Talkie Slideshows

Talkie-1930 is a 13-billion-parameter language model trained only on text written before 1931. Give it a subject and it writes a slide deck as a list of commands: find a picture, colour the background, show the picture, add a line of text, next slide. The deck runner carries them out with public-domain and openly licensed pictures from Wikimedia Commons and Openverse, renders each slide, and a local voice reads the deck aloud. From the HG Wells computational creativity project.

A note on Talkie’s views. Talkie writes as someone of its time: it can repeat the attitudes and prejudices found in writing from before 1931, including racist, antisemitic and colonial views, stated as fact. Its decks are shown as it wrote them, for research; they are not the views of this project.

What should the deck be about?

Talkie will choose its own subject first.

Type it however you like. Talkie learned only from text written before 1931 and has never heard of a slide deck: it makes one by copying a single worked example, Make a slide deck about the sea., so your words are put into that form. Leave it empty and Talkie picks its own subject.

Model Fast model: 4-bit, on the GPU: Talkie writes in under a minute; about 2 to 3 minutes per deck with pictures and video. Full model: The original full-precision weights, on the CPU: about 8 to 10 minutes per deck.

Both are the same Talkie. The fast model is a compressed (4-bit) copy running on the graphics card; the full model is the original weights on the processor. Two fast-model decks can be made at once; a full-model deck runs alone.

Narrated video (about half a minute more)

Sampling settings

Seed

Temperature

Longest reply (tokens)

The project's defaults. The same seed, temperature and request give the same deck again on the same model. About 1,600 tokens is roughly 20 slides; the deck is cut there if Talkie never writes end-deck().

Decks Blind review

Windmills: the millers who worked them Talkie · fast model · 5 slides · 27 Oct 2026