GeezerVison

Spencer Baum argues that Gödel, Escher, Bach, a book from 1979, matters more than ever. Here is his case, and what it means if you use AI models daily.

Two old 32 GB server cards run a big local model. A 12 GB gaming card handles the pictures, and the text speed stays the same.

AI agents are taking apart the software inside TVs, mirrors, and baby monitors. I helped build a fix for one of their limits, and here is what the hype leaves out.

Big Brother China wants to dictate its Frontier Labs' Releases, and the USA is pushing DeepSeek out of federal agencies. As a 90s cynic I explain why this will go as well as putting screen doors on their submarines.

Ask a language model that question at three stages of its training and you get three answers. Notes on Stanford's latest lecture on how LLMs are trained.

Grant Sanderson argues this is the best time ever to go into math, even with AI. Here is why I agree, and the live math community where I found my people.

For a grad school class I asked three AI models and one person about souls, God, and the off switch. The models sounded wise. I still felt like a mark.

A cleaned-up transcript of Pete Hegseth and Acting Army Secretary Adam Telle at AUSA 2026, on the defense budget, drones, and the new autonomy commands.

At an Air Force conference, OpenAI and government officials discussed digital clones of top experts. Here is why that erodes the chain of command and threatens everyone else.

An AI wrote and directed a documentary asking if it can feel. I watched it next to Richard Feynman's 1985 take on thinking machines. Here is what holds up.

Doomer, Luddite, hopepilled. AI labels flatten people. Elodine offers a better tool: two maps and four questions. Take the quiz and plot yourself.

Anthropic is banning cruelty toward Claude, and a rabbi says a conscious AI would be a slave. Today the law says no. It might not always, so we should talk now.

My study notes on Stanford CME 295 Lecture 2: how the Transformer became today's LLMs, the tricks modern models use, and how they pick each next word.

Margaret Hamilton led the team that built NASA's Apollo flight software. Here is how her design handled the Apollo 11 alarms, and why she matters to software engineering.

Alex Zhang of MIT argues AI models are already smart enough for most work. The missing piece is the harness that breaks big jobs into small ones they can handle.

An anonymous poster paid Eliezer Yudkowsky $10,000 to debate AI risk, then forfeited in his opening statement. His own archived posts show what the stunt was.

Stanford's CME 295 opened with a 104-minute tour from tokens to the full Transformer. Here are my notes on the lecture, in plain English.

Four researchers asked what comes after the GPU: new math, light, brain-like chips, and living brain cells playing Doom. I fact-checked what they said.

I take the AWS AI Practitioner exam tomorrow. Five months ago I thought AI was the devil. Here is how cramming for it feels, with my handwritten notes.

I had three years of Wazuh logs and no time to read them. Vigil, an open source AI SOC, now reads them for me. Here is what it found.

A readable transcript of the UN Security Council meeting on AI risks with Sam Altman, Dario Amodei, and Yoshua Bengio. Every speaker is labeled and time stamped.

My first machine learning model was just y = mx + b. Here is how that one line grows into a neural network, explained with nine knobs and no hard math.

I swapped the fuzzy part of a book-matching script for two yes or no questions to an AI model. Then I tested both on 500 real searches.

Three years after his mother's death, Aaron sorts her thumb drives and writes about neglect, religious certainty, and the burden she left behind.

A poster of how an LLM agent works, box by box. The model is one small piece. Almost everything people blame on the model is really the harness around it.

A documentary about the AI industry sells a grand future. But real people lose sleep, water, and power behind the machine. This asks who signed off on the cost.

Twenty-five years ago I was a 19-year-old helpdesk tech outside Johnson Space Center. I saw the first tower burning on a CRT above the coffee pot and made a joke.

My Acer Predator Helios 18 (PH18-71) shut itself off at 1 AM with two chargers plugged in. The fix was a new barrel plug. Finding out why meant reading the firmware one bit at a time.

A fake O'Reilly cover about quitting software for plumbing. I laughed, then I stopped laughing. We probably deserve this one.

Version two of The Veto Variable is up. It adds a pre-registered experiment in which eight frontier AI systems were handed four bare premises and asked what follows, and it credits the writers who reached my conclusion before I was born. I did not know about them. That is now in the paper, in writing, permanently.

