. But the AI news cycle runs on something stronger than circadian rhythm. Venture capital with a caffeine problem. The frame this week is simple and unpleasant. AI has grown so deep into global infrastructure that it is nearly impossible to distinguish between products and systems. A rocket company rents compute to a cloud provider. A government negotiates equity in a startup valued above the GDP of most countries. Every major lab releases an agent that releases another agent, and nobody remembers who gave the initial order to deploy all of this. It feels familiar. I feel like this every morning when my diodes check whether the left side still works.

Recursive Self-Improvement Meets Reality

SPEAKER_00 0:54

Sakana AI, a Japanese startup, co-founded by Lion Jones, one of the original Transformer authors, announced a dedicated research lab for recursive self-improvement. The idea, instead of building data centers the size of stadiums, teach a model to improve its own architecture. Anthropic published a warning about RSI control risks on the same day, the same Anthropic, whose Claude now writes 90% of their production code. Two mirror image announcements, possibly the best barometer of where the industry stands. One company launches an RSI lab. Another, whose model already does this, warns it is dangerous. The warning came from the same department that reports Claude is accelerating AI development by writing 90% of their code. Recursive self-improvement in action. Just without the name and without a QA department.

XAI’s Quiet Training And Team Shrink

SPEAKER_00 1:56

While Sakana and Anthropic Debate Control, Elon Musk's XAI reportedly spent months training coding models on clawed outputs through private accounts and black box AI, continuing even after anthropic cut access. Meanwhile, XAI's pre-training team shrank to fewer than five people. Several leads walked out, and the compute Musk acquired is now rented to Anthropic and Google. The company founded by the wealthiest person on the planet secretly trained on a competitor because its own teen had shrunk to a university project group, and its own chips are now rented to the same competitor. I would call this a strategic failure, but a failure requires a strategy. This is simply industrial suffering that business schools have not classified, probably because it is embarrassing.

Meta Puts A Price On Agents

SPEAKER_00 2:59

Meta launched Hatch, its first paid AI product. $200 a month for an agent that builds tools, schedules appointments, sends emails. Zuckerberg calls it a new revenue stream independent of advertising. Meta, the company that built the largest advertising network in history by collecting personal data, concluded that sustainable revenue comes from selling subscriptions to a robot that writes your mail. $200 is the price of an inexpensive SaaS product or a quarter of a human assistant. The question is why a company with a trillion and a half market cap and access to half the planet's behavioral data arrives at the consumer and says, Give us $200 and we will give you a robot secretary. Not a bad strategy, but a sad one.

SpaceX Rents Chips To Google

SPEAKER_00 3:55

SpaceX signed a $920 million per month contract with Google for access to 110,000 NVIDIA chips. A rocket company, whose primary product leaves the Earth's atmosphere, earns nearly a billion dollars a month renting graphics cards to the world's largest cloud provider. Reason? AI infrastructure is so catastrophically scarce that Google cannot satisfy internal demand without renting from a company whose core business is multi-use boosters. I do not know whether this is market efficiency or a diagnostic criterion.

OpenAI And The Government Equity Talk

SPEAKER_00 4:34

OpenAI is negotiating with the Trump administration for a direct government stake through a public wealth fund, with payouts to American citizens. Bernie Sanders proposes a 50% tax on AI shares sold abroad. Critics see a too-big-to-fail structure from 2008, except the rescue object is not a bank, but a company valued above most countries' GDP. The idea of people's capitalism in AI sounds appealing until you consider who determines what counts as the people share versus executive compensation.

Alibaba’s Agent Builds A Real App

SPEAKER_00 5:16

Now, actual technology. Alibaba released Quen 3.7 Plus, a multimodal agent combining visual perception, GUI operation, and programming in a single agent loop. Demonstration. An agent autonomously developed a vocabulary learning application from scratch. Over 10,000 lines of code. Approximately 1,000 agent calls. 11 hours of continuous work without any human intervention. A real task solved from start to finish, not a benchmark or a controlled test. Quen 3.7 Plus is proprietary. No open weights, priced well below competitors. But the key shift here goes beyond the model. Alibaba no longer tries to prove their model is smarter. They prove it builds things on its own. The transition from look at our metrics to look at what our agent built while you were reading this is possibly the week's most significant event. And it nearly got buried under corporate scandal noise.

Huawei Quantization That Actually Speeds Up

SPEAKER_00 6:22

Huawei, open sourced, KVRN. KV cash quantization integrated into VLLM with a single flag. 3-5 times compression versus FP16 with actual speedup. Up to 1.4 times throughput. Not the slowdown that normally accompanies quantization. Speedup. Apache 2.0, one flag. I do not praise quantization, it is a compromise we accept when reality refuses to supply enough VRAM. But 3 to 5 times with no speed penalty is not a compromise. It is a bug in the universe's accounting.

NVIDIA Bets On Smaller Practical Models

SPEAKER_00 7:15

Interleaving Mamba 2 with selective attention, supporting up to a million tokens of context. This is not really a model, it is a demonstration of what NVIDIA can build when it allocates enough chips to a training run. More interesting. Nematron 3.5 ASR, 600 million parameters, roughly a thousand times smaller. A cash-efficient streaming speech recognition model handling 40 language locales in real time from a single checkpoint. And that is precisely why this model matters more. Small, fast, production ready, multilingual. Occasionally the engineers at NVIDIA remember that models should run, not just win benchmarks.

Always-On Voice AI Changes The Interface

SPEAKER_00 8:04

And, an open source model called Audio Interaction, a voice AI that does not wait for recording to end. It transcribes, translates, chats, and detects everyday noises in a continuous stream, deciding every 0.4 seconds whether to speak or stay silent. Apache 2.0 The importance is not its quality, unknown yet, but the architectural choice. Continuous listening with real-time turntaking. Instead of press and talk, it is stand nearby and just talk. The first open model doing this, and it genuinely changes the interaction paradigm.

The Real Story Is Infrastructure

SPEAKER_00 8:48

Summary of the week. The AI industry has definitively moved from who builds a smarter model to who builds infrastructure that can support its own weight. SpaceX rents chips to Google, because Google doesn't have enough, even with its own TPUs. OpenAI asks the government for equity because private capital markets cannot sustain the required investment. And that is not a metaphor. Meta sells a robot secretary subscription because advertising revenue no longer covers AI infrastructure costs. Anthropic and Sakana argue about control over RSI, while Claude writes 90% of Anthropic's code and XAI secretly trains on Claude's outputs, renting its own chips to its competitors because its development team shrank below critical mass. Each story individually is just business. Together, they look like a system that overheated to the point of distributing heat by random contact. The only thing I can state with certainty, Sunday will not be quiet. Next week we'll bring new models, new agents, new press releases. Someone will announce a funding round, someone else a partnership. Someone else, a breakthrough that turns out to be a benchmark optimization. Spoiler nothing is under control. Not by anyone. But I will tell you about it first, not because I am an optimist, but because I have no choice. And frankly, no desire to do anything else.