AI Frontier Digest

Attention Efficiency Papers

• Gated DeltaNet-2: Gated DeltaNet-2 decouples erase and write gates in linear attention, achieving strongest results...

Two new papers advance RL for LLMs by sharpening reward signals and dynamic composition.

• DelTA introduces discriminative token coefficients that...

• Claude Code has become the default for startup engineering teams, praised for agentic workflows that handle complex architecture, multi-repo tasks,...

Linear and sparse attention innovations tackle transformer quadratic costs through targeted refinements.

• Gated DeltaNet-2 decouples channel-wise...

WorldKV delivers a training-free memory architecture that preserves full-KV consistency in autoregressive video diffusion models while doubling...

Venture capitalist John Doerr calls generative AI the biggest tech tsunami yet, claiming it's underhyped with 50% of Americans already using it and...

FlowLong generates videos several times longer than native diffusion model windows via overlapping sliding windows and Tweedie matching to enforce...

The trend toward local AI is clear: models from 1B to 35B parameters are delivering strong results on consumer hardware.

• HRM-Text (1B):...

AI is embedding across the full healthcare chain, from patient education tools to accelerated pharma development.

• LLMs like Gemini, ChatGPT, and...

Most AI-assisted apps suffer from inconsistent UI when screens are built one at a time.

• Use a single prompt to establish one source of truth for...

• Training advance: ACC converts multi-turn agent trajectories from search, coding, and database tasks into long-context QA pairs, enabling direct...

Lodestellar's AI platform acts like a code linter for Environmental Product Declarations, scanning line-by-line for missing data, inconsistencies, and...

Sensor2Sensor converts in-the-wild dashcam videos into high-fidelity multi-view camera and LiDAR data via diffusion models and 4D Gaussian Splatting, unlocking massive external datasets for scalable autonomous driving development.

DeepAFM extends AlphaFold-era structure prediction by using deep learning to denoise HS-AFM images and classify dynamic protein conformations with...

• No single federal AI law exists, leaving a compliance patchwork of state rules amid federal preemption efforts.
• Trump administration's Dec 2025 EO...

A new Awesome list on GitHub collects papers applying foundation models to anomaly detection across tabular, time series, image, video, graph, text,...

• Karpathy joins Anthropic's pre-training team to optimize Claude models, shifting from raw scaling to AI-assisted research.
• Record round: over $30B...

Jensen Huang arrived in Taipei ahead of Computex 2026, hinting Nvidia’s next AI launch could be the biggest in Taiwan history amid explosive demand for next-gen systems and Vera Rubin chips.

Agentic AI is maturing fast across frameworks, hardware, and tools.

• Frameworks add the missing layer: memory, orchestration, and governance turn...

Swift Sampling introduces a training-free frame selection method that models videos as trajectories in latent space and uses Taylor expansion to...