Recurse Center: return "statement";

— 4 minute read

Boy did that go fast...

RC has a motto to "Never Graduate". An encouragement to always persue learning and growth. To build your volitional muscles. To learn generously.

A worry I had before joining my batch was ~unemployment~. Being laid off earlier last year, hearing about all the other layoffs in the industry, and going through some loops before the new year, I knew it was going to be hard to focus 100% on RC.

But during my interview with RC and during the batch, I was encouraged to take care of myself and do what I needed to do. And I found support from some of my batchmates in similar situations.

It's a weird time right now. Whether people blame it on "ZIRP", pumping up stock prices, or looming AI automation, the tech job market is not poppin'.

Anyways, back to RC.

What were your goals when you applied to RC? How did you expect to spend your time here? What did you actually work on?

To roll back the tape one blog post.

  1. Contribute regularly to PyTorch with the goal of contributing to a major feature. ➡️
  2. Learn the in’s and out’s of ML / AI and share learnings along the way. ➡️
  3. Join a team building cutting-edge ML / AI solutions to solve pressing and meaningful problems. ➡️

All on-going. "Never graduate" - right?

PyTorch

For PyTorch, I got two small PRs in. A fellow Recurser recommended I take a look at the MPS implementation of PyTorch as an area that could use some improvement and that was a great idea. I love the idea of making ML more accessible for the "GPU-poor". A Mac Mini M2 with 16GB of RAM is $700. For some rough comparison of LLM inference performance, checkout this repo.

While I loved building my PC back in the day, it's not something I'm drawn to anymore and don't think most people will either. A low power server that I can SSH into via Tailscale is awesome and I highly recommend if you already have a laptop and don't want to shell out $2K+ for a new M-series Macbook Pro.

I'm looking forward to diving deeper into the MPS implementation.

Learning ML / AI

My two slides on Transformers and DPO were well received on LinkedIn. I enjoy teaching and sharing a lot so that refilled the tank. They both did take longer than expected but that's the price of learning. It's also nice to have some artifact to post on the blog. Check out the slides.

The field doesn't stop so also looking forward to doing more of this.

The Latent Space Discord has a weekly journal club on Wednesdays at 12pm Pacific if you want more of this.

Job

This one is difficult. I still think of this blog post from DeepMind about AI modeling global weather forecasting. HOW COOL IS THAT?!

That's the direction I want to move in. The journey is likely a 5-10 year endeavour. I just need to keep moving in that direction.

So for now, I'm looking to do what I'm good at - building backend and infra systems, while continuing to learn and apply my learning to create cool ML / AI powered solutions.

RC Thoughts

Being remote

TBH, I think to experience RC the fullest, you'll want to visit regularly.

Being in LA, I got to meet up with a local batchmate twice and that was really nice. I just imagine that times 100 being in Brooklyn.

RC does put a lot of effort into setting up coffee chats and pairing sessions. Those were always nice too. I just get too much Zoom fatigue and after Zoom, I go back to staring at my monitor.

A cool thing is that RCers are welcome to apply again to later batches. I'd love to go in-person one day.

Career team

The career team has been very helpful in getting me intros and interviews with companies. They're so encouraging and I really appreciate that. I love that they have this to lean on.

Conclusion

I'm running out of gas writing this blog post. I look fondly back at my time in RC and recommend tinkerers from all backgrounds to check it out, and I look forward to building connections with alumni in the area too.

Take care.