← Home

Week 1155

I want to be a better writer. This desire was one of the motivating factors that led me to start writing these weekly updates. I was recently reminded of this when I found out that a friend from Retool was quitting to become a writer. He sent me one of his articles and I was blown away by the creativity and style in his writing. One highlight was when he described the color of a bird as an “unripe banana's green.” I would have never correlated bananas with the color green although it seems obvious after thinking about it. The hard thing is to even consider thinking about it and be open minded to these connections that don’t stand out at the surface level. It just reminded me that I have a lot to improve on and that there is a long journey ahead.

--------

I went out to some bars in the Marina District on Friday night. It was my first time going out in a long time and it just reminded me of why I won’t be doing it again for a while. You sacrifice your health by sleeping way too late, drinking way too much alcohol, and getting your eardrums blasted by music that’s way too loud. I felt terrible the next day; I was fatigued and couldn’t think straight for the whole day. I did meet some cool people and caught up with some friends, but there’s many other non-destructive ways to do that.

--------

I watched this video from Peter Whidden where he trained an reinforcement learning agent to play Pokemon Red. At first, I thought the concept was kind of dumb because you can just program a deterministic agent with a preset list of rules to beat the game. The RL agent was making mistakes that I thought was could be easily avoided with just a bit of human intervention. However, I soon realized how powerful the RL paradigm can be and how it can surpass human performance if incentivized correctly. After several training iterations, the RL agent discovered a specific way to walk around Pallet Town that would trigger a guaranteed capture on the first wild pokemon encounter. This was due to the pseudo-randomness of the game which relied on the user inputs. This is an insight that would be near impossible for humans to discover unless you deeply understand how the game’s pseudo-random generator works (probably on the scale of tens of people in the world?). It’s inspiring to see these type of projects and it would be really cool to do my own experiments like this in the future.

--------

I started digging more into stable diffusion this week. I’ve been going down a rabbit hole of github projects, reddit threads, and discord servers to get stable diffusion running locally on my Mac laptop. It’s been super fun and interesting to learn about. The quality of current open source models is surprisingly good and there are so many ways you can customize the image generation pipeline to suit your needs. I think I’ll soon hit a point where the bottleneck is my own creativity. It’s actually quite difficult to come up with good image ideas. Maybe it could be useful to build a creative exploration tool for the images similar to Google’s TextFX.

--------

Some random stuff I did this week:

--------

Content consumption: