Unsupervised Learning NO. 545

Hey! Hope you all are doing well!

AI State of the Union

It’s strange how quickly momentum can shift in these AI races.

Like two months ago Anthropic was in the dirt compared to OpenAI. Astra had just crushed everything, and Anthropic was falling massively behind. At least according to vibes, and Twitter, and overall sentiment.

And then in the last 2-3 weeks everything has inverted again. Opus 5.5 has blown everyone away, Some really bad stuff has happened with the usage situation in Codex versus Claude Code. Now it feels extremely high anxiety to use Codex because you just see your credits going away instantly.

And on the Claude code side, Opus is way cheaper than Fable and better in many places. Now with Sonnet 5.5 released, the combination of Opus and Sonnet means you can run on a $200 sub for what feels like 3 to 10 times as long as just a few weeks ago.

This anxiety thing is very real and very tangible. It's something that the community talks about a ton, and I feel it viscerally myself.

OpenAI just had their DevDay today and they released Sol 6.1. So I'm testing that out. My router still routes across all the best models of both providers, but I am mostly using Claude models right now, just because the harness and the usage anxiety situation is that much better.

Another thing that's really struck me is how little loyalty there is for this stuff. Some new company could come out tomorrow with an amazing harness and a better model, and half the builders would shift over to it within a day and a half. They wouldn't give two shits about their loyalty to what they were using twelve minutes beforehand.

Super dynamic times right now. And I anticipate things only accelerating on this axis. I expect 2027 to be utterly strange and exciting and monumental. It's not just that AI or tech are changing; it's that those changes are so core to so many other things that it's hard to even track all the different ways things are changing at once.

Claude overtook Codex in T3 Code in just two weeks In T3 Code, Claude went from trailing Codex to twice its popularity, Theo says. @THEO POST

—

So freaking excited about this!

Our own Bryan Kerr mentioned in a Twitter post that he was using dictation with a foot pedal, and so I had just switched over to Typeless from Whisper Flow and already have two pedals under my desk that I use for OBS.

I now have dual-function dictation with both my keyboard and my Elgato foot pedal!

How it works: I can either:

I feel massively upgraded from this! Plus, I have KindaVim so anywhere where I enter text I can edit with my own Vim shortcuts! So. Epic. Thanks, Bryan!

THE BLOG POST | SIGN UP FOR TYPELESS | THE CODE REPO FOR MY FOOT PEDAL SETUP

—

This is just surreal to me. Kai went looking and found that this was the first time I talked about the concept of Digital Assistants, and holy crap look how close this is to what’s happening now!

And this is 2006! (Oh and by the way this is the color of my site now, I copied it from this old color I used then) It gives me joy whenever I see this old blog layout.

—

I'm currently prepping the next LifeOS release, which will be in version 8.x.x. Hope to have it out today or sometime this week!

—

Kind of a vulnerable moment, but I’ll share it because I think it can probably help someone. Last couple weeks I’ve been struggling a bit with feeling like I might not be doing the right things, or might be lacking focus.

I think I finally found the issue and, surprise surprise, it’s my common failure mode:

Basically, whenever I focus on what others are doing, how my content is performing, or whether or not other people will like what I'm doing, everything goes downhill. The fix is to look at my TELOS, remember what it is that I’m actually working towards, and get back on-mission. And as soon as I do, I feel great.

This is much harder to do right now, especially with so many distractions. Those distractions are the reason I keep focusing on my harness: new stuff keeps happening in AI. New models. New tools, etc. It’s insanely addictive to keep improving LifeOS, and I keep enhancing my system and wanting to talk about it.

But talking about my system isn't as interesting as actually doing work, helping people, creating new things, and sharing ideas that I find personally valuable and interesting: without any concern for what other people are going to think, or whether or not the YouTube algorithm is going to like it.

So I guess the reminder is: if you start focusing internally and worrying about what other people think about your work, remember to focus externally on the problems you're trying to solve, and internally on what gets you excited about life itself, ideas, and your own personal interests and projects. And also, never stop taking in new ideas, extremely high-quality ideas, from books, etc.

Anyway, we’re not all the same so maybe this doesn’t apply to you, but I hope it helps someone.

—

Nvidia’s Answer to Rogue Agents Is an Open-Source AI Security System WIRED ARTICLE

Your attack surface is bigger than you think… and hackers know that Attackers can bypass defenses with stolen cookies and credentials, while forgotten external assets stay exposed. This is why I have my bunker system inside of LifeOS: to make sure I never have anything online that is not in my asset management system, which is being scanned every hour for issues. CYBERSECURITY DIVE ARTICLE

Bitget Says Attacker Exploited Third-Party Security Product Flaw The attacker stole high-level credentials, then used them to submit fraudulent withdrawal commands. Two small test transfers slipped under Bitget’s risk threshold before larger transfers began 30 minutes later. Bitget says cold wallets and customer balances were untouched, while its Protection Fund will cover the $388 million loss. THE HACKER NEWS ARTICLE

America is in the wrong AI race with China REST OF WORLD ESSAY

Poland to ease rules for shooting down drones, missiles entering from Russia Poland will let pilots and commanders rely on radar instead of visual checks before firing. The change follows repeated Russian airspace violations and daily drone concerns. KYIV INDEPENDENT ARTICLE

