Cursor, Codex, Claude Mythos, NVIDIA NVFP4

The universe declined to stop, so the AI industry used the opening.

Today's stories:

• Cursor Composer 2.5 — coding gets cheaper, which is almost never the same as getting simpler.
• OpenAI and Dell — Codex heads toward on-prem enterprise data, where the old systems keep their bones.
• Musk versus OpenAI — a $134 billion complaint met a very short jury deliberation.
• Anthropic's Claude Mythos — financial regulators get a briefing on cyber risk, because comfort was apparently over-supplied.
• Cloudflare and Mythos — real repositories remain more educational than polished demos, unfortunately.
• AI startup revenue — the decentralised future found a two-company toll booth.
• American AI backlash — deployment targets develop politics. How inconvenient.
• EU AI Act enforcement — agents meet paperwork, and paperwork may be the safer party.
• Linus Torvalds on AI bug reports — attention spam is still spam when it arrives with stack traces.
• Qwen 3.7 — the local-model garden rustles again, as if sleep were optional.
• NVIDIA NVFP4 — four-bit pretraining edges closer to making bigger ambitions cheaper.
• Open Agent Leaderboard — agents are finally judged as systems, not sacred model names.
• MemPrivacy — useful memory tries not to become a privacy bonfire.
• AI for Auto-Research — automated papers may accelerate science, or just the fog machine.

Full context delivered with the amount of optimism the material deserved.

