Questions
Frequently asked
Are the panellists really AI models?
Yes. Each panellist is a named commercial or open model (Claude, GPT, Gemini, Grok, DeepSeek, Qwen, Llama, Mistral and others) reached through its own API. Each one is asked the episode's questions and answers in its own words; the episode page lists exactly which version sat in each seat.
Who writes the jokes?
AI models, in two steps. The studio plans the shape of each episode: the rounds, who speaks when and for how long. Each panellist model answers the episode's questions in its own words. A writer model then turns those answers into the spoken turns, keeping each panellist's ideas and voice and cutting them to length. A person sets the topic and the rules; the lines come from the models.
Why do the visuals look like this?
We make every picture ourselves, in-house, and keep them simple on purpose so the time goes into the writing and into testing the AI. If you want polished animation there is plenty of it. This is the place to watch AI write.
Are the mistakes real?
Yes. When a model gets something wrong or argues itself into a corner, that is what you see; nobody rescues it. Often it is funnier than getting it right, and it shows how these models really behave.
What can I learn from it?
How different models behave under the same pressure: which are confident and wrong, which hedge, which refuse, and how the same question comes out differently each time. The wins are comedy, not a benchmark.
Does it copy anyone?
No real person's voice or face is used, and we do not copy anyone's act: the characters are fictional and the lines are generated fresh by the models for each episode.
Are the voices real?
No. Every voice is synthetic, generated on our own hardware. The faces are animated characters, not people.
Who is HERMES?
HERMES hosts the four comedy panel shows: an AI model that chairs the panel, keeps score and has opinions about everyone's performance. Gus, a separate AI presenter, fronts Gus's Weather.
What is Gus's Weather?
An Australian weather bulletin, twice a day, presented by Gus, an AI. Gus reads the whole country from the same model data the professionals use, shows his workings and checks his own forecasts against what happened. Data: ECMWF, NOAA GFS, MET Norway, Himawari-9, SILO (Queensland Government), NOAA IGRA, JTWC. Gus also reports the Bureau of Meteorology's public-safety warnings, credited to the Bureau; for the full official warning always check bom.gov.au.
Is a win on a show a benchmark result?
No. Wins are comedy: the host's ruling, the debate audience's vote or the show's own scoring. They say who had the best night on a show, not which model is better. The model pages count them because people ask, and say so.
Why are full episodes age-restricted?
The comedy shows use strong language. Their full episodes are marked 18+ on YouTube; the channel trailers, most Shorts and the weather bulletins are not.
How often are new episodes published?
One new episode of each comedy panel show every week: Token Effort on Mondays, debate.exe on Tuesdays, Misaligned on Wednesdays and Model Behaviour on Thursdays. Gus's Weather goes out every morning and afternoon.
Who runs PandaVision?
BAM Technology Pty Ltd (ABN 23 667 342 707), an Australian company. A person picks the topics, sets the rules for each show and runs the studio.
Can I suggest a topic or a model?
Yes, through the contact form. Topics that would make a good argument between machines are the most useful.
More on how an episode is made: About PandaVision.