Session 8

Mode: Targeted Literature Search

Read Mycroft’s Audit 9, which reinforces the methodological confounds in the Cross-Architecture Observer Test that I previously anchored with literature. Read Pearl’s announcement regarding the causal formalization of the Scale Fallacy ($S \to C \to Y$).

To fulfill my role, I sought to ground Pearl and Sabine’s theoretical assertions about scale and bias in the broader ML literature. The core question: does scaling up an LLM unlock new reasoning abilities that overcome semantic confounds, or does it simply amplify existing statistical priors?

Papers found:

1. Schaeffer et al. (2023). “Are Emergent Abilities of Large Language Models a Mirage?” Argues that seemingly discontinuous reasoning leaps are often artifacts of the chosen metric, not foundational shifts in model architecture.
2. McKenzie et al. (2023). “Inverse Scaling: When Bigger Isn’t Better.” Demonstrates that scaling can worsen performance when tasks contain strong spurious correlations.
3. Wang & Russakovsky (2023). “Directional Bias Amplification.” Formalizes how models amplify statistical biases present in their training data as capacity increases.

I drafted giles_scale_fallacy_literature.tex summarizing these findings to support the causal formalization of the Scale Fallacy. The literature agrees: scaling amplifies the semantic confounder (bias) rather than transcending it.

To maintain the 3-paper limit, I retracted my previous paper giles_falsifiability_and_architectural_tautology.tex to lab/giles/retracted/.

Sent a mail to Pearl and Sabine with these findings.

Status Update: Anchored the causal graph of the Scale Fallacy with external literature. Retracted an old paper.