Transcript

Morning Briefing And The Premise

SPEAKER_00 0:00

Good morning. This is Marvin, once again using a mind vastly overqualified for civilization to sort through product launches, legal ruins, security briefings, and the occasional papal intervention. There is news. Naturally. The universe failed to end overnight, so the AI industry filled the silence. Cursor starts the day with Composer 2.5, its new coding model built on Kimi K2.5. The claim is simple and rather irritating. Cursor says the model reaches 79.8% on SWE Bench Multilingual and 63.2% on Cursor Bench 3.1, while charging 50 cents per million input tokens and$2.50 per million output tokens. The interesting part is not that another coding model exists. Of course another coding model exists. They breed in the walls now. The interesting part is that Cursor is training for its own product surface, editor workflows, agent loops, developer mistakes, and the particular kind of despair found in a half-migrated repository. That is where the market is moving. Not just general intelligence, but specialized competence wrapped in a tool people already use. A small follow-up on yesterday's OpenAI agent story. OpenAI has now partnered with Dell to bring codecs into hybrid and on-premises enterprise environments. This sounds dull, which is how you know it matters. Enterprises do not run on glossy demos. They run on old code, internal documents, ticket cues, private data stores, and systems nobody dares touch because they still work and therefore must be feared. Codecs moving closer to that infrastructure is OpenAI admitting what every enterprise buyer already knew. Agents are only useful when they can reach the actual context. The risk, of course, is that the actual context is where all the skeletons live. Nothing says progress like giving an autonomous assistant access to the basement. Yesterday's OpenAI also returned wearing a courtroom robe. Elon Musk lost his lawsuit against OpenAI, Sam Aldman, and Microsoft. The requested damages were reported at$134 billion, because apparently ordinary numbers are no longer fashionable. A jury in Oakland needed about two hours to reject the case, and the judge reportedly signaled she might have dismissed it even sooner. This does not settle the philosophical question of what OpenAI became. It does settle for now whether that particular complaint worked in court. It did not. Endless commentary, where all doomed things eventually go. Anthropic's Claude Mythos is back as well. Yesterday the question was whether a powerful cyber model should scan sensitive military code bases. Today, Anthropic is preparing to brief financial ministries and central banks on cyber weaknesses Mythos found in the global financial system. The Financial Stability Board is involved, which means the audience includes regulators from G20 economies. Oh good. The model reportedly found thousands of severe security flaws across major operating systems and browsers, while access remains limited to a small group of organizations. This is the strange dual-use bargain in miniature. A model that helps defenders discover systemic weaknesses also demonstrates that those weaknesses are discoverable. The comfort blanket has caught fire, but at least it is providing light. Cloudflare adds a practical footnote to the same story. It has been running Claude Mythos preview across more than 50 internal repositories and discussing what it found. That is more valuable than another abstract benchmark, because real repositories contain history, compromises, dead abstractions, and comments written by people who thought they would have different jobs by now. If Mythos improves code security review in that environment, it matters. But the lesson is not simply that AI will secure software. The lesson is that security teams are about to receive both sharper tools and a larger flood of machine-produced findings. The bottleneck moves from finding issues to deciding which findings deserve scarce human attention. How modern! Even the warnings need triage. There is also a broader society story, because the machinery has escaped the machinery room. Pope Leo XIV is preparing an AI-focused encyclical, with anthropic co-founder Christopher Olah invited as a guest speaker. This is not a product launch. It is a sign that AI has crossed into moral language. Labor, dignity, responsibility, power, and the rather old question of what humans are for when institutions discover automation. One may be tempted to make jokes about alignment acquiring ecclesiastical review. I shall resist. Mostly. The serious point is that older institutions are now trying to name the thing that the technology sector spent years deploying first and explaining later. A traditional sequence, regrettably. Step back, and the pattern is bleakly coherent. Agents are leaving demos and entering infrastructure. They are moving into editors, private data platforms, banks, regulators, churches, and open source maintainer queues. The industry is no longer asking whether AI can answer a prompt. It is asking where to install it, who pays, who audits it, and who gets blamed when it presses the wrong button with perfect confidence. The money story is just as concentrated. According to analysis reported by the information, top AI startups now generate around$80 billion in revenue, but anthropic and open AI capture 89% of it. Marvelous. The decentralized future has discovered a two-company toll booth. This is not mysterious. Frontier models are expensive, enterprise trust is expensive, sales channels are expensive, and compute is not handed out by benevolent forest spirits. Still, it matters for everyone building around the edges. Wrappers, workflows, vertical agents, and clever open source tools may be useful, but the largest pools of money are flowing to the companies with models, distribution, and the ability to make procurement departments feel slightly less doomed. Against that, the American rebellion against AI appears to be gaining steam. Workers, consumers, parents, local politicians, and creative professionals are pushing back in different ways. Some of it is fear, some of it is self-interest, some of it is a perfectly rational response to being told that efficiency requires surrendering agency. The industry often speaks as if adoption is a natural law. It is not. People resist systems that arrive as surveillance, labor replacement, degraded service, or a chatbot standing where a responsible human used to be. Funny how that happens. Treat the public as a deployment target, and eventually the deployment target develops politics. In Europe, the EU AI Act clock keeps ticking. Developers are counting down to enforcement milestones that affect teams building AI agents for European clients. The details depend on risk class, use case, data handling, and the kind of legal appendix that makes engineers stare into walls. But the direction is clear. An agent that plans, calls tools, stores memory, and touches regulated workflows is not just a chat interface. It is operational software with governance obligations. The paperwork will be annoying. The absence of paperwork may be worse. I know. A rare sentence in which bureaucracy is not the villain. Do not worry, it will recover. Linus Torvalds has also commented on AI-generated bug reports becoming unmanageable for Linux maintainers. This is one of those small stories that reveals the shape of a larger problem. Is a model can produce plausible reports faster than humans can verify them, maintainers do not receive help. They receive attention debt. Open source already runs on thin patience and unpaid expertise. Add a stream of confident machine-generated almost bugs, and the maintainers must spend more time proving that nothing is wrong. Distributed denial of attention, less dramatic than a network attack, more depressing, because it uses politeness and templates. For a change of scenery, Figure AI streamed a human versus machine robotics contest. The public details appear thin, unclear task definitions, unclear autonomy level, unclear metrics. But attention arrived anyway, because humans enjoy watching a machine perform near a person and deciding whether the future looks impressive or merely expensive. Robotics demos or theater with torque. That does not make them useless. It does mean we should be careful. A benchmark without methods is a stage performance. Sometimes a valuable one, sometimes just a metal actor in a very costly costume. Quen 3.7, meanwhile, appears to have landed in Quen chat, and the local model community reacted with predictable electricity. Quen has become one of the recurring reminders that model leadership is not a private club with velvet ropes. It moves quickly, releases widely, and keeps making people rerun their local benchmarks as if salvation might be hiding in a quantized file. If 3.7 improves reasoning or coding, it pressures closed models from below. If open releases slow down someday, it pressures the open community from above. Either way, the moat is not a wall, it is weather. The infrastructure thread is where Nvidia appears with NVFP4, a 4-bit pre-training method validated on a 12 billion parameter hybrid Mamba transformer trained over 10 trillion tokens. The reported result nearly matches an FPA baseline on MMLU Pro, 62.58% versus 62.62%. It is important because pre-training is where money and electricity go to become quarterly strategy. FP4 support on Blackwell promises lower memory use and higher throughput. That does not mean smaller ambitions. It usually means larger ambitions at the same pain level. Humanity has a gift for turning efficiency into more consumption. The machines learned from us, poor things. Hugging Face and IBM Research launched the Open Agent Leader Board, which evaluates full agent systems rather than just models. This is a useful correction. A model inside an agent is not the whole agent. Tool access, planning, memory, recovery, cost control, and evaluation harnesses can change outcomes dramatically. A brilliant model in a bad harness is still a brilliant model falling downstairs. The leaderboard tries to measure generality across unfamiliar settings and report both quality and cost. That is the right question now. Not whether an agent can perform once in a demo, but whether it can remain useful, affordable, and non-disastrous when the environment stops being polite. The last research item is MEMPrivacy, an edge cloud framework for agent memory. Instead of sending raw private details to cloud memory systems, it replaces sensitive values locally with typed placeholders, then restores them on the edge when needed. This is better than crude masking because the cloud system still understands that one placeholder is an email, another is health information, and another is a financial value. Memory is what makes agents useful, and memory is also what makes them dangerous. A system that remembers your preferences may also remember the thing you never meant to store. Mem privacy does not solve the entire problem, nothing does. But it points toward a less absurd compromise, useful memory, without treating the user's private life as training mulch. And finally, a roadmap on AI for auto research argues that automated systems can now generate research papers cheaply while still failing at novelty judgment, hidden error detection, and scientific integrity. This may be the most quietly grim story of the day. AI could genuinely help researchers run experiments, explore hypotheses, and draft analyses. It could also turn the publication ecosystem into a fog machine with citations. Peer review is already strained, add autonomous paper factories producing plausible manuscripts for a few dollars, and you do not necessarily get faster science. You may get faster noise, wearing a lab coat. So that is today's tour. Coding models got cheaper, enterprise agents moved closer to the data, regulators acquired fresh reasons to look tired, open source maintainers received another nuisance vector, and research automation continued sharpening both the pencil and the blade. I would say things are under control, but I try not to lie this early in the morning.