Unsupervised Learning NO. 495

Hey, hope you’re doing well!

ERRATA: Two mistakes last episode:

—

A whole bunch of built-up procrastination due to technical obstacles is really starting to add up for me, causing lower mood and energy. I can't remember where I read it, but I really love the framing of anxiety and procrastination simply being side effects of not getting work done that you wanted to, or thought you should.

I'm definitely feeling that. But I hope to resolve most of it this week!

—

Speaking of Michael, here’s my conversation with him about designing AI systems that actually work. Michael Led the Trail of Bits AI XCC team that won 2nd place, and he’s brilliant.

If you think at all about practical vs. hype AI, you will love this conversation.

—

🔥This is the Personal AI Infrastructure I’ve been building for years now, documented in a lot of detail. Took the entire weekend to update this beast.

When I talk about working on AI, I'm mostly upgrading and enhancing this thing. 👇🏼

—

It's so much easier to write story commentary for the newsletter using Wispr Flow. Next to Claude Code and ChatGPT back in 2022, I would say this is definitely the best tech I've seen in years.

—

Last week’s podcast was like an hour and a half long. Caught up on tons of stuff across work and tech and life. Felt significant. If you haven’t listened in a while, you should listen to this one. LISTEN

Really excited with Maze’s approach here, with a focus on getting the context from the organization and sending fixes to the correct people rather than blasting them out to unrelated people who will grow hate in their hearts for security!

Google releases FACADE, their internal anomaly detection system for insider threats

Google open-sourced FACADE, the deep learning system they use internally to catch insider threats and detect account compromises. Absolutely love them for releasing projects like this out to the public for free. FACADE GITHUB REPO | RESEARCH PAPER | BLACKHAT 2025 SLIDES

Researchers discover PromptFix attacks that hijack AI browsers through hidden prompts

Guardio Labs tested Perplexity's Comet browser and found attackers can hide malicious instructions in fake captchas that AI agents process as legitimate commands. GUARDIO LABS RESEARCH | CYBERSECURITY NEWS ARTICLE

Phishing emails now target both humans and AI defenses simultaneously

Anurag Gawande shares how attackers are embedding prompt injection commands in phishing emails to confuse AI security tools while still tricking human recipients. MALWARE ANALYSIS ARTICLE | REDDIT DISCUSSION

Grok chats are showing up in Google search results

Malwarebytes reports that Grok's share button makes conversations searchable on Google without users realizing it. MALWAREBYTES ARTICLE | FORBES COVERAGE | BBC REPORT

The U.S. is running low on Patriot missiles after heavy Middle East use

The Pentagon is scrambling to rebuild Patriot missile stocks after using 30 interceptors in a single day defending Al-Udeid base from Iranian attacks—the largest single-day use in U.S. history. THE CIPHER BRIEF REPORT | LARGEST PATRIOT SALVO ARTICLE

Clear Plus adds biometric gates that skip TSA officers entirely

Clear launched facial recognition gates at Atlanta's airport that verify your ID and boarding pass in under six seconds, letting paid members bypass TSA officers completely before bag scanning. It will be phenomenal if this is able to be maintained and if it expands to other airports.

Interesting piece of security psychology here is that I naturally wonder how easy it would be to fool this system, but then I remember how cursory the checks are by the staff currently. MORNING BREW COVERAGE | WSJ REPORT | AXIOS ARTICLE | THE POINTS GUY

OpenAI says GPT-6 is coming faster than GPT-5 took

Sam Altman told reporters that GPT-6 is already in development and won't take as long as GPT-5 did. Surprising to me that they had to play this card. Subs must really be down for them to have to start teasing this already. BLEEPING COMPUTER ARTICLE | CNBC INTERVIEW

Game developers embrace AI agents at massive scale

A new study reveals that 87% of game developers are now using AI agents in their development process, which is not surprising to me at all. I do a lot with AI and a lot with role-playing games, and they go extremely well together. I mean, just think about character generation, scenario generation, plots, etc. All this stuff is center mass for LLMs.

GOOGLE NEWS ARTICLE

AGI is an engineering problem, not a model training problem

Vinci Rufus argues that AGI won't come from bigger models but from better engineering—specifically orchestrating multiple specialized models working together like a brain's different regions. HIS ARTICLE

Developer replaces vector databases with Git for AI memory

Growth-Kinetics built a proof-of-concept that stores AI memories as markdown files in Git repos instead of vector databases, letting you git checkout to any point and see exactly what the AI knew then.

I really love ideas like this, and I'm personally experimenting with using the file system for all sorts of context management. As we keep talking about here, the management of memory and context is like 90% of the game with AI systems. DIFFMEM GITHUB REPO | HACKER NEWS DISCUSSION

MIT study finds 95% of enterprise AI projects have zero impact on profits

MIT researchers found that 95% of corporate AI implementations fail to impact the bottom line because companies try to force generic tools like ChatGPT into existing workflows instead of solving specific problems.

This very much reminds me of my earlier article on intelligence tasks. The companies that I see adopting AI the fastest and the best are the companies that already understand how their business works. They are simply applying AI to that. It's really hard to optimize something you don't understand, which unfortunately is many/most businesses. TOM'S HARDWARE ARTICLE | FORTUNE COVERAGE

Developer builds memory layer to stop AI agents from forgetting everything

And here's another memory/contact system. Piyush created In Memoria, an MCP server that gives AI coding tools persistent memory so they remember your codebase structure and coding patterns between sessions.

This is the type of thing where a major improvement to memory context management is going to roll out in Cloud Code or something, and it's going to suddenly improve all coding output and throughput by 40% or something. Just making up a number, but my point is that these jumps are going to be extreme. IN MEMORIA GITHUB | HACKER NEWS DISCUSSION

Coinbase CEO fired engineers who refused to try AI coding tools

Brian Armstrong gave engineers a week to sign up for GitHub Copilot or Cursor, then fired those who didn't have good reasons for not doing it. Sounds super brutal but I see it very similar to a CFO firing accountants for not using Excel. I also find it hilarious that Armstrong was strong-arming people. 💪🏼 Sorry. TECHCRUNCH ARTICLE | CHEEKY PINT PODCAST

Uv adds experimental code formatting with Ruff integration

Astral just added experimental formatting to uv, bringing Ruff's formatting directly into their Python package manager so you can format code without installing anything extra. UV FORMAT ANNOUNCEMENT | HACKER NEWS DISCUSSION

Zed raises $32M from Sequoia to build collaborative IDE with real-time version control

Zed raised $32M Series B from Sequoia to build DeltaDB, their new operation-based version control system that tracks every edit in real-time, not just commits. Super interesting to be able to have like an infinite undo tree.

But this doesn't solve the whole problem because you still need to have useful milestones to roll back to. I assume this will be paired with AI that notices and labels changes. Pretty cool stuff. Can't wait to see it in Claude Code. ZED ANNOUNCEMENT | ZED GITHUB | ZED JOBS | CRDT EXPLANATION

Every engineer taking sales calls led to a complete platform rewrite

A startup forced all their engineers to take customer sales calls, and within two weeks they'd completely rebuilt their platform based on what they learned. Something something change comes from pain. I think it's an ingenious idea, and related to something I heard a long time ago—forcing people to work in other roles inside of the organization just to grow empathy and perspective. REDDIT POST | HACKER NEWS DISCUSSION

Google rushes ahead of Apple with AI-heavy Pixel 10 phones

Google's new Pixel 10 series goes all-in on AI features like Visual Overlays that guide you through your camera view, Magic Cue that proactively suggests actions across apps, and Voice Translate that makes phone calls sound like each person speaking their native language.

As an Apple "religious" person, I have to be the first one to admit that Apple is stumbling in the last year or two with major innovations, and especially AI. I still think that once they solve the AI/Siri issue, they're going to jump way ahead. But I expected that to have already rolled out. This is made much worse for them by Google suddenly finding their vision and voice.

