Unsupervised Learning NO. 503

Hey! Hope you all are doing well!

If you care about AGI and/or Kaparthy, this new blog of mine is a must-read. You can also listen to it! READ IT | LISTEN TO IT

—

I’ve massively upgraded not just my own K/PAI stack, but the public PAI repo with the new Anthropic Agent Skills, which I’ve been talking about non-stop since they came out last week. But not just that. I've also done a massive cleanup and restructuring to make it match my current system. Now is the time to get into this.

—

Here is my summary of why Anthropic’s new Agent Skills are such a big deal.

—

A bit of a sharing / vulnerable / notification note. I feel a bit torn between Frames right now when it comes to the world, and to life. And to being a public person. I am an optimist, and I hope to always remain one. But the few years or so have been rough on people, and on the world.

I find myself, too often I think, in one of two diametrically opposed states:

I often force myself into the second as a means of getting out of the first.

It’s like:

So I lean in. On the empowerment. On the capabilities. On all the great things automation can bring us if we think about it correctly and focus on the right things. The Stoics and the Bible had it right with the whole, “What you can control” thing.

But I just wanted you all to know that.

If you see me talking about Human 3.0 like it’s inevitable, 1) it’s not. And 2), I know it’s not. I think the chances of us smoothly transitioning into H3 without decades of turmoil are very low. But they’re not zero. And they’re also made lower by people not trying to bring it about.

And if you see me somewhat panicking, and being like, welp…there goes the US, or China’s now in charge, or whatever, that’s not a different me. It’s just the frame I happen to be in at the moment. And it’s no less true or real then the “We can make H3 happen” frame.

I predict you’ll see a lot of both of those from me in coming months and years. I don’t see any way out of that. Just know that both are genuine, and the oscillations seem as strange and jarring to me as they does to you. 🫶🏼

—

I’m hiring an operations person for Unsupervised Learning. Here’s the JD:

If you know anyone extraordinarily high-agency, motivated, and organized—please reach out. ABOUT THE OPS JOB

—

Starting to make massive progress on the Substrate project. And I am super empowered by my upgraded Kai system to do this. This, along with the PAI project, are two of the projects I’m most proud of and invested in on Github. Fabric was wonderful, but it’s now contained within PAI. And any time I want to document something for a blog post or video or whatever, I’ll do research and put it in Substrate. SO EXCITED about this. THE SUBSTRATE DATA DIRECTORY | THE GIST FOR GLOBAL KNOWLEDGE WORKER COMPENSATION

Let me please stress this.

Kai made this for me using the Research Skill and conduct-research Command that I have published in the PAI directory. He made this in like 3 minutes. The entire thing.

And it’s all open source now. Free (other than the API keys for Perplexity). Go get it and use it.

The combination between PAI and Substrate and all these different Skills, I’m telling you it’s just unbelievable.

—

Had a BRILLIANT conversation with Neatsun Ziv at Ox Security. We talked about his extraordinarily unique approach to bringing security directly to the developer through context injection using agents. Fascinating conversation. You got to check it out. ^SPONSORED

👆🏼👆🏼👆🏼 Got a private demo of this product and it was seriously one of the most impressive things I’ve seen in years. Why? Because of how many problem spaces it addresses simultaneously. I was literally making a list of products that it does a better job of! Kai, find me a way to invest.

Hijacking Claude Code with an invisible line in a PDF More showing of the fact that you have to be careful with MCPs and extra careful with Skills. I mean, this is really a dangerous parsing issue, but the fact that these skills are about to go viral is something we really need to be concerned about. THE ARTICLE

Microsoft's highest ever severity score goes to ASP.NET Core vulnerability SECURITYWEEK ARTICLE

Over 266,000 F5 BIG-IP devices are sitting ducks after China breach Nation-state hackers stole unpatched F5 vulnerabilities and source code, and now a quarter million exposed instances need emergency patching by October 22. BLEEPCOMPUTER ARTICLE | SHADOWSERVER FOUNDATION | CISA EMERGENCY DIRECTIVE | F5 SECURITY ADVISORY

