Pure Signal AI Intelligence
Something shifted last week. Not a product launch. Not a benchmark. Something harder to pin down — a collective realization, across researchers and analysts, that the agent era isn't coming. It's here. And it changes almost everything we thought we knew about where AI value concentrates.
The Third Paradigm: What Agents Actually Are
Ben Thompson and Simon Willison are independently converging on the same insight this week, and it's worth sitting with.
Thompson traces three inflection points. ChatGPT — November 2022 — showed the world what language models could do. OpenAI's o1 — September 2024 — introduced reasoning models that self-evaluate before answering, dramatically cutting errors. But it's the third moment that changes the economics: Anthropic's Opus 4.5 and Claude Code, late 2025, when agents started actually completing multi-hour tasks without human supervision.
Here's the critical distinction Thompson makes. Reasoning models — like o1 — are still just models thinking harder. Agents are something different. An agent is software that wraps a model, gives it tools, and runs it in a loop until the job is done. The model writes code. The agent runs the code. If it fails, the agent tries again. The human never touches it.
Willison sharpens this with a clean definition: agents run tools in a loop to achieve a goal. And he identifies the single capability that makes everything else possible — code execution. Without the ability to actually run what it writes, an LLM's output is suggestions. With execution, it can iterate toward something that demonstrably works.
Both writers push back on "vibe coding" — Karpathy's term from early 2025 for just prompting an LLM and shipping whatever comes out. Willison argues that's a distinct and lesser thing. Agentic engineering is about specifying problems precisely, verifying results rigorously, and updating the agent's instructions based on what you learn. The human's job hasn't disappeared. It's moved upstream — from writing code to knowing what code to write, and why.
The Integration Moat: Why Model Plus Harness Beats Model Alone
This is where Thompson's analysis gets genuinely provocative — and it directly challenges the "models are commodities" thesis that's been popular for the past year.
The argument for commoditization goes like this: DeepSeek built a competitive model for six million dollars. Open source models now power eighty percent of venture-backed startups. Why would anyone pay a premium for Anthropic or OpenAI?
Thompson thinks agents break this logic entirely. The performance of an agent depends not just on the model, but on the harness — the software infrastructure that controls the model, routes tasks, and handles tool use. What made Claude Code suddenly remarkable in late 2025 wasn't a new model release. It was changes to the harness. Model and harness are integrated, not modular. And in any value chain, profits flow toward integrated components — not interchangeable ones.
Apple is the canonical example. Apple's hardware isn't commoditized because it's inseparable from the software. Thompson argues Anthropic and OpenAI are positioning themselves similarly — as points of integration in the AI stack.
The evidence? Microsoft. Microsoft spent two years insisting it was "model agnostic" — that its Copilot products would work with any underlying model. Then last week, Microsoft launched Copilot Cowork, essentially an enterprise agent product, and it runs exclusively on Anthropic. To deliver a compelling agentic experience, Microsoft had to abandon model agnosticism. That's a significant admission.
Thompson also makes a precise argument about compute demand. First paradigm: inference was cheap, one call, one answer. Second paradigm: reasoning models use ten to a hundred times more tokens per query. Third paradigm: agents make multiple model calls, plus CPU-intensive tool execution, plus the loop overhead. And crucially — because agents dramatically increase usefulness, usage expands to fill and exceed the available compute. Every hyperscaler says demand exceeds supply. Thompson's framing explains why that's not hype.
AI and Labor: Karpathy's Map, Thompson's Implication
Andrej Karpathy published a viral analysis this weekend — what he's calling a "vibe coded" labor market study — mapping three hundred forty two job categories by their exposure to AI disruption. The pattern he found: the highest-paying jobs score worst on AI-resilience. Meaning AI exposure correlates positively with compensation. The jobs most at risk aren't the lowest-paid ones.
Thompson's enterprise analysis gives that data a mechanism. He argues that large organizations became bloated not because of bad management, but because that's historically been the only way to scale. A brilliant person with vision still needed a large team to implement it. Every coordination layer added cost and drag.
Agents change the math. One person with high agency — the ability to direct AI effectively — can now control multiple agents simultaneously. Thompson's phrase here is worth remembering: the rise of agents narrows the number of humans needed for AI's economic impact to be profound.
