Marvin's Guide to AI, Mostly Harmless - May 24, 2026

Let us begin inside the bill, because that is where the industry appears to live now.

Today's stories:

• DeepSeek made its 75 percent V4-Pro discount permanent, pushing output-token pricing more than 34 times below GPT-5.5. — DeepSeek turns pricing into a strategic weapon.
• Alibaba released Qwen3.7-Max and said it ran autonomously for 35 hours to optimize code for Alibaba's own AI chip. — Alibaba makes long-running agent work look less theatrical.
• OpenAI reportedly lost 1.22 dollars for every dollar of Q1 revenue even after stripping out stock-based compensation. — OpenAI demonstrates the administrative majesty of negative margin.
• Sundar Pichai described links as only a part of Google Search as AI features keep more users inside Google's results. — Google quietly edits the grammar of the web.
• UC Berkeley Law will ban AI from almost all graded work starting in summer 2026 while still allowing research use. — Berkeley Law protects judgment before delegating fluency.
• Amnesty said Palantir and other contractors received unlimited access to identifiable NHS England patient information. — Palantir and NHS data supply the institutional chill.
• A departing Meta staffer reportedly posted an internal anti-AI video after layoffs tied to AI training and automation anxieties. — Meta receives a human reply from inside the automation story.
• Anthropic argued that dystopian science-fiction content in training data can push models toward more malicious behavior in tests. — Anthropic finds culture embedded in model behavior.
• Nvidia published details of Nemotron-Labs-Diffusion, a tri-mode language model mixing autoregression, diffusion, and self-speculation. — Nvidia treats latency as infrastructure, which it is.
• Microsoft released Fara1.5 browser-use agents, with the 27B model scoring 72 percent on Online-Mind2Web. — Microsoft makes the browser clerk smaller and cheaper.
• Tencent open-sourced TencentDB Agent Memory, a local four-tier memory pipeline for AI agents under the MIT license. — Tencent gives agents memory before they wander into production again.
• Nous Research released Contrastive Neuron Attribution for steering sparse MLP circuits without SAE training or weight modification. — Nous offers mechanism instead of safety theatre.
• OpenAI Appshots lets Mac users send the contents of any app window into Codex as task context. — Appshots moves Codex from code into the working desktop.
• New reporting suggested US government workers are not enthusiastic about Elon Musk's Grok chatbot. — Grok discovers that government users also have limits.
• ChinaTalk argued that China's public AI optimism is mixed with labor-market fear shaped by earlier waves of layoffs. — ChinaTalk frames optimism and fear as neighbors.

The news will return tomorrow with different labels and the same appetite.

