Hacker News Daily: The Sharpest Tech Stories & Debates, Curated Fast

A college instructor is making computer science students take their programming exams on a typewriter. This is happening right now, in April 2026, and it’s the perfect lens for the most urgent conversation in tech today: how we’re all scrambling to build systems of trust in a world overflowing with AI. That story, and the reactions to it, are just one piece of the puzzle we're piecing together from the heart of the tech community this week.

Let's start with the big picture. The social network Bluesky had a major outage, and its post-mortem included a fascinating comparison. A commenter on Hacker News nailed it: "The proper analogy here would be like Yahoo going down in 1999." Think about that. Bluesky, the supposedly decentralized future, is having failure modes that look like the pre-Google, chaotic, federated web. It’s a reminder that new architectures don't erase old problems—they just give them new names.

The dream of decentralization is about resilience, but the reality, for now, is that when a key node or service gets overloaded, the whole distributed system can still feel very, very centralized. Meanwhile, the latest "Ask Hacker News: What Are You Working On?" thread just exploded. Over eleven hundred comments. And buried in that avalanche of side projects is a clear signal of where developer energy is going. It's not just about building with AI anymore.

It's about building tools to manage AI. One developer, Tomasz Tomczyk, released a tool called crit-dot-md. It’s a command-line interface for reviewing code that an AI agent wrote, treating it just like a pull request from a human junior developer. Because that’s what these agents are, right? They’re prolific, they’re fast, but they’re also prone to making plausible-but-completely-wrong changes. Tomczyk said it's already his most successful side project ever, which tells you how big the pain point is.

We're moving from the "wow, it can code!" phase to the "okay, how do we clean up its mess?" phase. And that's not the only one. Other tools popped up in the same thread. There's 'Etch', which detects breaking API changes by watching real traffic. There's 'GhostDep', which finds phantom dependencies in your code—the kind of subtle bug an AI might introduce without realizing it. And there's 'ClaimCheck', which, as the name implies, tries to verify the claims an AI makes about the code it just wrote.

This is the new cottage industry: AI quality control. It's the emergence of a human-AI immune system, built by developers, for developers. Slightly different angle, but same theme of control and verification... a project called nofunds-dot-org is being built. The creator, going by Kokkis, describes it as a financial censorship monitor. The goal is to document and track how money is weaponized to silence activists and journalists—things like frozen bank accounts, blocked payments, deplatforming from financial services.

The project is actually using the Claude API to process data, but its mission is fundamentally about creating a check on power. The creator argues that tools like Bitcoin can be a form of resistance here. It’s a controversial take, but it’s another example of tech being used not just to build products, but to build counter-systems. To verify who is being silenced and create alternative rails. And all of this is happening against a backdrop of a much broader, more existential debate.

A thread about predictions for 2026 got right to the heart of it. One user, tomwojcik, warned of "Digital Dementia 2.0." His fear is that our constant use of LLMs will make our minds lazy, that our attention spans will crater, and that we'll lose the ability for deep, critical thought. It’s a powerful, dystopian image. But in the same thread, another user, handoflixue, offered the counter-argument. "You can still use pencil and paper for the difficult things," they wrote.

Their point is that AI doesn't have to replace deep thinking; it can enable it. By automating the simple, tedious tasks, AI could free up more time and cognitive energy for the truly hard problems. So which is it? Are we outsourcing our brains, or are we just outsourcing the boring parts? The community is completely divided. And the answer, almost certainly, is yes. It's both. So let's go deeper on that typewriter. Because it feels like a joke, but it’s not.

It’s a strategy. And it connects directly to this new wave of AI-wrangling tools like crit-dot-md. The instructor who brought back typewriters for programming exams explained their reasoning perfectly. They said, "When I did my Computer Science degree... programming exams were hand written, proctored by TAs... We already had AI proof education." Their point is that being forced to write code by hand—or by typewriter—forces a different kind of thinking.

