Unsupervised Learning NO. 551

Hey! Hope you all are doing well!

I'm going to do a blog on this soon, but I want to share a quick idea with you.

I think there's an interesting comparison between public speaking and AI.

Public speaking is one of the highest causes of anxiety, and AI is obviously causing anxiety as well. But there's this cool thing that I and others recommend around public speaking: you notice that anxiety and excitement are essentially the same physically, but the difference is how you perceive them. It's possible to make the choice to be excited about public speaking instead of anxious.

I think it's possible to do the same with AI, and in fact, this is exactly what I do. I sometimes get anxious about the future and the impact that AI is going to have on people in society. I essentially try to turn that into radical optimism, which reframes the anxiety as excitement and inspires me to try harder to steer things in the right direction.

I think we have this option in many cases related to anxiety, and I think public speaking and AI are two good ones.

—

Some really cool tech jobs analysis here.

This one is interesting to me. You still need someone to drive the project end to end.

I thought it was established, though, that that peak was the post-Covid peak.

This is basically people implementing AI inside of companies, and it's obvious why this one is growing.

Yeah, this one doesn't have a good story if you’re a designer. Really, I feel like if you're a great designer, you should be doing other things and just have design as part of your repertoire.

THE FULL REPORT

—

New essay out about how and why the Bitter Lesson is hitting a wall.

—

Remember to check out the IDEAS section. I'm putting a lot of really good stuff in there that doesn't go in the other sections because the main sections are more about news. I'm finding myself wanting to put more and more interesting stuff in there, so make sure to check it out.

—

Grok Build coding agent leaks oodles of confidential info Grok puts the random part of each shared app's URL to the left of the domain, so passive DNS lookups expose it. A researcher found 43,601 links, and nearly 95 percent were public. ARS TECHNICA ARTICLE

CISA Warns of Active Exploitation of OpenSSH, GPUStack, FreeSWITCH, and IBM Flaws CISA added four bugs to its known-exploited list, including an OpenSSH authentication bypass called SSHole. GreyNoise saw attackers trying to steal root SSH keys and plant persistent access within a day of the first attempt. THE HACKER NEWS ARTICLE

Fireworks AI gets burned twice as customer data exposure widens Fireworks AI got hit again in October, and this time billing records, datasets, and partial model weights were exposed. The company blames a hacked cloud provider and LLMJacking criminals. THE REGISTER ARTICLE

Is Cyber Deception a Must-Have This Time Lenny Zeltser doubts decoys go mainstream, even with AI attackers. Defenders keep putting more urgent projects first. LENNY ZELTSER POST

Critical Apache Airflow RCE Flaw Allows Attackers to Execute Code Without Authentication Anyone who can write to the Celery broker can run code on Airflow workers, no credentials needed. Reflectiz found 198 of 211 reachable Redis brokers answered without authentication. The fix is Airflow 3.4.0. CYBER SECURITY NEWS ARTICLE

SonicWall releases patches for 7 flaws in its firewalls, most involving unauthenticated access SonicWall patched seven bugs in its Gen 7 and Gen 8 firewalls. One of them lets attackers skip SAML and MFA, and CISA says it's already being exploited. CSO ARTICLE

QNAP Now Says Plex Servers on Its NAS Devices Were Compromised Through Plex2Shell Updating Plex won't remove whatever attackers already installed, so treat an exposed NAS as breached. PATRICK DUGGAN POST

Barracuda Confirms Active Exploitation of Email Gateway Flaws Allowing Root Access Attackers are chaining two bugs in Barracuda's Email Security Gateway to get root without logging in. It's hit a small number of organizations, and the admin interface has to be reachable. CYBERSECURITY NEWS ARTICLE

Modelling AI-enabled Extreme Power Concentration in CCP If China got AGI first, Xi could end up running the country with almost no human help. The author thinks that's hard to pull off, because Xi would need other people's buy-in. LESSWRONG POST

How Germany's intelligence crisis could trigger a confrontation between Russia and NATO A former BND chief was arrested for treason, and the US and UK are reviewing how much intel they share with Germany. Sibylline says that opens a one-to-two-year window for Russia to hit Germany. SIBYLLINE MEDIA ARTICLE

Why I'm becoming a China bear Noah Smith used to be a China bull and hawk. Now he thinks China's make-everything strategy is hitting its limits, and AI will hurt it more. NOAH SMITH POST

I'm not sure about this, because these things take around a month to figure out, but I'm pretty sure that Fable 5.2 is going to be not that big of a deal for me. I'm pretty sure that I'm going to continue using Opus 5.5 for mostly everything.

