SF Hill Climbers: an interactive site turning San Francisco into a loss landscape
CatAstro_Piyush · x · 2026-11-13
SF Hill Climbers treats San Francisco terrain as a loss landscape (loss = -elevation) and lets optimizers like SGD, Momentum, Adam, RMSprop, Muon, Nesterov and Lion climb from any starting point, visualizing each one's basins of attraction. Fun findings: at default learning rates SGD's map almost exactly matches the terrain's own basins, making it a good baseline; Adam rescales east-west and north-south slopes independently, so it's sensitive to map rotation — rotating 30° sends it to Alamo Square instead of Lafayette Park from Hayes Valley, a quirk the author notes matters in real LLM training. Interactive: click to race optimizers, tweak learning-rate sliders.
More from Research
- SteerAct lifts robot action-model success ~10% with just 24 rollouts per task — pliang279 · 2026-11-13
- PhD thesis explores intra-turn interaction: AI interfaces beyond turn-based dialogue — OfirPress · 2026-11-13
- CMU trains memory writing with RL: Qwen3.5-4B rivals GPT-5.4 on continual learning — aviral_kumar2 · 2026-11-13
- NTU paper: decision info lives in middle layers, last six layers removable loss-free — Minh-Anh Nguyen · 2026-11-13
- Inkling 2 Small's width-4 sconv forces a tiny rolling 3-token conv cache per layer at inference — andrew_n_carr · 2026-11-13
- Open-source legal conflict detector runs on encoder models only, 133-clause contracts in under 20s — ladyofcode · 2026-11-13