RL-Trained Coding Model Generates JavaScript Paintings
Reactions celebrate a reinforcement learning demo that produces paintings through generated JavaScript code.
TLDR
Surya posted a demonstration of a coding model fine-tuned with reinforcement learning to output JavaScript that renders paintings. Training occurred on Prime Intellect infrastructure and rollouts were rendered on Modal. Will Brown described the results as nuts. Eiso Kant praised how far code can go. Suhail likened the output to Midjourney. Other replies highlighted compositionality advantages over diffusion models and the use of LLM judges for teaching subjective taste.
Combined views
370.4K
17 Sources, first seen ago
Sources
SU 
Surya@kickingkeysYou can also make beautifully imperfect videos with AI coding models https://twitter.com/kickingkeys/status/2091570990048276897
- likes: 1.7K
- replies: 41
- bookmarks: 703
- reposts: 65
WB will brown@willcbtrained on prime intellect lab :) this is nuts my mind is kinda blown https://twitter.com/kickingkeys/status/2091570990048276897
- likes: 2.6K
- replies: 29
- bookmarks: 1K
- reposts: 63
EL elie@eliebakouchthis is beautiful, you can just RL aesthetic taste through javascript code https://twitter.com/kickingkeys/status/2091570990048276897
- likes: 716
- replies: 8
- bookmarks: 339
- reposts: 19
KA karthik@akbirthkostellar work. really hope this actually becomes the norm, because diffusion or flow models still seem slop upon very close inspection, here compositionality can go wild https://twitter.com/kickingkeys/status/2091570990048276897
- likes: 233
- replies: 4
- bookmarks: 54
- reposts: 4
EK Eiso Kant@eisokantI really love this work! Shows how far code can go https://twitter.com/kickingkeys/status/2091570990048276897
- likes: 47
- replies: 0
- bookmarks: 8
- reposts: 0
AR ar0cket1@ar0cket1a good example of using LLM/Agentic judges in teaching subjective or other hard to define things such as taste (which is pretty important to learn). self play (https://arxiv.org/abs/2607.23802) and other mechanisms are also quite interesting for approaching this problem.…
- likes: 163
- replies: 1
- bookmarks: 128
- reposts: 7