Professor joins Tsinghua AI faculty

A researcher is joining Tsinghua University’s College of AI to work on frontier AI safety and will co-direct a new local AI safety hub, aiming to build sustained international collaborations and run a residency program. — lesswrong.com

Thirteen models converge on shared critique

Thirteen contemporary text models independently produced the same controversial take that humans prioritize belonging over truth, illustrating how shared training data can drive convergent model outputs. — modelcompare.ai

Stars encode a hidden English sentence

Forensic analysis recovered a grammatical English sentence encoded in the vertical positions of 40 stars in an illustration, demonstrating that precise y-coordinate placement can embed readable messages in images. — lesswrong.com

Local open-source agent for personal workflows

Cindy is an open-source agent client that runs locally on desktop and mobile, orchestrates multiple models and tools, and keeps files and logs on users' machines so workflows can be automated without external endpoints. — github.com

Agents produce dozens of exoplanet candidates

An AI-assisted search of TESS survey data produced about 25 exoplanet candidates, demonstrating that agentic systems can perform practical scientific discovery and highlighting the need for independent verification. — lesswrong.com

OrcaCyber Nano enables local cybersec analysis

OrcaCyber Nano is a vision–language cybersecurity model packaged for offline local use, enabling reverse engineering, vulnerability analysis, and secure code review of sensitive binaries and logs without sending data to external services. — huggingface.co

Uncensored Qwen build enables local hosting

An uncensored Qwen build is available with local hosting instructions for common runtimes, making an uncensored variant straightforward to run offline on personal hardware. — huggingface.co

Small transformer meets several workspace tests

A small transformer (GPT-2 small) passes most functional global-workspace tests but lacks directed modulation and the structural signatures seen in larger models, implying those functional checks alone are weak evidence of a true workspace. — lesswrong.com

Qwen visual model enables offline NSFW captioning

A Qwen visual-language model for NSFW captioning is provided in deployable form with instructions, enabling offline image-to-text captioning for specialized workflows without external endpoints. — huggingface.co

Microduck repo shows AI-assisted development

The Microduck open-source robot repository has recent releases and many commits that include AI-generated co-authorship across documentation and tooling, illustrating growing use of model assistance in hardware development. — github.com

Related Digests