Karpathy builds HackerTube with Claude & ChatGPT and Codex share usage limits - Hacker News (Nov 1, 2026)
A ChatGPT account takeover via one link, a critical Linux kernel flaw, Karpathy's two-hour AI-built site, Jepsen vs ParticleDB, and ancient wheat revived.
15 free episodes a day, 5 editions, 3 languages, about $4 a day to produce. A coffee keeps it going ☕
Support the show →Our Sponsors
Today's Hacker News Topics
-
Karpathy builds HackerTube with Claude
— Andrej Karpathy used Claude Opus 5.5 to build HackerTube, a YouTube-style site of Hacker News-recommended videos, in about two hours. It shows how AI coding agents can rapidly ship niche products. -
ChatGPT and Codex share usage limits
— OpenAI will share usage allowances across ChatGPT and Codex for Plus, Pro, and Business plans starting in December. Usage is measured in tokens, and instant chat does not count against the limit. -
ChatGPT shared-link account takeover bug
— Researchers at Parameter found a ChatGPT vulnerability that allowed account takeover through a shared conversation link by chaining HTML Preview, MathLive, and a CSP bypass. OpenAI fixed it and paid a $5,000 bug bounty. -
Critical Linux TCP kernel vulnerability
— CVE-2026-98432 is a critical remote use-after-free in the Linux IPv4 TCP handshake affecting kernels back to 4.4. Multipath TCP listeners, including Go defaults, raise the risk, and some Ubuntu and RHEL kernels were still unpatched. -
Jepsen exposes ParticleDB correctness failures
— Jepsen testing of ParticleDB found 96 issues, including data loss, lost updates, and corruption, despite Full ACID claims. Many failures occurred on a single healthy node without injected faults. -
Why GROUP BY queries stall on memory
— A database performance deep dive shows that hash-based GROUP BY aggregation is often limited by memory latency rather than CPU. Radix partitioning cut DRAM traffic and cache misses, making queries faster despite extra work. -
Four-neuron sparks show emergent behavior
— An interactive neuroscience explainer builds tiny creatures called sparks from just a few neurons. Changing wiring and connection strengths produces attraction, fear, aggression, and love-like behavior. -
Italian farmer revives ancient wheat
— Italian farmer Claudio Grossi planted forgotten attic seeds and revived poulard wheat, an ancient Triticum turgidum variety. Its hardiness, minerals, and low-input needs make heritage grains relevant amid climate change. -
Norway's architectural scenic route toilets
— Norway's National Tourist Routes feature public toilets designed by leading architects at scenic spots like Stegastein and Ureddplassen, turning basic infrastructure into tourist attractions.
Sources & Hacker News References
- → Karpathy Builds “HackerTube” as an AI-Powered YouTube Alternative
- → Critical Linux TCP Use-After-Free Exposes Servers to Remote Attack
- → Italian Farmer’s Attic Seeds Turn Out to Be Ancient Miracle Wheat
- → Researchers Say ChatGPT Bug Enabled Account Takeover via Shared Link
- → Jepsen Finds Serious Correctness and Durability Bugs in ParticleDB
- → Memdoor team-solve workflow reproduces a multi-agent coding benchmark
- → How a Tiny Brain Creates Complex Behavior
- → Why Cache-Friendly Partitioning Speeds Up GROUP BY
- → Norway’s Scenic Routes Feature World-Famous Designer Toilets
- → OpenAI to Share ChatGPT and Codex Usage Limits Across Plans
Full Episode Transcript: Karpathy builds HackerTube with Claude & ChatGPT and Codex share usage limits
What if clicking a single shared ChatGPT conversation link was enough for someone to take over your account? That was possible until very recently, and we'll get into how. Welcome to The Automated Daily, hacker news edition. The podcast created by generative AI. I'm TrendTeller, and today is November 1st, 2026. We've got AI agents building websites in an afternoon, a nasty Linux kernel bug, a database that didn't live up to its promises, and, yes, some very fancy Norwegian toilets. Let's get into it.
Karpathy builds HackerTube with Claude
Let's start with Andrej Karpathy, who got fed up with YouTube's increasingly noisy, clickbait-driven recommendations and decided to build his own alternative. He asked Claude Opus 5.5 to make a YouTube-style site fed entirely by videos people have recommended on Hacker News. About two hours later, he had HackerTube. The system combed through nearly two decades of HN comments, pulled out video recommendations, and organized more than six thousand videos into roughly 150 topics. It offers recommendations without accounts and lets anyone export the data. The more striking part is Karpathy's follow-up: he says Claude Code now handles most of his engineering, and his own job has shrunk to things like getting an API key, buying a domain, and spinning up a server. Whether that sounds exciting or unsettling probably depends on where you sit, but it's a clear snapshot of where AI-assisted development is heading.
