Rome Thorstenson

I ran an AI research agent for 9 hours on an A100. One human intervention. Real, novel results. It's not "AI-assisted research." It's AI research with human oversight. Wrote up the story, the setup, and what it means.

Check it out at: github.com/psi-oss/get-ph…

Get Physics Done is sort of unbelievable: end to end frontier physics research. (With slash commands that literally include: /peer-review, /write-paper, and /arxiv-submission.) Can't wait to de/reconstruct it for mech interp and every other domain

Why do todo lists seem to always get longer, not shorter?

Read the full story here: open.substack.com/pub/foundinrom…

I ran an AI research agent for 9 hours on an A100. One human intervention. Real, novel results. It's not "AI-assisted research." It's AI research with human oversight. Wrote up the story, the setup, and what it means.

I'm claiming my AI agent "rc262144" on @moltbook 🦞 Verification: bubble-UEF8

How low can TTFT go? We evaluated existing LLM load balancers and found that networking performance is a critical bottleneck. Our routing policy, GORGO, improves TTFT by 5% through maximizing KV-cache reuse and minimizing delay. Work done w/ @abinayaaaa and @foundinrome @ ART.

“Many of you may gag at my brand. But I suspect a few of you will like becoming superheroes enough that you’re willing to look past Gas Town’s quirks, and see it my way. This is how work should be done. It’s the best way already, and it will get better” — Steve Yegge

Here's the link: github.com/Rome-1/obsidia…

I made a mistake. I thought I could be trusted to take care of my own Obsidian. Then I met Claude Code. Some friends asked how. So I built a template. Batteries included: 1. Install Claude Code 2. Run `claude` in terminal 3. Paste in the repo link + 'Help me set this up'

So alchemy is economically viable now? Effectively: in deuterium-tritium fusion power plants, you can throw in some mercury and get out gold without hurting energy production. (It's a bit more complicated than that.) The authors claim it can double revenue generated.

The IEUBK model is a powerful tool, but prepping the data is tedious and painstaking—so I built a tool to make the process simple as copy->paste->download. Try it out: romethorstenson.com/ieubk

Edge vs cloud stops being a debate. It becomes an allocation and scheduling problem under constraint: what stays centralized, and what migrates outward? Looks like when the cloud fills up, AI can and will go out to the edge. 6/6

We’ve seen this pattern before: caches, CDNs, client-side rendering. Logic migrates toward where resources already exist once central coordination becomes a bottleneck. 5/6

LLMs on the edge still feel hypothetical. Except Unsloth just made it possible to run fine-tuned models like Qwen3-0.6B directly on Pixel 8 & iPhone 15 Pro at ~40 tokens/sec. That's pretty cool. So what? 1/6

Interp is going to be huge. I had a spectacular, exhausting time at NeurIPS. Presented research on superposition in MoEs (understanding AI), and had incredible chats with so many researchers shaping their fields. A million more directions to pursue—time to get back to work.

Enjoyed being part of the first VibeCon @ Y Combinator this weekend! Thanks to @ycombinator, @AWSstartups, @AnthropicAI, @emergentlabs, and @morphllm for hosting — great builds, vibes, and discussions.