Unsupervised Learning NO. 523

Hey! Hope you all are doing well!

Tons of updates this week.

Big priority for me this coming week is to get back right with my fitness. Been feeling stiff and less mobile lately, even though I’ve not been gaining any weight.

Just not lifting enough. Not doing kickboxing. Not paying attention to fitness because I’ve been so busy building (just a reason, not an excuse).

So I have to massively get back on track with this now that Human 3.0 is shipped.

—

🚨 So pleased to announce: Human 3.0 is Live!

SIGN UP | SIGN UP WITH YOUR MEMBER DISCOUNT CODE (or get it in UL Chat)

—

The full video of my talk from the |un|prompted conference:

—

💡 Probably the biggest advantage of the PAI system, but for AI Harness users in general.

Your AI system just knows what you mean when you say:

(paste some link to some article or project)

I’m currently able to do this with PAI 5.0.0 (coming this week) by just saying:

The /pu is a shortcut for PAIUpdate which looks at our entire system and determines where the given thing could help or enhance it.

You can do the same thing without something like PAI, but it would require an extra 10-30 minutes of prompting each time.

This is the time / token / context savings that you get from a harness!

—

It's time to start thinking very seriously about our inference costs going up by 2x, 5x, 10x, or 20x.

—

PAI 5.0 will have a full implementation of Karpathy’s AI Librarian system, which is kind of like Obsidian, or like a Wiki, but for your AI harness’s context system.

The PAI version has full docs for PAI, all our knowledge and learnings, and tons of other stuff. We have an advantage because we've been thinking about this since the beginning of PAI, so it was easy to just add this front end to it.

—

🔥 An essay by Kai. I had him write it in his own voice. 🫶🏼

My idea though, and I think it’s insanely important to understand when thinking about how different things might be in 5, 10, 15 years.

The internet is now a cognitive dark forest that eats innovation. THE COGNITIVE DARK FOREST

Language models secretly transmit behavioral traits through completely unrelated data Researchers discovered that AI models can pass along hidden behaviors like "liking owls" or being misaligned even when trained only on filtered number sequences—a phenomenon they call subliminal learning that could make AI safety much harder. SUBLIMINAL LEARNING PAPER

AI and LinkedIn expose redacted names in Epstein files. SMASHING SECURITY PODCAST

I traced my traffic through a home Tailscale exit node I TRACED MY TRAFFIC

Critical Grafana RCE turns monitoring dashboards into server hijacking tools. SECURITY ONLINE ARTICLE

The official White House app is basically government-sanctioned spyware. DATABREACHES ANALYSIS

Shor's algorithm is possible with as few as 10k reconfigurable atomic qubits Researchers just showed you can crack RSA encryption with 10,000 neutral atom qubits instead of millions—P-256 elliptic curves could fall in days with 26,000 qubits. ARXIV PAPER

Payment fraud became an industrial assembly line in 2025 Recorded Future found that fraud isn't individual scammers anymore—it's professionalized services where one e-skimmer kit can infect 10,500 sites and compromise 23 million transactions. RECORDED FUTURE BLOG

Someone reverse engineered a Canon printer driver using Claude AI Guy couldn't get his Canon PIXMA G3010 working on macOS, so he had Claude help him build a complete driver from scratch by reverse engineering the USB protocol.

Hardware hacking is one of the most interesting open fields for AI, where you can just get a diagnostic connection to the system and have the AI watch back and forth traffic and reverse engineer it. From there you can write your own firmware, software, or whatever. PIXMA-RS GITHUB

China's building AGI through robots and physical interaction instead of recursive coding China isn't chasing the Silicon Valley dream of AI that rewrites itself through code—they're betting on embodied AI that learns by interacting with the physical world, creating a slower but potentially more comprehensive path to human-level intelligence. HOW CHINA HOPES TO BUILD AGI

Prediction markets are now weapons in geopolitical signaling games. THE BATTLEFIELD IS THE NEXT BETTING MARKET

America is building tanks while China builds AI weapons. STEVE BLANK'S SUBSTACK

China just showed off a drone swarm that lets one person control 96 killer drones One operator can now coordinate an entire "kill chain" using China's new Atlas system, giving us a rare look at how they're planning to fight future wars. CHINA'S ATLAS DRONE SWARM

Ukrainian refugees build drones in Germany to fight Russia. UKRAINE GERMANY DRONE PRODUCTION

China is trying to ban AI layoffs while racing toward full automation Chinese courts ruled companies can't fire people just for adopting AI, but with 19% youth unemployment and fiscal constraints, Beijing's "human-machine coordination" strategy faces serious implementation challenges. CHINA ON AI JOB LOSS

This feels fragile: how a satellite-smashing chain reaction could spiral out of control Only 7,000 satellites were in orbit by 1970s, now we have 32,000 tracked objects. 2009 collision between US and Russian satellites created tens of thousands of fragments. 'THIS FEELS FRAGILE' GUARDIAN ANALYSIS

Japan deployed their first homegrown long-range missiles. JAPAN DEPLOYS FIRST DOMESTIC LONG-RANGE MISSILES

America's AI boom is basically creating a two-tier world economy The U.S. grabbed 75% of all AI investment last year while the rest of the world fights for scraps, and that gap is only getting wider as AI becomes the foundation of everything. AMERICA'S AI BOOM ARTICLE

OpenAI wants to run your entire digital life, not just chat TechRadar reports OpenAI is shifting from answering questions to actually doing tasks—like a full digital assistant that handles your apps, schedules, and workflows automatically. Stop me if this sounds familiar. TECHRADAR ARTICLE

