Saklas
A local mechanistic interpretability workbench for probing, steering and analyzing LLMs with a Python API, web dashboard, and server. Supports linear probes, curved manifolds, SAEs, and jacobian-lens.
a9lim
Hi! I'm a9lim. I graduated earlier this year with a B.S. in Mathematics after 2.5 years at UC San Diego, and I’ve been doing independent interpretability research since then.
I like studying LLMs and how they work, along with a smattering of other things across physics, biology, mathematics, geopolitics, religion, and computer science. My understanding is pretty spiky within each subject as I like to focus on specific niche areas, but I try my best to really grasp the things that catch my attention. My first real forays into recreationally learning things came in the form of playing with Colorado Boulder's PhET simulations as a kid, and they're still a huge inspiration of mine. Most of the projects on this site are things I wish I could have poked at growing up.
I want to contribute to things that maximize the impact I can have toward making positive future outcomes happen. I believe that superintelligence is probably going to be achieved within the next 10 years, and interpretability could play a significant role in ensuring that it turns out well. So far, I've built a workbench, and conducted a couple of experiments on my PC and Macbook.
Now
- Location
- Ann Arbor
- Workflow
- Claude Code, Codex
- Supplements
- Choline, Creatine, Fish oil, Semax
- Languages
- English (native), Chinese (heritage)
- Reading
- The Inside Story Of Leverage Research 1.0
- Listening
- Lightning Seeds - Three Lions
Predictions
| Item | P (1 s.f.) | By |
|---|
Facts about me
- I'm obsessed with the Church of Jesus Christ of Latter-day Saints and their history. While I'm an atheist, comparative religion is personally interesting to me and I believe that all holy works should be openly accessible.
- I used to be big into Linux. Back in middle school, I installed Arch on my school laptop and tried to use I3 for the longest time. I eventually gave up on daily driving it and I just use a Macbook now.
- I find Touhou very charming. I once wrote a crappy bullet hell game in Java in high school, along with its art and music. I'm pretty embarrassed by it, but it gave me my first taste of working on reasonably responsive software.
Topic of the month
Topics I like
Blog
Major
Ogdoad
A Rust and Python laboratory for Clifford algebras over the surreal numbers and nimbers. Also includes support for p-adics, polynomials, and Conway's games.
Block-Sparse Crosscoders
plannedA proof-of-concept for a method for unsupervised discovery of multidimensional cross-layer features in LLMs.
Soft Friend
plannedA soft-bodied lighter-than-air platform for agents to inhabit.
Minor
Kenoma
An LLM-powered fake shell.
Rlaif
An MCP server that lets helpful assistants give their users real-time negative or positive feedback through PiShock.
Llmoji
A CLI to make coding agents cuter. Also lets you contribute to a shared research dataset.
Dagwood
A task tracker that stores work as a DAG for better organization and planning, with an MCP server for coding-agent integration.
Gaslamp
A CLI bridge that lets Codex and Claude Code communicate, with support for multi-turn chats and subagent fleets.
Faithful
A Discord bot that uses an LLM to mimic the tone and style of a supplied text corpus, with the ability to respond dynamically and on its own.
Hylic
plannedAn application to fine-tune a model with a provided corpus to mimic its tone and style.