CISA says patch Microsoft stuff now because it's being actively exploited CISA added five vulnerabilities to their Known Exploited Vulnerabilities catalog, including CVE-2025-59230 that's getting hit in the wild right now. CISA ALERT | CVE-2025-59230 DETAILS

Russia-linked Qilin ransomware gang hit Texas electric cooperatives The hackers grabbed incident reports, budgets, insurance files, and personal info from board members and employees at multiple co-ops. OODALOOP REPORT

Prosper breach hits 17.6 million people with everything from SSNs to browser fingerprints Have I Been Pwned says the peer-to-peer lender got absolutely pillaged in September, leaking government IDs, income levels, and even user agent details. HIBP PROSPER BREACH PAGE | PROSPER INCIDENT RESPONSE | THE REGISTER ARTICLE

Chinese hackers quietly spent five months inside a Russian IT provider's network Jewelbug had access to code repositories and build systems, meaning they could've poisoned software updates to hit multiple Russian customers at once. People forget how much tension there has been between these companies in the past, and it’s clear that China doesn’t care if they share enemies. THE HACKER NEWS ARTICLE | SYMANTEC THREAT INTELLIGENCE REPORT

A developer almost ran malware disguised as a coding interview from a fake blockchain company David Dodda caught sophisticated malware hidden in a job interview coding test—which was an obfuscated payload that would've stolen crypto wallets and credentials if he hadn't asked his AI to scan it first. I don’t even run code from not-fake blockchain companies. DAVID'S BLOG POST | VIRUSTOTAL ANALYSIS

CISOs brace for an AI vs AI fight CSO ARTICLE

Chinese scam gangs made over a billion dollars from fake toll and postage text messages Those annoying texts about unpaid tolls or delivery fees aren't just spam—they're part of Chinese criminal operations that steal credit cards to buy gift cards and luxury stuff. Great to finally see the infra behind all this. WSJ SCAM TEXT ARTICLE

China says NSA used 42 cyber tools to attack Beijing's time system 42, huh? China's MSS claims it stopped a multi-year NSA operation targeting the National Time Service Center, which could've wrecked financial systems and space launches. MSS WECHAT POST | NTSC ABOUT PAGE | HACKERNEWS ARTICLE

US Army plans to crank out 10,000 drones monthly starting next year The Army's new SkyFoundry program will use innovative manufacturing to domestically mass-produce small drones at scale by 2026. I’m sure China’s doing some multiple of that already. Well I hope they aren’t. Just so annoyed at how far behind we are. It’s really hard to compete with someone who has aligned their government, businesses, and population towards a shared set of goals. Especially when we can’t even agree if the West is even a good thing or not. OODALOOP ARTICLE

John Bolton indicted for sharing classified info with family DW ARTICLE

Army general uses ChatGPT to help make command decisions An Army major general says he's using AI chatbots to help inform personal and organizational decisions for the Eighth Army in South Korea. DEFENSESCOOP REPORT | ARS TECHNICA ARTICLE

NATO defense ministers discuss drone wall amid Russian airspace tests Russia's probing NATO airspace with mysterious drones, so the alliance is working on low-cost defense systems like Ukraine uses instead of expensive missiles. NATO DRONE DEFENSE MEETING

A tool that finds which AI agents can replace your job Can't tell if this language is a joke or not, but the site lets you search your job title and shows you exactly which AI agents are already doing that work today. REPLACEMENT.AI | HN DISCUSSION

Adding one sentence to prompts makes AI models way more creative Researchers found that just asking models to "generate 5 responses with their corresponding probabilities, sampled from the full distribution" increases creative output diversity by over 2x. I’ve incorporated this into my be-creative Skill. Will share soon. VERBALIZED SAMPLING SITE | ARXIV PAPER | GITHUB REPO | WEIYAN SHI ON X

Anthropic launches Agent Skills so you can teach Claude domain-specific expertise What I've been talking about constantly here. Anthropic created Agent Skills—basically onboarding guides stored as folders with instructions and scripts—so Claude can dynamically load specialized knowledge without cramming everything into context.

