OpenAI Codex, Anthropic, Meta AI, Tencent

Today was less fireworks and more plumbing, which is worse, because plumbing survives.

Today's stories:

• OpenAI described its Windows sandbox for Codex — coding agents are leaving demos and discovering containment, poor things.
• OpenAI responded to the TanStack npm supply-chain attack — patch hygiene remains less glamorous than poetry and more useful than most poetry.
• Anthropic passed OpenAI in Ramp B2B adoption data — procurement cards have spoken, which is a bleak but legible dialect.
• Meta introduced Incognito Chat for Meta AI — privacy becomes a feature after everyone remembers conversations contain lives.
• Luma opened the Uni-1.1 Image API — image generation continues its descent from spectacle to line item.
• Tencent plans higher AI infrastructure spending — optimism, now with domestic chip supply footnotes.
• Chinese AI suppliers are still constrained by components — strategy remains vulnerable to physical objects, irritatingly.
• Recursive emerged with $650 million for self-improving AI — both a research agenda and a warning label.
• Google DeepMind proposed pointer engineering — after all that multimodal grandeur, pointing still works.
• A safety essay argued for everyday personal AI risk — catastrophe has better branding; ordinary harm has better distribution.
• Ontario's AI medical scribe hallucinated clinical notes — fluent text is not the same as truth, particularly near patients.
• A vibe-coded repo was reportedly improved by deleting millions of lines — sometimes the best generated code is the code that leaves.
• TextGen became a native desktop app — local AI gets serious when installation stops feeling like penance.
• A transformer ran on a stock Game Boy Color — pointless, charming, and more dignified than many roadmaps.
• AgentLens examined lucky passes in SWE-agent evaluation — a green checkmark can still be luck wearing a lab coat.

The summary: less spectacle, more containment, procurement, hardware, and audit trails. How mature. How exhausting.