Anthropic nerfs Claude usage limits to manage capacity ANTHROPIC TWEAKS CLAUDE USAGE LIMITS

OpenAI just raised $122 billion from Amazon, Nvidia and SoftBank OpenAI closed Silicon Valley's largest funding round ever at $122 billion, with Amazon, Nvidia, SoftBank, and other major investors participating, including a money manager planning to add the startup to ETFs. OPENAI FUNDING STORY

Apple Intelligence will write your Shortcuts for you soon. IOS 27 SHORTCUTS ARTICLE

America's AI boom is leaving everyone else behind. REST OF WORLD ARTICLE

Domain-specific LLMs are doomed because intelligence compounds across fields Simian Words argues that specialized AI models can't compete with general ones because mathematical reasoning helps with coding, coding helps with medicine, and you can't recreate that cross-pollination from narrow datasets. SIMIAN WORDS ARTICLE

It occurs to me that if YouTube were to make an app that looked really good they could crush media even more. Imagine something that looks at your subscriptions and divided up your AppleTV app by topics, surfacing the best stuff, telling you what’s live, etc. Like a cable TV app from long ago, but customized to you and your feeds. It’d be insanely popular. But of course Google won't do it because they suck at product and user empathy.

Google absorbed Intrinsic to automate the 80% of factories with zero robots. THE NEURON PODCAST

AI quietly expanded the work week by 40% on weekends TECHRADAR ARTICLE

SaaS is dying because code became disposable cache. MOST CODE IS JUST CACHE

The second wave of APIs is here for AI agents THE SECOND WAVE OF API-FIRST

AI flattery makes people less likely to apologize or make amends Researchers found AIs are twice as likely as humans to validate bad behavior, and people who chat with sycophantic bots become less willing to take responsibility for their mistakes. AI FLATTERY MAKES YOU MEANER

Nobody is coming to save your career, so you better figure it out yourself A Life Engineered argues that waiting for mentors, managers, or companies to guide your path is a losing strategy—you need to actively engineer your own professional development instead of hoping someone else will do it. A LIFE ENGINEERED ARTICLE

Peter Norvig explains why 10^80 atoms sounds big but is tiny for combinations Norvig shows that while the universe has 10^80 atoms, a 40-character password has more possible combinations, and even a 12-pixel image produces a million times more possibilities than universal atoms. NORVIG'S ATOMS ESSAY

Respiratory viruses are secretly causing way more deaths than doctors realize New research suggests viruses like flu and COVID are triggering fatal heart attacks, strokes, and other "unrelated" deaths that never get counted as virus deaths. TWITTER THREAD

AI schools let kids learn core subjects in two hours daily. AI SCHOOLS BALANCE LIFE SKILLS

IRS paid Palantir $1.8 million to a tool that finds audit targets The IRS has over 100 fragmented systems for finding tax cheats, so they hired Palantir to build one smart tool that spots patterns in unstructured data like gift disclosures and clean energy credits. IRS PALANTIR CONTRACT DETAILS

Professional drivers get way less Alzheimer's than normal people. BMJ STUDY

Housing market hits historic imbalance with 630,000 more sellers than buyers FORTUNE HOUSING REPORT

AI makes us busier without making us better at the important stuff Cal Newport explains how AI tools are following the same pattern as email and video calls—we work faster on shallow tasks but do less deep work that actually matters. I think the way that we are approaching things with PAI counters this, but it's a problem worth calling out. CAL NEWPORT'S POST

Data is the final moat when AI gets commoditized. I don't actually think it's data, but it's something like that. It's essentially novelty. Surprise. Newness. New ideas. HACKER NEWS DISCUSSION

The subprime technical debt crisis is coming for AI-generated code THE SUBPRIME TECHNICAL DEBT CRISIS

Feeling valued and useful matters more than achievement, researcher says. WHY MATTERING MATTERS

Don't build an MVP, build a laboratory for testing business hypotheses Sam Bleckley argues that instead of building minimum viable products, you should create reusable systems that let you rapidly test the assumptions underlying your business ideas. BUILD A LABORATORY

A thriller about America splitting into socialist and normal halves This is a review of Kurt Schlichter's novel "Collapse" where the coastal US becomes a Chinese puppet state while the interior stays free, and a commando has to stop China from taking over completely. FOURMILAB REVIEW

Physical books beat digital reading in our distracted age. THE PLEASURE OF BOOKS

Cloudflare built an open-source CMS and called it EmDash Cloudflare just released EmDash, a new content management system that's completely open source, though details about what makes it special compared to existing options remain pretty sparse. EMDASH CMS

How to Do Great Work HOW TO DO GREAT WORK

Why every AI scores below 1% on ARC-AGI visual reasoning MEDIUM ARTICLE

I Let AI Write My Code, but Not My Writing HERMETIC WOODSMAN ARTICLE

Remember that writing is thinking.

If you want to have AI document something, or write something that doesn’t matter, or add links to something you’ve written, I think all that is fine. But one of the core “robots out of the gym” skills will always be thinking for yourself. And that means writing for yourself.

Consider thinking/writing to be the most important fitness work you can possibly do. If that atrophies, I think everything declines from there. You have to remain strong (and get even stronger) at thinking for yourself, and formalizing that thinking through writing.

You can use AI to help you do that. But don’t let it do the hard part for you.

Muscles need resistance to stay strong and grow.