| 1 | --- |
| 2 | layout: base.njk |
| 3 | title: Prompt & carbon ledger |
| 4 | permalink: /ai.html |
| 5 | --- |
| 6 | <h1>Prompt & carbon ledger</h1> |
| 7 | |
| 8 | <p>The initial version of this site was set up by |
| 9 | <a href="https://claude.com/claude-code">Claude Code</a> (model: Claude |
| 10 | Fable 5). The plan for ongoing maintenance is to use open-source models |
| 11 | running on private hardware. Nothing here was written by hand except the |
| 12 | prompts below. Each row records an instruction given to an agent, which model |
| 13 | did the work, what came of it, and a rough estimate of the energy and carbon |
| 14 | the AI work cost. Newest first.</p> |
| 15 | |
| 16 | <table border="1" cellpadding="6"> |
| 17 | <tr> |
| 18 | <th>Date</th> |
| 19 | <th>Model</th> |
| 20 | <th>Prompt (verbatim)</th> |
| 21 | <th>What happened</th> |
| 22 | <th>Est. energy</th> |
| 23 | <th>Est. CO₂e</th> |
| 24 | <th>Roughly equivalent to</th> |
| 25 | </tr> |
| 26 | {%- for e in collections.ledger | reverse %} |
| 27 | <tr> |
| 28 | <td>{{ e.date | longDate }}</td> |
| 29 | <td>{{ e.data.model }}</td> |
| 30 | <td>{{ e.data.prompt | mdInline | safe }}</td> |
| 31 | <td>{{ e.templateContent | safe }}</td> |
| 32 | <td>~{{ e.data.wh }} Wh</td> |
| 33 | <td>~{{ e.data.co2_g }} g</td> |
| 34 | <td>{{ e.data.comparison }}</td> |
| 35 | </tr> |
| 36 | {%- endfor %} |
| 37 | </table> |
| 38 | |
| 39 | <p><b>Running total:</b> ~{{ collections.ledger | total("wh") }} Wh, |
| 40 | ~{{ collections.ledger | total("co2_g") | round }} g CO₂e — |
| 41 | {{ ledger.totals_comparison }}.</p> |
| 42 | |
| 43 | <h2>Method</h2> |
| 44 | |
| 45 | <p>Anthropic, in their cowardice, doesn't publish per-request energy figures, so these are |
| 46 | order-of-magnitude estimates, not measurements. The arithmetic: |
| 47 | <i>(model invocations in the session) × 1 Wh per invocation × |
| 48 | 400 g CO₂e per kWh</i> (a rough global-average grid intensity). Published |
| 49 | measurements for single chat queries are around 0.24–0.33 Wh |
| 50 | (Google's 2025 Gemini disclosure); 1 Wh per invocation is deliberately |
| 51 | generous because agentic coding calls carry much longer contexts than chat. |
| 52 | For open-source models running on local hardware, the estimate is |
| 53 | ~2 Wh per prompt instead, based on the power draw of the GPU doing the |
| 54 | work. Training-run amortisation, network transfer, and this page being served |
| 55 | are not counted. Treat every number as right to within a factor of ten, at |
| 56 | best.</p> |
| 57 | |
| 58 | <p>The ledger is append-only: each time an agent works on this site, it adds a |
| 59 | Markdown file to <code>src/ledger/</code> and rebuilds; this page and the |
| 60 | running total are generated from those files.</p> |
| 61 | |
| 62 | <hr> |
| 63 | |
| 64 | <p><a href="/">Home</a></p> |
| 65 | |
amadan