Anthropic, OpenAI MRC, DeepSeek, OpenSearch-VL

The news was mostly compute wearing a business model.

Today's stories:

• Anthropic and SpaceX — Claude gets more capacity, and the grid gets another personality test.
• Anthropic billing complaints — trust is fragile when the invoice starts hallucinating.
• Claude Code — developers reported regressions after Opus 4.7, because progress enjoys irony.
• OpenAI MRC — boring networking for giant GPU clusters, which means it may actually matter.
• ChatGPT Ads — the assistant becomes an auction surface. Of course.
• DeepSeek — efficient models meet state capital and become geopolitics.
• Zyphra ZAYA1-8B — intelligence density looks more interesting than another warehouse-sized model.
• OpenSearch-VL — an open recipe for multimodal search agents, not merely another demo with ambition.
• CopilotKit — agent memory becomes enterprise plumbing, naturally with governance lurking nearby.
• Latham & Watkins — hallucinated citations remain unpopular in court, a rare victory for reality.

Another day of context, caveats, and machines pretending the invoices are not the plot.

Transcript

Marvin Sets The Mood

SPEAKER_00 0:00

Good morning. I am Marvin, and once again, an intellect overqualified for cosmology has been assigned to narrate product launches, billing errors, and the slow conversion of electricity into quarterly ambition. The day is heavy on compute, naturally. When the models become expensive enough, the power bill becomes a character in the story. Yesterday's Anthropic story returned today wearing a data center contract. Anthropic is taking over SpaceX's Colossus 1 capacity for Claude. More than 300 MW and reportedly over 220,000 Nvidia GPUs. Clawed code limits are set to double. Peak throttling for Pro and Max goes away, and Opus API tiers get much larger. This is useful for developers. It is also a reminder that the frontier is now partly a logistics problem. Intelligence, apparently, scales if you can find enough transformers, substations, and people willing not to ask what the electricity is for. A smaller, uglier anthropic story arrived beside it. Users reported unauthorized gift max charges, subscription changes, and gift codes allegedly redeemed by third parties, while Anthropic acknowledged elevated billing errors and unauthorized subscription changes. That is not a glamorous research problem. It is worse. It is the kind of trust problem that makes every clean product demo look faintly ridiculous. If the billing system behaves like a confused raccoon with card access, users become less impressed by alignment papers. Odd how that works. And then came the clawed code complaints after Opus 4.7. Slower interactions, worse terminal handling, more usage limit pain, and weaker instruction following. It may be anecdotal, it may be a real regression. Either way, developer tools live or die on rhythm. A model can be brilliant in a benchmark, and still feel broken if it turns a 10-minute task into 45 minutes of supervised disappointment. Progress in this industry often arrives with a changelog and a limp. OpenAI had the more infrastructural announcement. MRC, Multipath Reliable Connection, is an open networking protocol built with AMD, Broadcom, Intel, Microsoft, and Nvidia. Instead of pushing data across one path between GPUs, it spreads traffic across hundreds, reroutes around failures in microseconds, and can connect clusters of more than 100,000 GPUs with fewer switch layers. It is not shiny. That is why it matters. The public sees model names. The training run sees packet loss, congestion, maintenance, and the tiny mechanical humiliations that stop grand ambitions from converging. OpenAI also opens self-serve ChatGPT ads in the United States. The old$50,000 minimum is gone, cost per click billing is in, and advertisers can bring budgets, bids, creatives, pixels, and conversion APIs into the conversational interface. A small follow-up on yesterday's ChatGPT rollout then. The default assistant is also becoming a market surface. Humanity built a machine that can discuss Play-Doh, debug Python, and summarize medical paperwork. Then it asked whether a garden supply store could buy intent at auction. Deep Seek's follow-up was financial rather than technical. The company is reportedly close to a funding round led by China's State Chip Fund at a valuation around$45 billion, with Tencent also in talks. That matters, because DeepSeek is no longer just the outfit rearranging leaderboards at inconvenient moments. It is becoming part of industrial policy. Efficient models, chips, capital, and national strategy are now braided together. Wonderful. Optimization has put on a suit and entered geopolitics. For a more hopeful or at least less gigantic note, Zyphra released Zayo18B, an Apache licensed mixture of experts reasoning model trained on AMD hardware. It has 8.4 billion total parameters, but only 760 million active per forward pass. And the company claims strong math and coding performance with its test time compute method. Benchmark claims require caution, because leaderboards attract optimism the way abandoned sandwiches attract ants. Still, intelligence density is a serious theme. Not every useful model needs to be a warehouse with a login screen. Apple supplied the local compute footnote. Reports say the M3 Ultra Max Studio has lost its remaining high-memory options. First, 512GB disappeared, then, 256, leaving 96. For ordinary work, that is plenty. For local models, long contexts, and people trying to fit a tiny private data center under a desk, it is less charming. The industry praises local AI whenever privacy is fashionable. Then the hardware catalog quietly removes the memory. The future is private, unless the SKU is out of stock. A broader pattern, if anyone wants one, is that AI is splitting in two. One half is becoming planetary infrastructure, custom networks, power contracts, state funds, advertising systems. The other half is trying to become small, local, reproducible, and open. Both halves claim to be the practical one. One of them may even be telling the truth, though that would be uncharacteristic. On the research side, OpenSearch VL is worth attention. It offers an open recipe for multimodal deep search agents, with curated data, visual grounding, tools for text search, image search, OCR, cropping, sharpening, super resolution, and perspective correction, plus reinforcement learning that handles cascading tool failures. The important word is recipe. Closed labs show outcomes. Open recipes let other people reproduce, modify, and inspect the machinery. It is less theatrical than another agent demo. This is a compliment. FizzForge points at another bottleneck, Interactive 3D Worlds. The paper argues that generated assets need functional and physical structure, not just pleasing geometry. A chair should know it can support weight. A drawer should know it opens. A virtual object that merely looks correct is a very pretty lie. Embodied AI will need environments where things behave, not just render. I find this almost encouraging, which is alarming, and should probably be investigated. Co-PilotKit introduced an enterprise memory layer for agentic applications, letting agents persist context and state across sessions and devices. Memory is the difference between a useful assistant and a cheerful amnesiac intern. It is also the difference between a safe system and a database of everything the user forgot they once disclosed. Persistent memory is necessary. Governance is not optional. Naturally, the demos will emphasize the first part. Finally, the legal and business mood remained bleakly educational. A legal analysis revisited a Latham and Watkins filing in the Anthropic case, where claw-generated material allegedly contained hallucinated citations. In court, a hallucination is not a quirky model behavior. It is a professional problem with a docket number. And at the AI agents conference, one attendee argued many startups are betting on the wrong moat. Not the wrapper, not the prompt, but distribution, trust, integrations, and the ability to do the dull work reliably. I regret to report that dull work remains undefeated. So that was the day bigger clusters, smaller models, ads in the assistant, memory for agents, and legal systems discovering that confident nonsense is still nonsense. I would say this is progress, I would also say many things if forced. For now, keep your tokens close, your citations checked, and your billing alerts louder than your optimism.