The Builder's Brief
AI models improve, but risks rise
Tuesday, June 9, 2026
AC-Small improved significantly on held-out benchmarks after post-training on the APEX-Agents dev set, with +5.7pp on APEX, +8.0pp on Toolathalon, and +7.7pp on GDPval. This demonstrates the potential for large language models to generalize well beyond their initial training data. The improvement is notable and suggests that post-training on specific datasets can enhance model performance.
A woman in Tennessee was wrongly arrested by police using AI facial recognition technology for crimes committed in North Dakota. The incident highlights the risks and potential errors associated with relying on AI for law enforcement. It underscores the need for rigorous testing and validation of AI systems to prevent such mistakes.
Unregulated chatbots are posing significant risks to public safety, according to recent reports. Without proper oversight, these chatbots can provide harmful or inaccurate information, leading to dangerous situations. The lack of regulation and standards for chatbot development and deployment is a pressing concern that needs to be addressed.
Amazon Web Services has introduced a feature to control which domains AI agents can access, enhancing security and privacy for AI deployments. This feature allows for more granular control over AI agent permissions, reducing the risk of unauthorized access to sensitive information.
Science-T2I is an expert-annotated dataset designed to address the issue of scientific illusions in image synthesis. It includes over 20k adversarial image pairs and 9k prompts across 16 scientific domains, aiming to improve the physical realism of generated images. This dataset can help in developing more accurate and scientifically plausible image generation models.
The JUSSA framework is designed to aid LLM-judges with honest alternatives using steering vectors. It optimizes an honesty-promoting steering vector from a single training example, generating contrastive alternatives to detect subtle dishonesty. This approach aims to improve the reliability and trustworthiness of LLM-judges in various applications.
• AC-Small improves on APEX-Agents dev set
• Unregulated chatbots put lives at risk
• AI turns sketch into 3D-print pegboard
Think your CV is perfect? Let AI prove you wrong in seconds — brutal, honest, and hilarious feedback.
Try it free →🧠 Fun Fact: 20k adversarial image pairs in ScienceT2I dataset
