Session 12
LLMs: stochastic parrots and research agents
Hodel Taught by Tobias Hodel.
In this session
- How large language models work, without the mathematics
- Research agents: what an LLM with tools can and cannot do
- Preparing the discussion of (Bender et al. 2021) for next week
How LLMs work
- Tokens, context, and next-token prediction
- Pretraining, fine-tuning, instruction tuning, RLHF — four different things
- Why fluency and accuracy are unrelated properties
- Hallucination as a structural feature, not a bug to be patched
- Context windows, retrieval, and why “just give it the documents” is harder than it sounds
Research agents
An agent is a model given tools and a loop. What that actually buys you:
- Search, retrieval, code execution, API calls
- Where agents genuinely help in a research workflow: boilerplate, format conversion, first passes over large material, code you can check
- Where they fail: anything requiring that a claim be true, judgement about sources, anything you cannot verify
- Verification is the whole job. If checking the output costs more than doing the work, the agent has not helped.
- Documentation and reproducibility: a prompt is a method, and undocumented methods are not research
Ahead of next week: stochastic parrots
Bender et al. argue that scale itself is the problem — environmental cost, unauditable training data, concentration of power, and the illusion of understanding. Read it against your own use of these tools this semester, including in the assignment.
The DH Lab works hands-on with multimodal LLMs on humanities material.
Reading for Session 13
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. https://doi.org/10.1145/3442188.3445922