Coursera Spring Sale
40% Off Coursera Plus Annual!
Grab it
Explore the hidden political dimensions of tokenization in generative AI through this 38-minute conference talk from 39C3. Discover how generative AI models process language through computational fragments called tokens, which break down human language into subword units stored in large dictionaries that encode political ideologies, corporate interests, and cultural biases before model training even begins. Learn why social media handles like "realdonaldtrump" and brand names like "louisvuitton" exist as single tokens while other words remain fragmented, and understand how this tokenization process determines what can be represented computationally. Examine artistic and adversarial experiments demonstrating 51 different ways to spell "giraffe" using token combinations, from single tokens to complex splits like "gi|ra|ffe" or "g|i|r|af|fe". Investigate how researchers hijacked the prompting process by feeding token combinations directly to text-to-image models, revealing that token beginnings and endings hold particular semantic weight in image generation. Analyze experiments using genetic algorithms to reverse-engineer prompts from images and explore how respelling words in token language changes generative outcomes. Delve into critical examinations of token dictionaries to uncover edge cases where vocabulary breaks down entirely, creating speculative languages at the intersection of English and token nonsense. Understand how token dictionaries encode stochastic worldviews shaped by training data dominated by popular culture, brands, platform-speak, and non-words, making tokenization a fundamentally political act that defines computational representation of the world.
