mirror of
https://github.com/harvard-edge/cs249r_book.git
synced 2026-05-07 10:08:50 -05:00
64c940c363
Companion commit to c5f90022b (the YAML migration). Pulls in the rest of
the periodic-table work that was sitting in the working tree.
- periodic-table/paper/: LaTeX paper draft (paper.tex, references.bib),
Makefile with hero-figure and SVG->PDF rules (uses rsvg-convert), the
Puppeteer capture_table.js script used to screenshot the table for the
paper, generate_periodic_svg.py which builds the hero SVG from
table.yml (the same source of truth used by the React app), and the
figure sources (SVGs) + derived outputs (PDFs/PNGs) + compiled paper.pdf.
- root .gitignore gains two entries following the existing convention
(cf. the !interviews/paper/fig-*.pdf line just above) so the
periodic-table paper PDF + figure PDFs are not swept up by the blanket
*.pdf LaTeX-artifact rule.
- periodic-table/paper/.gitignore excludes the LaTeX build artifacts
(aux, bbl, blg, log, out, fdb_latexmk, synctex, toc) that make paper
regenerates.
- periodic-table/{iteration,refinement,debate}-log.md: research
provenance from the 100-round LLM iteration loop and the 5-expert
debate simulations that produced v0.2 of the table.
- periodic-table/scripts/archive/: historical iteration scripts
(iterate.sh, debate.sh, debate-continue.sh, run_100_rounds.sh, plus
the Python helpers append_log.py, get_elements.py, patch_informal.py,
patch_website.py, run_claude_loop.py, run_iterations{,_13,_16_20}.py,
update_log.py) moved out of the repo root into an archive subdirectory
with a README documenting their provenance and caveats. These scripts
are preserved for reproducibility and are not part of the active build
pipeline -- the source of truth is now periodic-table/table.yml.
- root package.json + package-lock.json pin puppeteer ^24 for
capture_table.js.
31 lines
2.7 KiB
Python
31 lines
2.7 KiB
Python
import re
|
|
|
|
log_path = 'periodic-table/iteration-log.md'
|
|
with open(log_path, 'r') as f:
|
|
content = f.read()
|
|
|
|
new_log = """
|
|
---
|
|
|
|
## Loop Iterations 21-50 — The "30-Round Expert Panel Debate"
|
|
**Date:** 2026-04-06
|
|
|
|
To stress-test the framework beyond standard deep learning, we instantiated a 30-round simulated debate using the Gemini 3.1 Pro Preview model, casting personas of David Patterson (Hardware), Chris Lattner (Compilers), Jeff Dean (Distributed Systems), Claude Shannon (Information Theory), and Dmitri Mendeleev (Taxonomy).
|
|
|
|
Over 30 violent iterations of critique, the experts identified critical "leaky abstractions" where the periodic table ignored the harsh realities of physics, memory management, and information theory.
|
|
|
|
### Key Breakthroughs & Element Injections:
|
|
1. **The Thermodynamic Floor (Rounds 1-7, 30):** The table assumed infinite energy. We formally introduced **Landauer Limit / Thermodynamics (Td)** at the Hardware/Measure layer and **Bandwidth (Bw)** as absolute physical constraints capping arithmetic intensity and cluster scale.
|
|
2. **Information Theory Bounds (Rounds 14, 25):** The experts noted that aggressive quantization and sparsity cause "representation collapse." We added **Entropy (En)** at the Data/Measure layer to represent the strict Shannon limit of compressibility.
|
|
3. **The OS/Compiler Memory Crisis (Rounds 2, 8, 22):** Tensors are not just mathematical shapes; they have lifetimes. We introduced **Virtualization (Vr)** (representing PagedAttention/OS-paging) to solve KV-cache fragmentation, and **Materialization (Mz)** to represent the compiler's choice between spilling to HBM vs. kernel fusion.
|
|
4. **Semi-Parametric Memory (Rounds 3, 16, 27):** Parametric weights are a bottleneck for long-tail knowledge. The panel mandated the addition of **Indexing (Ix)** (representing HNSW/Vector DBs) as an architectural representation primitive to shift dense O(N) compute to sparse O(log N) memory traversals.
|
|
5. **The Death of Static Graphs (Rounds 17, 29):** To support Mixture of Experts (MoE), we added **Routing (Ro)** as a dynamic, data-dependent control flow primitive that shatters static cache locality.
|
|
6. **Macroscopic System Decay (Rounds 18, 26):** At 100k GPU scale, Mean Time Between Failures (MTBF) is a constant fire. We added **Resilience (Rs)** (checkpointing, elastic recovery) and **Asynchrony (As)** (stale gradients) to handle inevitable hardware death and stragglers.
|
|
|
|
### Verdict: Convergence Reached
|
|
The panel converged on these final adjustments. The table now fully accounts for physical thermodynamics, Shannon entropy, compiler memory lifetimes, and macroscopic cluster decay. The matrix is stable, and we are ready to formalize this into the final research paper.
|
|
"""
|
|
|
|
with open(log_path, 'a') as f:
|
|
f.write(new_log)
|