Skip to content

History

Hi 👋 I'm Matt and I made Frisky. This story tells you why.

Dask, rise and fall

Most people know me because I initially made Dask. Dask was a lot of my life for a while. I loved that work until I made a company around it, and then it became kind of a grind 😔.

Dask felt great because it brought tons of new of people into distributed computing. Dask was flexible enough to be used in new fields and was accessible enough that people without distributed systems experience could use it. But mostly, Dask was fun.

Then I started making money off of Dask, first at NVIDIA, and then in my own company. I like money, but it resulted in me shifting mostly to management and sales, which I didn't like (people are hard). So after making enough money and no longer having fun, I stepped back from community open source and commercial activity.

AI and Consulting Work

I built an early prototype for Frisky as a learn-how-to-use-AI experiment. It was super fast and cool but I didn't want to pay the effort to publish it out to the world (open source is emotionally expensive). Frisky hung out on my hard drive for a while.

Months passed and I'm doing some consulting work setting up an AI-forward devops stack for an early stage company. Agents could inspect and reason about the entire system, or at least the entire system except Dask. I found myself looking at Dask dashboards manually to understand performance, and resented the manual work.

Why am I looking at this dashboard like a chump?
AI should reason about Dask performance for me.

This brought me back to the Frisky prototype. Frisky had way better telemetry (telemetry is cheaper in Rust). I wanted a system that AI agents could inspect directly, without me being in the loop.

AI Feedback Loops

So I unboxed Frisky and pivoted the project to AI engineering feedback loops. I wasn't using AI to develop Frisky. I was teaching Frisky to teach AI to develop Frisky.

This worked. I would feed Frisky new workloads, point an agent at the frisky CLI, and good performance enhancements emerged.

Moreover, this process didn't produce slop. Frisky's codebase remains compact (more compact even than the manually built Dask scheduler codebase). The results here felt solid.

Frisky's primary innovation may be the agent-forward telemetry. I think of Frisky as a computational nervous system for distributed hardware, including both action and sensation. Frisky is designed to teach agents what's going on with their computations so that they can quickly adapt. Frisky helps agents use Frisky more intelligently.

Open Source and Ambitions

Frisky is free and available to use (see license), but it's not open source today. I may change my mind on this at some point, but right now my thinking is as follows:

  • I don't want to deal with a developer community (but love hearing from people)
  • I don't want to review pull requests (but please send issues)
  • I have more fun developing in private

I can appreciate people having concerns about this. Let's voice some concerns:

  • How do I know this isn't exfiltrating my data?
  • How do I know this is giving correct results?
  • Will you take this away in the future?

Those are all super-reasonable concerns. Some basic responses:

  • You don't have to use Frisky
  • You can inspect traffic in and out
  • My personal reputation is financially more valuable to me than your data

I suspect that in the future Frisky will either go big, at which point it'll be open sourced or something, or, more likely, it'll die like 99% of other software projects. I'm pretty ok with either case.

Have fun

I had loads of fun building Frisky and I look forward to using it whenever I get a chance. It's super fast, the dashboard feels beautiful, and working with it and agents feels really high-velocity to me.

I hope that you enjoy it too.

uvx --with numpy frisky demo