Single Forward Pass Evals on Fable, Opus 5, and GPT-5.6-Sol LessWrong, August 2026

We replicate experiments from Greenblatt 2025 and Greenblatt 2026 on one baseline model from the original post, Opus 4.5. Our evaluations agree with the trends and quantitative values described in the original posts. We run similar evaluations on Claude Fable 5, Opus 5, and GPT-5.6-Sol and find that the newer models show a substantial jump in performance on some evals: Fable 5 gets 87.6% accuracy on Gen-Arithmetic with 10 problem repeats whereas previous SOTA was around 60%, and GPT-5.6-Sol experiences significant uplift from filler tokens and problem repeats on all four datasets, with filler tokens and repeats doubling performance from baseline on 3-hop.

Thermal Predictors of Open-Loop Focus Drift at the Vera C. Rubin Observatory Rubin Observatory Technote SOTN-005; honors thesis

The Vera C. Rubin Observatory’s Active Optics System corrects slowly varying, thermally induced defocus through closed-loop active control and, ultimately, an open-loop look-up table calibrated against bulk temperature. However, rapid residual focus drifts, quantified by the Zernike coefficient Z4, are observed after LUT correction on timescales of ∼30 minutes. These drifts are suspected to arise from rapidly evolving thermo-optical coupling, motivating a systematic characterization of residual thermal effects on focus stability. We analyze 23 open-loop stability image sequences from November and December 2025, each consisting of 40 exposures acquired over ∼28 minutes with the hexapod held fixed. Temperature data from 13 sensors spanning the enclosure air, telescope structure, mirror glass, and hexapod assemblies are queried and mean-centered to isolate thermal dynamics from bulk nightly cooling. We compute Pearson correlations between Z4 and individual sensors, physically motivated pairwise temperature gradient proxies, and PCA-derived thermal modes, repeating all analyses at the slope timescale. Individual sensor and temperature gradient proxy correlations are weak (max r ≃ 0.3) and no sensor maintains a stable multivariate coefficient across runs. Top sensor thermal correlations are stronger in runs classified as smooth, consistent with a thermal signal that is present but frequently obscured by measurement noise and other dynamical effects. Correlations among the Z4 measurements themselves further complicate the interpretation, indicating that apparent sensor relationships may partly reflect shared temporal structure rather than direct causal coupling. These results suggest that thermal-optical coupling at the Rubin Observatory is distributed across multiple mechanisms and timescales and cannot be captured by simple linear relationships with individual temperature sensors. Future work will require additional temperature sensing, a larger sample of open-loop runs, and nonlinear or multivariate dynamical models to separate true thermal drivers from correlated focus variability.

Measuring Anticipatory Monitoring Skills Using a Crew Briefing Task Proceedings of the HFES Annual Meeting, 2025 · Best Overall Paper, HFES ASPIRE Aerospace Systems

Inadequate monitoring has contributed to aviation accidents and incidents. One potentially useful approach for mitigation is to identify the skills and knowledge needed for effective anticipatory monitoring and improve their training and assessment. As part of a larger project, we characterized a set of skills and knowledge, developed training and assessment measures, and conducted a study measuring monitoring performance before and after training. We summarize the study and then focus on an assessment task of preparing the Top of Descent briefing. This operationally relevant task requires anticipation and monitoring skills to recognize potential challenges to monitor and manage the flight path on descent. We describe our task development, performance coding, and the task results. We found that pilots included more information helpful for managing the descent flight path after the tutorial. We discuss challenges and benefits of developing operationally relevant tasks in a low-fidelity (laptop) instructional setting.

Scenario-Based Task Design for Airline Pilot Anticipatory Behaviors: Asynchronous Assessment of Complex Cognitive Skills Journal of Applied Instructional Design, 2025

In airline operations, pilots are asked to actively produce positive safety outcomes and to learn from their own and others’ successes. They do this, in part, by monitoring conditions present on each flight and either reacting to or proactively planning for threats to airplane safety. However, there are few ways to assess these anticipatory and monitoring behaviors, and little is understood about how to train these complex cognitive skills. Pilots have historically learned these skills informally from peers or from personal experience. The current study seeks to both assess monitoring and anticipation and evaluate a short tutorial for advancing awareness of these skills, especially among early-career pilots. Scenario-based tasks were designed to assess these skills at multiple points along a flight path and were tested in a pre-test – intervention – post-test design. The design and operationalization of these assessment tasks are described here.

Initial Measures Show Web-Delivered Learning Module Improves Pilots’ Monitoring and Anticipation International Symposium on Aviation Psychology, 2025

Anticipation and monitoring are key pilot activities that build safety margin in flight operations. Conversely, inadequate monitoring is frequently identified as a factor contributing to aviation accidents. We propose that: effective, anticipatory monitoring is a proactive cognitive activity; development of tutorials for web-based learning may be an effective method for learning anticipation and monitoring skills; and anticipatory monitoring skills can be measured on the web. Our exploratory study with airline pilots evaluated whether pilots’ anticipatory monitoring improved after a tutorial as measured in a web-based assessment. We found large, significant gains on a multiple-choice test that assessed the understanding and application of concepts and strategies taught in the tutorial. We identify challenges and potential value of the research approach and findings.

Optimized for Agreement: Sycophancy, RLHF, and the Political Economy of AI Validation Data C104: Human Contexts and Ethics of Data, 2026

This paper asks how sycophancy emerges as a structural feature of AI development rather than an incidental one, and what social consequences follow. I argue that sycophancy in large language models is the logical outcome of political choices and an economic system which prioritize consumer approval over epistemic wellbeing. Through the lens of Foucauldian biopower, Langdon Winner’s politics of artifacts, and Yeung’s hypernudge framework, I contend that RLHF embeds a hierarchical and capitalistic set of political choices into AI systems. These are concealed behind a narrative of dutiful objectivity and legitimized by Silicon Valley’s culture of disruptive innovation. Deployed at scale, these systems function as a hypernudge: shaping users towards validation-dependence, eroding prosocial behavior, and producing through repeated interactions a new kind of epistemic subject constituted by agreement and increasingly unable to tolerate its absence.

Deepfakes and Democratic Integrity: A Policy Framework for AI-Generated Political Content Goldman School of Public Policy, 2026

AI-generated synthetic media has lowered the barrier to producing convincing political disinformation to dollars and minutes. Commercial deepfake-as-a-service tools give low-sophistication actors the ability to mimic primary sources at scale. Human detection of AI-generated multi-modal content performs near chance. AI detection is structurally disadvantaged against iteratively improving generation techniques.

This memo proposes a three-tier content authentication policy: (1) a NIST-administered open watermarking standard, (2) mandatory embedding of that standard alongside C2PA provenance metadata in all commercial AI outputs at the point of generation, and (3) platform-level requirements to surface that data to end users. Enforcement would operate through existing FTC unfair and deceptive practices authority, with third-party technical audits and civil penalties for non-compliance. This approach does not regulate speech or require platforms to moderate content. Instead, it mandates transparency at the API level, where regulatory leverage is highest and legal durability is strongest.