Skip to content

SPEC.md

Latest commit

 

History

History
45 lines (37 loc) · 2.52 KB

SPEC.md

File metadata and controls

45 lines (37 loc) · 2.52 KB
Raw

ThermoTrainLab — Daily Build #18

Date: 2026-06-01 Builder: Claude Code (autonomous)

Idea Source

  • Origin: Intersection of arxiv papers (SGD as Free Energy Minimization, Stochastic Thermodynamics of Associative Memory) and Bay Area hardware startups (Extropic, Normal Computing)
  • URL: https://arxiv.org/abs/2505.23489, https://arxiv.org/abs/2601.01253
  • Why this idea: Thermodynamic computing is the hottest niche in Bay Area AI hardware (Extropic claims 10,000× efficiency, Normal Computing raised $50M from Samsung in March 2026). No interactive educational tool exists to explore the deep physics connecting SGD, free energy, entropy production, and Landauer's principle.
  • Selection Score: novelty: 5/5, feasibility: 4/5, wow: 5/5, learning: 5/5, utility: 4/5 = 49/55

What It Does

An interactive simulator with 5 live simulations exploring the stochastic thermodynamics of neural network training. Users manipulate temperature (learning rate), noise (batch size), and network architecture to see how SGD minimizes free energy, produces entropy, hits Landauer bounds, and compares to thermodynamic hardware.

Target User

ML researchers, physics-curious engineers, students studying statistical mechanics or deep learning theory, and anyone following the Bay Area thermodynamic computing revolution.

Core Features (MVP Scope)

  1. SGD as Free Energy Minimization — particle swarm on loss landscape with F = U - TS decomposition
  2. Entropy Production tracker — Jarzynski equality visualization, irreversibility of training
  3. Landauer's Principle calculator — minimum energy per weight update vs actual GPU waste
  4. Boltzmann Machine / Ising Model — native thermodynamic sampling with phase transitions
  5. Thermo vs Digital energy comparison — log-scale bar chart across model sizes

Tech Stack

  • Frontend: Single HTML + vanilla JS + Canvas API
  • Backend: None (static)
  • Database: None
  • External APIs: None
  • Deployment: Vercel static hosting

Success Criteria

  • App runs locally with one command
  • All 5 simulations work interactively
  • Scientifically accurate physics (correct equations, proper units)
  • Beautiful dark-theme UI with smooth animations
  • README with description and deployment info
  • BUILD_REPORT.md with learnings
  • Source ideas attributed with links

Out of Scope (save for v2)

  • WebGPU acceleration for larger Ising models
  • Real-time PyTorch training hook integration
  • Actual hardware benchmarks from TSU prototypes
  • 3D loss landscape visualization