Thoughts on AI, learning, and building things.
Three frontier models given the same six-word SVG prompt. Claude drew ten unique characters, GPT drew ten uniform figures with red crosses, and Gemini wrote the most code-efficient solution using <defs> and <use>.
500+ independent API calls to Claude Sonnet 4, asking for one fictional doctor each time. The same male cardiologist named James Whitfield appeared in 167 of 400 outputs — the model returning the peak of its training distribution.
Read more →A breakdown of Dan Koe's viral article — the one-day protocol for identity-first change, the Anti-Vision framework, morning excavation questions, and key takeaways.
Read more →A reflection on what three books taught me about the two traps of money — spending everything and spending nothing — and why uncertain times make it harder, not easier, to get it right.
Read more →An honest look at what Claude Code excels at, where it struggles, and the guardrails every developer should set when using AI coding assistants to build real projects.
Read more →Meta open-sources a model trained on 1,000+ hours of fMRI data that predicts neural responses to content — and combined with their wearable hardware, it starts to look like a closed loop on human attention.
Read more →What actually improves Claude's output, what's cargo-culted nonsense, and which popular prompting tricks you can safely drop.
Read more →Nine categories where small teams can build real AI businesses — from vertical SaaS and consulting to fintech and developer tools, with honest assessments of competition and defensibility.
Read more →What the numbers actually say about AI's energy consumption, CO2 footprint, and water usage — plus practical steps developers can take to reduce the impact.
Read more →Prompt injection, data leakage, API abuse, XSS, and privacy — the security checklist for embedding an LLM chat on your site, drawn from building my own AI Twin.
Read more →Prompt injection, deepfakes, model theft, data poisoning, agent autonomy risks, and supply-chain attacks — the AI security landscape as of early 2026.
Read more →A practical overview of Supabase — its components, common use cases, the security risks developers overlook, and how to prevent them.
Read more →How the RAG pipeline works, the key decisions when building one, and what I learned putting it together for my Financial Risk Copilot.
Read more →Practical techniques for fine-tuning large language models on limited hardware — LoRA, QLoRA, quantization, and smart dataset choices.
Read more →How I got started with AI, the resources that helped most, and advice for others beginning the same journey.
Read more →