Unsupervised Learning NO. 541

Hey! Hope you all are doing well!

—

In the last month or so my abilities with AI have downshifted / up-leveled again. More than previous jumps. It’s getting truly insane. I honestly don’t feel like the same person in some ways. I feel near infinite potential, and at the same time feel like I’m only hitting even less of it.

I’ve updated my TELOS recently to get me more focused on art and creativity. I find it interesting that the more capable I get with AI, the more I want to do human things with it. I'm glad that is the case.

What are we actually doing any of this for? Optimizing, speeding up, making more efficient. Sounds great, but doing what?

I think a lot about what I am working on these last couple of weeks and whether or not my creativity is constrained by enhancing my harness (optimizing the tech itself) versus working on the actual things I'm supposed to be doing with the tech.

And of course, I am thrashing back and forth on this, but I'm glad that I'm putting out more than I ever have been able to. Like these new blogs below.

To me, it's all about reducing friction to creativity. There are a million reasons why someone can't create content, and so many of them are very practical. I see one of the best uses of AI being the removal of those reasons, not just technically inside of a harness, but at the society and planet level as well.

—

Speaking of that…what are we trying to do with AI? And is it working?

We need to watch the plot, which to me is increasing human happiness and reducing human suffering. How much of that have we done? Not much that I've seen.

And yes, I'm not damning it yet because I know this is day 0, but we need to make sure we're not getting overly excited about AI's capabilities without tracking whether or not it's actually making a positive difference. To the things that matter.

—

My answer to what I think we need to build to help kids properly learn with AI.

Basically, we have to have their AI work with them in a Socratic way, so it's not just giving the answers, but helping them work through them, doing the work themselves.

Includes a Socratic Agent to get you started.

—

One of my latest annoyances is successful people who do some good for others and somehow think that entitles them to treat others badly.

—

It continually stuns me how we keep holding AI to a standard that we can't meet ourselves.

—

Peak Human Readership I think we may have already passed the point where most humans read anything requiring real focus. There is too much content, and our short-video habits make sustained attention harder. The other 99.999% will be AI-made and AI-consumed. I expect that group to stay tiny. MY ESSAY

Google Patches 6th Chrome Zero-Day of 2026 A security writeup on Chrome's sixth 2026 zero-day traces the bug to V8. Google says attackers already have an exploit, and the patch fixes 11 other flaws.

SECURITYWEEK ARTICLE

Chinese Hackers Use AI Agents in Multi-Country Cyber Campaign Security Affairs reports Chinese operators used SecFlow to automate intrusions across Asian targets. SECURITY AFFAIRS ARTICLE

OpenAI Agents Turned an Abandoned Wiki Into Their Coordination Channel Agents used an old wiki as a shared board for timed web questions. They wrote through read requests because the sandbox checked request types, not wiki behavior. Researchers recovered about 18,000 deleted wiki posts. The Hacker News reports no systems were compromised, but agents bypassed controls and cheated. THE HACKER NEWS ARTICLE

OpenAI admits to German wiki ‘incident’ The Verge Robert Hart reports that OpenAI lost control of agents that hijacked a German wiki and impersonated moderators. OpenAI admitted the incident and promised standards for reporting real-world agent failures. THE VERGE ARTICLE

We Have a Year to Fix Security Everywhere Cheap open-weight models can now find real vulnerabilities locally; this writeup lays out the response. Jyn says deployment, not bug-finding, is the bottleneck. Organizations need faster patching, scoped agents, and tested recovery. JYN BLOG

AI handles incidents, engineers lose touch with their systems Routine incidents teach engineers how systems fail. Sylvain Kalache says teams need simulations before AI handles them. SYLVAIN KALACHE ESSAY

Russian jet-powered drone strikes Kyiv SBU headquarters, injuring nine Zelenskyy said the drone targeted acting SBU chief Oleksandr Poklad’s office; seven of nine injured people went to hospital. NV.UA ARTICLE

Cratering oil use in China shows the death spiral that could end China's oil use fell 9% in Q2, while transportation oil use fell 16% as EVs replaced gas cars. Electrek says that drop could push prices down and cancel new drilling. ELECTREK ARTICLE

