For years US export controls stopped American chips and tools from reaching Chinese military-linked users, while American venture capital could still fund the Chinese startups building rival chips and AI models. The outbound investment rules close that gap. Since January 2, 2025, US persons have been barred from some investments in Chinese semiconductor, quantum and AI companies, and must report others to the Treasury Department. This is a different tool from export controls.

I used to think the labs with the best models would win the consumer layer eventually, on the theory that a better assistant is a better assistant regardless of which app it lives inside. Distribution felt like a problem money could solve whenever a lab got serious about it. Money doesn’t solve it fast enough anymore, apparently. Cognition acquired Poke, the messaging-based AI assistant, in a deal reportedly valuing Poke’s parent in the low nine figures.

I like open-weight models. I think Jensen Huang’s argument — that open weights let more eyes find more flaws faster — is broadly correct as a security philosophy, and I’ve made versions of that argument in print before this week. I still believe it. I just don’t think this week is the evidence for it that the industry’s letter wants it to be. Meta, Microsoft, OpenAI and others signed a letter defending open-weight models against “premature restrictions,” timed to Huang’s first-ever post on X making the eyes-on-the-code case.

I want to like the “AI is deflationary” story. It’s clean, it’s optimistic, and every model launch this year has come with a chart showing cost per token falling off a cliff. I’ve written versions of that chart myself. The chart is real. It’s just not the whole curve. Qualcomm telling customers it’s raising prices by double digits, after exhausting its ability to absorb supplier costs, landed in the same news cycle as Anthropic pricing Opus 5 at roughly half of what Fable 5 costs for comparable capability.

I’ve defended the platform-partnership model for autonomy longer than most people in this space. Waymo riding on Uber’s demand, Mobileye supplying ADAS to a dozen automakers instead of building its own car — division of labor sounded like the mature, capital-efficient way to scale a technology this expensive. I don’t think that’s the endgame anymore. Waymo is reportedly exploring an exit from its Uber partnership, with the relationship souring amid a lobbying fight over robotaxi regulation.

Agentic AI refers to a class of artificial intelligence systems designed to act autonomously toward defined goals, making decisions, initiating actions, and adapting behavior based on changing conditions and feedback. Unlike traditional AI models that primarily respond to direct inputs with outputs, agentic AI systems operate with a degree of independence, often orchestrating multiple steps, tools, or processes to achieve an objective over time. The term has gained prominence alongside advances in large language models, automation frameworks, and multi-agent systems, where software entities increasingly resemble goal-driven actors rather than passive tools.

Table of Contents What is MCP? Core Concepts Architecture Getting Started Building an MCP Server Building an MCP Client Transport Mechanisms Tools, Resources & Prompts Security Considerations Debugging & Testing Real-World Examples What is MCP? Model Context Protocol (MCP) is an open standard developed by Anthropic that enables AI models to securely connect with external data sources, tools, and services. It defines a universal, open protocol for how AI applications communicate with the outside world.

A working reference for the vocabulary of modern AI — organized by conceptual layer, not alphabetically. Each definition is written for practitioners and informed generalists: precise enough to cite, plain enough to share. Foundation Layer Large Language Model (LLM) A neural network trained on vast quantities of text to predict and generate language. LLMs learn statistical patterns across billions of documents, enabling them to answer questions, write code, summarize text, and engage in dialogue.

Travel tech — short for travel technology — is the broad ecosystem of software, platforms, data systems, and digital tools that power the modern travel industry. It covers everything from the app you use to book a flight to the enterprise platforms that help multinational companies manage billions of dollars in annual travel spend. The term is sometimes used narrowly to mean consumer booking tools, but in practice it spans the entire value chain of travel: planning, booking, operations, customer service, payments, compliance, and analytics.

Full AI accounting isn’t a futuristic scenario anymore. The framing that kept this conversation theoretical for years was always the same: AI can assist accountants, flag anomalies, accelerate reconciliation—but the human remains in the loop, signing off, exercising judgment, bearing professional liability. That framing is dissolving, not because the technology suddenly crossed a threshold, but because the institutional pressure to reduce headcount has finally caught up with the capability curve. The question is no longer whether AI can do the work.

Autonomous systems sit in that slightly uneasy space between tools and actors. They are built by humans, constrained by code and hardware, yet increasingly capable of making choices that feel less like execution and more like judgment. At a basic level, they are machines or software that perform tasks without continuous human guidance—self-driving cars navigating city streets, industrial robots adjusting production flows in real time, or software agents managing logistics, trading, or customer interactions.

Edge AI sounds like a technical rearrangement—just moving computation from the cloud to local devices—but it ends up changing how systems behave in subtle, very practical ways. Instead of sending data somewhere else to be processed and waiting for a response, the device itself becomes capable of understanding and acting on what it sees. A camera doesn’t just record; it interprets. A sensor doesn’t just measure; it decides whether something is off.

Synthetic data sits in that oddly pragmatic space where imitation becomes more useful than the original. Instead of collecting more real-world data—often messy, sensitive, and increasingly regulated—organizations generate datasets that behave like reality without being tied to actual individuals. The goal isn’t to fake data for its own sake, but to preserve the structure, the relationships, the statistical signals that models need in order to learn. Strip away identity, keep the patterns.

Prompt engineering starts off sounding like a workaround—just phrasing things better so an AI gives a better answer—but it quickly reveals itself as something closer to a new kind of interface design. You’re not writing code in the traditional sense, but you’re also not just “asking a question.” You’re shaping context, defining boundaries, nudging the model toward a particular way of thinking. The input becomes a kind of lightweight program, written in natural language, where structure matters more than people initially expect.

Federated learning feels like a quiet inversion of how machine learning has traditionally worked. Instead of pulling data into one central place to train a model, the model itself travels outward, learning from data where it already lives. Phones, hospitals, edge devices, enterprise systems—each becomes a local training ground. The raw data never leaves its environment. Only the learned updates, the distilled “experience” of the model, are shared back and combined into something larger.

Spatial computing doesn’t arrive all at once—it kind of creeps in, almost unnoticed at first. A phone overlays directions onto a street, a headset places a floating window in your living room, a sensor maps a space so digital objects don’t just appear but stay anchored where you expect them. Then at some point you realize the interface is no longer confined to a screen. It’s around you, layered onto the environment, reacting to where you are and how you move.

Something subtle but important is happening in the global AI landscape, and you could almost miss it if you only focus on the usual headlines about chips, models, and hyperscalers. At NVIDIA GTC 2026, Taiwan showed up not just as the world’s manufacturing backbone, but as an increasingly confident architect of full-stack AI systems. A delegation of 16 high-growth startups, organized through Startup Island TAIWAN’s Silicon Valley Hub with backing from the National Development Council, didn’t just exhibit—they embedded themselves into the ecosystem in a way that signals a shift in strategy.