A list of things that I came across this week which seem to be pretty cool.

  1. Why should we not use vector databases? I first came across this concept while going through vector sets on Redis. Current vector databases use a secondary index as the vector index, contrary to making the vector the main thing as done in vector sets. I intend to explore this idea in more detail.
  2. How flow models work
  3. Momentum in ML optimizers
  4. Deep double descent

This week I built

  1. A RL framework for optimization of traffic signals. Learnt a lot about deep RL, GNN and SUMO in the process. Check it out here. Wandb runs hosted here.
  2. Backend part of research assistant, to solve problems I faced while doing research last year. There’s a YC company doing the same thing, but I’d prefer a local solution.

What I’m reading

I’ve been looking into building websocket servers at scale as part of my tinychess project (local, simplified version of lichess). I know it won’t need nearly as much scale, still it’s really interesting to see how this works.

  1. Chat service in Go using Redis Pubsub
  2. Stateless WS server using Redis Pubsub
  3. Redis vector sets (new data structure)
  4. Reward hacking in RL: I discovered this blog from Saurabh Kumar’s twitter, and it explains one of the reasons why my RL agent was so awful.
  5. Reproducing hackernews writing style fingerprinting

TODO

  1. Learn how to jailbreak LLMs
    • https://unit42.paloaltonetworks.com/jailbreak-llms-through-camouflage-distraction/
    • https://jailbreaking-llms.github.io/
    • https://arxiv.org/abs/2403.12171
    • https://github.com/verazuo/jailbreak_llms
<
Previous Post
Why Data modeling
>
Next Post
Player flows
>
Reading List
My Reading List