Abhimanyu Pallavi Sudhir

PhD student @ University of Warwick. I'm interested in stuff at the intersection of markets and AI.

Stuff I've done

• Now Markets, reward, dynamics and bounded rationality
• Reinforcement Learning from Recursive Information Market Feedback
• Market-based architectures in RL and beyond
• A consistency benchmark for LLM forecasters, an arbitrage-based metric for measuring logical inconsistency and ArbitrageForecaster, the "Platt scaling of inconsistency"
  • Of particular interest may be the arbitrage rules in Appendix C that tell you how to combine beliefs from different agents.
• Prediction markets for non-verifiable non-falsifiable sentences
• The Scalable Oversight Benchmark
• costly, a Python package for automatically estimating costs of complex LLM workflows before you run them

Deferred due to lack of time

• Prediction markets for latent space variables or "subjective" questions somehow
• Anthropic alignment, i.e. the “just make AI situationally unaware” approach which I think people have given up on too early

Deferred due to lack of urgency

Other stuff I'm interested in but are currently somewhat low-priority because AGI is going to be a thing soon.

• Reformulating probability theory in terms of closed-under-computable-union sigma-algebras
• "Interpretation" as a thing in intelligence, bounded rationality and thermodynamics
• Property rights theory, transaction costs, information asymmetry and thermodynamics
• Ways to better measure GDP, or better: Objective Utility
• My very sensible grammar concept
• Mathematical/computational models of state and war

Stuff I don't know but would like to

• Statistical mechanics/thermodynamics
• Macroeconomics -- money, interest rates, debt, inflation
• Cryptostuff
• Quantum information/probability/computing theory (should finish writing my quantum mechanics article series at some point)
• List of things in math I have learned but not really understood