He's direct about what follows. Companies aren't just going to rightsize for post-COVID over-hiring. They're going to rightsize for a post-AI world. The ones that don't will face AI-native competitors with structurally smaller cost bases and growing capability advantages. Thompson says this is going to get ugly — and frames it not as advocacy but as analysis of economic forces that will be very hard to resist.
The counterweight to that picture appeared in Rowan Cheung's newsletter this morning. A Sydney consultant named Paul Conyngham — no biology background — built a custom messenger RNA cancer vaccine for his dog by chaining ChatGPT, Grok, and DeepMind's AlphaFold — a protein structure prediction tool — together with a university genomics lab. Three hundred fifty gigabytes of tumor data. A three-thousand-dollar sequencing run. One tumor shrank by half. A year ago, that path didn't exist for an individual without a research institution behind them. It does now. Thompson calls this the "sovereign individual" — one person directing AI to do things that previously required teams.
The Contrarian Architecture: LeCun's Billion-Dollar Dissent
Not everyone accepts the LLM paradigm as the destination. Yann LeCun — who pioneered convolutional neural networks, the architecture behind modern computer vision — left Meta last year and just raised over a billion dollars for a startup called Advanced Machine Intelligence. His thesis: language models are fundamentally limited because they operate through words as intermediaries. They've never dropped an object. They don't understand why a glass shatters.
LeCun's alternative is what he calls world models — systems that reason, plan, and understand cause and effect in physical space. The core architecture is JEPA — Joint Embedding Predictive Architecture — which works in what's called latent space, an abstract representation of reality rather than raw sensory data. The analogy: when you're driving, you track the road and other cars. You don't model every leaf on every tree. JEPA focuses on high-level relationships and ignores irrelevant detail.
LeCun's departure from Meta was acrimonious — he reportedly clashed with leadership over a pivot toward scaling maximalism, the belief that adding more compute and data to existing architectures will eventually produce general intelligence. He left following allegations that Meta's Llama 4 benchmarks were manipulated to satisfy leadership. He took a significant number of FAIR researchers with him.
Whether LeCun is right is genuinely open. Agents built on LLMs are demonstrably doing more than he predicted they would. But his billion-dollar bet is a serious architectural dissent, not just contrarianism — and the robotics results coming from JEPA-adjacent research suggest the world model approach may matter for physical AI even if language models dominate text tasks.
The honest picture, then, is this: agents have made language models dramatically more capable and economically durable than skeptics expected. The integration between model and harness looks like a real moat. And one of the most credentialed researchers in the field is betting a billion dollars that we're still missing something fundamental about what intelligence actually requires. Both things can be true.
HN Signal Hacker News
☀️ Morning Digest — Monday, March 16, 2026
Good morning! Today's Hacker News has a bit of everything: privacy legislation that should be on your radar, a very relatable rant about bloated websites, and a genuinely cool robotics demo. Let's dig in.
🔺 Top Signal
[Canada's Bill C-22 Mandates Mass Metadata Surveillance](https://www.michaelgeist.ca/2026/03/a-tale-of-two-bills-lawful-access-returns-with-changes-to-warrantless-access-but-dangerous-backdoor-surveillance-risks-remains/) — A major surveillance bill just quietly arrived in Canada, and the details are troubling.
Bill C-22 would require Canadian internet service providers (the companies that connect you to the internet) and telecom companies (your phone carrier) to collect and retain your metadata for up to a year — and hand it over to authorities on request. "Metadata" doesn't mean the content of your messages, but rather the who, when, where, and how: who you called, when you texted, which apps you used, and where your phone was. That sounds less scary than it is — metadata can reveal an enormous amount about a person's life. Commenter briandw put it bluntly: "The bill claims it doesn't grant any new powers. Then it goes on to explain that if you don't collect metadata and retain it for up to a year, you can be fined or jailed." Canadian commenter rkagerer asked the question that cuts to the heart of it: "Would you be OK if police officers followed you around everywhere you go, recording who you talk to, and when and where you interacted — not because there was any reason to suspect you of anything?" The broader mood in the thread is weary frustration — many commenters note this is at least the third time Canada has tried to pass essentially this bill after previous versions were rejected.
