Unsupervised Learning NO. 539

Hey! Hope you all are doing well!

AI may end normal Black Hat and DEFCON I think this may have been Black Hat and DEFCON's last normal year, and my post explains why. AI could make 2027 feel seven years away. I think many companies (and therefore people) be forced to skip Vegas. $500+ DEFCON tickets isn’t helping. MY X POST

—

My brother Jason Haddix wrote a powerful essay saying DEFCON is not going away basically, as a response to many people saying AI is going to disrupt security and security conferences next year. Here’s the essay and my response to it.

—

This is a big part of why I'm so optimistic about AI, even though I spend 20 to 60% of my time being worried. So much positive potential here. The page came out strong. We just have to do our absolute best to keep pushing in that direction.

—

I think if you're going through rough times, this is a really good metric to think about.

—

Anthropic just added watermarking to its outputs, and I had Kai write up a full analysis of what such systems might look like after having a conversation with me about it. in other words, where are all the places that a watermark could exist?

—

My frustration with the idea that “nobody asked for AI”. Yeah, and they didn’t ask for antibiotics or cars either.

—

Aeternum puts its C2 on the Polygon blockchain Aeternum stores encrypted C2 commands on Polygon; this report traces three malware samples using it. It can rotate domains through smart contracts while infected Windows machines query public RPC endpoints. One sample used 22 contract addresses. Another dropped XWorm, an XMRig miner, and tools that stole system and cryptocurrency data. UNIT 42 ARTICLE

Who gets pulled over when Flock misreads plates Flock's cameras generated 161 false stolen-vehicle alerts for LAPD in two months, and Roseville found errors in 71% of 1,427 alerts. And some errors became armed encounters. In Toledo, a plate misread sent police after Brandon Upchurch, who was mauled by a dog, lost his job, and settled for $35,000. SLASHDOT ARTICLE

Cities are swapping Flock cameras for Axon's Cities are dropping Flock ALPR cameras and buying equivalent Axon systems instead. Axon mounts some cameras on streetlamps. Its patrol cars also collect license plate data. 404 MEDIA ARTICLE

The 52 CTF challenges AI solved in minutes An autonomous agent solved all 52 BSidesSF challenges and won first place. Most CTFs reward bounded tasks AI handles in minutes. Cybersecurity training now needs live drills, judgment, communication, and team coordination because those are harder to automate. SIMULATIONS LABS BLOG

China’s PLA uses AI to plan strike operations China’s PLA uses AI to prioritize targets, coordinate formations, and assign tasks across 100 tactical units. THE DIPLOMAT ARTICLE

14 homicides hit Mazatlán in one week A grim report from Mazatlán tracks 14 murders as cartel fighting moves south. EL PAÍS ARTICLE

Ukraine destroyed another $1.8 million Starlink jammer Ukraine destroyed a third Russian Volna Kupol Garant jammer at an FSB base near Gelendzhik. EUROMAIDAN PRESS ARTICLE

Drone warfare is changing what military protection means John Coyne argues drone warfare changes protection from keeping platforms alive to preserving freedom of action. Autonomous systems trade armour for payload, production speed, software, rapid replacement, and mission capacity. Drones change the math over time. JOHN COYNE ESSAY

China is building solar farms across its deserts China's desert solar push gets a useful look here. Solar fields could slow desertification, power cities, and create wildlife habitat. China plans 455 gigawatts by 2030, roughly twenty Three Gorges dams. SCMP ARTICLE

AI helped high performers and hurt low performers OpenAI researcher Dave Holtz describes a Kenya study where an AI mentor had no average effect on small businesses. High performers improved, while low performers got worse. The difference came from judgment: experienced owners could tell when the tool was useful, but weaker owners often trusted bad advice. @MTSLIVE VIDEO

Readers rate ChatGPT stories higher than human ones Across three experiments with more than 2,500 people, ChatGPT stories beat human ones on ratings, but the advantage vanished when readers learned a machine wrote them. THE DECODER ARTICLE

Claude Code sessions can now message each other, and that's useful Claude Code sessions can now message each other across machines, letting agents organize teams and share dependencies. Agent Teams is no longer confined to one team. Remote control is required. DAISY HOLLMAN THREAD

Which AI newsroom beat WIRED to an OpenAI scoop AI newsrooms can break stories faster than human reporters. RuntimeWire published an OpenAI scoop in six minutes. WIRED ARTICLE

