Research paper to video: an AI weather model paper as an explainer
Research paper to video with an API: a new paper on an AI weather model became a narrated explainer. Watch it, copy the request, and see the cost and our check.
How this was made. Audio and video overviews, slide decks, infographics and quizzes are made with Google's Gemini Notebook (formerly NotebookLM), on accounts that AutoContent operates. AutoContent has no agreement with Google. Google can limit or stop this access at any time.
- Length
- 8 min 14 s
- Credits used
- 50 credits
- Cost on the Starter plan
- €1.45
- Generation time
- 14 min 56 s
- Format
- Explainer video
- Job
- Research briefs
- Language
- English
- Generated on
- 2026-11-03
This output used 50 credits, which is €1.45 on the Starter plan (€29 per month for 1,000 credits). Failed generations use none. Generation time runs from the request to the finished file; it depends on the sources, the options and the queue.
What came out, and what we would change
We sent the PDF of a new arXiv paper on HRRRCast, an AI model that emulates the operational HRRR weather forecast, and asked for an explainer for a general science audience: what the model does, how it was trained and tested, how well it forecasts storms and heavy rain, and what its limits are. The result is a 1280 × 720 MP4 of 8 min 14 s.
It took three runs. The paper's Figure 2 has six charts of rain skill that look almost the same, one for each rain threshold. The first run showed the 0.25-inch chart as the 2-inch chart and said that the AI beats the operational model on 2-inch rain; the second showed the 0.01- and 0.10-inch charts as the half-inch and one-inch charts. In the paper's own 0.50-, 1.00- and 2.00-inch charts, the operational model is ahead at the shorter lead times, and at 2 inches over all seven months. We rejected both runs. For the third we added one instruction: state the results only as the paper's text states them, and do not read values or rankings off charts with several panels.
The published run gets the paper right: nine years of training data (2016 to 2025) on the native 3-km grid instead of a 6-km grid that caused aliasing, 20 pressure levels, a sigmoid-weighted loss that gives rare intense events up to three times the weight, seven months of checks against observations, reflectivity skill "near parity at 6-km pooling and better at pooling radii ≥30 km" at 40 dBZ, and rain skill competitive through 1 in/h. Two lines go further than the paper: the model "knows exactly where the storm is", where the paper says it places storms better than it shapes them, and the hourly time step is given as the cause of the fragmented storms, where the paper names it as one of three candidate mechanisms.
For papers with multi-panel figures, ask for results in the paper's own words, as we did here, and check every chart that the video shows against its caption. The video took 14 min 56 s to generate and used 50 credits; the two rejected runs used 50 credits each.
Transcript excerpt
Speech-to-text of the output, not edited and without speaker labels. It is part of the sample output.
So they used what's called a sigmoid weighted loss. The crucial point here is that this weighted loss literally forces the AI to pay up to three times more attention to extreme storms, which are super rare, but obviously highly impactful. By applying this multiplier selectively to reflectivity and heavy rain, they made sure the AI cares about extreme, heavy-tailed events, without completely ruining its ability to predict a normal Tuesday.
The AI faces a displacement structure tradeoff. This means it knows exactly where the storm is, but the internal morphology of the storm looks overly fragmented. So while the AI gets the placement right, if you zoom into the storm cells themselves, the AI's predictions look kind of unnatural and disjointed compared to actual radar. The traditional physics models still produce a much more realistic internal structure of the storm cells. And the reason why is actually fascinating. Because the AI is only trained on snapshots taken one hour apart, it has to guess how the storm evolved in between, which causes that fragmented structure.
The exact request
The body that was sent to POST /content/Create, unchanged, and the same request as curl.
{
"resources": [
{
"type": "file",
"content": "https://arxiv.org/pdf/2610.36186v1"
}
],
"outputType": "video",
"format": "explainer",
"text": "Explain this paper to a general science audience: what the AI weather model does, how it was trained and tested, how well it forecasts storms and heavy rain, and what its limits are. State the results only as the paper's text states them, including where the physics-based model is still better. Do not read values or rankings off charts that have several panels. Do not name the authors.",
"language": "English"
}curl -X POST "https://api.autocontentapi.com/content/Create" \
-H "Authorization: Bearer $AUTOCONTENT_API_KEY" \
-H "Content-Type: application/json" \
-d '{"resources":[{"type":"file","content":"https://arxiv.org/pdf/2610.36186v1"}],"outputType":"video","format":"explainer","text":"Explain this paper to a general science audience: what the AI weather model does, how it was trained and tested, how well it forecasts storms and heavy rain, and what its limits are. State the results only as the paper'\''s text states them, including where the physics-based model is still better. Do not read values or rankings off charts that have several panels. Do not name the authors.","language":"English"}'
# then poll until "status" is 100:
curl -H "Authorization: Bearer $AUTOCONTENT_API_KEY" "https://api.autocontentapi.com/content/Status/<request_id>"How we checked it
- Transcript
- We transcribed the output with Whisper large-v3-turbo and compared the transcript with the source.
- Language
- Detected en, requested en.
- Loudness
- -19.7 LUFS integrated, true peak -0.9 dBTP.
- Notes
- The training data, the native grid, the pressure levels, the weighted loss, the months of checks, the quoted reflectivity and rain results and the displacement-structure trade-off match the paper. Overstatements: "knows exactly where the storm is"; the hourly time step given as the cause. The first two runs were rejected: they mislabeled panels of a multi-panel figure.
- Checked on
- 2026-11-04. Failed generations are run again and never published.
More examples
All examples
Explainer video · GermanA research paper turned into a German explainer video7 min 24 s · 50 credits (€1.45 on Starter)- Podcast · GermanKI-Podcast aus einem Fachartikel: die letzten Gletscher Bayerns4 min 5 s · 10 credits (€0.29 on Starter)
Infographic · SpanishInfografía con IA a partir de una nota de prensa: el uso de internet y de la IA en EspañaOne image · 10 credits (€0.29 on Starter)