Hacker News Daily: The Best Threads and Boldest Ideas in Tech

Google just open-sourced AX, a system designed to run billions of AI agents at once. Now, last week we were talking about AI breaking RSA encryption—that kind of raw, adversarial power. This week, we're seeing the complete other side of the coin: the architecture, the plumbing, the un-glamorous but absolutely critical infrastructure for managing AI at a scale we haven't even fully imagined yet. AX isn't about one agent doing one smart thing; it's about orchestrating a whole civilization of them without letting them burn the house down. And that shift, from a single powerful model to a managed population of agents, is where everything is heading. Okay, so let's do the sweep of the other big threads before we dive back into that.

First up, in the hardware world that powers all this AI, Samsung just announced it's planning to more than double its production of next-generation HBM memory in 2027. We're talking about HBM4 and even HBM4E—the stuff that will be in the GPUs of tomorrow. The interesting part here isn't just "more chips." It's the bottleneck they're solving. The plan involves a massive 2.5 times increase in their use of something called glass carriers. Think of these like hyper-precise, perfectly flat glass trays that hold the silicon wafers while they're being thinned down and stacked into the 3D structure that makes HBM so fast. Without enough of these glass trays, the whole AI hardware pipeline grinds to a halt.

So when you see Samsung betting this big on a single supply-chain component, you know the demand for AI compute isn't just hype; it's reshaping global manufacturing priorities. Then we have a story that’s… well, it’s a bit of a heartbreaker, especially if you've ever spent time in a machine shop. Sherline Tools, the American company that has been making these beautiful, precise little lathes and CNC machines for decades, is shutting down manufacturing. They're citing the usual suspects: rising costs, the pandemic's impact, and just… changing habits. Production is ending by the end of this year. For the maker community, this is like hearing that your favorite indie band is breaking up.

Sherline wasn't about massive industrial scale; it was about giving individual craftspeople the ability to make incredibly precise things on their own workbench. It’s a sad end of an era for a certain kind of American manufacturing, and it stands in such stark contrast to the massive scale-up we're seeing from a company like Samsung. From the end of one era to the quiet closing of another. An analysis this week confirmed something that's been sort of an open secret: no new documents from the Snowden archive have been published since May 29th, 2019. That’s over seven years ago. The major news outlets that broke the story—The Guardian, The Intercept—they all formally ended their projects years ago.

The Intercept's archive even went dark back in March of 2019. The reasons are complicated… budget constraints, editorial priorities, the documents getting old. But the bottom line is, the firehose of revelations about global surveillance that defined so much of the last decade has, for all intents and purposes, run dry. It’s a strange, quiet end to one of the loudest stories of the 21st century. It just sort of… faded out. And finally, for a completely different kind of story about changing habits. Singapore’s National Library Board has launched a five-year pilot program to literally pay people to read. I know. It's a system of micropayments designed to build daily reading habits in a population that, like everywhere else, is glued to their phones.

It's this fascinating experiment in behavioral economics. Can you use the same kind of incentive loops that make apps addictive, but to encourage something like deep reading? Most of the time we talk about tech causing a problem; here's a government trying to use a tech-like solution—micropayments—to solve it. We'll have to see if it works, but it's one of the most direct attempts I've ever seen to fight for our attention spans. So what does it all add up to? You have the end of a manufacturing icon, the end of a transparency era, and the beginning of a whole new infrastructure for AI. It feels like a week of handoffs—passing the baton from one age to the next. Okay, so let's dive deeper into two of these, because the patterns are just… they're perfect.

Let's start with the AI agents. On one hand, you have this blog post making the rounds on Hacker News declaring that a protocol from Anthropic called MCP is now basically obsolete. And on the other, you have Google releasing AX. These two stories are two sides of the exact same coin. So, let’s rewind a bit. MCP, the Multi-Channel Protocol, came out in late 2024. At the time, it was a really clever idea. AI agents needed a way to talk to the outside world—to use tools, to call APIs, to get information. But the Large Language Models themselves weren't quite smart enough to handle all that complexity on their own. They'd hallucinate API calls, they'd mess up the authentication. So MCP was created as a kind of… structured middleware.

A universal translator. It gave the agent a safe, standardized way to say "I need to use the weather tool with this location" and the MCP server would handle the messy details of making the actual API call and getting the data back. It was a crutch, but a necessary one. The problem is, the models got better. A LOT better. And way faster than anyone expected. As the author of that blog post put it, "They are now able to execute code on a computer, reason about large codebases, and generally act much more autonomously than ever before." Suddenly, the crutch wasn't just unnecessary, it was getting in the way. Forcing everything through this MCP translation layer added complexity and bloated the context window you had to feed the model.

Why bother with a translator when the model can now speak the language of APIs fluently? We're seeing things like Cloudflare's Code Mode, where the LLM doesn't just ask for a tool—it literally writes and composes the script to call multiple APIs itself, in a sandbox. The model is becoming the programmer. And this is a pattern we have seen over and over and over in technology. Remember CORBA? Or SOAP? These were complex, rigid protocols for getting different systems to talk to each other. They were a whole industry. And then REST and simple JSON over HTTP came along and just… ate their lunch. Because it turned out, if you make the endpoints a little bit smarter, you don't need this incredibly heavy, complicated middleware.