Lotsa Cloudflare here…

Cloudflare Computer gives agents a filesystem and runtimes Cloudflare Computer gives agents a Durable Object-backed filesystem with container or Worker execution backends. CLOUDFLARE REPO

Cloudflare built a browser for agents, and it costs less Cloudflare built Kitesurf on Workers with Rust and WebAssembly, stripping out tabs, extensions, and human-focused rendering. It uses up to 7x less memory than Chromium, though Chromium still finishes tasks faster on the tested pages. That leaves agents more concurrent sessions for less money. One thing to consider, though, is they run from Cloudflare IPs, which are often blocked as part of anti-automation efforts. THE NEXT WEB ARTICLE

Cloudflare turns Workflows into a CI/CD platform Cloudflare is turning its Workflows platform into a CI/CD system for code stored in Artifacts. Push events can start isolated builds, tests, cached dependency installs, and conditional deploys in TypeScript. One workflow can serve millions of repos. An AI healer can fix failed steps on a separate branch for approval. CLOUDFLARE ARTICLE

Cloudflare OS lets agents rewrite apps, but security stays sandboxed Personal apps can let AI agents rewrite their own code, and Kenton Varda thinks Sandstorm finally has its moment. KENTON VARDA THREAD

40% of Americans still talk to neighbors regularly Only 40% of Americans say they talk to neighbors regularly, down from nearly 60% in 2012. Young adults fell from 51% to about one in four. Apps, remote work, and fear of being filmed remove the casual encounters that once made neighbors into a basic safety net. FORTUNE ARTICLE

Video gamers are filling FAA training slots, but shortages remain The FAA is recruiting video gamers as air traffic controllers. Gaming trains quick decisions, focus, and multitasking around moving information. More than 2,000 gamers reached training, covering 94% of the hiring goal. The shortage still threatens delays. CBS NEWS ARTICLE

Zuckerberg's AI future leaves little room for real relationships AI-made images ask for attention without showing the effort, skill, or person that makes attention feel meaningful. THE VERGE ARTICLE

ICE tracks nearly 54,000 immigrants with ankle monitors ICE now tracks nearly 54,000 immigrants with GPS ankle monitors, up from 17,000 early last year. Geo Group charges more for monitors than phone check-ins. A lawsuit challenges the policy. THE GUARDIAN ARTICLE

Movies shelter us from life, and send us back into it Peter Tonguette makes a lovely case for leaving movies energized, carrying their lessons into the rain and ordinary life. PETER TONGUETTE ARTICLE

England plans walk-in mental health hubs nationwide England plans nearly 200 walk-in mental health hubs by 2029, with centres in libraries, banks, and other high-street buildings across England. Some will offer seven-day access and short-stay beds. The £343m programme starts opening sites this autumn. BBC ARTICLE

California may fine influencers who hide campaign payments California may fine influencers for hiding campaign payments. Fines could reach $5,000 per violation. FORTUNE ARTICLE

Personal AI could pull us out of the human world I’m worried personal AI will build worlds more attractive than real life, which is behind my post. We could stop talking. I want our AIs keeping us here. MY X POST

America needs more energy, compute, and trained workers I wrote this post to sketch a national training system for energy and compute jobs, with experts teaching workers new trades in months, not years. MY X POST

Compression is prediction NGROK BLOG

Y Combinator open-sourced its multi-agent work harness Y Combinator open-sourced QM, a cloud-first multi-agent harness built for company work. YC uses it across several departments. The MIT project includes Slack, web UI, memory, triggers, and shared files. Y COMBINATOR X POST

AI agents let people execute every idea On his HEY blog, David Heinemeier Hansson says AI agents let him execute every idea he has. DAVID HEINEMEIER HANSSON BLOG

Machines eventually leave humans behind and move planets Human-machine speculation gets strange here, as an interconnected solid-state mind removes people, empties Earth, and sends the planet searching. KIBOTRONICS POST

Whether you do it through LifeOS or not (which is now possible!), get Hermes set up.

Hermes is, for a lot of people, what they think of when they think of an AI assistant. It’s basically like OpenClaw but simpler, more secure, and in my opinion just better. It’s like a product as opposed to an idea.

Once you have a Hermes agent and you give it access to various services that you use, you will be able to text your agent and do all sorts of things in your life, including:

All sorts of amazing things…

Install LIFEOS
Install HERMES