About
BASIS, the Bangalore AI Safety & Alignment Study, is a small group for thoughtful, technically grounded discussions on AI safety and alignment, spanning papers, books, and open questions. We follow an active-participation discussion model that encourages deep engagement and free-flowing discussion.
We're a high-participation reading group: each attendee is assigned a role before the session, with a specific kind of preparation expected. The roles sum to a fuller reading than any one person could do alone. We meet in person in Bangalore, currently about 12–15 members with 6–7 attending each session.
Roles
- Discussion Generator
- The engine of the conversation. Comes prepared with 2–3 open-ended, high-level questions that force people to take a stand or think beyond the text. Avoids yes/no questions.
- Summarizer
- Provides the "minimum viable knowledge" needed for discussion to kick off. Distills the entire paper into a 3–5 minute narrative, the why, the how, the so what. Sets the stage so even fuzzy readers can participate.
- Highlighter
- Picks 1–2 passages that deserve to be read aloud, the parts so profound, well-written, or controversial that they merit anchoring the discussion in the actual text.
- Contrarian
- Finds the cracks in the armor. Thinks like the skeptical reviewer looking for reasons to reject. Surfaces methodological flaws, over-extrapolated conclusions, hidden assumptions. Asks: "Is there a simpler explanation?" or "What would it take to prove this wrong?"
- Concept Enricher
- Picks 1–2 dense concepts (e.g. mech-interp, KL-divergence, a specific cited methodology) and explains them clearly. Ensures no one leaves the room confused by jargon.
- Practical Applicationist
- The grounded realist. Focuses on the "what now?" Brainstorms how findings could apply to engineering work, policy, or personal habits. For theoretical papers, proposes a toy model or follow-up experiment the group could actually run.
- Bridge Builder
- Looks outward. Connects this paper to previous readings or current interests in the AI safety space. Answers: "How does this change our worldview?"
- Forecaster
- Takes the paper's results and projects them onto future, more capable models or agent ecosystems. Presents a high-impact scaling hypothesis based on the data. Helps the group move from specific results to the general trajectory of the field.
What we read
Papers across the technical AI safety landscape, mechanistic interpretability, AI control, evals, alignment, emergent misalignment. We browse the papers index to see the canon as it's grown.
Tone
We're trying to read carefully and disagree well. The high-participation format pushes people to defend specific positions rather than soft consensus. We are not interested in doom-takes, hype-takes, or status games. We are interested in being wrong out loud and updating.
Want to join?
See how to join.