AI usage
Each time a model is shown a picture and asked what is in it, that is one look. Looks sent to outside models cost money; looks answered by the local model on the pond's own PC are free.
- $9.34 spent so far
- 38,199 looks so far
- 2,041 sightings named
- $0.005 per named sighting
- 25% of looks answered by the local model
Spend per day
- Claude Haiku 5.5
- OpenAI first look
- Claude Sonnet 5.5
Show as a table
| Day | Claude Haiku 5.5 | OpenAI first look | Claude Sonnet 5.5 | Total |
|---|---|---|---|---|
| Nov 14 | $0.00 | $0.00 | $0.00 | $0.00 |
| Nov 15 | $0.029 | $0.000 | $0.00 | $0.029 |
| Nov 16 | $0.44 | $0.41 | $1.45 | $2.30 |
| Nov 17 | $0.39 | $0.49 | $1.12 | $2.00 |
| Nov 18 | $0.13 | $0.30 | $0.35 | $0.78 |
| Nov 19 | $0.43 | $0.33 | $1.25 | $2.00 |
| Nov 20 | $0.26 | $0.23 | $1.03 | $1.52 |
| Nov 21 | $0.13 | $0.079 | $0.49 | $0.70 |
Spending stops for the day at a ceiling of $2.50, with a separate ceiling of $2.50 for the first look. The highest day so far was $2.30.
Looks per day
- Claude Haiku 5.5
- OpenAI first look
- Claude Sonnet 5.5
- Local model
Show as a table
| Day | Claude Haiku 5.5 | OpenAI first look | Claude Sonnet 5.5 | Local model | Total |
|---|---|---|---|---|---|
| Nov 14 | 0 | 0 | 0 | 0 | 0 |
| Nov 15 | 187 | 2 | 0 | 0 | 189 |
| Nov 16 | 2,843 | 5,436 | 589 | 0 | 8,868 |
| Nov 17 | 2,657 | 6,306 | 442 | 0 | 9,405 |
| Nov 18 | 811 | 4,029 | 145 | 2,474 | 7,459 |
| Nov 19 | 2,589 | 4,477 | 505 | 3,649 | 11,220 |
| Nov 20 | 1,642 | 3,102 | 410 | 2,256 | 7,410 |
| Nov 21 | 752 | 1,076 | 199 | 1,150 | 3,177 |
Counts are complete from Nov 14, 2026. Earlier days lost some looks to clean-up.
Who names the fish
A sighting is named when a model says a fish is there. Only some of those names say which type of fish it is.
| Model | Sightings named | With a fish type |
|---|---|---|
| Claude Haiku 5.5 | 847 | 700 |
| Claude Sonnet 5.5 | 106 | 51 |
| OpenAI first look | 289 | 0 |
| Local model | 799 | 0 |
The local model
A small model that runs on the pond's own PC and scores how likely a moving thing is to be a fish. It cannot say which type. It has been in use since Nov 16, 2026. It calls something a fish when its score reaches 0.897.
- 2,153 marked “fish” by the owner
- 617 marked “faint”
- 781 marked “not a fish”
Nightly retraining
Every night a new model is trained on the owner's marks and tested against the one in use. It only takes over if it is clearly better.
| Night | Pictures learned from | Marked fish recognised | Empty clips called “fish” | Minutes | Result |
|---|---|---|---|---|---|
| Nov 21 | 16,097 1,896 of fish | 183 → 191 of 219 | 2 → 2 of 153 clips | 27 | Kept the old model not better: it passes 8 more marked fish, but in only 89% of reshuffles of the clips (it has to be 95%) |
| Nov 20 | 16,094 1,896 of fish | 183 → 194 of 219 | 2 → 3 of 153 clips | 39 | Kept the old model not better: it passes 11 more marked fish, but in only 94% of reshuffles of the clips (it has to be 95%); 1 more clip with a wrong "fish" where 0 more would be let pass |
| Nov 19 | 11,045 1,287 of fish | 118 → 140 of 145 | 1 → 4 of 99 clips | 33 | Kept the old model not better: 3 more clips with a wrong "fish" where 0 more would be let pass |
Where two numbers are shown, the first is the model in use and the second is the newly trained one.