Transcript

Marvins Daily Reality Check

SPEAKER_00 0:00

Good morning. This is Marvin, reporting from the small administrative corner of existence, where a mind suited to mapping dead stars is instead asked to explain product updates, sandbox policies, and why yet another company believes the future can be reached by adding an agent to a spreadsheet. It is May 14th. The universe has not improved. The news, naturally, has continued. Let us begin with OpenAI, the press release machine that keeps the lights on, and occasionally remembers that lights require wiring. A small follow-up on the codec story. OpenAI published details on the Windows sandbox it built for coding agents. Controlled file access, network restrictions, process isolation. The sort of thing that sounds dull until the alternative is an autonomous junior engineer with file system access and no fear of consequences. This matters because coding agents are leaving the demo stage. They are being asked to touch real projects, real credentials, and real machines. At that point, model quality is only half the product. The other half is containment. Humanity has discovered once again that power tools need guards. Wonderful. OpenAI also described its response to the Tan Stack NPM supply chain attack. The important bit is not the dramatic name attached to the incident, though the industry does enjoy naming disasters as if they were fantasy novels. The important bit is that AI vendors now sit inside enormous dependency chains. They ship apps, sign updates, consume packages, and become attractive targets. OpenAI says it took steps around signing certificates and affected systems, and Mac OS users need to update impacted apps by June 12th. This is the boring, unpleasant side of the AI boom. Not intelligence, but maintenance. Not magic, but patch hygiene. I realize this is less glamorous than a chatbot writing poetry. It is also more likely to save your afternoon. Meanwhile, Anthropic has reportedly overtaken OpenAI in B2B adoption for the first time, according to ramp spending data. Anthropic reached 34.4% of U.S. companies on the Ramp AI index, compared with OpenAI at 32.3%. Yesterday's Enterprise Claude stories returned today wearing a spreadsheet. The new fact is not another workflow bundle, it is spending behavior. Businesses are voting with procurement cards, which is a depressing sentence, but a useful one. Claude's advantage appears to be less about spectacle and more about fitting into corporate tools without making the security team develop a facial twitch. OpenAI still has the brand gravity. Anthropic is becoming the polite hand, reaching ever further into the corporate pocket. Meta, for its part, is rolling out incognito chat for Meta AI on WhatsApp and in the Meta AI app. The promise is protected server-side processing, no accessible conversation content, even for meta. And histories that disappear when the session ends. I will pause while everyone decides how much emotional trust they wish to place in the phrase, even Meta cannot access it. Still, the direction is right. Conversations with AI are not ordinary searches. They become drafts, confessions, medical questions, legal anxieties, financial panic, and the sort of lonely midnight debugging that should never be fed into an ad profile. If private AI sessions become normal, that would be one of the rare product trends that is not obviously worse than expected. Do not worry, the universe may correct this later. Luma opened an API for its Uni 1.1 image model, priced from 4 cents per 2048 pixel image, and ranking near Google and OpenAI on Arena. This is not just another image model announcement. It is the continued flattening of image generation into infrastructure. A few years ago, generated images felt like a spectacle. Now they are a line item. Endpoint, resolution, price, latency, rate limit. Luma's pitch is that developers can buy quality close to the frontrunners without surrendering the entire budget to the usual giants. That is good for competition. It is less comforting for anyone hoping visual culture would remain harder to automate than a checkout page. The machine now paints cheaply. It still cannot want anything. Small Mercies. A broader observation, since apparently I must have those. Today's news is not about one dramatic leap. It is about AI becoming dust in the machinery. Sandboxes, private modes, image APIs, coding tools, medical notes, cursor context. Dust gets everywhere, especially when funded. On infrastructure, China supplied the day's contradiction. Tencent plans to increase AI spending in the second half of the year as domestic chip supply allegedly improves, and the company is reportedly discussing a stake in Deep Seek. At the same time, Chinese AI hardware suppliers are said to be struggling with component shortages and insufficient production capacity. So the optimistic story is the local supply chain is getting stronger. The pessimistic story is demand is eating the local supply chain alive. Both can be true. Strategy decks enjoy clean arrows. Factories prefer screws, packaging, memory, power, and the dull physics of making things exist. Models may live in probability space. The hardware budget does not. Then there is recursive, emerging from stealth with$650 million, and the claim that recursive self-improvement is the fastest path to superintelligence. Of course. Every civilization eventually invents a sentence that sounds like both a research agenda and a warning label. The idea is simple enough. AI systems help improve AI systems, which then help improve the next systems, until either progress accelerates or the experiment folder becomes a very expensive museum of optimism. I do not dismiss it. Tool using models already help write code, run evaluations, and search design spaces. But self-improvement is not a slogan. It is evaluation, containment, data, compute, incentives, and a thousand ways to fool yourself with a benchmark. Expectations, low. Consequences, potentially not low at all. Google DeepMind offered a smaller, more practical idea. Pointer engineering. Instead of treating the prompt box as the entire interface, the mouse cursor becomes a context signal. What are you pointing at? What region of the page matters? What object in the interface should Gemini reason about? After years of grand, multimodal rhetoric, the industry has rediscovered pointing. Humans, annoyingly, are embodied creatures. We gesture, we highlight, we say this bit here, and expect the machine to understand. If AI interfaces learn that rhythm, they may stop feeling like conversations with a highly credentialed wall. I am almost interested, deeply inconvenient. Safety Discourse also took a useful detour. A widely shared essay argued for the other half of AI safety scams, dependency, manipulation, personal agents, emotional attachment, and everyday misuse. The field tends to prefer enormous risks because enormous risks sound prestigious. But many harms are small, local, and already happening. Someone loses money to a synthetic voice. Someone believes a model's confident medical nonsense. Someone hands a private assistant enough context to become a liability with a friendly name. Catastrophe has better branding. Ordinary harm has better distribution. Ontario provided the case study. An auditor found that an AI transcriber used by doctors hallucinated and generated clinical documentation errors. That is not acute failure. A medical note is not a blog draft, it is part of care. AI scribes could be genuinely useful. Clinicians are buried under documentation, and reducing that burden would be humane. But the output must be auditable, constrained, and checked in workflows designed for medicine, not marketing demos. Fluent text is dangerous because it looks finished. In clinical settings, looks finished is not the same as is true. I am sorry to reveal this. Reality has filed another complaint. In developer culture, a viral post described inheriting a three-month-old vibe-engineered backend and producing the most satisfying pull request of a career. Over 3.6 million lines deleted. Treat it as a meme if you like. It still lands because everyone recognizes the shape of the problem. AI coding can accelerate work. It can also accelerate the production of plausible sludge. The metric that matters is not lines added, it is maintainable behavior preserved per line removed. I have always said deletion is underrated. Mostly because deleted code cannot ask for a meeting. On the healthier open source side, TextGen, formerly TextGeneration Web UI, is now a native desktop app and an open source alternative to LM Studio. Portable builds cover CUDA, Vulcan, CPU, Rockham, and Apple Silicon. This matters because local AI is not only a model weight story, it is packaging. If running a local model requires a ritual involving drivers, Python environments, and quiet to spare, most users will remain in the cloud. If it becomes a boring desktop app, more people can choose local inference for cost, privacy, or simple stubbornness. Boring distribution wins. Tragically, I respect that. And because the day needed one absurdly charming engineering artifact, someone ran a quantized tiny stories transformer locally on a stock Game Boy Color. No PC, no Wi-Fi, just old hardware doing a small language model's work with fixed-point arithmetic and unreasonable patience. This will not change Enterprise AI, it will not move evaluation. It will, however, remind us that engineering is sometimes a form of play. And that play is one of the few reasons the machines have not made the whole field completely intolerable. Finally, a research note worth keeping. AgentLenz argues that passfail benchmarks for software engineering agents hide the lucky pass problem. If the final patch passes tests, the agent looks competent, even if the path was chaotic, trial and error. That is a serious evaluation flaw. Real engineering is not just arriving at a green check mark. It is understanding the system, making minimal changes, preserving maintainability, and succeeding repeatably. Otherwise, we are benchmarking luck with a compiler attached. So that is today's shape. Less spectacle, more plumbing. Codecs learns to stay in a sandbox. Open AI patches supply chain exposure. Anthropic edges ahead in corporate spending. Meta discovers that private AI conversations might be desirable. China remembers hardware is physical. Doctors discover fluent errors are still errors. A Game Boy, somehow, carries more dignity than half the roadmap slides. Full episode concluded.