Unsupervised Learning NO. 524

Hey! Hope you all are doing well!

Here’s my new video with David Bombal about PAI!

—

I’ve been experiencing a feeling/idea that’s kind of massive and hard to pin down, but here goes.

If you remember my whole piece on Constraints on Creativity, and now this new one on How to Avoid Aperture Collapse, the idea is the same:

I am constantly catching myself being far too cautious, too tentative, thinking too small, forgetting the big picture, and just generally not thinking about what is possible and chasing it aggressively…

- - - - - - - - - - - - - - - -

(snipped, ended up turning it into an essay here) 👇🏼

—

Biggest news in my circles is that Claude Code (or Opus, or both) seem to be massively failing people. So many of the people I know who were only Opus-based have now moved to Codex and GPT 5.4.

I've recently had some really bad experiences as well, which are much less obvious because of PAI, but I think I can actually tell the difference. This is causing me to maybe accelerate my work on P2 (PiPai) even more so that I can pivot PAI to different models as needed. Really would prefer not to, but it appears our hands are being forced between the limits running out so fast combined with the quality drop. : (

—

I finally completely replaced Zapier.

For those who care, it was a webhook and a bunch of AI flows that went through it. But one main one. Sad it took me this long.

—

A somewhat depressing / negative look at the current state of things, but I think accurate and more importantly: useful.

This is generally true for most:

Just keep in mind the reason humans are in such bad shape here is not because we lack the potential to be the 9 or 10. It's just that we have so many weaknesses because of our animal nature (bad incentives like money and power) that they inevitably get in the way of progress.

—

ClickFix now skips Terminal by tricking you into Script Editor ClickFix social-engineering attacks for macOS now use the applescript:// URL scheme to auto-open Script Editor with a ready script that downloads and runs malware, while the page pushes a fake disk-cleanup dialog. MALWAREBYTES ARTICLE | JAMF RESEARCH BREAKDOWN | ATOMIC STEALER DETECTION

Chrome’s device-bound cookies aim to kill session hijacking Google is rolling out Device Bound Session Credentials (DBSC) in Chrome 146 on Windows, tying session cookies to hardware keys so stolen cookies can’t be used elsewhere.

This has always been a major attack vector. Compromise person X, use their cookies to move all over the place and pull data. So I love seeing this. DEVICE BOUND COOKIES DETAILS

MITRE’s Fight Fraud Framework adds fraud-specific tactics to cyber defense. MITRE Fight Fraud Framework

AI governance gets real when pipelines explode and shadow AI shows up THE MESSY TRUTH

State propaganda can sneak into AI training data via Common Crawl A new DFRLab audit finds Pravda, Glassbridge, and RT content inside Common Crawl, and shows some can be reproduced by a large model. THE REPORT

US AI companies unite to stop Chinese model copying. BUSINESS TIMES ARTICLE

FBI tallied $893 million in AI fraud losses for 2025. FBI IC3 REPORT

China's AI harvests social media for US military intelligence during Iran war. WASHINGTON POST ARTICLE

Stripe engineers explain how to build real AI guardrails. UNPROMPTED 2026 TALK

BBC interview shows Iran’s slick Lego-style AI propaganda machine The BBC talks to a man running Explosive Media, who admits Iranian officials pay them to spread viral AI memes to Western audiences. IRAN WAR PROPAGANDA AI VIDEOS

Ukraine tips drone war in its favor. UKRAINE TIPS DRONE WAR IN ITS FAVOR

Russian SOF doctrine turns Iran satellites into faster missile hits. EUROMAIDAN PRESS ARTICLE

China’s AI labs move toward closed frontier models because open-source budgets don’t make enough money. CHINATALK

Wait-mode drones can stay lethal for over a day, advisor says. UKRINFORM PAGE

US banned Chinese spy cameras so China switched to surveillance drones Stephen Bryen argues that the dramatic rise in military base drone incursions directly correlates with America's 2019 ban on Hikvision cameras, as China pivots from fixed surveillance to mobile aerial spying. WEAPONS AND STRATEGY

Claude’s reasoning got throttled 67%, and devs noticed A leaked analysis of thousands of Claude Code sessions shows Opus 4.6 “thinking depth” fell about 67%, plus users report lazier code edits. CLAUDE 67% DUMBER ARTICLE

Anthropic might build its own AI chips to cope with shortages. DATA ECONOMY

OpenAI wants workers to share AI wealth in a new social contract OpenAI released a policy roadmap that argues AI should create shared rewards for workers, including taxes, benefits, and a public wealth fund. OPENAI POLICY ROADMAP | OPENAI NEXT PHASE AI | ALTman TALKS TO AXIOS

