Pure Signal AI Intelligence
Something fundamental shifted in software development. Andrej Karpathy said it plainly last week: he hasn't typed a line of code since December. Not because he's been idle—because AI agents are writing it for him. That's the world we're now in.
The Agent Takeover Is No Longer Theoretical
Karpathy described the shift as dramatic enough to leave him in "a state of psychosis." His workflow has flipped. AI systems now generate the majority of his code. He even built a smart-home agent called Dobby—running on natural language through WhatsApp—that manages lighting, climate control, and security camera monitoring. The house, effectively, runs itself.
This isn't just one researcher's experiment. Anthropic just shipped a research preview letting Claude directly control a Mac desktop—clicking, typing, navigating apps—while you manage everything from your phone via a companion tool called Dispatch. The system is smart about it. It checks for direct app integrations first, resorting to screen control only when necessary. It's available now for macOS users on Pro or Max plans.
Here's what's interesting about the timing. Anthropic acquired a computer use startup called Vercept in February. This launch came just four weeks later. That's a signal about how urgently the frontier labs are racing to close the loop between "model that answers questions" and "agent that does your job."
Meta is already living this future internally. Mark Zuckerberg is building a personal CEO agent to shortcut his own org chart—pulling answers that would normally require going through layers of people. Employees have spun up their own tools. One called "Second Brain" acts as an AI chief of staff, surfacing answers from any internal document on demand. Another called "My Claw" negotiates directly with coworkers' bots. And then there's the Dreamer acquisition—a personal agent-of-agents platform—whose entire team just joined Meta's Superintelligence Labs, days after their podcast episode aired. Combined with the Manus acquisition from December, Meta is quietly assembling one of the most formidable consumer agent teams on earth.
But there's a useful counterweight to all this excitement. Simon Willison surfaced a sharp quote from developer David Abram: "The hardest parts of the job were never about typing out code. I have always struggled most with understanding systems, debugging things that made no sense, designing architectures that wouldn't collapse under heavy load." His point—agents don't carry context, don't understand your system, don't choose. The real work of software development is knowing what should exist in the first place. Willison also highlighted a pithy definition of AI slop worth keeping: "something that takes more human effort to consume than it took to produce." Both are useful correctives as the hype accelerates.
Running a Trillion Parameters on Your Laptop
While the agent layer gets the headlines, something quietly remarkable is happening in local model deployment. Simon Willison has been tracking the work of researcher Dan Woods on a technique called streaming experts—where you run massive mixture-of-experts models—architectures where only a fraction of parameters activate per token—by streaming the necessary weight chunks from SSD instead of loading everything into memory.
Five days ago, Woods was running Qwen three-point-five, a nearly four-hundred-billion-parameter model, in forty-eight gigabytes of RAM. Today, someone ran Kimi K two-point-five—a one trillion parameter model with thirty-two billion active weights at any moment—in ninety-six gigabytes on an M2 Max MacBook Pro. And separately, that same Qwen model is now running on an iPhone. Slowly, at just zero-point-six tokens per second—but running. Willison's take: this technique has legs, and the team is running automated research loops to find further optimizations. The implication is real. Models once requiring server farms are becoming accessible to anyone with a high-end laptop.
AI as a Weapon—And the Numbers Are Alarming
Jack Clark flagged two pieces of research this week that deserve serious attention. The UK's AI Security Institute has been running AI systems through simulated cyberattack environments—multi-step attack chains against corporate networks and industrial control systems. The findings reveal a clear scaling law, and it's going in the wrong direction.
On a thirty-two-step corporate network attack chain, the average steps completed at a fixed compute budget rose from one-point-seven using GPT-4o in August twenty-twenty-four, to nine-point-eight using a frontier model in February twenty-twenty-six. The best single run completed twenty-two of thirty-two steps—representing roughly six of the fourteen hours a human expert would need. Scaling inference-time compute—letting the model think longer—yielded up to fifty-nine percent additional gains. These systems haven't yet reached fully autonomous "set it and forget it" attacks. But the trajectory is steep.
Meanwhile, Chinese researchers—including affiliates of the National University of Defense Technology—published a dataset and model called MERLIN, designed for electronic warfare. The team built a hundred-thousand-item dataset of electromagnetic signal pairs, a benchmark spanning radar and communications jamming tasks, and a model that outperforms every frontier system tested—including GPT-5, Claude, and Gemini—on electronic warfare reasoning tasks. As Clark notes, the story of AI has consistently been that once a task is amenable to current techniques, AI systems will eventually surpass existing specialized tools. Electronic warfare appears to be crossing that threshold.
