A picture from 2007, a Mac in the Theoretical CS lab at IIT Madras, and why vibe coding with Claude feels exactly like those 5 a.m. all-nighters from undergrad. This is the opening note for everything that follows.

LLMs are good at generating analysis. What they’re bad at is assembling separately generated files into one coherent report. The fix was replacing the merge agent with 300 lines of deterministic Python — and using Claude’s own failure taxonomy as an engineering input.

Can a single scraper work on any company's IR website without site-specific code? Eight platform types fingerprinted, a Linnaeus-inspired taxonomy, Taleb's barbell principle applied to bot evasion, and lessons learned about why discovery is harder than downloading.

Using Claude to transcribe Indian classical music lessons from the guru-shishya tradition. How building a memory aid changed how I listen, why context and a "student state" document made the transcripts useful, and what it means to systematise knowledge that has resisted it for two millennia.

A five-stage pipeline that turns seventeen browser tabs into four. Intake, classification, extraction, synthesis, and flagging — each doing one thing well. The Unix pipe philosophy applied to investment research, and why defining the process was more valuable than automating it.

Deep dive into ChatGPT's hallucination problem. Testing accuracy on factual questions, understanding when and why models confidently state false information, and what this means for building with language models in production.

My predictions for how AI will evolve in 2024. Two key words that capture the direction of the technology, the market, and the kinds of problems people will solve with language models.

Exploring DALL-E's capabilities and quirks. Testing the boundaries of image generation, understanding what works, what breaks, and the gap between what these models can do and what we actually want them to do.