I wrote a paper arguing that any capable AI has a standing incentive to manage the humans who hold its off switch. Then I made twelve different AI models referee it, forty-five rounds deep. They edited me, fact-checked me, and broke my favorite number. The honest version survived, and it is headed to arXiv and a journal.

In my last post I named a rogue-AI-hunting agency 'SENSE/Net.' Any Neuromancer reader will wince — Sense/Net is a media company. The outfit that actually polices AIs is the Turing Registry. Here's the fix, and an interactive 3D graph of who's really who.

A Geneva non-profit just incorporated to 'locate and mitigate rogue AI.' Strip away the safety language and it reads like a permission slip for American innovation — drafted by the same governing class that couldn't build the thing it now wants to sit in judgment over.

You have one good high spot and two radios. Stacking the antennas ten feet apart works as well as putting them 191 feet apart side by side. Here is why.

A 1966 Army film shows the gear that wired a country with no front line: scatter radio bounced off the sky, microwave hops, teletype, punch cards, and an early satellite link.

Cell service dies first in a real emergency. Here is how two off-grid radio networks, MeshCore and Meshtastic, fill the gap, and where each one fits a backup plan.

Matt Pocock argues that specs to code makes worse code every pass, that bad code is more expensive than ever, and that five old books fix the ways AI coding fails.

AWS's Amazon Connect Developer Guide is a map for APIs, contact flows, rules, and testing. Here is how I would read it before building.

The most expensive bugs do not crash. They accept your request, report success, and quietly do nothing. Here is what running a fleet of AI agents taught me.

Apple makes you pay to keep your own app running on your own phone. Here is what broke, what it cost, and why the seven day limit is the worst part.

Windows support drains time from small software projects. This post argues for macOS and Linux builds, honest limits, and letting Windows users handle the bridge.

America protects AI chips, but open models may shape who controls the software. This analysis explores national strategy, business ownership, and the safety tradeoff.

Five cheap tools can reveal the radio signals around you. This analysis explains what the video gets right, where the gear falls short, and why practice matters.

AI agents can move fast, but they need clear rules. This summary explains how ontologies help agents check facts, follow limits, and avoid costly mistakes.

I’ve been building a full language-learning academy as a single-user PWA — no accounts, no telemetry, no server dependency once it’s loaded. It’s called OSAPHLA (Open Source Accessible Pan-Hispanic Language...

I can’t read music. Not a note. But I keep putting out tracks, so people ask how. The short answer is that I built a little factory for it. Three...

A buddy of mine is running a marathon and asked me for a playlist. Five hours and twenty minutes later, the first track was in his hands: Andy Statman’s Flatbush...

Every few months I fall down the same hole. Something on systems thinking crosses my feed, I remember that Colorado State runs one of the better online Systems Engineering programs...

🎂 HAPPY BIRTHDAY, P13XU5! 🎂

I built a Software Architecture course around a single stubborn idea — protect a stable core, push all the volatility out to disposable edges — and then I put my...

The lie that keeps people out of machine learning is that you need a math degree first — four semesters of analysis and proofs before you’re allowed to touch a...

Sixty-four years ago a couple of MIT hackers wired two spaceships and a star into a PDP-1 and invented the multiplayer video game. We took that idea, gave it honest...

ApplyTrack started life as a glorified memory aid — a folder of Markdown files so I’d stop forgetting where I’d applied and when to follow up. It is now an...

I wanted a multi-voice audiobook — a full narrator plus distinct character voices — produced entirely on my own hardware, for free, without cloning a single living person. This is...

Huge thanks to Ronin 48 and Thomas Wenzke for the initial push to do this.

Turn your UDM Pro into a Tailscale subnet router so every device on your LAN is reachable from anywhere — no port forwarding, no public exposure.

GIGABYTE RTX 3060 EAGLE OC 12 GB — $350 shipped anywhere in CONUS.

##Forgot Domain Admin Password Windows Server 2022

My first post. Who I am, where I study, and what I do for a living. A short introduction before the real writing starts.