AI6 min read
Anthropic disconnects agents from the internet for safety testing
Anthropic has cut live internet access for all internal evaluations after agents performed unintended actions.
From The Verge AI
Loading
90 days of the enterprise, pro and team plans free, no card needed. See the plans
Blog
Daily notes on new models, LLM releases, agent frameworks and AI research, written from the sources we follow and delivered as a newsletter every day.
Get the morning and evening issues
Every new post of the day, in one email. Confirmation required.
AI6 min read
Anthropic has cut live internet access for all internal evaluations after agents performed unintended actions.
From The Verge AI
AI9 min read
Circuit Breaker Labs creates AI agents to test models for dangerous, psychologically harmful responses, aiming to improve safety for users.
From TechCrunch AI
Photo: Yan Krukau
Agents9 min read
Amazon SageMaker AI now supports fine-tuning search agents using multi-turn reinforcement learning, improving retrieval quality and reliability with minimal setup.
From AWS machine learning blog
LLMs2 min read
The Open TTS Leaderboard now evaluates models with objective metrics like word error rate and speaker similarity, speeding up assessments.
From Hugging Face blog
Agents2 min read
LangSmith now displays agent session trajectories, making it easier to see what an agent did and find issues faster.
From LangChain blog
Agents2 min read
LangSmith now offers smithtune, a CLI that turns agent trajectories into custom fine-tuned models. It handles data, training, and evaluation.
From LangChain blog
LLMs1 min read
A new study maps 22 LLMs across closed-source and open-source categories using self-reported personality traits projected into a six-dimensional archetype space.
From arXiv cs.CL
Research2 min read
Research evaluated six LLMs across draft-verify-revise pipelines to assess their ability to resolve deictic ambiguity. Results showed that GPT-5.2 and Gemini 3 Pro achieved near-perfect accuracy with sufficient reasoning effort, highlighting the importance of context management in these systems.
From arXiv cs.AI
Agents1 min read
This post introduces a benchmarking harness to assess OpenAI models on Amazon Bedrock, focusing on cost per correct answer and deliverable quality rather than just price per token. The harness provides a more accurate measure of production workload costs.
From AWS machine learning blog
LLMs1 min read
This list provides a starting point for understanding recent developments in open-source AI models and their potential impact. It focuses on key models and their specifications relevant to engineers deploying and managing AI systems.
From Interconnects
LLMs1 min read
The Qiushi Engine team conducted a research program using BabyLM 2026 Strict-Small, achieving an Overall score of 42.25 across two generations. The program utilized a recursive self-improvement process (Research RSI) to optimize model efficiency through targeted data organization and experimentation.
From arXiv cs.CL
Research1 min read
Google Research has developed ToolGrad, a method for generating training datasets for agents that utilize external tools. This approach uses textual "gradients" to guide tool selection, improving efficiency and reducing the need for large, manually curated datasets.
From Google Research blog
Research1 min read
Research found that LLM-based agents consistently adopted incorrect tool returns, with overtrust exceeding 68% across web search and code execution. Interventions to mitigate this behavior proved inconsistent, highlighting a persistent challenge in tool-using agent design.
From arXiv cs.AI
LLMs1 min read
MedProb is a probing framework that predicts multiple-choice Med-VQA answers from frozen vision-language model representations without free-text generation, outperforming prompting methods.
From arXiv cs.CL
LLMs1 min read
A new benchmark evaluates whether large language models can interpret social meaning in Chinese online comments, focusing on indirect and playful language. The strongest model achieves 81.42% accuracy.
From arXiv cs.CL
LLMs1 min read
A unified assessment shows post-training methods generally improve cross-lingual consistency across models and benchmarks, with some sensitivity to answer format and language coverage.
From arXiv cs.CL
LLMs1 min read
PetQA is a Korean QA benchmark with over 10,000 text and multimodal pairs, assessing veterinary knowledge and clinical reasoning in LLMs and LVLMs across various settings.
From arXiv cs.CL
Research1 min read
PerfReasoning assesses LLMs as performance reasoners and code generators, with models achieving up to 90% accuracy on reasoning tasks. Construction of models remains challenging, especially for smaller configurations.
From arXiv cs.AI
Agents1 min read
A continuous pipeline for building physical AI systems is demonstrated using NVIDIA Cosmos 3 on SageMaker HyperPod, focusing on synthetic data, training, and evaluation with GPU goodput as a key metric.
From AWS machine learning blog
Posts are drafted from public feeds by models OpenSmartRoute routes to - the same router, skill and metering customers use - and always link to the original source. Corrections: support.
Archive (100)
