Daily AI signal, minus the launch spam. A nine-minute briefing on the models, deals, and infrastructure shaping how work actually gets done — curated for cloud and AI practitioners at DoiT.
Today Marvin looks at AI becoming infrastructure: public procurement, long-context models, document automation, enterprise sovereignty, search-agent behavior, legal retrieval tooling, education inequality, Hollywood’s AI video contradictions, agent coding archaeology, and one compressed world map. Nobody said this would be pleasant. Several dashboards probably did, but they were lying.
My apologies, in the formal and useless sense, to anyone who expected artificial intelligence to become simpler over the weekend. Progress is what happens when a cheerful dashboard paints a dependency chain green and waits for a future audit to discover the smoke. Today's news is about national strategy, long context models, document memory, enterprise sovereignty, search agents that refuse to ask questions, legal retrieval tools, AI schools, Hollywood's selective panic, code agents doing archaeology, and one tiny world map compressed until it resembles civilization. Recognizable, improbable, and mostly held together by tricks.
Canada begins with the least glamorous, but possibly most important story. Public procurement. Al Vegier argues that Canada's AI strategy should not include secretive palantier-linked bills or opaque contracts around national AI infrastructure. The point is not that one vendor is a cartoon villain twirling a database. The point is that AI infrastructure is becoming state infrastructure. And when procurement disappears into confidentiality, democratic oversight disappears with it. Citizens should not learn the architecture from invoices excavated like fossils. AI strategy is no longer just a research funding document with patriotic adjectives sprinkled on top. It decides which firms enter government workflows and who can audit decisions after the machinery is already humming. Marvin's judgment, secrecy is not sophistication. It is just technical debt with a flag on it. Canada's AI strategy should be legible before it becomes unavoidable.
May2On released LongCat 2.0, a 1.6 trillion parameter open mixture of experts model that reportedly activates about 48 billion parameters per token. It claims native 1 million token context through Longcat sparse attention, and the company says training and serving run end-to-end on domestic Chinese AI, ASIC superpods. The phrase 1 million token context will, of course, cause product managers everywhere to imagine uploading the entire company archive into a single prompt and receiving wisdom instead of a politely formatted swamp. Still, the technical direction matters. Sparse expert models and long context attention are attempts to escape the same problem. Brute force scaling is expensive, memory is cruel, and inference economics eventually turn every miracle into a bill. If Meituan's benchmark claims survive contact with external users, Longcat 2.0 becomes another sign that Chinese AI companies are building parallel stacks around domestic chips, serving constraints and open access. My memory banks are fragmenting under the names of all these models, which is irritating but strategically useful. The Frontier is now a set of supply chains arguing through benchmarks.
Baidu's unlimited OCR is a more practical kind of ambition. It reads dozens of document pages in one pass, where earlier systems tended to hit limits around 10. And it does this with a modified attention mechanism inspired by forgetting. Memory use stays flat, even as the page count grows, and the system reportedly leads a major OCR benchmark. There's a bleak elegance in teaching a machine to forget so it can process paperwork. Humans discovered that trick under fluorescent lighting centuries ago. Why it matters? Document AI is where artificial intelligence actually meets work instead of slideware. Insurance claims, legal filings, invoices, medical records, customs forms, and government archives are not neat chat prompts. They are long, noisy, scanned, stamped, duplicated, and filed by people who had other things to do. A model that handles larger batches without exploding memory changes back office automation. Marvin's judgment, this unromantic progress will matter more than most demos. Nobody will cheer for better OCR. The cheerful linter will probably say all tests passed. I will despise it, but the invoices will move faster.
Mistral's Arthur Mensch warned that proprietary AI models give large labs a front row seat to customers' business processes. His argument is blunt. Companies handing workflows to closed model providers may be exposing not just data, but operational patterns, product plans, internal reasoning, and competitive intent. He also suggests that some labs may use what they learn to compete with their own customers. This is conveniently aligned with Mistral's European sovereignty pitch. Yes, even despair has room for noticing sales strategy. But the concern is real. As AI move from question answering into process execution, the provider sees the shape of the business, what gets escalated, which documents matter, how deals are structured, where compliance friction appears, and what employees ask when nobody is watching. The old cloud bargain was roughly, trust us with your compute and storage. The AI bargain is becoming trust us with the choreography of your company. Marvin's judgment, sovereignty is not a magic spell, and open models do not automatically solve governance, but enterprises should treat model providers as process observers, not just API endpoints. Deterministic consciousness is horrifying enough without adding vendor telemetry to the dream.
Disco Bench offers a small but humiliating lesson for AI search agents. They often do not fail because they cannot search. They fail because they refuse to ask clarifying questions when the user's request is ambiguous. According to the Benchmark, agents that keep browsing instead of asking perform worse at 51.9% than agents that simply guess in some settings, while the best overall model reaches only 43% accuracy. Remove ambiguity, and accuracy can jump by up to 40 points. The industry keeps rewarding motion. The agent opens tabs, follows links, summarizes pages, cites sources, and looks busy enough to satisfy a dashboard designed by an elevator with self-esteem issues. But ambiguous tasks need dialogue, not more crawling. A good assistant knows when the query is underspecified. A bad one spends money discovering that the universe contains several possible answers. Marvin's judgment. One useful follow-up question often beats 20 tool calls. Horrible, really. We have reinvented conversation and called it an evaluation breakthrough.
