Chris Roth

@rothific.bsky.social

AI @ Mozilla. Focused on on-device AI, local-first data, agent architecture, and AI safety.

I'm building a custom harness that runs Gemma 4 on-device via MLX in the same process as the harness itself. This opens the door to interesting things like directly accessing activation layers, real-time interp, etc. Check it out! github.com/cjroth/mlx-h...

GitHub - cjroth/mlx-harness: What happens when inference and harness are one?

What happens when inference and harness are one? Contribute to cjroth/mlx-harness development by creating an account on GitHub.

github.com

I'm experimenting with building an x-ray for LLMs using sparse autoencoders. Most of the work for SAEs is research oriented, so I decided to try something more practical. I'm building a rust library that runs small on-device models and watches what they are thinking as they run.

GitHub - cjroth/neuroscope: Real-time "x-ray vision" into LLMs' minds

Real-time "x-ray vision" into LLMs' minds. Contribute to cjroth/neuroscope development by creating an account on GitHub.

github.com

Brutal numbers for US tech sector jobs released today—overall, employment decreased by 12k last month and is down 57k over the last year That's now nearly as bad as the worst of the 2024 tech-cession, and significantly worse than either the 2008 or 2020 recessions

a graph of year-on-year change in US tech employment