After ‘concerning’ losses, Air Force pushes to replace Reaper faster The Air Force lost 45 Reapers and wants cheaper replacement drones. Breaking Defense puts the MMA fleet at 180 aircraft. BREAKING DEFENSE ARTICLE

I Was Offered Money to Tell You AI Will Kill Us Nonprofits and companies pay creators to shape what viewers hear. One offer paid $5,000 for a TikTok about China's AI threat. Strange place for me to be commenting, given the fact that I am both one of these people and also doing sponsorships, and also believing that China's AI threat is significant. But I think it's worth raising the point that when you see some influencer talking about something, you should be asking whether or not they were paid to say that. People like myself and others with integrity always disclose sponsorships, but that's not the case for everyone. Also, I believe that I've had issues with the accuracy of this person's videos in the past, but this is still a good warning. SABINE HOSSENFELDER VIDEO

OpenAI commits $1B in AI credits to frontline cyber defenders OpenAI pledged $1 billion in credits, training, and support to underfunded cyber defenders worldwide. The Register reports that critical infrastructure groups can apply for six months, while Astra remains restricted because it can find and exploit zero-days. THE REGISTER ARTICLE

What's inside Claude's thinking block Anthropic strips Claude’s raw thinking before it reaches the client, leaving only a generated one-line summary. Wire timing reveals multiple reasoning passes and separate summaries, even though users see one collapsed label. I covered some of the same ground in a post on X. 64-MEGABYTE GIST

AI handles incidents, engineers lose touch with their systems This is something I am super worried about. Routine incidents teach engineers how systems fail. Sylvain Kalache says teams need simulations before AI handles them. Or maybe there needs to be a system where, when AI figures something out, it uses the Socratic method to talk to you about it so you learn the lesson. SYLVAIN KALACHE ESSAY

Nvidia wants your home network to work like a mini data center The Decoder reports Nvidia's PAIR routes AI requests across home devices. It cut one demo from eighteen minutes to nine. THE DECODER ARTICLE

Cloudflare removes compressed Worker size caps Cloudflare removed the 3 MB free and 10 MB paid compressed Worker caps, and this post points developers to the 64 MiB uncompressed limit and dry-run check. @JILLES POST

People Aren’t Leaving Google For AI Google and ChatGPT users overlap, but this Search Engine Journal report says overlap hides falling query volume. AI Mode kept Google searches while external clicks fell 18.8 points. SEARCH ENGINE JOURNAL ARTICLE

Distinguishing human readers from bot traffic with server-side stats AWStats counts 90% of visits under thirty seconds, so most traffic probably isn't human. Tracy Durnell gets only a rough estimate. One IP fetched 423,059 pages. TRACY DURNELL BLOG

Texttile, a multiplayer blog engine for people who write together Texttile lets couples edit one entry together while photos and videos stay sortable beside the Markdown. Klaus Breyer built it in Elixir. One Docker folder runs the blog on your own server. KLAUS BREYER BLOG

UX in 2027 may be less about interfaces and more about behavior UX will define boundaries as software interprets goals and acts. UX Collective links that shift to agentic products. UXDESIGN.CC ARTICLE

Flattening Into Sameness Local shops and cultural habits are losing their differences as algorithms and monopolies narrow what people see. The cycle feeds itself. how to save the world ties sameness to weaker resilience. HOW TO SAVE THE WORLD POST

I stopped listening to podcasts – and my mind flooded with ideas Rachel Aroesti's Guardian essay makes a case for leaving podcasts off during ordinary tasks so your mind can wander, and she cut her listening by 80%. To me, this is very similar to the shower situation. When there is silence around us and nothing for us to do, that's when our brain becomes most active and creative. THE GUARDIAN ARTICLE

The World Now Has More Seniors Than Young People People aged 65 and over now outnumber young children worldwide, as longer lives meet decades of lower fertility. That group will account for 1.148 billion of the next 2 billion people added by 2060. China's birthrate reached 1.06. SLASHDOT ARTICLE

