RuCCS-Drs. Jacob Feldman & Susanna Schellenberg- Double Feature Dialogue-Tuesday, September 15, 2026

Value of the Information & Strength of the Argument

The video offers substantial value by presenting a concrete, empirical comparison between human and AI learning on a shared task, avoiding speculative claims. The argument is built on a well-designed experimental paradigm: the Game of Hidden Rules allows for controlled manipulation of rule types (feature-based, space-based, sequence-based) and compositionality. The key argument—that human and AI transfer from simple to composite rules are almost uncorrelated—is supported by a transfer design with congruent and incongruent conditions. The reasoning is clear: if AI learned like humans, prior exposure to component rules should facilitate composite rule learning, but the data show no such facilitation. The discussion of ‘insight’ in human learning (irregular performance curves) versus the AI’s gradual parameter-space descent is a strong conceptual contrast. However, the argument would benefit from more statistical detail (effect sizes, p-values) and a fuller discussion of alternative explanations (e.g., AI’s lack of prior knowledge, different exploration strategies). The second talk on self-reference and metacognition is more programmatic, arguing that AI lacks a general capacity for self-reference, but it is less empirically grounded in the transcript.

Scientific Rigor, Source Quality, Title Accuracy

The scientific rigor is moderate to high: the study uses a controlled experimental design, pre-registered exclusions, and a standard RL model (A2C with transformers). The methodology is transparent, though the transcript lacks full statistical reporting. Sources are not explicitly cited in the description beyond the abstracts; the talk references Paul Gander’s work and the history of compositionality debates at RuCCS, but no URLs or formal citations are provided. The title accurately reflects the content: a double-feature dialogue between two faculty members on cognitive science and AI. The format is informal, which may reduce perceived rigor, but the substance is solid. No comments were provided for analysis.

The title accurately reflects the event: a double-feature dialogue between two RuCCS faculty, Jacob Feldman and Susanna Schellenberg, on cognitive science and AI. The content matches the announced format.

7/10

The talk presents original experimental data from a controlled human-vs-AI comparison on the Game of Hidden Rules, with clear methodology (Prolific participants, RL model details, transfer design). However, the presentation is informal, lacks full statistical reporting in the transcript, and the second half (Schellenberg) is cut off. No external sources are cited in the description beyond the abstracts.

Markers derived by PSI from the transcript: the creator did not define chapters.

• 00:00Introduction to the RuCCS fall events and the double-feature dialogue format.
• 05:00Jacob Feldman introduces the Game of Hidden Rules (GOHR) and the goal of comparing human and AI performance empirically.
• 10:00Demonstration of the GOHR game: sorting shapes into bins based on an unknown rule.
• 15:00Explanation of rule types: feature-based, space-based, sequence-based, and compositional rules.
• 20:00Human performance data on 18 fiducial rules, ordered by difficulty.
• 25:00Human learning curves showing irregular progress and 'insight' moments.
• 30:00AI (RL) performance data: over 8,000 moves on average vs. 68 for humans.
• 35:00Correlation analysis: no relationship between human and AI performance across rules.
• 40:00Transfer experiment design: simple rules A and B followed by composite rule A+B.
• 45:00Results: humans show transfer, AI shows almost none; discussion of implications.