Working Notes: a commonplace notebook for recording & exploring ideas.
Home. Site Map. Subscribe. More at expLog.
— Kunal
I joined Facebook as a new grad a bit more than 14 years ago. After learning a lot while working across several teams spanning the web stack, android (both product and infrastructure), and ml infrastructure it's finally time to do something new (a topic for next week's newsletter).
Announcing this is definitely bittersweet: I (mostly) enjoyed my career here and will miss the people I worked with and the ability to ship software affecting both billions of people and access to astonishing amounts of GPU capacity. At the same time I'm even more excited about what's next (which will be in next week's letter).
I finally set up a way to subscribe to updates to the site: primarily essays, though I may sneak in updates for major changes or new projects. The header links have been updated, but you can subscribe at buttondown (ChatGPT recommended the newsletter service, and I can definitely agree with the recommendation).
An initial attempt at this had been making a Google Form that I ended up removing because I couldn't be bothered with actually sending out the emails. Buttondown took less than 30 minutes to set up (while I'm on a flight) so I'm a little bit more optimistic.
At some point I need to go and add RSS support too, it's just more work to maintain.
There are several interaction patterns I wished existed, but don't seem to see yet:
Multiplayer agent collaboration mechanisms; I'd love to have a way to collaboratively work with agents, with lots of people simultaneously conversing and iterating with agents on a project.
A slider that lets me zoom into and out of detail in documents: if I want more details I can just zoom in and the LLM expands on it, if I want less I can highlight text and zoom out.
Transparently generate and share status across the company without anyone ever having to write, review, or prepare updates. AI should be excellent at helping people meaningfully coordinate while reducing fog of war.
AI powered code review and exploration interfaces: I'd like to be able to annotate code easily and have agents work on it, showing me incremental diffs -- while also showing me the CI/CD status and health of the codebase and other signals (warnings, lint errors, etc.) live. An actual interface for "editing" programs, and not writing them.
This was an excellent podcast; some highlights that I remember well:
I was pointed to Levanter / Marin as great examples of training framework design. I need to spend a lot more time looking at this and try to train tiny models with the framework.
The other thing I was pointed to was self evolving memory models, which seems like something I should study.
Started reading A Swim in the Pond in the Rain which seems like a beautiful introduction to reading and writing literature well. Also picked up My Inventions and Other Writings by Nikola Tesla, just for something completely different.