Can LLMs Discover Interesting Math? New Research Explores AI-Driven Mathematical DiscoveryTime-Series Foundation Models Face Data Revisions That Leak Future InformationIndependently Trained RL Policies Can Be Audited via Rule Overlap and Behavioural AgreementTAM-Chain Cuts Thyroid Cytology False Negatives via Absorbing Markov ChainsSGA Method Quantifies Uncertainty in Multi-Step Time-Series Foundation Model ForecastingCARE: Condition-Aware Representation Regularization for Diffusion ModelsLLM Watermarking Imposes 'Provenance Tax' on AI Agent Behavior, Lasso Security FindsMistral CEO: AI Is Software, and It Can Be ControlledOne Month Without AI: A Developer's Experiment in Going Cold TurkeyOpenAI Codex Releases 0.158.0-alpha.15.2 and Multiple Alpha BuildsOpenClaw Ships v2026.8.33, Removes Low-Value Tests in Plugins, Model Picker, TerminalCan LLMs Discover Interesting Math? New Research Explores AI-Driven Mathematical Discovery

A new arXiv paper highlights that historical observations are not fixed—statistical agencies revise published values as new evidence arrives. Models forecasting from a contemporary data download may access information unavailable at the purported prediction date, causing leakage. The work examines how time-series foundation models should understand and handle data revisions, with direct implications for economics, weather, and other revision-heavy forecasting domains.

Reinforcement learning (RL) policies are often distributed as opaque neural checkpoints, with training logs showing a run occurred but not what the policy learned. This paper studies whether independently trained policies can be represented and compared, proposing that auditability is not one property but spans rule overlap, behavioural agreement, and composition. The key finding: different policies may share overlapping rules and exhibit partial behavioural agreement, but composed behaviour requires separate verification. For developers, this suggests black-box RL policies can be audited through behavioural comparison against expected rules rather than relying solely on training logs.

Thyroid fine-needle aspiration biopsy (FNAB) cytology based on the Bethesda System is critical for early thyroid cancer detection, but deep learning faces high false-negative rates and domain shift. This paper proposes TAM-Chain, a framework using absorbing Markov chains for multi-scale classification and Shannon entropy for uncertainty quantification, suppressing false negatives and adapting to domain shift. The key effect: lower miss rates while maintaining classification accuracy, plus improved cross-domain generalization. For clinical AI developers, this means cytology-aided diagnostic systems can run more stably across data from different hospitals, reducing misjudgments caused by equipment or staining variations.

A new arXiv paper introduces SGA, a method for uncertainty quantification in multi-step forecasting with time-series foundation models (TSFMs). It notes that while TSFMs have advanced long-horizon prediction, they suffer significant inherent uncertainty. SGA models predictive distributions to help users assess risk when relying on TSFMs for long-term decisions in energy, finance, and beyond.

A new arXiv paper proposes CARE (Condition-Aware Representation Regularization) for diffusion models. Existing regularization methods often overlook built-in conditions such as labels or text; CARE incorporates condition information into the regularization process to improve sample quality and training efficiency. It has potential to enhance conditional diffusion tasks like image generation.

Lasso Security published a blog post examining how LLM watermarking affects AI agent behavior. It introduces the concept of a ‘provenance tax’—the performance and behavioral costs watermarking imposes on agent tasks while enabling content traceability. The post drew 56 points and 70 comments on Hacker News, sparking developer discussion on the trade-offs of watermarking.

Mistral AI CEO Arthur Mensch told Le Monde that AI is fundamentally software and therefore can be controlled. The statement addresses growing concerns over AI safety and regulation, framing AI as something governable rather than an uncontrollable black box. The piece drew 87 points and 153 comments on Hacker News, fueling debate on AI governance.

Blogger bustikiller documented a month-long experiment of going without AI tools. The post reflects on how deeply AI has permeated daily development and life, and what changed after stepping away. It drew 169 points and 209 comments on Hacker News, signaling growing developer self-reflection on AI dependency.

OpenAI Codex shipped 0.158.0-alpha.15.2 alongside a rapid series of alpha builds including 0.159.0-alpha.6, 0.159.0-alpha.5, 0.159.0-alpha.4, plus 0.158.0-alpha.2.1 and 0.158.0-alpha.15.1. The fast version cadence indicates Codex is under heavy active development with ongoing feature and stability changes.

OpenClaw released build release-publish/8d8bd43120f3-1790449772 and version v2026.8.33. The update focuses on test cleanup, removing low-value test cases across plugins, model-picker, and terminal modules. The move aims to streamline the test suite and speed up CI, reflecting ongoing engineering refinements during rapid iteration.