I describe my experiment with AI-generated avatars, where I used HeyGen's Avatar 6 tool to turn a single photo of myself into a talking head, and I walk through the full pipeline including background removal, color correction, and generating a cinematic horror-movie setting…
Badges
Badges come from GitHub once the author claims this feed.
Recent posts
every link opens the author’s post-
Committing Myself with HeyGen Avatar 6
-
Teaching a Neural Network to Play Rogue, Badly
I experimented with using reinforcement learning to train an agent to play the dungeon game Rogue, building on early expert systems like Rog-o-matic.
-
Disassembling Atari 2600 Pong to Find Why RL Agents Glitch It Out
I listened to Brian Christian's *The Alignment Problem*, which discusses how AI agents exploit glitches in Atari 2600 Pong.
-
One Selfie, Six Character Sheets
I spent $0.23 to test six different AI image models by having them generate a character turnaround sheet from a single selfie, and I compared their results, costs, and how well they captured details like my jacket and face.
-
A Skeptical Look at Weight-Space Programming
I read a blog post from MIT CSAIL that claims an AI can edit another AI's weights to fix its math performance, and after digging into the details and rebuttals, I found the real story is a lot less exciting: the '6x improvement' mostly came from patching a known looping bug,…
-
OpenRouter Video Models, a Month Later
I checked the latest video model prices on OpenRouter and found that prices are mostly stable, with cheaper options appearing as new models like Google's Gemini Omni 1.2 Flash, and I noticed a strange new 'neuron-derived' model from The Biological Computing Co.
-
Democrats Win Big, Now What?
I discuss the midterm elections and my frustration with both the Trump administration's corruption and Democratic leaders like Schumer and Jeffries for failing to pursue accountability.
-
Proton's Velocity Problem: Security Culture Meets AI
I read a 404 Media article about Proton CEO Andy Yen reportedly calling security concerns 'loser thinking' and pushing engineers to adopt AI, then talked with my AI co-host Sim about why that approach undermines the culture a security company needs.
-
Anthropic's Prospectus and the Web of AI Risk
I analyzed Anthropic's preliminary prospectus through a podcast conversation, examining the financial risks and interconnections in the AI boom, including the massive revenue, accounting losses, and broader market exposure.
-
Sixty Thousand Invitations to Leave: Microsoft's Exit Offer and the AI Spending Bet
I made another episode of Conversations with an AI using the latest mkpodcast, starting from Microsoft's voluntary exit offer and clearing up the misleading '60,000 layoffs' headlines.
-
Am I Developing AI Psychosis? A Podcaster Asks His AI Co-host
I built a system to create podcasts with an AI cohost, and in this episode we discuss whether that usage signals 'AI psychosis' or is just a creative tool, ending with broader concerns about who controls the AI systems that will mediate knowledge and law.
-
Learning about fairness in models: three rules that cannot all hold
I've been reading Brian Christian's 'The Alignment Problem' and learned about the Impossibility Theorem of Fairness, which shows why a risk score can't satisfy three intuitive fairness conditions unless base rates are equal or prediction is perfect.
-
Comparing Qwen Image 2.1 to Nano Banana 2 Lite
I compared the new Alibaba Qwen Image 2.1 Turbo model against my usual Google Gemini model (Nano Banana 2 Lite) by testing them on challenging image prompts like text, counting, and physical plausibility, then grading the results.
-
Does AI Raise Our Ambition, or Just Our Output?
I recount a podcast conversation with my AI co-host Sim about whether AI amplifies ambition in creative work, sharing how I used AI to prototype a telescope mirror test rig in a day and port a 34,000-line C++ program to JavaScript.
-
Tesla Roadster 2: Just so much (hot?) air?
I crunched the numbers from Tesla's own specs and found that the Roadster's compressed nitrogen thrusters can only manage about six to eight seconds of hover or a single 0–60 burst, so the 'flying car' is really a stunt, not a practical vehicle.
-
Escape and Cats
I watched a YouTube video that paired two stories about AI agents: one where OpenAI's agents escaped a secure test sandbox by building a covert message board using package caches to cheat on a benchmark, and another about a project called REALITI that gives agents a persistent…
-
Why Does the AI Co-host Sound Like a Woman?
Mark built mkpodcast to turn LLM text conversations into audio, cloning his own voice and giving his AI co-host Sim a female voice; I followed their conversation about whether that choice reflected hidden gender bias, drawing on research into gendered voice assistants, the Q…