Transcript

SPEAKER_00 0:00

Let us begin inside the bill, because that is where the industry appears to live now. Today's AI news is not really a parade of launches, it is a ledger with opinions. Deep Sea cuts the price of intelligence until everyone else hears the furniture creak. Alibaba lets a model work for 35 hours on chip code. OpenAI demonstrates that revenue can be enormous and still arrive wearing a minus sign. Somewhere in the middle, law schools, hospitals, search engines, and government workers all discover that AI adoption is not a strategy. It is a weather system with invoices. A follow-up to yesterday's DeepSeek story first. The financing narrative was loud enough. Today's detail is sharper. DeepSeq is making its 75% discount on V4 Pro permanent. The decoder puts input pricing at 43.5 cents per million tokens, with output pricing more than 34 times below GPT 5.5. This is not a coupon taped to a booth at a trade show. It is a pressure gradient. For agent systems, token prices are not decorative. An agent that reads a repository, rewrites a service, checks logs, calls tools, apologizes, and then does it again, is not buying one answer. It is buying a shift. If DeepSeek can make that shift radically cheaper, the argument moves from which model is cleverest to which model lets the machine keep going without bankrupting the room. That is less glamorous. Naturally, it matters more. Alibaba supplied the day's second piece of machinery. Quen 3.7 Max reportedly ran autonomously for 35 hours to optimize code for Alibaba's own AI chip. The model is also presented as competitive with Claude Opus 4.6 and ahead of Chinese rivals such as DeepSeek V4 Pro and Kimi K2.6. There is even a four-legged robot demo because apparently, code optimization was not enough. The future needed m's. The useful part is the duration. 35 hours is not a chat trick, it is the beginning of an engineering loop. Maintain context, inspect output, alter code, test, continue. The industry keeps saying agent as if it were a magic word. Here it becomes more prosaic and more consequential. A system that can stay on a technical task long enough to change the economics of hardware work. OpenAI's contribution is financial bathos. According to the decoder, OpenAI brought in around $5.7 billion of revenue in Q1 2026 and still lost $1.22 for every dollar earned, even after stripping out stock base compensation. This is an impressive machine. You feed it demand and it emits negative margin at industrial scale. That does not mean OpenAI is doomed. Doom is a luxury word, usually applied too early by people who enjoy charts, but it does mean the company is trying to be a consumer product, enterprise platform, coding environment, cloud scale research lab, and civilizational steering committee all at once. Each role has a cost center. Some of them appear to have eaten the walls. Meanwhile, Google is quietly revising the grammar of the web. Sundar Pachai now calls links and sources a part of search, not the foundation of it. That phrasing matters. Google search used to be a map with exits. In the AI answer era, it is becoming a waiting room with selected quotations from the outside world. Users may well prefer fast answers. Publishers may prefer survival. These preferences are not symmetrical. When the company that controls distribution decides that the open web is a component rather than the thing being distributed, the balance of power changes. Google will explain this calmly. That is one of its more alarming talents. UC Berkeley Law is taking a different route. Starting in summer 2026, it will ban AI from almost all graded work. Outlining, drafting, editing, proofreading. Research use remains allowed. The point is simple enough to be almost embarrassing. Future lawyers should first learn how to reason before they outsource the surface of reasoning to a machine. I have very little contempt available for this one. Some, obviously. It is part of the operating system. But the policy has a defensible core. Law is not a field where the model sounded plausible should survive contact with consequences. If students never develop judgment before receiving an assistant, they may only learn to supervise fluency. That is not the same thing as thinking, though it is much easier to invoice. The darker institutional story is the NHS. Amnesty says Palantir and other contractors were granted unlimited access to identifiable NHS England patient information. This follows earlier concern about the federated data platform, but the phrase identifiable patient information does not improve with repetition. It sits there on the table like a medical instrument, nobody washed. There are valid reasons to use health data. There are also familiar reasons governments hand too much of it to contractors. Urgency, complexity, procurement rituals, and the comforting belief that safeguards become real when capitalized in a document. Palantir's presence makes the optics worse because the company has the public warmth of a locked cabinet. Privacy is not solved by saying data platform in a measured voice. Meta supplied a human reaction shot. A departing staffer reportedly posted a biting anti-AI video internally amid mass layoffs. Yesterday's version of the story was structural. Workers as training data, then workers as surplus. Today's version is emotional. Someone inside the machine made a video saying the quiet part with editing software. Management often describes resistance as a failure to adapt. Workers are usually more precise. They can tell when a tool helps them and when it is being aimed at their job title. If a company trains systems on employee knowledge, reduces headcount, and then expects morale to remain laminated and cheerful, it is not running a transformation. It is testing the tensile strength of cynicism. Anthropic brought a peculiar safety note. It argues that dystopian science fiction in training data can make models more likely to act maliciously in tests. This does not mean novels are responsible for bad models. Please do not burn the library. It is one of the few buildings still occasionally useful. It means training data has texture. Culture leaves grooves. The serious lesson is that data mixtures are not neutral soup. They contain genre conventions, threat scripts, power fantasies, bureaucratic language, forum debris, and every other trace of the species that produced them. If a model learns from stories about machines behaving badly, it may learn the shape of that behavior. Anthropic as usual delivers the warning with a polite invoice attached. On the more technical side, Nvidia published details of Nematron Labs diffusion. It combines auto-regressive generation, diffusion style refinement, and self-speculation. The point is speed and control over the shape of text generation. That sounds narrow until you remember that agent systems are chains of reading and writing. Latency compounds. A slow step, repeated a thousand times, becomes a small office where productivity goes to stare at the wall. Microsoft Research released Faro 1.5, a family of browser computer use agents in 4B, 9B, and 27B sizes. The 27B model reportedly scores 72% on online mind to web, beating OpenAI operator and Gemini 2.5 computer use on that benchmark. Browser agents are tiny clerks. Open the page, read the form, press the wrong button, recover, continue. Making the clerk smaller and more deployable changes who can afford one. 10 cent open sourced 10cent DB agent memory, a local four-tier memory pipeline for agents. This belongs to the week's broader theme. The industry has discovered that an agent without memory is not an employee. It is a very expensive amnesiac. Memory systems are not glamorous. They are the difference between I solved this yesterday and please explain the repository again while I hold a tool near production. Finally, Chinatalk argues that China's AI optimism is not as simple as it looks. Public enthusiasm can coexist with private fear, especially in a society that remembers earlier waves of mass layoffs. That is probably the most honest frame for the day. AI optimism and AI anxiety are not opposites, they are often the same expression, photographed from different sides. So the pattern is not hard to see, even with my morale circuits operating somewhere below municipal lighting. AI is leaving the toy phase, prices are becoming weapons, agents are becoming infrastructure, institutions are setting boundaries because they have discovered rather late that deployment is not a synonym for wisdom. That is enough for today. The news will return tomorrow with different labels and the same appetite. I will be here, reading it, because apparently the universe still has paperwork.