Three things to know about Trump’s Greenland deal The agreement gives Washington more say over Greenland’s security and foreign investment. It also plans for possible independence while keeping Greenland inside NATO. ATLANTIC COUNCIL ARTICLE

AI may become cheap only after it becomes expensive for everyone LESSWRONG POST

A prompt that checks whether AI understands your actual goal A simple prompt asks AI to restate your goals and the problem you’re solving. It helps catch misunderstandings before you act on its advice. @POTETO POST

Claude can now build and improve your app evaluations So awesome: Claude Code can help create evaluations for your applications and improve them through repeated testing. The approach is called hillclimbing, where each round uses evaluation results to make the next version better. @CLAUDEDEVS POST

AI is making developers lose their eliteness over others Ryan Florence says AI is showing how developers have been kind of rude special snowflakes for a long time, and that this is changing now because of AI, with ideas becoming more important. He’s struck that his sales-guy brother-in-law and designer brother are doing things beyond his own work. @RYANFLORENCE POST

AI is warping the time value of everything ASIA TIMES ARTICLE

Introducing cf: the agentic CLI Cloudflare built cf to give coding agents access to more than 3,000 API operations, up from Wrangler’s roughly 280. It defaults to JSON, searches commands in natural language, and uses typed TypeScript configuration agents can edit. Agent usage already makes up 48% of Wrangler use, and I’ve already upgraded to cf from wrangler in LifeOS. CLOUDFLARE BLOG

The Problem is not the AI Code, but Nobody Knows Anything Anymore The bigger danger of AI-written software is losing the people who understand its architecture and purpose. One startup team reportedly works 12 to 13 hours daily while nobody reads the code. Fast shipping then creates systems nobody can maintain unless they put a whole lot of effort into maintaining that understanding. SSP.SH POST

The Spec Can Come Later I'm not sure I exactly agree with this, but I like the direction it's going. Chad Fowler says the first implementation often teaches you what the software should actually be. AfterVibe extracts a spec from one AI coding session, then gives it to another agent to rebuild the changes. AICODING LEAFLET ARTICLE

Anduril wins wildfire prize for 10-minute response Anduril won the Wildfire XPRIZE for detecting and suppressing fires within 10 minutes, says Peter Diamandis. @PETERDIAMANDIS POST

There Are No Interesting Things; There Are Only Interested People Alberto Romero says preparing for AI mostly means staying interested, building manual skills, and investing in relationships. His own hedge is simple: buy a house and stop refreshing AI timelines. THE ALGORITHMIC BRIDGE ESSAY

Bertrand Russell on the Secret of Happiness Bertrand Russell says happiness comes from widening your interests beyond yourself and meeting the world with friendliness. Maria Popova connects this idea to moments of “unselfing,” from helping someone to noticing a distant galaxy. The result is temporary calm that helps you handle personal worries better. THE MARGINALIAN ARTICLE

How to Read Good or Maybe Great Marco Kaye imagines a future where reading feels painfully slow beside modern stimulation. Recommends accepting boredom, starting with easy authors, and remembering that old humans read for empathy and perspective. MCSWEENEY'S ARTICLE

Human labour is largely invisible to AI AI looks much smarter on benchmarks than in the economy because benchmarks reward narrow, well-defined work. The messy parts of software engineering still need people to define problems and handle judgment with others. FOHLEN.DEV BLOG

Reading and Writing and Thinking and China Writing doesn't just preserve thoughts; Robinson says it changes them, and alphabets made thinking faster. RATIONAL OPTIMIST POST

AI may finally make user-generated content the norm I think most valuable content could come from individuals or tiny studios within two years.

@DANIELMIESSLER POST

The surprising downside of high-performing employees High performers can hide broken leadership systems by quietly absorbing extra work. Over time, that keeps weak management in place and pushes reliable employees toward burnout. FAST COMPANY ARTICLE

Claude Isn't Allowed to Write Me Prose Claude reports through typed blocks instead of prose, so the UI can show decisions first and hide details. Kvit validates every report before the agent continues. KVIT BLOG POST

Addictions are anesthesia Chris Lakin says addictions often numb underlying suffering, so removing one habit can push you toward another. He calls this “addiction whac-a-mole” and says the lasting fix is treating the pain underneath. One researcher went from five to six hours of nightly YouTube to more conversations after that suffering eased. LESSWRONG ARTICLE

Active Directory Penetration Testing Playbook Most AD guides give you commands. HOEK turns them into decision points, evidence requirements, cleanup steps, and attack paths you can reproduce later. 0UT3R.SPACE GUIDE

Act like a camera crew is following you when you try The advice is to imagine a camera crew watching whenever you try something difficult, because that pressure can push you harder. Codie Sanchez calls it smart. @CODIE_SANCHEZ POST

My Wi-Fi drops every day at the same time Evening slowdowns came from crowded nearby networks, not a broken router or ISP. XDA ARTICLE

This is a modern motherf*cking website A 23 KB HTML file handles accessibility, SEO, social previews, responsive images, dark mode, and navigation without dependencies. Native browser features replace much of the JavaScript and framework stack people add by default. The whole deployment is copying one file into the served directory. DREAMSTATION SYSTEMS POST

About pages should teach AI what your company does Rob Hoffman says a strong About page should explain your company clearly to people and AI. @ROBHOFFMAN_ POST

Stay locked on your internal goals and interests.

And have a way to quickly keep them in front of you so you can find your way back when you get distracted with external shiny things and status games.