• No warm-up today!

• Final Presentation: Brigette, Lindsey & Olivia
• Topic Models
• LLMs and the future of social science

• Office Hours: Today, Friday 1:30pm-3:00pm (Tyler)
• Tuesdays, 10:40am-12:00pm (Yao)

• PSet 8: deadline extended to Thursday 11:59pm
• Rate limit errors - wait, run, save
• PSet 9 is optional!
• If no submission: full credit
• If submission: full credit + replace an earlier assignment grade
• A reminder to use GitHub commits

• A reminder to engage / ask your peers questions! (part of rubric now)

• Term frequencies
• Document structure
• Semantic similarity

• The second family, document structure analysis, assumes one can extract from word co-occurrence statistics what any given document is “about” (i.e., what the appropriate keywords or themes are) and represents text as observations that vary on this feature.

• Each document has “themes” or “topics”
• Words combined with each other can comprise topics (e.g. “march” with “january” vs. “soldier”)

• Documents are comprised of topics
• Topics are comprised of words
• Researchers must length of document, proportion of document expected to come from each topic
• Algorithm gives us words related to each topic

1. Each topic \(\theta\) has a distribution
2. Each word \(\phi\) has a distribution within each topic
3. Together, words and topics make up the document

• Examples:
• What topics have newspapers discussed?
• Academic journals?
• College application essays?

• Topics in college essays are related to household income (Alvero et al, 2021)

• Large Language Models (LLMs) are neural networks with billions of parameters trained on large amounts of data
• Model: a statistical tool that uses data to make predictions
• Language: text data, such as words and topics
• Large: Lots of data!
• Increasingly, LLMs refer to generative models

• Large Language Models (LLMs) are neural networks with billions of parameters trained on large amounts of data
• Model: a statistical tool that uses data to make predictions
• Language: text data, such as words and topics
• Large: Lots of data!
• Increasingly, LLMs refer to generative models

• A single LLM can’t write a sociology paper
• But a multi-agent workflow can (?)

• Sociologists Engzell and Wilmers automated the entire process of writing quantitative sociology papers
• Their model: a multi-agent workflow

• In pairs:
• What are some of the benefits of automating social science research? The risks/drawbacks?

• Codifying research processes and “hidden curriculum”
• Open and reproducible science
• Freeing up time for non-trivial tasks

• “Mediocrity at scale”
• Destruction of peer review

• Silicon Sampling refers to “sampling” LLMs rather than humans
• Argyle and colleagues (2023) argue that we can learn much from surveying LLMs (trained on human backstories)
• “The information contained in GPT-3 goes far beyond surface similarity. It is nuanced, multifaceted, and reflects the complex interplay between ideas, attitudes, and sociocultural context that characterize human attitudes.”

• Silicon Sampling refers to “sampling” LLMs rather than humans

• Messeri and Crockett (2026): “The proliferation of AI tools in science risks introducing a phase of scientific enquiry in which we produce more but understand less.”

• Reduced administrative burdens?
• Differences in valued skills?

• Existential risks

“If somebody builds a too-powerful AI, under present conditions, I expect that every single member of the human species and all biological life on Earth dies shortly thereafter.” - Eliezer Yudkowski (2023)

• We’ve covered a lot of methods for computational social science research
• Data viz
• Spatial analysis
• Networks
• Prediction and algorithmic modeling
• Text analysis

• Anything that’s cool?

• In the era of LLMs, division between social scientists and data scientists may be sharper than ever

• What do you think the future of computational social science be?
• What would you like it to be?