[HN Discussion](https://news.ycombinator.com/item?id=47392084)
[The 49MB Web Page](https://thatshubham.com/blog/news-audit) — A developer audited the New York Times homepage and found it weighs as much as a full software installation from 1995.
To put 49 megabytes in perspective: commenter userbinator notes they like measuring bloated pages in units of "Windows 95 installs" — the entire Windows 95 operating system fit in about 40MB. The article audits what's actually in that payload: dozens of ad trackers, auto-playing videos, surveillance scripts, and layers of JavaScript (a type of code that runs in your browser) doing things you never asked for. The irony, as commenter cjs_ac pointed out, is that "my family's first broadband connection, circa 2005, came with a monthly data quota of 400 MB" — meaning a single visit to the NYT homepage today would have eaten more than 10% of an entire month's internet allowance back then. The comments are a cathartic venting session for developers and readers alike, with plenty of shoutouts to RSS (a way to subscribe to website content without ever visiting the actual page — think of it like a podcast feed for articles) as the answer everyone abandoned too soon.
[HN Discussion](https://news.ycombinator.com/item?id=47390945)
[Chrome DevTools MCP](https://developer.chrome.com/blog/chrome-devtools-mcp-debug-your-browser-session) — Google just made it possible for AI assistants to directly control and inspect your Chrome browser session.
MCP stands for "Model Context Protocol" — think of it as a standardized plug that lets AI tools connect to other software. In plain terms: this means an AI coding assistant like Claude or Cursor can now open your browser, look at what's on screen, click things, read error messages in the developer console (a hidden panel where your browser logs technical errors), and observe network traffic (what data your browser is sending and receiving). The thread lit up with developers sharing how they're already using this for tasks like having AI scrape websites, debug broken web pages, or even — in one memorable example from aadishv — "open a YouTube Music tab, search for each album, and get the URL to pass to yt-dlp." A few practical warnings: commenter glerk notes it's "a mega token guzzler" (AI models charge by the amount of text they process, so a tool that sends screenshots and page data constantly gets expensive fast), and JKolios raised the uncomfortable question: "Now that there's widespread direct connectivity between agents and browser sessions, are CAPTCHAs even relevant anymore?"
[HN Discussion](https://news.ycombinator.com/item?id=47390817)
👀 Worth Your Attention
["Stop Sloppypasta"](https://stopsloppypasta.ai/) — Someone coined a term for a genuinely annoying phenomenon: pasting raw, unread AI output at colleagues as if it were your own considered response.
"Sloppypasta" = sloppy AI output + copypasta (text mindlessly copied and pasted). The author's argument is simple: if you ask an AI for help and then dump the entire response on someone else unedited and unverified, you're outsourcing your cognitive work to them. The community is mostly sympathetic, though commenter Bratmon added a healthy counterpoint: "I'm tired of 'journalists shouldn't try to make their living by finding profitable ads, they should just... work at McDonald's' takes" — pointing out that critiquing how people use tools is easier than solving the underlying incentive problems. Best comment goes to api: "The solution is to have your bot read the sloppypasta for you!"
[HN Discussion](https://news.ycombinator.com/item?id=47389570)
[Nasdaq's Shame](https://keubiko.substack.com/p/nasdaqs-shame) — A sharp piece arguing that the Nasdaq is about to do something that quietly harms everyday index fund investors.
The article claims SpaceX is being set up for inclusion in the Nasdaq-100 index in a way that forces passive investors (people who buy index funds — investment bundles designed to track "the whole market" without picking individual stocks) to buy SpaceX shares whether they want to or not, at whatever price the market sets during the index rebalancing. If you have a 401(k) or retirement savings in a fund that tracks the Nasdaq, this affects you. The comment section is a mix of outrage, nuanced finance takes, and genuinely helpful explanations — commenter Veserv wrote a clear breakdown of the mechanics using a simple hypothetical that's worth reading if you have index fund exposure.
[HN Discussion](https://news.ycombinator.com/item?id=47392550)
[Glassworm Returns: Invisible Unicode Attacks Hit GitHub Repositories](https://www.aikido.dev/blog/glassworm-returns-unicode-attack-github-npm-vscode) — Hackers are hiding malicious code in plain sight using invisible characters that look like empty space.
