Daily AI

SkillEvaluator

Checks AI agent skills for quality and safety, then compares task outcomes with and without a skill to measure its actual impact. Its pipeline covers validation, overlap detection, and live agent evaluation.

claude-skills

Packages reusable workflows and instructions as installable Claude skills. Primarily designed for Claude.ai, with distribution through a Claude Code plugin marketplace.

TensorFold

Runs and serves language models locally on Apple Silicon and GPUs using native Metal, CUDA, and other backends. Its Zig engine performs inference without requiring Python or MLX.

SAM 3D Body: Robust Full-Body Human Mesh Recovery

Recovers a full-body 3D mesh and body, hand, and foot pose from a single image, with optional 2D keypoint or mask prompts. Its DINOv3-H+ variant uses the Momentum Human Rig representation for challenging poses and occlusions.

Multimodal Knowledge Agent (MKA) Showcase Examples

Tests whether AI agents can interpret complex visual files, synthesize information across documents, and produce finished professional deliverables. Its 10 public tasks span distinct domains and require multimodal perception and multi-step tool use.

LAION-Audio-XL

A corpus of roughly 366 million audio-text clips, each up to 30 seconds, curated for contrastive language-audio pretraining. Clips are organized by licence, with downloadable audio where redistribution is allowed and links for the rest.

Physical-Engineering’s Last Exam (PLE)

Evaluates whether AI agents can complete long-horizon engineering deliverables across structural, civil, electrical, and mechanical disciplines. Its 29-task development set combines automated checks with rubric-based grading.

Actor, Judge, or Teacher? Understanding Decision Models in Long-Horizon Agents

Tests decision models as actors, trajectory judges, and reinforcement-learning teachers for long-horizon agents. It finds that a model can be a weak actor yet still improve policy learning by evaluating actions after seeing their consequences.

On the Design of Hyper-Connections: Simplification, Stability, and Scalability

Studies how to simplify Hyper-Connections and reduce their training overhead. SimpleHC removes residual mixing and shares state access, matching mHC performance on a 30B MoE model.

Duo: A Reciprocal World Model for Full-Duplex Interaction

Continuously perceives live audio and video while deciding whether to respond, interrupt, or keep listening. It couples interaction reasoning with audiovisual generation for real-time, full-duplex exchanges.

BirdDreams: Joint Perspective and BEV Video Generation for a Plausible, Reactive Driving World Model

Generates driving video and a bird’s-eye view from one camera image, with optional ego-trajectory control. The paired views help model surrounding traffic reactions without requiring vehicle-state inputs or an HD map at inference.