Gadfly

What it does

Gadfly’s contributions are short, contextual, and designed to provoke discussion instead of provide answers. This last quality is essential to how we think about the tool. In Plato’s Apology, Socrates calls himself the “gadfly” of Athens, “arousing and persuading and reproaching” citizens so that they may be “stirred to life.” He was a good nuisance. In this tradition, our AI agent will ask pointed questions or make provocations that invite us humans to think critically and reflect deeply. 

Gadfly can contribute in the following ways:

• Challenge an implicit assumption made during the discussion.
• Surface an overlooked perspective.
• Ask whether an argument’s logic holds (Particularly looking at syllogisms “If X=Y, does it follow that…?”, etc. ).
• Surface a relevant passage from class texts.
• Surface a classmate’s point from a previous class.

Philosophy

Terrence Sejnowski writes that, although LLMs seem to exhibit intelligence, “it’s not human intelligence.” Indeed, he notes, it is frequently “superhuman”. Gadfly is an opportunity to embrace the alien qualities of an LLM’s intelligence. We focus on: 

• Its lack of social ‘skin in the game’ and subsequent willingness to ask uncomfortable questions
• Its perfect recall and subsequent capacity to process entire conversations at once, catching threads that a human might miss. 

We posit that its usefulness can come from being a distinctly nonhuman interlocutor that has read everything, forgets nothing, and has no stake in the relational fabric of the conversation. 

Here, we’re inspired by Suyash Fulay’s work on “The Empty Chair” and the premise that LLMs can effectively surface perspectives absent from a conversation. We also refer to Valdemar Danry et al.’s “Don’t Just Tell Me, Ask Me,” which finds that AI systems that pose questions rather than provide answers improve human reasoning. As Danry et al. write, AI may serve us better as a “critical thinking stimulator” rather than an “information teller.”

Technology

Gadfly is a React/TypeScript web app. It transcribes live audio (AssemblyAI), processes discussion against ingested readings via the Anthropic API, and provides a UI cue (a ‘raised hand’) when it identifies a useful contribution. A student presses a button or says a keyword to invite Gadfly to speak (ElevenLabs TTS). 

A key challenge has been calibrating Gadfly’s  tone, contribution length, and criteria for raising its hand. Knowing when to speak during a group conversation is challenging even for emotionally intelligent humans. Teaching a machine to do so requires thoughtful testing and iteration in real-world deployment.