The intelligence moved closer to the edge, and the "protocol" became simpler. MCP is having its SOAP moment. It was a brilliant solution to a temporary problem. The problem was just way more temporary than its creators imagined. And THAT is why Google's AX is so important. AX, or Agent-X, isn't another protocol for agents to talk to each other. It's an orchestrator. It's the operating system for the agent data center. Google's documentation says it best: "Agents are a new kind of workload. They are neither microservices nor batch jobs." Think about that. A microservice just sits there waiting for a request. A batch job runs, does its thing, and dies. An agent… it lives. It has state. It learns.

It can get stuck in a loop and burn a million dollars in API calls if you're not watching it. So AX provides the guardrails. It has four main ideas, or primitives. A 'Task' is the agent's actual code, running in a strict sandbox so it can't mess with anything else. A 'Workspace' is its little environment—the files it needs, the tools it has access to. A 'Gateway' is its network connection, with strict rules about what servers it's allowed to talk to. No rogue agents calling random APIs. And a 'Model' just defines which generative AI model it's allowed to use. What this lets you do is run billions of these agent sessions on a single cluster. It's designed for dense, multiplexed execution with sub-second resumption time.

That means an agent can be paused, its entire state saved, and then brought back to life almost instantly. So while the community is busy declaring a two-year-old protocol obsolete because the models got too smart for it, Google is building the city planner for the civilization of super-smart agents to come. It's not about how they talk anymore. It's about how they live together, at scale, without causing chaos. It’s the shift from inventing a language to building a government. Now, let's switch gears completely. From the purely digital to the intensely physical. The story of Sherline Tools closing down alongside the news of Samsung’s massive HBM memory expansion. On the surface, these are unrelated.

One is a small, beloved American manufacturer of machine tools going out of business. The other is a South Korean tech titan doubling down on the future of AI hardware. But if you put them side-by-side, you get this incredibly sharp picture of the tectonic shifts happening in manufacturing and value creation. Sherline is… or was… something special. For almost fifty years, they've made these miniature lathes and milling machines. They were famous in the hobbyist and model-making communities. If you wanted to build a tiny, perfect steam engine, or machine a custom part for a robot, or make jewelry, a Sherline was often your first choice. They were made in Vista, California. They represented this ideal of accessible, high-precision, domestic manufacturing.

The value wasn't just in the tool itself, but in the skill it enabled in its owner. It was a force multiplier for individual creativity. You bought a Sherline to make things. And now they're closing. The official statement talks about rising costs and changing consumer habits. And you can read between the lines. It's hard to be a small manufacturer in the US. Global supply chains, labor costs, and a market that increasingly favors cheap imports over long-lasting, repairable tools. It's a story we've seen before, but it hurts every time, especially for a company with such a dedicated following. It feels like losing a piece of tangible, physical capability. Then, on the same day, you read about Samsung.

They're not just making more chips. They're planning to consume 50,000 sheets of hyper-specialized glass carriers per month in 2027. These aren't just panes of glass. They are perfectly flat, thermally stable substrates that are essential for a process called "wafer thinning." To make HBM memory, you have to stack multiple DRAM dies on top of each other. To do that, you have to grind the silicon wafers down until they're incredibly thin. The glass carrier is what supports the wafer during this violent, high-precision grinding and drilling process. If the glass isn't perfect, the multi-billion dollar wafer is ruined. Look at the contrast. Sherline made tools that empowered a single person to turn a block of aluminum into a precise object.

The value was in the craft. Samsung is scaling up a supply chain for a component that empowers a process—a process so complex and automated that very few single humans understand it end-to-end. The value isn't in an individual's hands; it's embedded in the system, in the capital expenditure, in the robotic precision of the factory. One is about the craft of the atom. The other is about the industrial-scale manipulation of the atom to serve the bit. The pattern here isn't as simple as "America loses, Asia wins." It's a deeper shift. The kind of value Sherline represented—durable, repairable, empowering tools for individual makers—is becoming a niche, a luxury. The kind of value Samsung represents—unimaginably complex, high-volume, disposable-by-design components for the global tech ecosystem—is where the economic gravity is.

We're not just losing a company in Sherline. We're seeing a tangible demonstration of which kind of making the world is choosing to prioritize. And right now, it's prioritizing the invisible, impossibly complex components that power our digital lives over the tangible tools that let us shape our physical one. So this week sets up a fascinating, and maybe a little unsettling, foundation for what's next. We're watching the scaffolding of the recent past get torn down at an incredible speed. A protocol like MCP, which seemed essential just twenty-four months ago, is now legacy code. A manufacturing company that represented a fifty-year tradition of American craft is becoming a memory. And in their place, the next layer of abstraction is being poured.

It's not just a slightly better version of what we had. It's a system like Google's AX, built to manage a type of computing that barely existed a few years ago. It’s a supply chain for components like glass carriers, being scaled up to meet a demand that is completely unprecedented. The claim I'm willing to make is this: the next year is going to be defined by this process of radical replacement. We’re going to see more foundational assumptions—about software architecture, about hardware bottlenecks, about what's valuable to manufacture—get thrown out the window. The pace of change is no longer just about the top-level applications. It's moving down the stack. The very ground beneath our feet is being rebuilt, and this week gave us a clear look at both the blueprints for the new foundation, and the rubble of the old one.