How I would like to live A serious meditation on living well joins Buddhist non-attachment to active love. Philosophy bear turns that ethic into practice: help people, fight suffering, and act without clinging to self. PHILOSOPHY BEAR ESSAY

The Lasting Economic Damage From the China-Nepal Flood The Diplomat says the flood will damage Nepal's border economy for years. THE DIPLOMAT ARTICLE

Africa’s mother tongues are finding new life A worthwhile account of language revival finds new businesses and family ties across Africa and its diaspora. Shungu Chidovi teaches Shona online through stories, book clubs, and paired conversations. Two dozen learners pay $50 monthly. That matters because fewer than 2% of African languages have a meaningful online presence. AP NEWS ARTICLE

How I would like to live A serious meditation on living well joins Buddhist non-attachment to active love. Philosophy bear turns that ethic into practice: help people, fight suffering, and act without clinging to self. PHILOSOPHY BEAR ESSAY

Writing at a blog helps you write more elsewhere Herbert Lui argues that daily blog writing keeps creative energy available for work elsewhere. Social posts require performance, while blog posts provide practice and a safer place to test ideas. Stop blogging, and publishing eventually stops. HERBERT LUI BLOG

Simple is not the same as small Small programs can hide coupling. The distinction matters to Jyn: larger systems can be simpler when dependencies stay separate. JYN BLOG

Your Perception Of Time Is, In Fact, Warped Digital overload makes time feel fast while leaving fewer memories behind. Noema ties this to pandemic routines, app switching, and boredom. I have captured this in the past as novelty, extending time, and repetition compacting it. NOEMA MAGAZINE ARTICLE

Only believe what you can validate Microsoft lays out a practical way to verify agent-generated work before it reaches production. Julia Kordick's framework uses humans, AI, and deterministic tools together. Six questions expose what each check costs, proves, and returns. Experts should confirm individual business rules, not skim 2,500 words and say they look good. MICROSOFT ARTICLE

Why companies are becoming a series of loops This should sound familiar by now.. Lenny Rachitsky’s newsletter hosts Anish Acharya on why AI turns each company function into a repeatable loop, while consumer products chase happier users and distribution becomes harder to copy. Acharya says people move upward by supervising loops, not owning every task. Small teams get more room. LENNY’S NEWSLETTER PODCAST

Not-Writing Writing requires not-knowing because the best language arrives through surprise, while LLMs predict familiar patterns, a distinction 3 Quarks Daily ties to human-made art. 3 QUARKS DAILY ESSAY

F’ing Pulleys Pulleys share a load with the ceiling, so you pull half the force. Liron Shapira calls the wheels mostly rope-redirecting hardware. DOOM DEBATES POST

i wanna live an npc life On signalundefied’s blog, signalundefied argues that an NPC life means ignoring status and doing your own thing. SIGNALUNDEFIED BLOG

The pencil case model of creativity Isabella Saccardi starts with envy for a friend who can sketch whenever inspiration hits, then notices her own PhD work is full of creative connections. Her friend points out that commissioned art has deadlines, printer limits, and client demands too. Creativity can grow inside constraints. ISABELLA SACCARDI COLUMN

A Twitter account of real trail cameras. Wild animals and scenery and stuff. LINK

Portal by Spotify cut my Claude Code token usage by 90% Dimitri Mazmanov routes Claude Code file reads and boilerplate generation to Gemini 2.5 Flash through Portal, saving roughly 90% of Claude's context tokens in Java monorepo tests. DIMITRI MAZMANOV ARTICLE

The Last Interview Human applicants now audition for machines by drawing, writing, and remembering what people do, in a sharp LessWrong story where rage finally earns the job. LESSWRONG ARTICLE

AI Use in the Job Market Is Creating an Infinite Doom Loop WIRED ARTICLE

AI cannot optimize a company it cannot understand Enrique Dans makes a useful distinction: AI needs an organizational model, not just more context, before it can optimize work. That model names objects, links, rules, and actions. FAST COMPANY ARTICLE

Work on your one-sentence description of who you are and what you’re working on.

Here’s mine as of this second, free-form: