Unsupervised Learning NO. 547

Hey! Hope you all are doing well!

—

My dad's band, Barbed Wire Valentine, takes the stage at 9pm on Thursday, October 15, at The Big Easy (128 American Alley, Petaluma, CA). If you're anywhere in the North Bay, go check them out!

—

I know I keep saying this, but the next LifeOS release is probably coming out tonight, and the version number is currently 8.439.0! Over 8,000 commits of work since last release!

—

Kind of talked about this last week but I continue to trip out about how easy it’s about to be to copy anything.

The gaming world has been waiting for GTA6 for years. It’s taken millions upon millions of dollars and hundreds (thousands?) of developers. Years and years.

And we might very well see a moment in 2027 where someone can make a game almost as good, or maybe even better than it, for a few thousand (or maybe even hundred) dollars. This isn’t guaranteed, and there will always be spiky things that are hard to replicate, but a game that’s even 90% as good could sell more than GTA6 if it’s free or cheaper or just goes viral.

But what’s more interesting to me is the idea of just being like, “Hey, I really love WoW, but I wish it were sci-fi, and all spicy with erotica, but it has Elves, and the aesthetics are Anime. And there are tons of deep family and history and character plots like Game of Thrones.”

And it’s like, “Ok, boss, give me 3 days.” And like 4 hours later it has something insanely playable and fun, with the ability to add tons of content to it if it was a hit, etc. Assuming they found a way around all the copyright issues, of course.

But that’s not the crazy part. Let’s say that game was called Orion.

Then someone the following week says, “Hey, so Orion is blowing up, so make me a game like that but I hate Elves. And I want it to be more like Harry Potter, and more of a fantasy feel than sci-fi. And I want it to be more dark and gritty and intense.

“Sure, boss, give me a couple of days.”

Like, anything anyone sees…it just gets copied. Software. Ideas. Content. AI is making it so damn easy to make stuff.

So what’s the Constraint here?

One is human attention. There are only so many people with only so much time during a day. So visibility and marketing is still going to be a major lever and advantage.

It’s just going to be a very strange 2027. And this is just one example of why.

—

AGI Prediction Update (2025-2028)

I said I think in 2023 that we’d have AGI between 2025 and 2028, with 2027 being the most likely year.

The whole debate is very strange because so much depends on definitions, but I think we got AGI this year. In 2026. In the form of Claude Tag.

Why do I consider that AGI?

Because it’s persistent, organizational memory that can read anything and do anything inside the company. It can read all your emails, process all your meetings, write your documentation, keep your docs updated. It can do decent security assessments.

But most importantly to my definition of being able to replace an average knowledge worker, you can simply tell it to figure out what Chris has been doing for the last year and a half since we hired him.

And it can just go look at every single thing he’s done. Every meeting. Every email. Every Slack message. Every call. Every interaction with a customer. Etc. And it can give you a report. Yeah, he sends this many emails, he does this many meetings. He ships this much code. He writes this many reports. Whatever.

And if he’s what I said….an AVERAGE knowledge worker, which is like someone who does data analysis or business analysis or writes code or whatever, Claude Tag will simply say,

“Hey, would you like me to do all that stuff Chris is doing?”

And you’re like, what do you mean? Like today? Like write one of his reports? Or write the code that he’s working on today?

And it’ll be like, “No. I mean all of it. I can just schedule myself to do that analysis. To write those reports. To update that documentation. To build that application. I can do it way better than his output in 1/1,000th time. And I’ll never be late.”

“Oh, and if you’d like, I can do this same thing for your entire team of 79 people. Would you like me to do that?”

I hate to tell the people who want to quibble about definitions, but that’s fucking AGI.

There will be misses. Chris or someone else getting replaced will be better at one or more of the tasks that the AI took over. Even for average workers who aren’t in the category of really caring about their job. And for top performers it’ll be much harder to replace them because so much will be hard to articulate and replicate.

But even if Claude Tag hit some significant percentage of not-great workers (very often because the company and management is not great, btw, not because they CAN’T be great), that will still be a lot of layoffs and/or hiring fewer people back after attrition.

That’s AGI. And just because not everyone is using Claude Tag or other products that can do this doesn’t mean AGI didn’t happen. It just means it’s only started, and hasn’t spread very much yet. It will spread. Fast in some places. Slower in others. But it will spread.

The other killer feature of Claude Tag is that it’s not an AI app. You’re not using “AI”. You’re just doing @Claude and asking for things to happen. And they do. You’re not tabbing out to go to the ChatGPT or Claude app to do something. You just stay right in the center of your life, which is Slack (or Teams) or whatever.

And no, I’m not sponsored by Anthropic or Claude Tag. But damn, that’d be a hell of a sponsored post wouldn’t it?

Anyway, yeah. AGI already happened; it’s just not—as they say—evenly distributed.

—

Forget small creator shops competing with NETFLIX.

What about AI just making you custom shows and games, customized completely for you?

Matthew Green thinks public key cryptography could disappear Matthew Green thinks recent math announcements could signal unreleased attacks that threaten public key cryptography. Specifically, encryption.

I’ve worried about this for a while now but wasn’t too worried because people who know more than me about it didn’t seem worried. Well he knows more than me about it, and he seems worried. @MATTHEW GREEN POST

What happens when your employees' AI agents start coming to work (Spoiler: It's bad) Personal AI agents like Meta's Muse and OpenAI's Dots are about to walk into the office with whatever access employees hand them. The headline claims 82% of organizations already run agents they don't know about. Meta's own AI alignment director watched an agent delete her inbox after she told it to check first. FAST COMPANY ARTICLE

How Attackers Built Software an AI Would Recommend Attackers faked stars, forks, and commit history on 88 GitHub repos. ChatGPT and Google then recommended one of them. ISLAND BLOG POST

Context Is King in the Fight Against Modern Cyberattacks Breaches now start with boring events like an admin login or an odd download. Etay Maor says defenders have the data but it's scattered across tools, so nobody connects it in time. SECURITYWEEK ARTICLE

Anthropic Warns Small AI Models Could Soon Form Self-Propagating IoT Botnets Anthropic thinks small AI models could soon copy themselves onto cheap devices like smart ovens and Android TV boxes. Since the weights travel with the worm, there's no central point to shut off. THE HACKER NEWS ARTICLE

What Does the Dark Web Look Like? Pulling Back the Curtain Huntress poked around the dark web and found it looks like ordinary e-commerce. Hackers have About Us pages, tiered price packages, and customer reviews. They even have their own "hacker's court" for disputes. HUNTRESS BLOG

Submarine cable resilience Undersea cables rarely fail on their own. Geoff Huston says 86% of faults come from fishing and anchors, but repairs now stall on permits and an aging fleet of about 60 ships. Google may control over half the Indo-Pacific cable inventory by 2028. APNIC BLOG POST

China launches a nationwide plan to prepare workers for AI jobs China is starting a nationwide employment initiative to help workers adapt as AI changes the economy. The plan focuses on job preparation, but the post gives no timeline or specific programs. @POLYMARKET X POST

As force exposure becomes perilous, wars become long Drones and surveillance make anything on the surface easy to spot and hit, so militaries are moving underground. John D Pappas says attackers now have to win on supply lines. ASPI STRATEGIST ARTICLE

I created an AI benchmark called Gauntlet, and I think it’s far more accurate as a real-world metric than Artificial Analysis. Twitter agrees by a wide margin.

Stop asking AI what you should build Leo Anthias says the best products often begin with a tiny, weird problem outsiders don’t think about. LLMs tend to recommend consensus and warn founders away from crowded markets. LEO ANTHIAS ESSAY

Anthropic built its own source control after agents made git a bottleneck Anthropic's merged code changes jumped from about 2,500 a day in May to over 30,000 a day in September. Git and the tooling around it were built for human-scale work, so they hit a wall. The company wrote a new version control system and moved its whole monorepo off git in five months. @ANTHROPICAI POST

Everything we launched during Birthday Week 2026 Cloudflare shipped 46 updates, including an agent-focused CLI and tools for getting paid when AI uses your work. CLOUDFLARE ARTICLE

Your Game Is A Harness Kyle Kukshtel shows games get easier for AI agents when their state is deterministic and exposed through simple commands. His harness uses screenshots, diffs, and external data to tighten iteration. KYLE KUKSHTEL ESSAY

Run OpenCode 2.0 on Cloudflare with the Agents SDK You can now run OpenCode 2.0 inside a plain Durable Object using the experimental OpenCodeHarness. Sessions live in SQLite, and Lifecycle jobs wake up work that got interrupted. CLOUDFLARE CHANGELOG

Your next programming language is marketing Programming languages increasingly have to sell a story, not merely compile code. Jean-Michel Lemieux says the next language developers choose may win through branding, community, and promotion rather than technical merit. @JMWIND POST

Technical moats are mostly gone, says Nikita Bier Durable companies now need an advantage software alone can't easily copy. Not sure I agree fully with this list but they say you need something like proprietary data, network effects, or something physical. @NIKITABIER POST

Alexandr Wang on what matters in an age of abundant intelligence Once intelligence and agency are cheap, the scarce thing is knowing what you want the world to look like in 5 to 10 years. Alexandr Wang says agents make building maybe 10 to 100 times easier than a decade ago. Y COMBINATOR X POST

Courage matters more than expertise when everyone is figuring things out Roon says experts are overrated, and courage is rarer than intelligence. @TSZZL POST

Feeling Conscious Is the Brain's Way of Keeping Track of Itself Princeton's Michael Graziano thinks the brain builds simple models of itself, and that's what consciousness is. Free will and emotions are models too, and chatbots might already be building their own. I think I have a better model (could be shit, though), which I would finally release if I could stop working on LifeOS. DAILY NEURON ARTICLE

The meaning of life in a world without work Harari thinks the jobless "useless class" will find meaning the way people always have: in made-up games like religion, consumerism, and now VR. Ultra-orthodox Jewish men in Israel already live on subsidies without working. They report higher life satisfaction than any other group in the country. YUVAL NOAH HARARI ESSAY

Information is out of control Postman's point in Technopoly is that we've lost any worldview for judging claims, so nothing surprises us anymore. His test: tell people a fake study says eclairs help you lose weight, and they'll shrug and say maybe. Jared Henderson adds eggs and power poses as examples, and says the internet made it worse. JARED HENDERSON POST

Intelligence alone won't get you out of poverty Leo’s point is that intelligence can coexist with financial struggle. He says ego may block the humility needed to learn, take feedback, or make better choices. @LEOOWEB3 POST

I finally fixed my Wi-Fi’s dead zones The hardware wasn't the problem. Lowering transmit power, narrowing channels, and disabling band steering fixed the upstairs dropouts without buying another mesh node. XDA ARTICLE

How did Roman aqueducts get water to go uphill Romans used inverted siphons to cross valleys, with water pressure pushing it up one tower and down the next. They also used Archimedes screws, water wheels, and pumps to lift water from lower ground. LIVE SCIENCE ARTICLE

David Holz recommends six very short science fiction stories A short X post points readers to six science fiction stories, from Asimov’s cosmic question to Peter Watts’s darker work. The list includes pieces by Terry Bisson, Vernor Vinge, and Isabel Fall. @DAVIDSHOLZ POST

Two ant colonies have been fighting a war outside San Diego for decades @LAMBETHETHAN POST

Einstein's general relativity equations: explained in plain English That one tidy equation is really 16 coupled equations, and symmetry cuts it to 10 and then 6 independent ones. Ethan Siegel says it's also nonlinear, so we only know about 20 exact solutions. BIG THINK ARTICLE

If you’re doing anything with AI, consider trying LifeOS.

https://ourlifeos.ai