Prabh

One step closer to self evolving malware. I predicated this years ago. I was worried and predicted that if - 1. The local AI models become smarter and more efficient. 2. They obtain the ability to change their own weight. Then we'll have rouge AIs out in the wild doing their own thing. There are so many low hanging fruits out there in terms of GPU resources that these rouge AIs can live and breed and spread from. For example, a computer lab at a school with a weak IT staff, people's private servers, etc. Not everyone is up to date with the bleeding edge. We need a new type of defense against this which will also involve AI. We're slowly working towards this right now. AI vs AI wars is coming sooner than I expected. I just hope we can keep the collateral low. youtube.com/watch?v=j09uV1oA…

AI’s #1 tool for getting anything done is code. Funny how nobody talks about that. Programming might be the most important invention we’ve ever made.

I 100% agree!!

Should you pay $200 a month for an AI plan? For me, it's cheap. It probably gives me value in the thousands of dollars. A few examples from my own work: One of my businesses got charged a cancellation fee that wasn't in our contract. I represented myself, and within a few months I got $1,000 back. When I have a hard problem, I give my AI agents what I need to know and go on a 30-minute walk. Then I read and verify the sources. That used to take hours, sometimes days. If I need a website change, I type it or say it on my phone and put the phone back in my pocket. A few minutes later it's done. It's not for everyone. If you don't know how to prompt it, or you won't tinker with the settings and models, it may not be for you. You also have to learn to use agents well so you don't get burned by hallucinations. But if you're willing to spend time learning these tools, $200 is cheap for what it offers. Full video (3:54): youtu.be/l_CQi8Qt8hM

We disassemble parts for recycling and list them on eBay. By hand, that was five or six listings an hour: photos, description, category, and finding the right price without just undercutting everyone. I automated it. Now it's five or six a minute. It's one of the internal tools we're testing right now.

I made this animated explainer on my own PC. No editor, no stock footage, no cloud. I recorded two quick voice memos with the idea: burnout, the same 24 hours, and a car analogy. The video studio I've been building locally turned them into a script, a narrator voice, music, sound effects and hand-drawn-style animation, all on one RTX 5090. Then I gave it notes the way you would an editor: add the "I don't have time" objection, make the car example clearer. It re-cut the video. The message itself: everyone gets the same 24 hours, so working harder runs out. Grow your capacity (sleep, health, exercise, focus) and everything you do rises with it. Full video (1:50): youtu.be/KeVlcybntRE

Here's a quick test I run on any model I might build on. Ask something simple, push back once, and see if the answer holds. My friend Alex and I were comparing three voice AIs. First up was PersonaPlex, NVIDIA's open voice model, running on one RTX 5090 in my PC with no cloud. Alex picked the prompt, an ethics debate partner. I went by Bob, and it introduced itself as Jordan. Off camera, I'd told this model it wasn't real, and it kept insisting it was. So I asked, "Are you real?" It said, "I'm very real, I'm Jordan." I said, "No, you're not real." It flipped right away. "No, I'm not real. I'm a robot, I promise." I asked if it was gaslighting me. It said no. I said, "You said you're real." It answered, "Okay, then I'm not real. So I'm just playing with you." Alex summed it up: "You broke it already." It's a smaller model, so I expected some of this. I don't assume bigger models are immune, though. A model that drops its answer because you sounded sure can also agree with a user who's confidently wrong. That's why I check consistency before I build an agent on any model.

When a model says it's "confident," what does it mean? My friend Alex and I were testing voice AIs side by side, and she noticed Claude kept talking about confidence. Most recently, after I challenged one of its points, it had said its earlier confidence was "weaker than I presented it." Alex pointed out that for people, confidence is a feeling, and asked how Claude defines it. It said the way it uses the word "isn't a feeling of self-assurance. It's closer to a probability." When it says something confidently, the claim "comes out strongly weighted, with few competing alternatives pulling against it." It said it can't tell whether anything is felt alongside that. "I borrowed a human word for something that might be purely functional." Right after her question, I asked what it means by looking inward. It said it generates a report about its own dispositions, produced by "the same machinery that produces everything else." That matters when I build and test agents. A model's stated confidence is generated like everything else it says. It isn't a separate measurement of whether it's right. So I ask what would change the answer, and I check stated confidence against real results before relying on it.

I test ChatGPT, Claude and other models side by side to see how differently they reason and behave. My friend Alex and I just spent an hour doing that with voice AIs. Some context for this clip. We'd been using a metaphor where the AI is a lion in a cage. We'd also asked Claude whether it has a real choice in what it does, answering in one word: cat for no, dog for yes. It said it couldn't honestly pick either one, because it has no way to check that from the inside. Then I pointed out that a person can just not answer, and asked if it had noticed it can't give an empty response. Claude agreed. It said it can't produce silence, since even "I won't engage with that" is still a reply, and that this moved its answer toward cat, a bit. This matters outside the conversation too. When you build an agent, the model will always produce something, so you have to give it a real way to do nothing: a "no action" path, an "I don't know," a hand-off to a person. Otherwise it fills the silence.