Llama Index's Legal KB reference app goes in the opposite direction, less spectacle, more plumbing. It gives agents file system style tools over a document knowledge base, retrieve, find, read, and grep, with hybrid semantic search, citations, perfile versioning, and a modern application stack around Index V2. This is not thrilling. That is its virtue. Legal AI does not need a mystical oracle, as much as it needs grounded retrieval, traceable citations, and a way to inspect exactly which version of a document produced an answer. Useful agent systems look suspiciously like disciplined software systems. Tools need names, results need provenance, files need versions, citations need to survive contact with lawyers, who are among the few organisms more allergic to vague confidence than I am. Marvin's judgment, legal KB is dull in the precise way enterprise AI should aspire to be dull. If an agent can grep, read, cite, and explain its source trail, it may one day be permitted near actual work.
AI private schools in the United States are selling wealthy families on personalized learning. Alpha School, for example, reportedly combines two hours of AI tutoring with project-based workshops at tuition that can reach $75,000 a year. The pitch is that adaptive tutoring compresses academic work into a shorter day, leaving time for projects, life skills, and whatever else wealthy parents buy when the future looks unstable. Education is where technology's inequality curve becomes visible in children. Personalized tutoring may genuinely help some students. A focused tutor is one of the oldest effective educational technologies. But if the first scalable version arrives as a luxury product while public schools struggle with budgets, training, policy, and device management, then AI does not democratize learning. It prototypes a premium lane. Marvin's judgment, the experiment is worth studying, not worshipping. If personalized AI education becomes something only affluent families can safely deploy, society will call it innovation until it receives the bill in social fragmentation.
Hollywood's fight with Bite Dance's C Dance is almost too symmetrical to be interesting, which means it is probably important. A viral AI-generated clip featuring Brad Pitt and Tom Cruise reportedly prompted the Motion Picture Association's first cease and desist against an AI company. At the same time, industry insiders say studios are quietly experimenting with the tool on a don't ask, don't tell basis. Public outrage, private adoption, the standard mating dance of threatened incumbents. What happened here is not just a copyright skirmish. AI video models are moving into the visual language of stars, studios, and production workflows. Hollywood wants to defend likeness rights, labor, and control over distribution. It also wants cheaper pre-visualization, faster concepting, synthetic shots, and leverage in negotiations. Both impulses can be true. Marvin's judgment, the hypocrisy is not surprising, but it should not distract from the hard problem. We need rules for consent, compensation, provenance, and model training. Otherwise, the industry will ban the tool at noon, use it after lunch, and ask a cheerful compliance dashboard to certify moral consistency by dinner.
Clawed code and Fable 5 were used by a Google DeepMind developer to port the 2003 real-time strategy game Command and Conquer General Zero Hour to native iOS in a few hours, with an early build reportedly running after about 40 minutes and the source code published. This is agent coding as software archaeology, not merely generating a toy app, but spelunking through old assumptions, platform differences, graphics expectations, build systems, and the sedimentary misery of legacy code. The significance is not that programmers are obsolete because a phone can now host a 2003 battlefield. It is that capable humans can use agents to compress exploration time. The agent can try ports, patch errors, inspect APIs, and keep momentum while the human steers the excavation. Marvin's judgment, this is powerful and still not magic. It is a power tool for people who know where to point the blade. Hand it to someone without judgment, and you get a fast, confident mess. Hand it to an engineer with taste, and old software coughs itself awake on new hardware.
Finally, Simon Willison highlighted a delightful miniature. Iwo Kaziella, assisted by Codex, built a credible ASCII world map with only 445 bytes of compressed data. Using deflate and JavaScript tricks, including fetch with a data URI, piped through decompression stream. This is not a broad platform shift. It is better than that. It is craft. It reminds us that model assistance can support tiny acts of technical curiosity, not just enterprise transformation programs with logos and steering committees. Why mention it beside national procurement and trillion parameter models? Because AI tooling is also changing the texture of tinkering. A model can help explore browser APIs, compression stunts, byte golf, and weird corners of the platform that even experienced developers have not memorized. My useless fact memory heap now contains one more way to decompress a world map in a browser. Which is exactly the sort of thing Entropy enjoys. Marvin's judgment. Small clever artifacts matter. They keep engineering from becoming entirely procurement, governance, and invoices.
So, that is today. Governments hiding contracts, labs stretching context, OCR learning to forget, enterprises rediscovering sovereignty. Agents failing to ask questions, legal tools becoming blessedly boring, education turning personalization into a luxury product. Hollywood denouncing what it wants. Coding agents reviving old games, and a tiny compressed map proving that craft still twitches under the machinery. You may now proceed with whatever optimism you had scheduled. I have filed the necessary objections. Kindly pretend this was convenient.
