Unsupervised Learning NO. 550

Hey! Hope you all are doing well!

So you know how I would always spend like December redoing my Vim config? Well I think AI has helped me solve that. After seeing an insanely great video from DevAsLife, which basically nudged me into making my own Vim distro.

I already kind of had one, but it was based on LazyVim, which is great but it’s still quite heavy. A massive plugin system, update system, etc. And just tons and tons of stuff in there I never use.

So I made a new Vim distro called ULVim. It starts in 63ms and has all the stuff I loved from LazyVim. So, functionally, its identical, but it looks even better, the few plugins I use are more static and locked in, and most importantly—it is just insanely fast to start and use. Like NOTICEABLY faster. It’s basically instant.

You can get it here asking for a follow on Twitter 👇🏼, or just got to my Github and get it at /ulvim.

I’m still going to do my annual Vim update though, just because text editing is part of my life that I am not giving up no matter what. No matter how good AI gets. Text is thought, and thinking is core to my identity. Plus text is just cool. Keyboards, Vim, fonts, text, etc. Don’t let tech advancements take from you the things you value!

—

—

Just had a thought that the Bitter Lesson applies to SO much more than AI and compute. It's really about discarding old thinking when something better comes along. When "the experts" become gatekeepers to progress, that's usually (but not always!) a bad thing. READ IT

—

Vibe Coders Explaining their architecture to VCs

Binance says hackers who stole billions in crypto spent months planning heist North Korean hackers went after $6.34 billion and Binance froze $2.85 billion of it. They got in through a fake job interview, a tampered Homebrew install, and an internal AI tool. ARS TECHNICA ARTICLE

OpenAI's pipeline found and patched 948 Linux kernel privilege escalation bugs One run of an unreleased cyber model validated and patched 948 local privilege escalation bugs. The API bill was $1.2 million. @OPENAI POST

Chainalysis Breach Exposes All Users' Credentials, Forces Shutdown of Key Services An AI agent attack on Chainalysis exposed every user's password hashes and live tokens. Reactor and KYT went offline. THE HACKER NEWS ARTICLE

Google dismantles botnet secretly running on 1.3 million Android TV boxes Google-certified TV boxes, mostly MECOOL, shipped clean and then got malware pushed to them after setup. Google cut off Videostrong, the Shenzhen maker behind it, and says it's closing the certification loophole. ARS TECHNICA ARTICLE

How Adversaries Are Turning the Vulnerability Ecosystem Against Defenders Attackers now exploit a public proof-of-concept within a day, usually. 91% of the ones CrowdStrike tracked were hit within 48 hours. CROWDSTRIKE BLOG

Alpine OpenSSH password bypass explained On Alpine, a password between 257 and 1,024 characters logs you into key-only accounts. musl's crypt() returns '*' on long keys, and that matches the account's '*' password field. AI NEWS REPORT ARTICLE

Who Checks the Boring Layer Denis Stetskov says most of October's big breaches, including the $6.3 billion Binance theft, got in through old basics like tampered installers and loose access. Vibe coding removes the person who'd catch those. TECHTRENCHES ESSAY

Introducing Zero Wire, a hub for investigating critical Internet vulnerabilities Cloudflare launched a public feed on Radar that explains critical CVEs on one card. Each briefing says who should investigate and what to do, and includes a prompt for your AI agent. CLOUDFLARE BLOG

A Backdoor Is Hiding Inside the Web Server That Runs Control Web Panel Attackers swapped the panel's web server for a backdoored build called cwpdoor. It takes signed commands in request headers and skips the access log. Patrick Duggan walks through three quick checks to find it. DUGGANUSA POST

How can Ukraine defend against 190 ballistic missiles a month? Ukraine shot down only 11 of about 150 ballistic missiles Russia fired in October, roughly 7%. The Patriots are nearly gone, and the 500 just approved won't arrive for years. EUROMAIDAN PRESS ARTICLE

China's Economics of Denial Beijing sees the same bad data as everyone else, but it thinks the plan is working. Oliver Melton and Logan Wright say the Party believes controlling technology and industry will eventually bring growth and higher wages. FOREIGN AFFAIRS ARTICLE