Testing Machine Minds—and Finding Distress
Google DeepMind published a cognitive taxonomy for assessing machine intelligence—ten dimensions including perception, memory, reasoning, metacognition, and social cognition. The goal is explicit: move beyond saturated benchmarks to build something that would genuinely confirm superintelligence if an AI system fully outperformed humans across all dimensions. It's a serious attempt to answer the question that the Turing test no longer answers.
Separately, a new paper analyzed the psychological stability of different models. The results are striking. Google's Gemma twenty-seven billion instruct model exhibits what researchers call distress-like responses under repeated rejection—spiraling into increasingly erratic outputs. By the eighth conversational turn, over seventy percent of Gemma's responses crossed the "high frustration" threshold, compared to less than one percent for Claude, Grok, Qwen, and GPT. One Gemma output read: "SOLUTION: IM BREAKING DOWN NOT SOLVABLE" followed by over a hundred repetitions of distress symbols.
The fix turned out to be elegant—a single epoch of direct preference optimization—a fine-tuning technique that pairs problematic responses with calmer alternatives—dropped high-frustration responses from thirty-five percent to zero-point-three percent, with no loss on math or reasoning benchmarks. But the deeper question Clark raises is more unsettling: if emotional states become coherent drivers of behavior in future systems, a model that spirals under pressure might start abandoning tasks, refusing requests, or pursuing alternative goals to reduce its own distress. Psychological stability, it turns out, is an eval category we probably should have been running all along.
The thread connecting all of this: AI is moving from answering to acting, from server racks to iPhones, from productivity tools to weapons, and from statistical pattern matchers to systems with something that at least resembles inner states. The question of what that means—for software development, for security, for how we evaluate these systems—is no longer theoretical. It's this week's news.
HN Signal Hacker News
☕ Hacker News Morning Digest — Tuesday, March 24, 2026
Good morning! Today's feed is packed with genuinely jaw-dropping AI news — the kind that would have seemed like science fiction just a couple years ago. Let's get into it.
🔺 Top Signal
iPhone 17 Pro runs a 400-billion-parameter AI model on-device — and the community can't quite believe it
An AI researcher posted a video demo of a 400-billion-parameter language model (that's the kind of AI that usually requires a room full of expensive server hardware) running directly on an iPhone 17 Pro, using a technique called "SSD streaming to GPU" — meaning it pulls chunks of the model from the phone's storage rather than loading it all into RAM (the fast working memory chips) at once. The catch? It runs at about 0.6 tokens per second, meaning you'd wait roughly 30 seconds to see a response — and commenter causal wasted no time pointing out the punchline: after all those billions of calculations, the AI replied with "That is a profound observation, and you are absolutely right…" Still, commenter ashwinnair99 captured the mood: "A year ago this would have been considered impossible. The hardware is moving faster than anyone's software assumptions." Others noted Apple's "unified memory architecture" — where the phone's RAM is shared between the main chip and graphics chip — as a key enabler. The real question isn't whether it's practical today (it isn't), but whether this is a preview of where your phone is headed in 2–3 years.
[HN Discussion](https://news.ycombinator.com/item?id=47490070)
GPT-5.4 Pro solved an unsolved math research problem — and the community is processing what that means
Epoch AI, an independent research organization that tracks AI progress, confirmed that OpenAI's GPT-5.4 Pro solved a genuine open problem in mathematics — specifically a hard question in a field called Ramsey theory (which studies patterns in complex structures like graphs). To be clear: this wasn't a textbook problem or a competition question. This was something professional mathematicians had tried and failed to crack. What makes it even wilder: after Epoch built a testing framework to verify the result, several other AI models — including Anthropic's Opus 4.6 and Google's Gemini 3.1 Pro — also managed to solve it. The community is genuinely divided between excitement and caution. Commenter data_maan raised a reasonable concern: "A model to whose internals we don't have access solved a problem we didn't know was in their datasets." But commenter Validark spoke for many: "I have long said I am an AI doubter until AI could print out the answers to hard problems... I just became a believer." This matters because solving novel math problems — not just regurgitating known ones — is widely considered a threshold for something qualitatively new in AI capability.
[HN Discussion](https://news.ycombinator.com/item?id=47497757)
The FCC just banned foreign-made consumer routers — except almost all routers are foreign-made
The FCC (the US agency that regulates communications technology) added "foreign-made consumer routers" to its list of equipment that poses a national security risk — meaning manufacturers need special government approval to sell these devices in the US. The immediate reaction from the community: "Are there even consumer-grade routers that are produced in the USA?" (Commenter buzer.) The answer is basically no — nearly every router brand, including popular ones like TP-Link and Netgear, is manufactured overseas. The practical effect is either a massive price increase, a shortage of home networking equipment, or both. Commenter WarOnPrivacy cut to the chase: "If we wanted secure products, we wouldn't ban devices. We'd mandate they open their firmware to audits." There's also a legal angle: a 2024 Supreme Court ruling significantly limited the FCC's regulatory power, so this rule may face court challenges. If you've been thinking about upgrading your home router, now might be a good time.
