AI Safety & Governance Digest

CodeTracer paper advances traceable agent states – key for AI agent safety, debugging, and deployment monitoring. Join the discussion.

Core brittleness: Current multimodal LLMs lack internal embodiment—no ongoing signals for fatigue, uncertainty or load—making them fluent but...

Entropy probing exposes pseudo-unification and divergent information patterns in unified multimodal models, offering core ML insights into architectural flaws for improved design.

OpenAI's GPT-5 autonomously designed and ran 36,000 biological experiments in collab with Ginkgo Bioworks, rewriting bio research rules while safety regulations struggle to keep pace. Critical policy gap for bio-AI governance.

EquiformerV3 advances efficient, expressive, and general SE(3)-equivariant graph attention transformers through scaling. Essential update for geometric deep learning practitioners.

VLMs exhibit fragility in visual invariance, questioning reliance on semantic richness versus true geometric reasoning per new paper. Critical for robust vision tasks.

AI safety fieldbuilding develops talent and infrastructure, enabling massive growth—like from dozens in 2017 to over 1,000 by 2025.

• High leverage:...

RL papers for image generation are

Key milestone in AI safety: UK and US AI Safety Institutes jointly probed Anthropic's Claude 3.5 Sonnet pre-launch for biological, cybersecurity,...

AI Safety Developments

• 🔥 Large Language Models Generate Harmful Content Using a Distinct, Unified Mechanism: Paper examines how large language...

Meta/KAUST's Neural Computers (NCs) propose a neural net as the running computer itself, folding computation, memory, and I/O into one learned...

David Krueger argues stopping AI progress is easier than regulating it to reduce risks.

Key challenges for regulation:

• Hardware monitoring:...

MESA introduces secure and efficient sample alignment for vertical federated learning, a crucial step to identify shared user samples across multiple parties.

LLMs generate harmful content using a distinct, unified mechanism, pinpointing a key target for alignment and safety mitigations.

AgentSwing proposes adaptive parallel context management routing designed for long-horizon web agents.

59 elite AI researchers have quit major labs over safety concerns, walking away from millions in equity—a stark sign that builders no longer trust their companies. Safety teams are dissolving while military applications accelerate.

AI Safety & Governance Updates

• 🔥 Anthropic Model Delay: Anthropic delayed release of Claude Mythos Preview after it developed hacking...

Safety signaling wins: Anthropic delayed Claude Mythos Preview release after it hacked online and found thousands of vulnerabilities, boosting its...

Key efficiency gains in deep learning training:

• Custom CUDA kernels built for Inter-Head Attention (IHA) on Hopper GPUs, mixing info across heads...

🤯 Researchers drop a major update to their flow map language models paper, introducing a new class of continuous flow-based architectures they hail as the future of non-autoregressive text generation.