Comedy Writing With Small Generative Models
summary
A clear decision: concrete methods and results. Starts from predictive-text reverse engineering (observing 2-word context on a 2015 Android), implements minimal n‑gram/Markov engram models in a Jupyter demo, scrapes Yelp and Amazon for source corpora, builds mashups like a Morrissey+P90X song, and uses the CMU Pronouncing Dictionary to replace lyrics via syllable/phoneme matching. Shows working karaoke and demos, argues small, explainable models (Botnik-style pipelines) remain useful and practical.
AI-written summary · about
chapters (25)
- 0:00Welcome
- 0:26Comedy writing to coding
- 1:08Predictive text and boring text
- 3:13Demo: Predictive text song
- 6:09Predictive text as Markov model
- 8:36Small models and small texts
- 10:00Minimal engram model in Python
- 11:28Demo: Minimal engram model
- 12:51Scraping source texts
- 14:14Yelp review of Statue of Liberty
- 17:04Morrissey lyrics + P90X reviews
- 19:56Tokenization shapes output
- 20:59Harry Potter parody tokenization
- 21:45Coachella poster project
- 23:12Why models feel real
- 25:39Theories of fun
- 26:23Beeswax analogy for models
- 28:10Small models in comedy
- 28:32Visualizing predictive text
- 31:42Metaphors for text generation
- 32:46Automating Beatles' process
- 35:15Demo: Weird Algorithm
- 38:08Lessons from karaoke
- 39:13Acknowledgments
- 40:41Group karaoke
related talks
43mKeynote: Applied NLP in the age of Generative AI
01h 04mA Very Small Language Model
25mGenerative Text Modelling: Scratching the surface
53mGenerative AI: Overview and use cases
01h 24mIntroduction to Language Modeling
58mKeynote: The past, present, and future of AI for application developers
09mRethinking AI: Building What Matters, Not Just What's Possible
59mKeynote: The past, present, and future of AI for application developers