star-bits

• Initial singularity → Light elements → ⭐ Stars → Heavier elements → 🌏 Earth → 👫 Humans → 👣 Giants → ⚡🧠 Intelligence

slides and code from presentations i’ve given:

• nn.md: on neural networks
  • with focus on how to calculate the partial derivative, ∂L/∂w, at each layer
• nlp.md: on natural language processing
  • with topics ranging from word2vec to RLHF, multimodal, and CoT
• it.ipynb: on how to instruct-tune gemma-2-2b-it with lora

blog posts:

• thought-token: (token-level) latent space에서 thought token으로 CoT 하기 (Coconut)
• concept-model: sentence-level latent space에서 diffusion으로 reasoning 하기 (LCM)
• entropy-patch: byte끼리 entropy를 기준으로 묶어서 patch를 만들고 token 대신 사용하기 (BLT)
• socratic-learning: language space에서의 pure recursive self-improvement
• thought-templates: 해당 reasoning task에 적합한 thought template으로 RAG 하기 (BoT)
• testtime-compute: 같은 FLOPs에서 test-time compute와 pretraining cost의 tradeoff 관계
• selftaught-reasoner: reasoning trace 데이터셋의 self-feeding amplification loop (STaR)
• rstar-math: rounds of MCTS-assisted code-augmented reasoning traces (rStar-Math)
• latent-diffusion: latent space에서 noise predictor를 steering 하기
• decoupled-cross: decoupled cross-attention으로 텍스트와 이미지 프롬프트를 넣어주기

side projects:

• ui-time: add a 15-minute countdown timer in your macOS menu bar
• prettyURL: click the extension icon to copy human readable URL
• giftMacro: when you need to smash that button in your phone at superhuman speed