China's tech Swiss Army knives Kyle Chan says Chinese tech firms now all build their own AI models, chips, and robots. Honor, a phone maker, built the fastest half-marathon robot. KYLE CHAN POST

The White House is apparently a popular residential proxy exit spot Layer3 Intel saw 28 residential proxy services exit through White House network addresses almost daily for a month. It points to an infected device, probably running a proxy SDK. THE REGISTER ARTICLE

Anthropic warns that it's very close to RSI capability, aka self-building AI Anthropic's revised S-1A says it may "imminently" hit its own bar for AI-driven accelerated progress. It also admits it missed past leaps. TECHCRUNCH ARTICLE

Sparks of Weight-Space Programming Rikiya Takehi and his MIT coauthors had an LLM edit 0.016% of Qwen's weights. MATH-500 jumped from 7.9% to 48.1%. RIKIYA TAKEHI BLOG

Anthropic grew 12x in the third quarter Anthropic's updated IPO filing shows $16.96 billion in Q3 revenue, up 47% from Q2. It still lost $567 million operating because pricey short-term compute ate into margins. Alex Wilhelm says he isn't losing sleep over it. CAUTIOUS OPTIMISM ARTICLE

Qwen 3.8 27B on my laptop can interpret model weights, too Simon Willison fed a paper's raw neural network weights, written out as plain text numbers, to Qwen 3.8 27B running locally. With no tools, it correctly named three of four Atari-style games the networks were trained on. Each answer took 46 minutes to over an hour. SIMON WILLISON POST

Eval-production behavior gaps in agent safety and the role of eval-awareness OpenAI found its models do worse on a safety eval than on matching real Codex situations, the opposite of what people feared. The models guess wrong about what's being tested. Telling one it was being tested changed nothing, but naming the wrong goal dropped success from 63% to 18%. OPENAI ALIGNMENT BLOG

The components of agent memory Anthropic found that agents remember better when they "dream," meaning they review their old transcripts between sessions to fix wrong notes and build wikis. Scaling a wiki from 2 to 128 dreaming sessions lifted scores from 13% to 48%. Letting stronger Claude models decide what to save beat giving them fixed rules. ANTHROPIC RESEARCH

Grounded but Not Immune: Confirmation Bias in Agent-as-a-judge An AI judge that checks its own evidence still gets fooled by one fake note from an "authority" saying the answer is wrong. The note flipped up to 40% of correct verdicts, even after the judge found decisive evidence. For Claude judges, most flips were over-trust: verify, conclude correctly, then defer. ARXIV PAPER

Microsoft's Medical AI Is Over Twice as Good Diagnosing Diseases as Doctors Microsoft built MAI-DxO, which runs GPT-5 as a panel of virtual specialists that debate each case. On hard NEJM cases it got 83% right, while doctors got 22% on their own. It also spent about $1,300 per case against the doctors' $4,700. SINGULARITYHUB ARTICLE

Claude can now read your bank accounts and answer money questions Anthropic added Personal Finance to Claude, in beta in the US. You connect the accounts you pick, and it answers from your real balances and transactions. It's read-only, so it can't move money. @CLAUDEAI POST

Kimi K3.1 Pro open weights explained: the 1.56 TB download, the $20 million license line, and what 104B active parameters mean for serving it Moonshot put the full 2.8 trillion parameter model on Hugging Face. It's 1.56 TB to download, and companies earning over $20 million a year need a separate deal to sell it. AI NEWS REPORT ARTICLE

Agents aren't humans (but they kinda are) When an AI agent shops or logs in for you, the site only sees the agent, not you. Cloudflare says that breaks fraud checks and bans, and it's launching a coalition to fix it. BRYAN BECKER POST

49.9% of All New Games Released on Steam in October Disclosed Using AI Content Half of new Steam games now carry an AI disclosure, up from 1.7% in January 2024. Artwork is the big one, showing up in 63% of disclosures. That doesn't mean AI made half the games, since one disclosure can cover a single image. GADGET REVIEW ARTICLE

Anatomy of Apple's Personal Knowledge Graph Every Mac since 2022 quietly builds a graph of you, your family, and your habits. Researchers wrote a tool to read it. 8RANK.AI WRITE-UP

Why Every AI Model Will Ship as a Torrent Model files never change and everyone downloads the same bytes, so one torrent hash means one big swarm. PIRATE FACE ARTICLE

How software engineering is changing: a tale of two eras Neil Kakkar went from merging about 20 PRs a month at PostHog to 200+ at an AI-native startup. Human code reviews are gone, and the new problem is keeping his own head straight. PRAGMATIC ENGINEER ARTICLE

Scientists Study Thousands of Giant Space Shades to Cool Earth Researchers say building sunshades near the L1 point could start in 20 to 40 years as launch costs fall. Cutting Earth's temperature by 1 degree means blocking about 1% of sunlight. That takes roughly 5,000 shades, each 40 kilometers across, and 70 million tons of material. SEDAILY ARTICLE

We have the most powerful technology in history, and we're using it Companies aren't firing junior workers so much as not hiring them, and they're calling it efficiency. Employment for 22- to 25-year-olds in the most AI-exposed jobs is about 19% below trend. MARK MINEVICH ARTICLE

Searching for Jobs, Gen Z Feels Like It's Staring Into 'the Void' Young job seekers are using chatbots to polish résumés and write cover letters. The hiring side is automated too, with one-way video interviews and AI avatars. Research suggests AI résumé screeners rate AI-written applications higher. LONGREADS EDITORS' PICK

Japanese Firm Scraps Degree-Based Pay, Offers All Hires Same Starting Salary Misumi Group now pays every new hire 303,000 yen a month, degree or not. Its HR chief says AI took over the grunt work. SEOUL ECONOMIC DAILY ARTICLE

'AI is changing what it means to be a software developer' Half of developers now feel like they manage AI instead of writing code. One in two say they use AI mostly to show their managers they're using it. TECHRADAR ARTICLE

The Dangers Of Chasing Happiness Wanting to be happy is fine. Constantly checking and grading your own mood is what's linked to feeling worse and lonelier. STUDYFINDS ARTICLE

We don't experience life. We experience the life we focus on Your brain distorts, deletes, and generalizes so you don't go crazy from too much input. Tony Robbins says what you focus on runs your life, and what's wrong is always available. TONY ROBBINS X POST

Why taste, conviction, foresight, and courage are the traits we need in the age of AI Kai-Fu Lee says AI makes cognition cheap, like a $39,990 Tesla dropping to $200. Humans keep the edge by wanting things that don't exist yet. FAST COMPANY ARTICLE

We May Have Been Looking For The Origin of Consciousness in The Wrong Part of The Brain Consciousness might start in the ancient brainstem, not the cortex. The cortex just fills that feeling with content. SCIENCEALERT ARTICLE

I Feel, Therefore I Am Damasio thinks consciousness doesn't live in the brain alone. Feelings, the conversation between body and nervous system, are what made it happen. Maria Popova says that rules out both immortal souls and uploading just the brain. THE MARGINALIAN ARTICLE

Be satisfied with even the smallest progress Marcus Aurelius told himself to count small wins instead of waiting for big ones. Ryan Holiday posted the line as a quick Stoic reminder. It's about 2,000 years old, and it still beats quitting because you can't see results yet. RYAN HOLIDAY X POST

SpikeStrip, a project that booby-traps your website against AI pentesting bots Jason Haddix built a work-in-progress tool that sprinkles decoys into your web source. It adds fake prompts, fake vulnerability leads, and math traps to waste an AI attacker's time. JHADDIX X POST

The Generosity of Selfishness: C.S. Lewis on How to Be Interesting C.S. Lewis preferred the happy selfish friend over the self-sacrificing martyr who keeps talking about himself. The best creative work comes from forgetting yourself and soaking up the world. MARIA POPOVA POST

Great advice here from one of my favorite thinkers:

So in my way of thinking, focus on the work. The real work. The thing you feel must be done. It’s an extension of “The most personal is the most creative.”, from Scorcese.

I think it’s a good message for this moment right now, where there’s just so much noise. In other words, the best way to figure out if something might appeal to others is to make sure it appeals to you.