Touch Grass
Paris · 2026-10-09 · forecast on this laptop, no account, no cloud. Model: TabPFN-v2 regressor, Prior Labs License v1.1 (Apache-2.0 + attribution).
Best hours to be outside
- 10:00–12:00 · score 86
- 12:00–14:00 · score 84
- 14:00–16:00 · score 83
Daylight hours, predicted
| Hour | Temp | Rain | Wind | Cloud | Score |
|---|
| 09:00 | 15° | 0.0 | 6 | 74% | 76 |
| 10:00 | 17° | 0.0 | 7 | 62% | 85 |
| 11:00 | 18° | 0.0 | 7 | 75% | 88 |
| 12:00 | 18° | 0.0 | 8 | 92% | 85 |
| 13:00 | 19° | 0.0 | 8 | 79% | 83 |
| 14:00 | 19° | 0.0 | 8 | 79% | 83 |
| 15:00 | 19° | 0.0 | 10 | 78% | 82 |
| 16:00 | 20° | 0.0 | 10 | 76% | 81 |
| 17:00 | 21° | 0.0 | 8 | 76% | 78 |
| 18:00 | 20° | 0.0 | 8 | 76% | 79 |
| 19:00 | 20° | 0.0 | 6 | 63% | 83 |
Score = 100 − 4·|temp−18| − 25·rain − 1.2·max(0, wind−15) − 0.15·cloud. The model predicts the weather; this rule is applied afterwards and is yours to change.
Was it worth running a model?
Same pipeline, re-run on the last 3 fully observed days, each predicted only from the days before it. Mean absolute error against what actually happened, next to the dumbest baseline: "same hour 2 days earlier".
| Variable | TabPFN | Same hour 2 days earlier | Model better? |
|---|
| Temperature (°C) | 2.83 | 1.67 | no |
| Rain (mm/h) | 0.10 | 0.11 | yes |
| Wind (km/h) | 2.38 | 3.43 | yes |
| Cloud (%) | 51.49 | 63.79 | yes |
Inputs per hour: hour of day, weekday, and each variable 2, 3, 4 and 7 days earlier. Training window: 41 days, 4 ensemble members, CPU. Data: Open-Meteo. Built for the DEV Hacktoberfest Open-Source AI Challenge, Week 1 "Touch Grass". — Yvoo Lab