You can't just run it and see if it works. You have to trace the logic in your head. You have to think about readability, about structure. You have to be certain, because there's no compiler to catch your mistakes. It's an environment of extreme constraint. And in that environment, you can’t fake it. An AI can't help you. It’s just you and the logic. Now, where have we seen this before? This pattern of retreating to an older, more constrained technology to preserve a specific human skill.

It feels a bit like photographers going back to film to escape the infinite tweakability of digital. Or musicians recording to tape to force themselves to commit to a performance. In each case, the limitation isn't the enemy; it's the tool. It’s a deliberate choice to enter a space where the firehose of modern tech is shut off, so you can focus on the fundamentals. But here’s where the analogy gets really sharp. The typewriter isn't just about nostalgia.

It's an authentication mechanism. It verifies that the student sitting there actually possesses the skill being tested. It’s a low-tech solution to a high-tech problem: the problem of provenance. How do I know you wrote this? How do I know you understand it? And that’s the EXACT same problem that a tool like crit-dot-md is trying to solve, just from the opposite direction. The typewriter creates a sterile environment where AI can't exist. Crit-dot-md creates a review process for an environment where AI is everywhere.

One is a firewall, the other is a vaccine. But both are part of this emerging immune response to a world saturated with synthetic text, synthetic code, synthetic everything. We are desperate for ways to tell what's real, what's competent, and what's just a plausible hallucination. This brings us back to Bluesky. To that "Yahoo in 1999" comment. Because the struggle for a trustworthy foundation isn't just about content; it's about the infrastructure itself.

The promise of decentralization, with protocols like Bluesky's atproto, was a more resilient, censorship-resistant social graph. No single company could pull the plug. But the recent outage shows the messy reality. A decentralized protocol still relies on a handful of very large, very centralized services to actually be useful—things like the main Bluesky app server, which aggregates all the data. When that aggregator goes down, the network effect for most users is identical to a centralized service like Twitter having an outage.

The lights go out. The pattern twin here, as the commenter noted, is the early web. Before Google, search engines like Yahoo weren't crawlers in the modern sense. They were giant, human-curated directories. They were a portal, a map to the chaotic, distributed mess of the early internet. They were a centralized guide to a decentralized world. And when Yahoo had a problem, a huge chunk of the web became effectively invisible to millions of people.

It was a single point of failure for discovering a distributed system. That's precisely the stage Bluesky is in. It's building the decentralized roads, but most people are still using one giant, centralized map to navigate them. The analogy holds because it reveals a fundamental tension: convenience and user experience will almost always pull decentralized systems back toward centralized choke points. It's just easier. But here's where the analogy breaks.

Our expectations have completely changed. In 1999, the internet being broken was normal. It was slow, pages failed to load, GeoCities sites were down. It was part of the experience. In 2026, we expect global, instantaneous, five-nines reliability. So when a next-generation decentralized network stumbles like a first-generation web portal, the disappointment is way, way bigger. It feels less like a growing pain and more like a broken promise. The challenge for Bluesky and others isn't just technical—it's about managing expectations that were set by the very centralized giants they're trying to replace.

So you have these parallel struggles happening all at once. On one level, you have an instructor with a typewriter, trying to verify one student's mind. On another, you have developers building tools to verify the output of a global AI's mind. And on a third level, you have architects trying to build a global, decentralized network that you can actually rely on. They all seem like different problems, but they're not. They are all quests for the same thing: a stable, trustworthy foundation in a world that feels increasingly synthetic and unstable.

We are building verification layers for everything. For human skill. For AI-generated code. For our social infrastructure. It’s a massive, collective effort to answer a single question that haunts every corner of the tech world right now: how can we be sure? How do we know what’s real? This week’s conversations on Hacker News show that nobody has the final answer. What they have are typewriters, and command-line tools, and post-mortems. The most important work in tech right now isn't building the next big model; it's building the tools to believe what the last one told us.