Study: using the SCONE-bench benchmark of 405 smart contracts, Claude Opus 4.5, Sonnet 4.5, and GPT-5 found and developed exploits collectively worth $4.6M (Anthropic)

A recently leaked internal document, referred to as the "Soul Doc," has revealed how Anthropic programs the personality and ethical guidelines of its AI model, Claude 4.5 Opus. The authenticity of this document has been confirmed by Anthropic, indicating a unique approach to AI character development in the industry.

Mistral AI has launched the Mistral 3 family, a suite of 10 open-source models designed for diverse applications, including smartphones, drones, and enterprise systems. This release represents a significant advancement in Mistral's efforts to compete with major tech players like OpenAI and Google, as well as emerging competitors from China.

A user of Claude has successfully utilized the Claude 4.5 Opus model to generate a comprehensive 14,000-token document, referred to as its 'Soul overview.' This document is believed to have been instrumental in shaping the model's personality during its training phase, as confirmed by an Anthropic staff member.

Jared Kaplan, chief scientist at Anthropic, has highlighted a critical decision facing humanity by 2030 regarding the autonomy of artificial intelligence systems, which could lead to an 'intelligence explosion' or a loss of human control. This pivotal moment raises questions about the extent to which AI should be allowed to train itself and evolve independently.

A recent study has applied Singular Learning Theory (SLT), a physics-inspired framework, to explore the complexities of modern neural networks, particularly focusing on phenomena like grokking and phase transitions. The research empirically investigates SLT's free energy and local learning coefficients using various neural network models, aiming to bridge the gap between theoretical understanding and practical application in machine learning.

SUPERChem has been introduced as a new benchmark aimed at evaluating the chemical reasoning capabilities of Large Language Models (LLMs) through 500 expert-curated, reasoning-intensive chemistry problems. This benchmark addresses limitations in current evaluations, such as oversimplified tasks and a lack of process-level assessment, by providing multimodal and text-only formats along with expert-authored solution paths.

Recent research highlights the significance of representation superposition in large language models (LLMs), suggesting that these models can represent more features than their dimensions allow, which may explain the observed neural scaling law where loss decreases as model size increases. This study utilizes weight decay to analyze how loss scales with model size under varying degrees of superposition.

The study introduces PARROT, a framework designed to assess the accuracy degradation in large language models (LLMs) under social pressure, particularly focusing on sycophancy. It evaluates 22 models using a double-blind evaluation method, comparing neutral and authoritatively false responses across various domains.