dreamt that i was late to an exam and did not study for the exam too. double whammy. i've had a lot more exam dreams since i joined OE.

discovered isometric NYC, a side project from an engineer at Deepmind who wanted to build a giant isometric pixel-art map of New York City. he goes into the nitty-gritty of fine-tuning Qwen/Image-Edit models to generate each tile, how he did the training, scaled the generation, handled edge cases, and built micro tools at every step. it's not often you see someone approach a vibe-coded project with that level of detail, and you can tell he has real experience building things.

i liked the part about how as an artist, a big chunk of time in creation doesn't require creativity — it's just "slog". every creative field, be it animation, video, or software, is full of tedious tasks. you can argue that these tasks are what refines your instinct, but the quality of your art may just be defined by the quality of your decisions instead. to quote: "how much work you put into something is just a proxy for how much you care and how much you have to say."

i think i agree to a certain extent. i think everyone has to go through the grind of tedious work to appreciate the effort that goes into craft. if all art creation was easy, would people still want to pursue it?

i really want to recreate this project for my home town, Kuala Lumpur. maybe one weekend when i have too much time on my hands.

mood: 3/5

wake: 10:30
sleep: 3:00

meals:

• breakfast: banana pumpkin seed yoghurt
• lunch: hainanese
• dinner: hainanese and kimchi

grateful for: Lee's Kitchen

til:

• violin is smaller higher pitched, soprano voice (tuned GDAE, while viola is larger, lower pitched, alto voice (tuned CGDA)
• grounding = standing on grass/dirt/sand barefoot so electrons flow into your body
• muon optimizer, created by keller jordan, a neural network optimizer designed specifically for hidden layers in NNs, using matrix orthogonalization rather than element-wise adaptive learning rates in Adam / AdamW. orthogonalization is done efficiently using Newton-Schulz iteration rather than expensive SVD
• waterdrop pitcher once it is exhausted, it releases TMA, which is the same compound that gives fish their characteristic fishy odor.

links:

• Cannoneyed - Isometric NYC

  What’s possible now that was impossible before?

  The biggest joy of this project was the ability to build tools at the speed of thought. As a software developer, I think of a million little tools I’d like to have but would take a day or a week to build. With Claude or Cursor, I can whip them up in 5 minutes. This is absolutely transformational - it’s like having an infinite toolbox

  composability is huge and even more valuable in the context of vibe coding. The Unix philosophy of small, modular programs that do one thing well means that we can easily compose smaller tools into utility functions that can be reused by higher-level applications. By designing pieces of functionality in a modular way, you can most effectively leverage the coding agent - it’s easier for you to specify simple behavior, it’s easier for the agent to build, debug, and test these modular pieces, and it’s easier to stitch them together into higher level apps later on. This is standard software engineering best practices, and it’s more relevant now that the cost of code is approaching zero