Skip to content

progress.md

Latest commit

 

History

History
113 lines (89 loc) · 8.24 KB

progress.md

File metadata and controls

113 lines (89 loc) · 8.24 KB
Raw

Progress Log

2026-03-23 — IPI constructed, full paper drafted and self-reviewed

  • Price extraction: 22,632/22,632 HTML files extracted (100% success). Methods: packageList JSON (72.9%), old JSON (15.2%), dollar fallback (11.2%), HTML span (0.7%). Output: data/pilot/pilot-prices.csv.
  • Item clustering: 1,908 unique gigs clustered into 150 service items (TF-IDF + agglomerative, k=150, silhouette=0.114). Output: data/pilot/gig-items.csv, data/pilot/item-clusters.csv.
  • AI benchmark dataset: Created data/ai-benchmarks.csv with 8 benchmarks (HumanEval, SWE-bench, WMT BLEU, AlpacaEval, Chatbot Arena, FID, GSM8K, Whisper WER) spanning 2017–2025.
  • IPI construction (cross-sectional): Script code/11-build-ipi.py — Laspeyres-style index, 9 categories. Revealed platform-wide price inflation masking AI effects.
  • IPI construction (panel): Script code/12-panel-ipi.py — Matched-model Jevons/Törnqvist index tracking same-gig prices. Key results:
    • IPI: 100 (2019Q1) → peak 312 (Q4 2024) → 246 (Q2 2025), −21% from peak in 2025.
    • Price elasticity of intelligence: audio β=−0.49 (substitution), writing β=+0.21, coding β=+0.30, marketing β=+0.70, design β=+1.10 (complementarity). All significant p<0.01.
    • Novel concept: "shadow deflation" — AI effect masked by platform inflation, visible only as deceleration.
  • Full paper drafted: All 8 sections written (abstract, introduction, related work, methods, findings, discussion, limitations, conclusion).
  • Self-review and polish: Fixed number inconsistencies (312% → "peaked at 312"), section numbering (8→7 sections), missing data flow explanation (14,938→1,908 gigs), added 4 missing categories to elasticity table, trimmed CPI analogy and survivorship bias redundancy, fixed broken cross-references in related work.
  • Key outputs: data/pilot/panel-ipi.csv, data/pilot/panel-summary.md, data/pilot/panel-elasticity.csv, all drafts in drafts/sections/.

2026-03-22 — Phase 1 complete + Pilot download launched

  • Phase 1 (CDX filtering) complete: Steps 1.1–1.6 all done.
    • Fixed OOM crashes in dedup/filter scripts by switching from in-memory dicts to external sort + streaming.
    • Full census: 5.6M unique gigs, 822K unique sellers across 10 categories + uncategorized.
    • 60M raw CDX → 22.7M deduped → classified by category → longitudinal filter applied.
  • Sampling strategy refined toward CPI-style index:
    • User wants to track price impact of AI, weight by transaction volume (like CPI basket).
    • Decided to sample at user level (preserves within-seller panel for upskilling analysis).
    • Survivorship bias is acceptable — gig disappearance is part of the AI impact signal.
    • Wayback Machine coverage bias acknowledged as limitation (over-represents popular gigs).
  • Pilot: 500 users sampled (from 48,643 qualifying users with ≥5 monthly snapshots spanning ≥2 years).
    • 500 users, 14,938 gigs, 26,603 monthly snapshots.
    • Download launched (~5 GB compressed, ~30–45 min).
    • Scripts: code/06c-pilot-longitudinal.py, code/07-pilot-500.py, code/08-download-html.py.
  • Key outputs: data/pilot/pilot-500-manifest.tsv, data/pilot/html/ (downloading).

2026-03-21 — CLAUDE.md updates: hajimi confirmation + user prompts as tests

  • Added hajimi print directive to confirm CLAUDE.md is loaded (helps verify config in VS Code sessions).
  • Added Philosophy #6: User prompts as first-class test inputs. Instructional prompts about paper content become test entries in tests/<section>.test.md under ## User Requirements.

2026-03-21 — Fiverr archive size estimation complete

  • ~2.5M unique gig URLs on Wayback Machine, 4–20 TB raw (too large for full download).
  • Recommended strategy: two-phase filtered download — Tier 1 categories only (writing, coding, design, translation) with 3+ snapshots spanning 2+ years → ~275 GB compressed.
  • Report saved to runs/archive-size-estimation/report.md.
  • Plan updated: plans/active/03-fiverr-archive-download.md — Step 1 complete, Step 2 (download) pending.