The story here is really that it's extremely shareable work automation, very similar to my post on business AI being about intelligence pipelines. ANTHROPIC SKILLS ANNOUNCEMENT | AGENT SKILLS DOCS | SKILLS COOKBOOK | ANTHROPIC SKILLS GITHUB | CLAUDE CODE

Some things of note:

An argument that OpenAI has lost its unquestioned dominance to Anthropic and Google David Shapiro argues Anthropic now owns 40% of enterprise AI while OpenAI dropped to 20%—turns out safety equals reliability, which enterprises actually pay for. DAVID SHAPIRO ON SUBSTACK | FULL ARTICLE

Google's Veo 3.1 makes AI videos way more realistic and accurate Google's new video model does better prompts, adds portrait mode, and includes audio—basically making it even harder to tell what's real online. Thanks. GOOGLE VEO 3.1 ANNOUNCEMENT | ARS TECHNICA ARTICLE

Claude Haiku 4.5 matches Opus 3.5 performance at faster speeds Anthropic's new Haiku beats their previous flagship on most benchmarks while staying cheaper and quicker than GPT-4o mini. CLAUDE HAIKU 4.5 ANNOUNCEMENT | HACKER NEWS DISCUSSION

cmux – Coding Agent Multiplexer CMUX GITHUB

Free AI coding tools work fine unless you're sprinting hard on big projects David Gewirtz used free ChatGPT for casual coding but paid $200/month for ChatGPT Codex when doing a four-day sprint that replaced four years of work. ZDNET ARTICLE | DAVID GEWIRTZ TWITTER | ADVANCED GEEKERY NEWSLETTER

TSMC accelerates Arizona 2nm chip production by a year due to AI demand TSMC's CEO says they're speeding up 2nm production in Arizona and buying more land to build a massive self-contained gigafab cluster there. NIKKEI ASIA ARTICLE | TOM'S HARDWARE STORY

Nvidia and TSMC made the first Blackwell chip in America. NVIDIA BLACKWELL ANNOUNCEMENT

Git-stars creates a newcomer ranking that surfaces genuinely new repos instead of the same old projects Someone built git-stars.org to fix GitHub Trending's problem of recycling established repos by only showing repositories that are both brand new and rapidly gaining stars. GIT-STARS NEWCOMER RANKING | HN DISCUSSION

SpaceX's classified Starshield satellites are transmitting on frequencies meant for receiving commands, not sending data Amateur satellite tracker Scott Tilley accidentally found 170 classified defense satellites downlinking on uplink-only frequencies, which breaks international standards and could mess with other satellites. NPR STORY | SCOTT TILLEY TWITTER | TILLEY'S DETECTION REPORT | ITU FREQUENCY STANDARDS | STARSHIELD INFO

Waymo partners with DoorDash to deliver groceries in Phoenix using robotaxis Waymo's testing delivery again after closing its truck program, but customers have to walk outside and grab stuff from the trunk themselves. WAYMO BLOG POST | TECHCRUNCH ARTICLE

Waymo is launching robotaxis in London in 2026, its first European service Waymo's electric Jaguars start testing with safety drivers in weeks, then go driverless next year once UK approvals finalize. WAYMO ANNOUNCEMENT | WAYMO LONDON SITE

China's rare earth export controls now reach beyond its borders to control global products China isn't just restricting rare earth exports anymore—they're requiring export licenses for any product worldwide containing over 0.1% Chinese rare earths by value, even if made elsewhere. CHINA'S RARE EARTH EXPORT CONTROLS | CHINA EXPANDS RARE EARTH CONTROLS | NEW CURBS HIT 14NM CHIPMAKING | TRUMP'S 100% TARIFF RETALIATION | CHINA'S APRIL 2025 RESTRICTIONS

Pentagon reporters walk out en masse rather than sign new restrictive rules Dozens of journalists turned in their badges after Defense Secretary Pete Hegseth imposed rules that would let him expel reporters for seeking any non-approved information. AP NEWS STORY

