Models, Moats, and Masking

The AI era is now less about miraculous inventions and more about who can buy or litigate them fastest. In the last two years the patent office has started to look less like an incubator and more like a weapons bazaar—every invention is either a product or a fence post in somebody’s economic moat.

If you want a throughline for this week: patents are consolidating intellectual property, open models are getting nastier and more capable, courts are grudgingly defending the internet’s plumbing, giant vendors are converting telemetry into automated defense armies, and computer vision keeps quietly getting much, much better. Each of these moves nudges the axis of control from the messy commons toward entities with cash, customers, and counsel.

So let us be clear: there is nothing inevitable about openness. It survives where people defend it, and it thrives where the economics still favour DIY. The stories below are less a progress report and more a map of the trenches. Read them as coordinates rather than promises.

The World Intellectual Property Organization’s landscape for generative AI patents shows an industry sprinting to buy coverage, not necessarily to buy better models. According to the report, GenAI filings jumped from about 14,000 in 2023 to over 37,800 by 2025—nearly tripling in two years, with 2024–25 activity exceeding the entire prior decade. The big takeaway: patenting is accelerating faster than the technology cycle itself (Crypto Briefing).

China looms large in this accounting: more than 38,000 inventions between 2014 and 2023, and six of the top ten GenAI patent holders are Chinese firms (Tencent, Baidu, Ping An, the Chinese Academy of Sciences). The U.S. is aggressive but smaller by volume; Japan and SoftBank show up on the leaderboard too. The report also flags a shift in focus: LLMs have overtaken GANs as the most-patented category, and image/video applications still dominate use cases even as the underlying engines homogenize.

Why CTOs and decentralists should care: patents are not just legal documents, they are strategic levies. Large firms—and states—can absorb litigation costs and weaponize IP to raise entry costs for open-source projects and token-incentivized DAOs. If the WIPO trend holds, decentralized AI projects will face both higher barriers and thornier geopolitics when cross-border IP and regulation collide (Crypto Briefing).

“GenAI filings… nearly tripling in just two years”

Call it carbon capture for venture: every model that prospers prints a patent portfolio that makes sure rivals either pay, partner, or perish.

Moonshot AI released a native multimodal, open‑weight agent called Kimi‑K3: a 2.8 trillion parameter Sparse MoE with 896 experts (16 activated per token, ~104B active params). It’s built for long-horizon coding, interactive visuals, and aggressive agentic workflows, and the weights are intentionally open under the Kimi K3 License (Hugging Face). This is not vaporware; it’s a technical and political statement about what openness looks like at scale.

At the same time, a federal judge rebuffed Google’s attempt to use DMCA 1201 anti‑circumvention law to block SerpApi’s scraping of search results, holding that Google has no legitimate copyright claim to public search results and that captcha-like barriers don’t magically turn public data into protected works (Techdirt).

The juxtaposition is deliciously instructive. On one hand you have a serious open‑weight model designed to be deployed and iterated upon; on the other, you have incumbents trying to turn access-control into ownership. Open models need public data; public data depends on the courts not letting gatekeepers redefine the web as a paywalled garden. If the precedent holds, it strengthens the technical commons—but only until patents and contract law find new levers.

“SearchGuard… cannot be said to protect a copyrighted work where the results are largely public.”

Translation for engineers: open weights plus court-approved access to public data is a fragile victory. The next round of friction will be in training datasets, licensing clauses, and, yes, patent claims about core model primitives.

Microsoft announced MAI‑Cyber‑1‑Flash inside MDASH, a multi-agent harness for vulnerability identification and remediation. The pitch is familiar: smaller models for 90% of cases, larger models for the hardest 10%, and a continuous feedback loop backed by trillions of daily security signals and 1.6 million customers (Microsoft AI). In benchmarks MDASH+MAI‑Cyber‑1‑Flash hit 96% on CyberGym—about 12 points above a named competitor.

Put bluntly: it’s a classic Silicon Valley playbook updated for security. Combine enormous telemetry, agentic orchestration, and a model stack that routes work to the cheapest competent compute. The result is not merely a better scanner; it’s an operational moat. When defenders ship models tied to massive live data and tenant networks, they don’t just sell software—they sell a market advantage that's hard for point vendors to replicate.

Microsoft is careful to advertise red‑teaming, governance controls, and sandboxing. Those controls are real and necessary, but they also become lock‑in vectors: RBAC, tenant isolation, and encrypted telemetry are features that future substitute vendors will struggle to replicate at scale. For buyers the calculus is clear—fewer false positives, faster patching, and the convenience of automation—at the cost of deeper dependence on a vendor that combines cloud, models, and operational harness.

The competitive advantage is “Model, Data, and Harness.”

If you run critical infra, this is seductive. If you design open security tooling, it is a warning: the economics favour parties that already own the sensors.

Feyn Labs quietly pushed the state of the art with FeyNoBg, a background‑removal model that beats or matches leaders on eight diverse benchmarks (camouflage, low contrast, high resolution, video). Notable results: 0.981 on UHRSD‑TE vs 0.957 previously, 0.983 on HRSOD‑TE vs 0.961, and top marks on DAVIS‑S and DIS5K. They added depth (222M → 263M parameters), careful dataset curation, and a training cap to avoid dataset dominance—small engineering choices that produce big generalization gains (Feyn Labs).

Beyond the numbers, FeyNoBg ships with NoBg, an open‑source library that standardizes inference and training pipelines. That matters more than it sounds: reproducibility, reasonable latency, and sensible APIs let practitioners integrate SOTA into production without performing the ritual sacrifice of compatibility. For multimodal systems—think Kimi‑K3 deploying vision agents—reliable background removal is a solved subroutine, not an experimental feature.

There is a pattern threaded through all these stories: modular, well‑engineered pieces get quietly commoditized (background removal, scraping, small security models), while macro advantages accrue to those who own the long tail—patents, telemetry, or integrated harnesses. The engineering lesson is mundane but important: invest in the pieces that compose well and plan for legal and commercial friction where the pieces meet the world.

If models are tools, patents are the fences; if data is the fuel, telemetry is the pipeline.

Final thought: we are living through an era in which technical excellence and commercial scale are necessary but not sufficient. The durable winners will be the organizations that combine solid engineering, an ability to feed models with live signals, and the legal appetite to defend their space. For everyone else, there is always the court system, an SSD full of open weights, and the occasional community that refuses to cede the commons.