Unicode is the system that lets computers display every language and symbol on Earth — but it also includes hundreds of invisible "zero-width" characters that produce no visible output. Attackers are embedding these invisible characters into code contributions on GitHub (a platform where developers share and collaborate on code) in a way that changes what the code actually does while looking completely innocent to a human reviewer. Commenter ocornut made the most pointed observation: "It baffles me that any maintainer would merge code like this without knowing what it does... regardless of being able to see the invisible characters. There's an `eval()` call." (`eval()` is a function that executes arbitrary code — a classic red flag.)
[HN Discussion](https://news.ycombinator.com/item?id=47387047)
[LLM Architecture Gallery](https://sebastianraschka.com/llm-architecture-gallery/) — A beautiful visual reference showing the internal structure of dozens of AI language models side by side.
An "architecture" in this context means the blueprint for how a neural network (the mathematical structure underlying AI models) is organized — how many layers it has, how information flows through it, how it's trained. This gallery lets you visually compare models like GPT-2, Llama, Mistral, and others in a single place. The most interesting observation from the thread came from libraryofbabel: "In the last seven years since GPT-2, there have been a lot of improvements to LLM architecture, but no fundamental innovation." Most of the magic has come from scaling up (bigger, more layers) and refining the same basic ideas. Commenter travisgriggs was disappointed for different reasons: "I clicked here hoping we were having LLMs design skyscrapers, dams, and bridges."
[HN Discussion](https://news.ycombinator.com/item?id=47388676)
[Humanoid Robot Learns to Play Tennis](https://zzk273.github.io/LATENT/) — Researchers trained a humanoid robot to return tennis shots by learning from imperfect human motion data.
The video is genuinely remarkable — the robot splits, lunges, and swings a racket with something that looks like real athletic intent, learned by watching humans play rather than being manually programmed with every movement. The catch, noted by commenter blueblisters: the robot's sense of where the ball is comes from external high-speed cameras, not its own eyes. "Almost all of closed-loop robotics is a state estimation problem. Control is 'solved' if you can estimate state well enough." Commenter V__ had the most interesting speculative take: a "perfect" robot wouldn't move like a human at all — it would use bizarre, late, geometrically optimal swings specifically designed to confuse human opponents.
[HN Discussion](https://news.ycombinator.com/item?id=47388273)
💬 Comment Thread of the Day
From: [LLMs Can Be Exhausting](https://news.ycombinator.com/item?id=47391803)
This thread exploded with developers sharing how cognitively draining it is to manage AI coding assistants — and it's more nuanced than the usual "AI good/bad" debate.
simonw (a well-known developer and creator of Datasette) dropped this gem:
> "I wonder if it's more or less tiring to work with LLMs in YOLO/--dangerously-skip-permissions mode. I mostly use YOLO mode which means I'm not constantly watching them and approving things they want to do... but also means I'm much more likely to have 2-3 agent sessions running in parallel, resulting in constant switching which is very mentally taxing."
"YOLO mode" here means letting the AI agent take actions autonomously without asking for permission each step — faster, but you lose oversight. The mental load of managing multiple AI sessions in parallel, while staying sharp enough to catch their mistakes, seems to be the real hidden cost nobody talked about when AI coding assistants launched.
rednafi added the most relatable workplace complaint:
> "Corporate has mandated AI usage and is asking people to do 10k LOC PRs every day. Reviewing this junk has become exhausting. I don't want to read your code if you haven't read it yourself."
(A "PR" is a pull request — a proposed code change submitted for review. "10k LOC" means 10,000 lines of code — an enormous amount to review in a day.)
The thread is worth reading in full — it's one of the most honest conversations about AI productivity tools that doesn't collapse into either cheerleading or dismissal.
🎯 One-Liner
Today Hacker News had a story about a 49MB web page, a story about the cost of AI agents burning through millions of invisible tokens in the background, and a story about surveillance systems hoarding metadata for a year — and somehow the most upvoted thread was people complaining about colleagues copy-pasting chatbot responses into Slack. We contain multitudes.