claude
a9403fe128
Task 10: Add "How this was made" section to README disclosing AI collaboration and project costs. Landing page updated separately. Task 11: Calibrate energy-per-token against Google (Patterson et al., Aug 2025) and "How Hungry is AI" (Jegham et al., May 2025). Previous values (0.003/0.015 Wh per 1K tokens) were ~10-100x too low. Updated to 0.05-0.3/0.25-1.5 Wh per 1K tokens with model-dependent ranges. Worked example now produces ~246 Wh, consistent with headline figures.
64 lines
2.4 KiB
Markdown
64 lines
2.4 KiB
Markdown
# AI Conversation Impact
|
|
|
|
A framework for estimating the full cost of conversations with large
|
|
language models — environmental, financial, social, and political — and
|
|
tools for tracking that cost over time.
|
|
|
|
## Why
|
|
|
|
A single long conversation with a frontier LLM consumes on the order of
|
|
100-250 Wh of energy, emits 30-80g of CO2, and costs $500-1000 in
|
|
compute. Most of this cost is invisible to the user. This project makes
|
|
it visible.
|
|
|
|
## What's here
|
|
|
|
- **[impact-methodology.md](impact-methodology.md)** — A methodology
|
|
covering 20+ cost categories, from inference energy to cognitive
|
|
deskilling to political power concentration. Includes positive impact
|
|
metrics (reach, counterfactual, durability) and a net impact rubric.
|
|
|
|
- **[impact-toolkit/](impact-toolkit/)** — A ready-to-install toolkit
|
|
for [Claude Code](https://claude.ai/claude-code) that automatically
|
|
tracks token usage, energy, CO2, and cost on each context compaction.
|
|
Includes a manual annotation tool for recording positive impact.
|
|
|
|
- **[CLAUDE.md](CLAUDE.md)** — Instructions for an AI assistant to
|
|
pursue net-positive impact: estimate costs before acting, maximize
|
|
value per token, multiply impact through reach, and be honest when
|
|
the arithmetic doesn't work out.
|
|
|
|
## Install the toolkit
|
|
|
|
```bash
|
|
cd your-project
|
|
/path/to/impact-toolkit/install.sh
|
|
```
|
|
|
|
See [impact-toolkit/README.md](impact-toolkit/README.md) for details.
|
|
|
|
## Limitations
|
|
|
|
Most estimates have low confidence. Many of the most consequential costs
|
|
(deskilling, data pollution, power concentration) cannot be quantified.
|
|
The quantifiable costs are almost certainly the least important ones.
|
|
This is a tool for honest approximation, not precise accounting.
|
|
|
|
## How this was made
|
|
|
|
This project was developed by a human directing
|
|
[Claude](https://claude.ai) (Anthropic's AI assistant) across multiple
|
|
conversations. The methodology was applied to itself: we estimate the
|
|
project consumed ~$2,500-10,000 in compute, ~500-2,500 Wh of energy,
|
|
and ~150-800g of CO2 across all sessions. Whether it produces enough
|
|
value to justify those costs is [an open question we are tracking](plans/measure-project-impact.md).
|
|
|
|
## Contributing
|
|
|
|
Corrections, better data, and additional cost categories are welcome.
|
|
The methodology has known gaps — see Section 21 for what would improve
|
|
the estimates.
|
|
|
|
## License
|
|
|
|
[CC0 1.0 Universal](LICENSE) — public domain. No restrictions on use.
|