Digital Humanities, University of Bern Digital Humanities, University of Bern Introduction to Digital Humanities
  • Course
  • Syllabus
  • Assignment
  • GitHub
  • Resources
  • Student Projects
  • About
  • DH Lab ↗
  1. Syllabus
  2. Session 12
  • Course
  • Syllabus
    • Session 1
    • Session 2
    • Session 3
    • Session 4
    • Session 5
    • Session 6
    • Session 7
    • Session 8
    • Session 9
    • Session 10
    • Session 11
    • Session 12
    • Session 13
    • Session 14
  • Assignment
  • Working with Git and GitHub
  • Resources
  • Student Projects
    • How to add your project here
  • About

  • DH Lab (companion course)

On this page

  • In this session
  • How LLMs work
  • Research agents
  • Ahead of next week: stochastic parrots
  • Reading for Session 13
  • Edit this page
  • Report an issue
  1. Syllabus
  2. Session 12

Session 12

LLMs: stochastic parrots and research agents

Author
Affiliation

Tobias Hodel

Walter Benjamin Kolleg / Digital Humanities, University of Bern

Published

1 December 2026

Modified

2 September 2026

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.

TipIn parallel in the Lab

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

Back to top

References

Bender, Emily M., Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. 2021. ‘On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?’ Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (New York, NY, USA), March, 610–23. https://doi.org/10.1145/3442188.3445922.
Session 11
Session 13
  • Edit this page
  • Report an issue