I'm about to count the total tokens in every logged conversation in the 7 months history I've been using the web UI of ChatGPT. This count will underestimate the true number, because I've deleted conversations in the web UI and there are unlogged conversations from before they started logging conversations.
I will count two quantities: The number of Mira-tokens and the number of ChatGPT-tokens. This market resolves to a weighting of the nearest entries to the count of Mira-tokens. See Answer Weighting below. If I haven't calculated the number in a week, resolves NA. The Mira-tokens and ChatGPT-tokens are the total count of tokens in messages tagged to Mira and ChatGPT respectively.
Counts will include all branching conversations. If I regenerated a ChatGPT response, each response sums to the ChatGPT-tokens but not to Mira-tokens. If I edit my prompt and resubmit, it sums to Mira-tokens.
Answer Weighting: This market resolves to a logarithmic weighting of the nearest answers below and above, if they exist. If there is no answer below or above, resolves to the bounding answer instead. k is a multiplier of size 1024. If the answer is 106,783 tokens, then log2(106,783) = 16.7043225 so I would resolve 70% "128k = 2^17" and 30% "64k = 2^16".
Every conversation I can get will be counted. So both GPT-3.5 and GPT-4 will be counted, and also plugins vs. no plugins. Does not include API requests, only ChatGPT. Only includes conversations before this market was created. I won't delete any conversations until this market resolves.
🏅 Top traders
| # | Name | Total profit |
|---|---|---|
| 1 | Ṁ4,057 | |
| 2 | Ṁ138 | |
| 3 | Ṁ54 | |
| 4 | Ṁ8 |
The final counts were:
Mira(total) | 499767
Mira(GPT-4) | 274865
Mira(GPT-3.5) | 163455
Mira(GPT-3.5-turbo) | 61447
gpt-3.5. | 392330
gpt-3.5-turbo | 69322
gpt-4. | 607556
tool | 210163log2(499767) = 18.930896116741497, so I have resolved 93% to 2^19 and 7% 2^18
The cost had I paid through the API should be:
GPT4: $44.70
GPT-3.5-turbo: $0.23
GPT-3.5 Legacy: $10.30 (OpenAI said they found a "90% cost reduction", so even though this model was never available through API I multiplied the later gpt-3.5-turbo cost numbers by 10)I've paid 5 months of ChatGPT at $20/month, so $100 excluding sales taxes. So I'm "overpaying" by about 2x compared to using the API. I probably still won't cancel my subscription though, since the responses on ChatGPT web are faster than the API and the rolling summarization is more convenient than doing it myself.