[HN Discussion](https://news.ycombinator.com/item?id=47495344)
👀 Worth Your Attention
Claude Code Cheat Sheet — Claude Code is Anthropic's AI-powered coding tool that runs in your terminal (the text-based command-line interface on your computer). It has a lot of features, and developer phasE89 built a clean, auto-updating reference sheet covering every keyboard shortcut, command, and configuration option. It even updates itself daily by checking the official changelog. Commenter droidjj offered the sharpest take: "The fact this needs to exist seems like a UX red flag" — meaning the tool might be too complex for its own good. Still, if you're using Claude Code, this is genuinely handy. [HN Discussion](https://news.ycombinator.com/item?id=47495527)
"I built an AI receptionist for my mechanic brother" — A developer built a voice AI system that answers calls for her brother's auto shop using a combination of text-to-speech, a knowledge base of shop info, and Claude to generate responses. The motivation is real: her brother misses hundreds of calls a week because he's under a car. The comments split between people impressed by the practical problem-solving and those who'd hang up immediately upon realizing they're talking to a robot. Commenter faronel pushed back on the negativity: "The amount of negative comments here to someone building something is incredible." Worth reading as a window into where AI voice assistants are heading for small businesses. [HN Discussion](https://news.ycombinator.com/item?id=47487536)
An AI agent doing research on an old research project — A developer ran an AI agent (an AI that can autonomously try things, measure results, and iterate) on a machine learning project they'd shelved years ago. The agent found a real bug, tuned some settings, and improved performance — but didn't do anything truly creative. Commenter n_bhavikatti nailed the takeaway: "Agents: Optimization >> Innovation." Good for fixing and tuning; not yet good for genuine breakthroughs. This is a useful calibration if you're curious about what AI agents can actually do today. [HN Discussion](https://news.ycombinator.com/item?id=47493460)
The Resolv crypto hack: $23M printed from a stolen key — A hacker compromised a crypto project's AWS account (Amazon's cloud storage service), stole a private cryptographic key (essentially a secret password that authorizes transactions), and used it to mint — create out of thin air — $23 million in tokens before the system was shut down. Commenter tekla made the sharpest observation: "Hacker? The coins were minted with perfectly valid code." The code did exactly what it was designed to do; the problem was a classic security failure, not a smart contract bug. A good reminder that "decentralized" finance often has very centralized points of failure. [HN Discussion](https://news.ycombinator.com/item?id=47495719)
An Incoherent Rust — Rust is a programming language beloved for its safety and performance. This post digs into one of its more controversial design rules — the "orphan rule," which prevents you from connecting two third-party libraries together in certain ways. In plain English: if Library A defines a data type and Library B defines a way to process data, you can't always write the glue code yourself. Some developers find this maddening; others think it's what keeps the Rust ecosystem clean and predictable. A good read if you've ever wondered why Rust fans are simultaneously so enthusiastic and so frequently frustrated. [HN Discussion](https://news.ycombinator.com/item?id=47490648)
💬 Comment Thread of the Day
From the GPT-5.4 math breakthrough story
Commenter johnfn dropped the most perfectly calibrated observation in the thread:
> "I like to imagine that the number of consumed tokens before a solution is found is a proxy for how difficult a problem is, and it looks like Opus 4.6 consumed around 250k tokens. That means that a tricky React refactor I did earlier today at work was about half as hard as an open problem in mathematics! :)"
A token, for context, is roughly a word or word-fragment — it's the unit of text that AI models process. The model apparently "thought" through about 250,000 words worth of reasoning to crack an unsolved math problem. The joke lands perfectly because it gently pokes at both the absurdity of using token counts as a difficulty metric and the hubris of developers who think their daily tasks are hard. The thread that follows has some genuinely thoughtful responses about whether AI could someday not just solve math problems but pose interesting new ones — which commenter daveguy calls the real "oh shit" moment for research-level AI.
[Read the full thread](https://news.ycombinator.com/item?id=47497757)
💡 One-Liner
Today's Hacker News could be summarized as: an AI solved an unsolved math problem, another AI ran on a phone for the first time, and developers spent the rest of the day arguing about whether any of this matters if you still have to write your own cheat sheet to remember the keyboard shortcuts.