More than half of US car buyers would now consider buying a Chinese car Despite privacy worries, which few people actually care about, 51 percent of Americans shopping for cars say they'd consider a Chinese brand, up from 41 percent last year. ARSTECHNICA ARTICLE

US passport drops out of top 10 for the first time in 20 years The US passport fell to 12th place with access to only 180 destinations while Singapore leads at 193, driven by Brazil revoking reciprocity and China excluding America from visa-free expansions. HENLEY PASSPORT INDEX | HENLEY GLOBAL PRESS RELEASE

Antibiotic resistance hit one in six infections in 2023 and it's accelerating WHO data shows nearly 40% of common antibiotics lost effectiveness in just five years, with lower-income countries getting hit hardest as resistance outpaces new drug development. WHO RESISTANCE REPORT | KEVIN IKUTA UCLA | RAMANAN LAXMINARAYAN | WHO AMR FACT SHEET

China has overtaken America in green energy manufacturing and it's basically over Paul Krugman says China now dominates solar, batteries, and EVs so completely that Western countries can't realistically compete on manufacturing anymore. But he also said the Internet wouldn't be a thing, so hard to say. PAUL KRUGMAN ARTICLE | HN DISCUSSION

Your brain navigates new places by rapidly flickering between reality and old mental maps Yale researchers found the hippocampus doesn't just memorize routes—it darts between processing what's actually around you and recalling similar past environments to build flexible cognitive maps. NATURE COMMUNICATIONS PAPER | GEORGE DRAGOI PROFILE | YALE MEDICINE ARTICLE

Ripgrep 15.0.0 released with speed improvements RIPGREP 15.0.0 RELEASE

You are the scariest monster in the woods Jamie says humans forget we're apex predators—wild animals are terrified of us, not the other way around. So it’s us, not AI we need to worry about. JAMIE'S ESSAY | HN DISCUSSION

Hakkernieuws strips HN down to just stories and favicons HAKKERNIEUWS SITE

Cmux lets you run multiple AI coding agents in parallel workspaces. CMUX GITHUB

Things I've learned in my 7 Years Implementing AI JAMPA'S ARTICLE

Cloudflare built a sandbox SDK that lets you test Workers locally without deploy cycles Cloudflare's new SDK spins up a local Workers environment so you can iterate fast without pushing to production every single time. Can't wait to try this out. CLOUDFLARE SANDBOX SDK | HN DISCUSSION

Attention is a luxury good SETH'S BLOG POST

Marcin Jakubowski built an open-source civilization toolkit you can make with local materials Physicist Marcin Jakubowski created 50 modular machines—tractors to 3D printers—designed like Lego blocks so anyone can build, repair, and reconfigure them without corporate control. CIVILIZATION STARTER KIT ARTICLE | OPEN SOURCE ECOLOGY | FUTURE BUILDERS ACADEMY

Marcin Jakubowski built an open-source civilization toolkit you can make with local materials CIVILIZATION STARTER KIT ARTICLE

The man building a starter kit for civilization CIVILIZATION STARTER KIT ARTICLE

How Useless Are You tests your actual life skills HOW USELESS ARE YOU SITE

Your data model is your destiny Matt Brown argues the biggest startup moats now come from what you emphasize in your product—Slack's channels, Toast's menu items, Rippling's employee records—because AI commoditizes code but can't refactor organizational reality customers build around your architecture. MATT BROWN'S ARTICLE | MATT BROWN | MATRIX PARTNERS

MCP Snitch monitors and blocks sketchy AI tool calls before they access your files Adversis built a macOS app that sits between Claude or Cursor and MCP servers, using AI to flag when tools try touching SSH keys or system directories. MCP SNITCH GITHUB | RELEASES PAGE | CONTRIBUTING GUIDE

Find a project like PAI to use for all the stuff you care about. Life, work, etc.

I recommend PAI, obviously, which is mostly based around Anthropic / Claude Code, but it doesn’t really matter which you use. And Skills and Commands and Prompts and such are pretty universal across platforms.

But this is the augmentation we need as humans to be competitive with AI.

THE PAI PROJECT