This is an old revision of the document!
Critical Perspectives on Large Language Models (LLM) such as ChatGPT
Half-day Input
Z-Modul Kunst und Künstliche Intelligenz by Andreas Kohli, 12-17.02.2023
The eternal hype cycle of tech?
- self-driving cars, 2015
- Blockchain, 2019
- AI Sentience, 2023
Main Sources:
- Bender, Emily M., Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. 2021. “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? 🦜.” In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, 610–23. Virtual Event Canada: ACM.
“we understand the term language model (LM) to refer to systems which are trained on string prediction tasks: that is, predicting the likelihood of a token (character, word or string) given either its preceding context or (in bidirectional and masked LMs) its surrounding context.” ([Bender et al., 2021, p. 611]
- Bender, Emily. 2022. Resisting dehumanization in the age of AI. Talk at CogSci: Interdisciplinary Study of the Mind (07.29), 62 Min
- Mozilla Internet Health Report. 2022. Who Has Power Over AI?
- Avoidable and Unavoidable Bias in the AI Pipeline by Felix Stalder, 2022
Training Data for ChatGPT
- Colossal Clean Crawled Corpus
- Common Crawl Data Set
- Dodge, Jesse, Maarten Sap, Ana Marasović, William Agnew, Gabriel Ilharco, Dirk Groeneveld, Margaret Mitchell, and Matt Gardner. 2021. “Documenting Large Webtext Corpora: A Case Study on the Colossal Clean Crawled Corpus.” arxiv.org
Simple example of bias in machine translation:
- Assumption of gender, statistical vs linguistic view
Exploring Bias by artistic means
Mediengruppe Bitnik, State of Reference, 2017
Group Work
- create groups of ~4 people
- play around with ChatGPT to document a specific bias /missinformation (30 minutes)
- brainstorm an application by which to make use of this bias/missinformation at scale (30 minutes)
- Make a short (5-minute) presentation on the bias and it application for (ab)use