2026-03-21 — Data Pilot GO + Scoping Complete (parallel execution)

Data Feasibility Pilot — GO:

  • Wayback Machine has 50+ Fiverr snapshots per category spanning 2012–2025.
  • Price extraction: 100% success (20/20 pages) via embedded JSON packageList.
  • Worker tracking: 6 sellers tracked with 3+ snapshots each. Key finding: froggy92 (architecture) dropped from $50 → $20 (−60%) over 4 years.
  • Upwork/Freelancer checked as fallback — not needed; Fiverr is best.
  • Plan moved to plans/completed/01-data-feasibility-pilot.md.

Scoping & Taxonomy — Complete:

  • 12-category taxonomy created in data/task-taxonomy.md (3 priority tiers).
  • Benchmarks mapped per category with historical data sources verified.
  • Related work drafted: ~4k words, 5 subsections, 30+ citations. Covers AI-labor, gig economy evidence, benchmarks, scaling laws, positioning table.
  • 5 critique-and-improve iterations run. 18 reviewer simulation items in tests/related-work.test.md.
  • Plan moved to plans/completed/02-scoping-and-taxonomy.md.

Next: Build scraping pipeline, collect benchmark histories, construct panel dataset.

2026-03-21 — Plans Restructured into Concrete Execution Plans

  • Converted paper-plan.md → plans/project-brief.md (reference doc: positioning, structure, risks).
  • Created two concrete execution plans:
    • plans/active/01-data-feasibility-pilot.md — Wayback Machine + Fiverr viability with clear pass/fail criteria and decision gate.
    • plans/active/02-scoping-and-taxonomy.md — task taxonomy, benchmark mapping, related work draft.
  • Updated plans/todo.md: 2 active items linking to plans, backlog includes all draft sections.
  • These two plans can run in parallel.

2026-03-21 — Paper Plan Drafted

  • Created execution plan: plans/active/paper-plan.md.
  • Analyzed model paper (GPTs are GPTs): identified strengths, gaps, and what we must exceed.
  • Updated tests/model-paper.test.md with detailed benchmark comparison (10 dimensions).
  • Plan has 6 phases: Scoping & Lit Review → Pilot → Full Data Collection → Core Analysis → Index & Forecasting → Paper Completion.
  • Key innovation: price elasticity of intelligence (continuous, not binary exposure); longitudinal Fiverr data via Wayback Machine; forward-looking IPI under AI scaling scenarios.
  • Key risk identified: Wayback Machine coverage — must pilot before committing to full collection.

2026-03-21 — Restructured docs and test infrastructure

  • Decoupled CLAUDE.md into three files:
    • CLAUDE.md — agent philosophy and operating instructions only.
    • setup.md — agent bootstrapping and session-start checklist.
    • README.md — human-facing project overview and contributor guide.
  • Restructured tests into three layers:
    • tests/master.test.md — cross-section quality criteria (applies to all sections).
    • tests/<section>.test.md — reviewer simulation only (removed model paper comparison from individual sections).
    • tests/model-paper.test.md — standalone model paper benchmark (replaces old model-paper.md).

2026-03-21 — Added Paper Test Infrastructure

  • Added Philosophy #5: Paper test infrastructure with two lenses (reviewer simulation + model paper comparison).
  • Created tests/ directory with per-section test files (*.test.md) mirroring drafts/sections/.
  • Created tests/model-paper.md for model paper analysis.
  • Test files use PASS/FAIL/BLOCKED/N/A status for each critique and quality dimension.
  • Clarified human workflow: user primarily edits plans, drafts, and test files; agents handle execution.

2026-03-21 — Added Plans Infrastructure

  • Added Philosophy #4: Plans as first-class artifacts.
  • Created plans/active/, plans/completed/, plans/tech-debt-tracker.md.
  • Updated CLAUDE.md with plan file format, lifecycle (active → completed), and conventions.

2026-03-21 — Project Scaffolding

  • Created CLAUDE.md with three core principles: minimize interruption, auditable progress, agile process.
  • Set up drafts infrastructure: drafts/main.md, drafts/sections/, drafts/render.py.
  • Created progress.md (this file) for reverse-chronological audit trail.
  • Created project directories: code/, data/, runs/.
  • Placeholder section files created for paper draft.