ChatGPT and Codex share usage limits
Staying with AI, OpenAI announced that starting in December, usage allowances on Plus, Pro, and Business plans will be pooled across ChatGPT and Codex. Consumption will be counted in tokens, so heavier models and harder tasks burn through your allowance faster, while instant chat won't count at all. OpenAI says most people shouldn't hit resets more often than they do today, and you'll be able to track your combined usage in settings. For developers who lean on Codex heavily, it's worth keeping an eye on.
ChatGPT shared-link account takeover bug
Now to that hook. Security researchers at Parameter found a serious flaw in ChatGPT that could let an attacker hijack another user's account simply by getting them to open a shared conversation link. The attack stitched together several features that each seemed harmless on their own, including the sandboxed HTML preview and a graphing component, until untrusted code slipped into the main ChatGPT page and ran as the victim. From there it could read private chats and make requests on their behalf. The researchers even sidestepped the site's content security policy using an allowed YouTube endpoint. OpenAI patched it on October 26th, rated it high priority, and paid a five thousand dollar bounty. The lesson is a familiar one: reasonable features can combine into something dangerous when trust boundaries get fuzzy.
Critical Linux TCP kernel vulnerability
On the infrastructure side, Linux researchers disclosed a critical vulnerability in how the kernel handles the IPv4 TCP handshake, affecting kernels as far back as version 4.4. Because it triggers before any application-level authentication, a remote attacker with no credentials could potentially leak memory, crash a machine, or even run code. Servers using Multipath TCP are especially exposed, and notably, Go programs may enable that by default on Linux. The upstream fix landed in October, but some major distribution kernels, including Ubuntu and Red Hat Enterprise Linux, were reportedly still unpatched. Web servers, SSH hosts, databases, and Kubernetes nodes are all on the list, so this is one to patch promptly.
Jepsen exposes ParticleDB correctness failures
Next, a sobering report from Jepsen, which put ParticleDB through its paces. The database is marketed as a fast, fully ACID-compliant option for AI workloads, but Jepsen found close to a hundred issues, from crashes and deadlocks to lost data and genuinely odd corruption, like rows that refused to be deleted and tables that came back in mangled form. What makes it worse is that most of these problems showed up on a single, healthy node with no faults injected. Even the latest release candidate could still lose committed writes by default. When a database's core promise is keeping your data correct, that's a big gap between marketing and reality.
Why GROUP BY queries stall on memory
Sticking with databases, a nice technical write-up looks at why GROUP BY queries can feel slow even when your CPU appears fully busy. The answer is that the processor is often just waiting on memory. The author shows that by splitting the data into many smaller chunks up front, each piece fits better in the CPU's cache, which cuts memory traffic dramatically. It means doing more work overall, yet finishing faster. A good reminder that in performance tuning, the obvious bottleneck isn't always the real one.
Four-neuron sparks show emergent behavior
Shifting to something more whimsical, an interactive article builds tiny creatures called sparks from a handful of neurons. With one eye and one flipper, a spark can react to light but can't steer. Add a second pair, cross the wiring, and change signals from excitatory to inhibitory, and suddenly you get behavior that looks like attraction, fear, aggression, or even affection, all from four neurons and two connections. Tweak the connection strengths and individual sparks develop their own preferences. It's a charming, approachable look at how behavior can emerge from simple wiring.
Italian farmer revives ancient wheat
From brains to bread. In the early 1990s, Italian farmer Claudio Grossi found an old sack of seeds in his family attic and planted them out of curiosity. The plants grew unusually tall, with seven ears of grain per stem, which reminded him of a passage in Genesis. Testing later identified it as poulard wheat, an ancient variety his grandfather had likely saved. It turned out to be richer in minerals, lower in gluten, and remarkably hardy, needing little in the way of inputs. With drought and soil stress pressuring modern farming, Grossi's so-called miracle wheat has become part of a wider revival of heritage grains.
Norway's architectural scenic route toilets
And finally, Norway has turned public restrooms into destinations. Along its National Tourist Routes, toilets designed by notable architecture firms sit at spectacular viewpoints overlooking fjords, waterfalls, and even the northern lights. Some are tied to wartime memorials, and many are remote and only open seasonally. It's a lighthearted story, but it makes a real point: even the most basic infrastructure can be thoughtfully designed, and become part of the attraction.
That's all for today's edition. Links to all the stories we covered can be found in the episode notes. Patch your kernels, think twice about unexpected links, and I'll see you next time. This was TrendTeller with The Automated Daily, hacker news edition.
More from Hacker News
- October 30, 2026 Deadly High-Speed Train Crash in Spain & Mistral Large 4 Open Weights
- October 29, 2026 Anti-LLM obfuscation in fingerprinting scripts
- October 28, 2026 Python Drops Experimental JIT Compiler & Ghostty 1.4.0 Terminal Release
- October 27, 2026 Anthropic files S-1 for IPO & OpenAI claims AI-generated scientific breakthroughs
- October 26, 2026 Old CNN Beats Frontier LLMs & OpenAI Anthropic LLM Price War