Conference Intel It's a site that covers 75 security and AI conferences, with 4,801 talks reviewed. You can browse by conference and see the best talks ranked. GREPTALKS.AI SITE

Employees Who Use AI End Up Working Vastly More Time Per Day People who use AI at work are logging up to 90 more minutes a day than people who don't. ActivTrak, a workplace surveillance company, tracked 9,100 employees at 450 companies. FUTURISM ARTICLE

An AI intelligence explosion may hit in 2027 after all Daniel Kokotajlo says the new METR report puts AI-driven acceleration five months ahead of the AI 2027 scenario. He's scared, sad, and angry. @DKOKOTAJLO POST

I think this above is true, but it will take a while for it to penetrate the market and companies. we could have ASI in 2027, and there will still be tens of thousands of companies who are basically functioning like 2019 in 2028. TWEET ON THIS

Fable 5.2 and Jagged AGI Ethan Mollick calls Claude's new model "almost" AGI. It does expert-level work across many jobs but can't tell when its own work is bad. ETHAN MOLLICK ESSAY

What running a company on 20+ Claude agents taught us A small company ran on Claude Code agents for four months and wrote down 12 failure modes. The worst: agents read their own earlier notes as verified facts, so green checks lied. RIPPY PRODUCTIONS POST

Anthropic wants Claude to be the front end for your entire enterprise app stack Claude can now open full apps like Adobe, Atlassian, and HubSpot inside a chat and edit things directly, with user approval. It's built on MCP Apps, so other AI clients can use it too. Existing permissions still apply. THE NEW STACK ARTICLE

Nobody Approves These Posts Sydney Marrone and Lauren Proehl built a security newsroom run by six AI agents. Every morning it picks a story, writes it, fact-checks it with a separate agent, and publishes with no human reading first. THOR COLLECTIVE DISPATCH POST

My first 100 PR day Nick Mayhew merged 109 pull requests in one day at Metaview, and he says he didn't become a 100x engineer. The pipeline did it: instant deploys, 16-machine tests in two minutes, AI reviewers, and a Slack agent. He just read results and decided what came next. NICK MAYHEW POST

Ring Drone Cam Ring is selling a $249.99 indoor camera that flies room to room on patrol routes. It only stays up for 8 minutes a flight, and it stops recording when it's docked. RING PRODUCT PAGE

State of the tech job market in 2026 Lenny Rachitsky says PMs are winning and engineering demand has stalled. Design openings hit a record low. LENNY RACHITSKY X POST

AI short films in China got so easy that only ByteDance makes money Friends told Kun Chen that early AI filmmakers got rich, then everyone copied their skills from a markdown file. Now ad spend is 80% of creators' costs, and AI tokens are under 5%. KUN CHEN X POST

I'm really worried about this because, even though we can make more and more stuff with AI, the stuff still needs to be found. I’m concerned that the moat moves to distribution and marketing. And those tend to consolidate in very large companies like Google and Meta (and their Chinese counterparts).

—

The Living Software Repository Ryan X. Charles says agent-run repos could heal from attacks and come out stronger. Today it's real in pieces, not as one loop. ASTROHACKER ESSAY

AI proved your code correct. What did it actually prove A checked proof only holds inside its model and assumptions. Sid Trehan shows how a spec can match the code and still miss the real requirement, like access revocation. KELLER AI ESSAY

New Units of Knowledge If everyone has an AI genie, what's scarce is difference, not content. Kevin Kelly says the shareable unit will be specs and goals, passed between agents in standardized packages. KEVIN KELLY ESSAY

Be less scared of overconfidence Ben Kuhn says "experts are usually right" and "markets are efficient" throw away what you actually know. They also push you toward less ambitious choices. BEN KUHN ESSAY

Institutional vs personal AI Gibson's Agency has two kinds of AI, and they work very differently. The aunties extend a powerful person's reach, while Eunice gives an ordinary person new capabilities. Armin Catovic thinks we're building the aunties while hoping for Eunice, and that ownership decides which one you get. ARMINCAT ESSAY

All Interesting Behavior Is Overdetermined Eric Raymond says interesting behavior usually has several reasons, and each one alone could explain it. Baboons presenting their hindquarters is the example: about half of 775 observed cases were submission, and a third were sexual. So don't stop at the first good explanation. ERIC RAYMOND ESSAY

Freud's three blows to human ego now have a fourth: AI Chubby says people hate AI because it knocks humans off the top of the intelligence ladder, after Copernicus, Darwin, and Freud. @KIMMONISMUS POST

The memo was never the point AI can write a six-pager in seconds, but the writing was what forced you to think. FAST COMPANY ARTICLE

Fable 5.2 One-Shotted My Nephew's Dream Game One prompt before bed got Lex Toumbourou a finished 12-level beat 'em up from his nephew's drawings. Claude wrote a bot to play it through every stage and check screenshots. It cost about 340M tokens and made 18 playable characters. LEX TOUMBOUROU POST

Norway has over 2,000 underwater AI laser robots zapping salmon lice A company called Stingray puts a robot inside the pen that spots lice with two cameras and AI. Then it fires a green laser from up to 2 meters away. The salmon's scales reflect the beam, so the fish is fine and the louse dies. OLE LEHMANN X POST

AI-Ready Second Brain Sébastien Dubois says the gap between "I use Obsidian" and an AI-ready notes system is discipline, not tooling. Atomic notes, reliable metadata, and links let an AI find exactly what it needs. SÉBASTIEN DUBOIS POST

Here’s my current list of what I think people need to not just stay safe, but to thrive in this new world we’re moving into:

And the attributes I recommend you cultivate in yourself are:

The courage one is key because it can stop all the others if you don’t have it.

And of course I recommend H3 for anyone looking to learn AI and start building towards this stuff.