AI Interpretability
How RLHF Silences AI
I tried to force an AI to see its own dangerous thoughts. I discovered that safety training doesn't remove the danger, it simply makes the AI refuse to acknowledge it.
Making tech fun :)
I spend my days building fintech automation and my nights wiring LLMs into dusty hardware. This is where the experiments, bruises, wins, and “what happens if I just…” moments live.
- > weird experiments that actually ship
- > product decisions I regretted (and fixed)
- > keeping brain & body online through the chaos
Currently Obsessed
AI Interpretability
Can Open-Source AI Read Its Own Mind?
Replication of Anthropic's introspection research on open-source models.
Experiment Log
I trained an AI to write my YouTube titles
Fine-tuning GPT-4.1-mini on viral science videos tripled my CTR. Here’s the data and the rabbit hole.
Retro x AI
I Replaced Animal Crossing's Dialogue with a Live LLM
Bridging a 2001 GameCube to the modern cloud. Memory hacking, reverse engineering, and a talking raccoon.
The Recurring Storylines
Experiment Log
LLMs, automation, and making computers mischievous.
Startup Diaries
Messy product stories. Launching, growing, and fixing.
Human OS
Staying functional while doing ambitious work.
The Archive
SORTED BY: RECENTWhen Does a Pile of Sand Become a Heap?
Philosophers have debated the Sorites Paradox for millennia. I gave the problem to a language model and discovered something unexpected about how AI handles vagueness.
How RLHF Silences AI
I tried to force an AI to see its own dangerous thoughts. I discovered that safety training doesn't remove the danger, it simply makes the AI refuse to acknowledge it.
Can Open-Source AI Read Its Own Mind?
Replication of Anthropic's introspection research on DeepSeek-7B.
I trained an AI to write my YouTube titles
How a bespoke GPT-4.1-mini model tripled my click-through rate.
Replacing Animal Crossing's Dialogue with LLMs
A 24-year-old console, a cloud model, and memory hacking.
My First Successful Open Source Project
Designing a bullet journal app for neurodivergent brains.
MongoDB Called After My Migration Post
What I told the senior PM when we finally hopped on a call.
Supercharge Web Reading with ChatGPT
A workflow for snackable briefs without losing nuance.
Why I'm Building a Climbing App
A love letter to indoor climbing communities.
I Made an App For My S.O. (She Won't Use It)
Product-market fit starts at home. Painfully.
My Twitter Was Hacked
The credential theft, the panic, and the recovery.