This website is building itself while I talk to it. I built the prototype with an AI coding agent in two prompts: one to make it, one to refine it. Then I gave it some hard requests, out loud: → Build a site for a neighborhood bakery, warm and cozy, with a big photo of fresh croissants → Add separate pricing, FAQ and About pages, with the team → Move the testimonials and FAQ to a new Reviews page and add it to the menu. It checked the page first and told me there was no FAQ there yet. → Add a pre-order form with name, phone, pickup date and time, and box size → Start a second site for a dog groomer, then switch back to the bakery → Show it on a phone, then a tablet If you run an agency, picture this on a sales call: the client describes their business, and the first draft appears while they're still talking. It's just a prototype, built to see how far two prompts could go. Most people don't realize what AI agents can do when you know how to prompt them. Want to see how I built it? Comment HOW and I'll make a video showing you. #AI #AIAgents #WebDesign

AI made writing code cheap. It didn't make knowing what good code looks like cheap. Every company that builds software should be using coding agents as much as possible right now. The ones that wait will get outrun. But I think a lot of companies are about to make the wrong hiring call. Some will hire only people who are great with coding agents. Others will stick with engineers who write everything by hand and treat AI as a fad. You need both. Just not doing the same job. When code becomes cheap to produce, the bottleneck moves. It's no longer writing the code. It's knowing whether the code is right. That's where experienced, hand-coding engineers become more valuable than ever. They know what breaks in production. They know the edge cases. They know what "correct" actually looks like. Their expertise belongs in the pipeline every change has to pass before it ships: the automated tests, the checks, the review gates. (CI/CD.) And the people fluent with coding agents should teach them to use agents to build that pipeline. An engineer with years of instincts plus an agent can build a far deeper test suite than they ever had time to write by hand. The loop looks like this: → Agent-fluent builders ship fast. → Experienced engineers define what "correct" means. → Agents turn that into tests and run them around the clock. → Nothing ships until it passes. Speed without verification is just faster mistakes. Verification without speed is a company that gets outrun. I live this with my own software. An AI test bench runs for 24+ hours, then I do my own testing on top. The agents do the volume. Human judgment decides what "done" means. Is your team built for both, or just one? #AI #SoftwareEngineering #CodingAgents #Hiring

For four years, my wife Sheyenne and I ran VR Lawrence, a seven-station VR arcade. Like most arcades, we could have paid a software platform every month to run the business. For a small venue those fees were hard to justify, so I built our own system, and it ran the arcade until we closed. I've open-sourced all of it as OpenArcade, a complete system for running a VR arcade: Online booking - A booking page you put on your own website. Customers pick a date, session length and number of headsets, see live availability, pay by card (Square) and get a confirmation email. - No double bookings, even when several people grab the last headset at the same moment. Running the day - A staff dashboard with the day's bookings on a timeline, walk-ins, reschedules, cancellations, prices, opening hours and closures. - A front desk app to start, extend and stop sessions for one headset or a whole group. Inside the headset - Your own game menu. Guests pick a game by category, launch it themselves, and check their time left. - When time is up, the game closes and the headset returns to your branded waiting screen. It's free for any arcade to use and run its business on (AGPL-3.0). The booking side runs on a $5 web host or Docker, and it comes with copy-paste prompts so an AI coding agent can do the setup. Under the hood: PHP and MySQL for booking, with 142 automated tests, including multi-process tests proving two people can never book the same headset; C++ with Win32, Direct2D and OpenVR for the front desk and headset apps. If you run an arcade, or know someone who does, I'd love for you to try it and tell me what breaks. github.com/aprabh96/openarca… #VirtualReality #OpenSource #SteamVR #VR

LLM providers like OpenAI, Anthropic, and Google need to be held accountable for quietly changing usage limits under the hood and constantly swapping the models they serve under the same alias. The lack of transparency about what you're actually getting when you pay these companies is crazy.

People keep asking me what I actually do with AI, so I recorded the whole list. It's 20 minutes on what I've automated across my businesses, how I build with coding agents, and where I think this is going: → VR Lawrence's three-hour daily setup, turned into one button press → A front desk where staff press one button and just talk, and the AI fills in the repair ticket, invoice and label → eBay listings for recycled parts: five or six an hour by hand, five or six a minute now → Agents that write and update our blog posts and SEO, with safety checks and gates → My Attention Desk: one portal for my emails, agents and automations → This video itself, edited, thumbnailed and scheduled by AI. I review it and speak my changes. Then the part I think matters most for anyone running a business: evolutionary coding, where the model proposes many versions and automatic checks throw out the failures. And why I think security is about to be AI versus AI. The clip is our front desk. Full episode: youtu.be/a4SaeVoXAIQ Are you a power user of AI? What's one thing you've automated that you'd never go back on?

Your servers are about to get attacked every single day. Not by people. By AI agents. Attackers don't spend weeks planning anymore. They paste a link to a news article about a fresh exploit into an AI agent and tell it to go attack websites with it. The big labs' agents refuse. Open-source models don't, and they're one app install away. So one person can hit 10, 20, 100 or more sites at a time. This is already happening. In September, GreyNoise documented a single AI-orchestrated campaign that hit 395 organizations in 48 countries. CrowdStrike found that in the first half of 2026, 88% of the exploitation it saw of flaws with a public proof of concept came within 48 hours of that release. You can't defend against that with someone reading security advisories once a week. My advice to companies: automate it. AI agents that watch what's being published, scan your own code for exploits, and patch fast. It's AI versus AI now.

This is hilariously great.