OpenAI cuts Codex-heavy Pro to $100 and reshuffles tiers OpenAI now offers a $100/month Pro plan aimed at people who run long Codex coding sessions, with Plus usage spread out weekly. THE DECODER AUTHOR | OPENAI CODING PRICING | OPENAI PRO PRICE UPDATE

AI is blasting China’s microdrama pipeline while regulators tighten the screws AI is speeding up microdrama production in China, while Beijing’s rules force approvals and curb risks from actor likenesses. HELLO CHINA TECH

OpenClaw only “works” until its memory breaks under real tasks. I don't know anybody using OpenClaw the way that people used it when it first came out. It's actually extremely expensive and not great for long-term use due to fragility / security (although those are improving). I do give the project massive respect, though, because it ignited something within the AI community. Nishant Soni

Inference costs push you toward multi-model + local routing In a few hours, Claude Code can burn a quarter of a MAX subscription, so Anthropic isn’t “hustling”—they’re pricing inference closer to reality, forcing hybrid model strategies and model-routing harnesses to survive Daniel Miessler | INFERENCE COSTS ESSAY

AI isn't just a feature anymore, it's becoming the entire operating system Private equity firms are no longer asking "what's your AI strategy" but "how far along are you in rebuilding the company around AI". Bolt-on AI features lose to companies that redesign everything from the ground up. AI IS BECOMING AN OPERATING SYSTEM LAYER

Ten principles that actually make engineering teams work better A practical field guide to software engineering principles that focuses on real problems like context switching, communication overhead, and technical debt accumulation. TEN PRINCIPLES GUIDE

Oxford AI reads invisible heart fat changes to predict failure risk Oxford trained an AI on routine cardiac CTs to spot subtle pericardial fat texture changes and predict heart failure up to five years later with 86% accuracy. UNIVERSITY OF OXFORD | MPPOST ARTICLE | JOURNAL LINKED OUTCOME | AI TAG | ARTIFICIAL INTELLIGENCE TAG | TRUST PROJECT GUIDELINES

Restaurants now want your phone along with your coat. MORNING BREW ARTICLE

Your most valuable work is invisible to employers. DEAD NEURONS ARTICLE

AI agents are just spam unless you make attention expensive Seth argues that door-to-door sales works because effort is scarce, but AI can scale pitches cheaply and burn trust fast. ATTENTION AND EFFORT

The median take is taken LESSWRONG

Why AI doesn’t instantly change everything. DAVE GRIFFITH’S SUBSTACK

Recording everything you do beats writing for AI immortality LessWrong argues the "write online to live forever through AI" idea is bogus, but recording all your computer activity to help current LLMs is already paying off. DON'T WRITE FOR LLMS

Thorsten doesn't love typing code anymore, just building things. JOY AND CURIOSITY NEWSLETTER

Check out UL’s close friend Tim Leonard’s community, DETSEC. Founded in 2017, Deep East Texas Security (DETSEC) is a collective of IT professionals, cybersecurity practitioners, makers, hackers, and enthusiasts. We started with a focus on information security and have grown into a community that spans cybersecurity, artificial intelligence, radio communications, home automation, hardware hacking, and physical computing platforms like Arduino and Raspberry Pi. We recognize that technology touches everything — and that understanding it, building with it, and securing it are skills worth sharing. DETSEC HOMEPAGE

slop-browser turns your real Chrome into a CLI-controlled agent It’s a Chrome extension + CLI that drives the browser from inside, so agents avoid CDP and automation fingerprints. This is made by a brilliant friend of mine, Ron Eddings. And along with agent-browser I’m actually running it in my PAI setup, unlike dozens of competitors. SLOP BROWSER

My friend Andrea responded when I asked why we’re using Obsidian when we can just build Markdown-based context, and she shared her Markdown-based context system that I LOVE the structure of. The repo gives you a plain-markdown memory store for agents: append daily logs, keyword-search with scoring, and consolidate logs down to fit context windows. Definitely check this one out. AGENT CONTEXT SYSTEM

Ghost Pepper is a fully local hold-to-talk transcriber for macOS Ghost Pepper GitHub repo

Competence is lonely and nobody talks about why SPACE DAILY ARTICLE

Advice to Young People, The Lies I Tell Myself ADVICE TO YOUNG PEOPLE

Keychron shared source files for keyboards and mice PCGAMER

Start preparing for what happens if/when intelligence becomes expensive again.

Start thinking about harnesses with maximum model flexibility, including local models, multi-vendor setups, etc.

Might not happen. But if it does, you’ll want to be ready.