Yonathan Arbel

We ran GPT-5.4 (xhigh) an additional ten times on Tier 4 to get a pass@10 score. This was 38%. In one of these runs, it solved another problem no model had solved before. This problem was by @nasqret.

Arbel, Goldstein, & Salib on Individuation of AI Agents, legaltheoryblog.com/2026/03/02/arb… - Yonathan A. Arbel (University of Alabama – School of Law), Simon Goldstein, (The University of Hong Kong – University of Hong Kong), & Peter Salib (University of Houston Law Center) have posted How to Count AIs: Individuation and Liability for AI Agents on SSRN.

Thank you for featuring this, Professor Solum (@lsolum) The timing is good: I had been working on a comment to Arbel (@ProfArbel), Goldstein (@simongold) & Salib from the perspective of law as extended phenotype and multilevel evolutionary game theory, and just published it on Zenodo this week. The framework raises two questions I try to work through in the comment. The A-corp's selection mechanism maps cleanly onto EGT replicator dynamics, and that mapping reveals the core structural problem: selection converges on well-governed A-corps over years; a misconfigured agent accumulates harm in hours. The ratio between those timescales is around ten to the seventh power. The mechanism prices harm retrospectively. It does not prevent it. The second concerns RLVR-trained agents specifically. Daniel Dennett's intentionality taxonomy distinguishes systems by their capacity for recursive social reasoning: * Level 0-1 systems optimize against objective functions; * Level 3 systems model others' beliefs about their own beliefs, feel shame, honor commitments. Legal standards presuppose Level 3 throughout. RLVR-trained agents do not operate there stably: during goal interpretation and configuration they reason at Level 2-3, but during execution, where most harmful actions occur, they run Level 0-1 optimization loops. The A-corp's incentives only reach the high-intentionality phases. I formalize this as Dynamic Classification Failure, the sixth mode of the Generalized Intentionality Mismatch Theorem. The constructive alternative is the Responsibility Ramp: dynamic intentionality classification by task phase, strict liability for execution-phase harms, and attribution tracing the configuration chain to the decision that mattered. Full Zenodo paper: doi.org/10.5281/zenodo…

Arbel, Goldstein, & Salib on Individuation of AI Agents, legaltheoryblog.com/2026/03/02/arb… - Yonathan A. Arbel (University of Alabama – School of Law), Simon Goldstein, (The University of Hong Kong – University of Hong Kong), & Peter Salib (University of Houston Law Center) have posted How to Count AIs: Individuation and Liability for AI Agents on SSRN.

How to Count AIs: Individuation and Liability for #AIAgents This Article is the first to diagnose the legal problem of identifying #AIs. Authors: Yonathan A. Arbel, Simon Goldstein, Peter Salib Read More: spkl.io/6018AvgjQ #AI