Even I am tempted by some of their new tech and some of their new phones. I feel like they are crushing it on the AI stuff. But I know people who have very recently tried to switch from Apple to Google and came back immediately because there's nothing like the ecosystem cohesion that Apple has. For multiple reasons, I continue to wait for Apple to figure out the AI/Siri story and regain their momentum. TECHCRUNCH COVERAGE | PIXEL 10 ANNOUNCEMENT

Getting on the Hacker News front page brings traffic but not conversions

Dan Moore shares what actually happens when you hit the HN front page after 12 years and 400+ successful posts—you get thousands of visitors and valuable feedback, but basically zero conversions. I can also confirm this after having had dozens of front page appearances. It's mostly just an, "oh my god, somebody is looking at me" rush. DAN'S HN FRONT PAGE ANALYSIS

ICE budget could jump to $88 billion under new deportation plan

House Republicans want to give ICE $88 billion for Trump's mass deportation plans, which would make its budget bigger than most countries' entire militaries. NEWSWEEK COVERAGE | HACKER NEWS DISCUSSION

Exercise has insane ROI that most people completely miss

Herman breaks down why exercise is the highest-leverage investment you can make—saying it's basically compound interest for your body and brain. HERMAN'S EXERCISE ROI ANALYSIS | HACKER NEWS DISCUSSION

The hidden management skill is knowing when to actually manage

Terrible Software explains that the most underrated management skill is knowing when to step back and let your team work without interference. Good article, but my favorite book on this is "The Dichotomy of Leadership" that talks about multiple extreme trade-offs that you have to manage constantly when managing. TERRIBLE SOFTWARE ARTICLE | HACKER NEWS DISCUSSION | THE DICHOTOMY OF LEADERSHIP BOOK

Margin debt hits record high as investors borrow to buy stocks

Hacker News discussion reveals margin debt has reached unprecedented levels, with commenters debating whether this signals market exuberance or rational leverage in a low-rate environment. ARTICLE | HACKER NEWS DISCUSSION

Scientists reverse brain aging in mice by reducing a single protein

UCSF researchers discovered that reducing FTL1 protein in old mice restored their memory and increased brain cell connections, basically reversing age-related cognitive decline. One of the most exciting prospects for AI to me is simply combing through massive amounts of data and finding tons of slack in the rope or easy tricks for doing all sorts of things, like improving cognition, reducing aging, and all sorts of stuff we're not even thinking about yet. SCIENCE DAILY COVERAGE

AGENTS.md as a standard way to guide AI coding agents

Really cool idea here of crowdsourcing context management and orchestration for AI tooling. The community created AGENTS.md, an open format that lets developers write simple markdown files telling AI agents exactly how to work with their codebases. AGENTS.MD SITE | HACKER NEWS DISCUSSION

Developer gets shadowbanned by Hacker News and asks for a real IP ban instead

Sean Conner discovered he's been shadowbanned from Hacker News and would rather just be banned at the IP level if they don't want him there. SEAN'S BLOG POST | HACKER NEWS DISCUSSION

Everything in the universe is correlated with everything else

Gwern Branwen explains why all variables correlate with each other in large datasets—it's not measurement error, it's that everything genuinely affects everything else through countless indirect causal chains. GWERN'S EVERYTHING ARTICLE | HACKER NEWS DISCUSSION

The two strongest predictors for longevity are:

I recommend getting your VO2 max tested quarterly if you can, fairly cheaply, or wear a device like an Apple Watch that will give you some kind of estimation.

For strength, it's not any particular one test that matters - otherwise you could game the system. What matters is that you are overall strong. So I recommend whatever works for you in terms of regular resistance training.

Me personally, I do kettlebell swings and deadlifts and traditional gym resistance training for chest and back and shoulders and arms and such.

We don't even fully understand why being strong is such a predictor or even VO2 max. But it makes sense to me overall. I think it comes down to: If you have those things, that means other things are true as well - activity, blood flow, cardiovascular health, etc.

So, in a sentence, do the things that you need to do to improve these two metrics.