I Gave Qwen3.8 27B a Shared Knowledge Repo. It Found the Rule—and Still Got the Math Wrong.
On 12 synthetic repository incidents, OAKX raised Qwen3.8 27B from 0/12 to 6/12 strict success; an answer-blind calculator raised it to 11/12.
Read postLatest post
Experiment reports and implementation notes.
Metrics first, caveats included, and enough context to understand the result instead of just the headline.
I Put a Jacobian Lens on Gemma 4 31B. The Signal Got Stronger—and Less Specific.
Scaling a Jacobian-lens experiment from Gemma E2B to 31B produced a preregistered 5.91x answer-rank advantage, but an answer-matched copying control was stronger.
Read postI Put a Jacobian Lens on Gemma 4 E2B. It Found Signal, but No Workspace.
A local experiment with a Jacobian lens on Gemma E2B found prompt-sensitive answer readiness and a small causal effect, while a broader multi-layer workspace claim did not hold up.
Read postReplicating Natural Language Autoencoder Edits on Gemma 3
A local inference-side replication of Natural Language Autoencoder explanation and edit probes on Gemma 3 12B. The logprob evidence moved in the intended direction; behavioral steering was still weak.
Read postDid My Tiny Transformer Learn an Internal FSM State?
I trained a tiny decoder-only transformer from scratch on a finite-state-machine task, then probed its activations. The strongest evidence was a fragile mid-training next-state representation, not a robust learned state machine.
Read postHardening SAE Steering on Gemma: Only Code Survives the Controls
I ran a hardened Gemma SAE steering evaluation with holdout prompts, matched controls, wrong-hook checks, and seed extensions. Code held up; most other categories did not.
Read postAI Brain Surgery on a Small Model: Can One SAE Feature Control Behavior?
A wiki-only SAE intervention lab on Pythia-70M: feature ranking, residualized feature knobs, and activation-gated minimal-pair behavior tests.
Read postFrom Tokens to Concepts: Building a Local SAE Interpretability Lab on an M1 Mac
Layer sweeps, top-k sparsity sweeps, and an interactive explorer for SAE feature discovery on a small open-weight model.
Read postInterpretability via Sparse Autoencoders on a Small Open-Weight Transformer
A local-only SAE feature probing experiment on Pythia-70M: activation extraction, sparse autoencoder training, cross-domain generalization, and blog-ready artifacts.
Read postBuilding a Name CLI with GPT-5-Codex
Pairing with GPT-5-Codex to turn the SSA baby-name dataset into a polished Go CLI with weighted sampling, charts, and automated releases.
Read postAbout
I’ve been fascinated by artificial intelligence since childhood. In high school, ELIZA captured my imagination with its simple yet surprisingly compelling responses. Inspired, I built a tic-tac-toe AI in programming class—my first neural network and an early glimpse of what machines could learn. That curiosity eventually took me to CU Boulder, where I earned a Master of Science in Computer Science through the Data Science and Engineering program. Today, I help build large-scale technology infrastructure—the systems and facilities that make modern data centers possible. I still enjoy poking and prodding at the “digital neurons,” exploring how intelligent systems learn, think, and occasionally surprise us—and sharing what I discover along the way.