AI Frontier Digest

US export curbs on AI tech are proving counterproductive, as China's new LineShine supercomputer bypassed GPU restrictions by scaling standard...

DeepSeek is developing a custom AI inference chip to cut dependence on Nvidia and Huawei, marking a shift toward full-stack control.

• Inference...

Warehouses serve as the essential proving ground where bipedal robots like Digit learn to navigate human spaces before entering homes.

-...

Waymo is launching fully autonomous rides with no human driver in Las Vegas, San Diego, Tampa, and Denver, expanding its network beyond 10 cities and...

SpaceX's enormous V3 Starlink satellites launching next week on Starship match the size of a new orbital AI data center design called AI1.

-...

Generative AI does not fit traditional enterprise architecture — it rewrites how architectures are designed, governed, and evolved.

• Living...

Current benchmarks fail due to visual overlap, saturation, poor diagnostics, and weak statistics.

• RoboLab enables robot-agnostic testing with rapid...

Large Model Breakthroughs

• 🔥 Colibrì local inference: Italian engineer runs 744B-parameter 1.5-TB GLM-5.2 model on CPU with 25 GB RAM via sliced...

GRASP introduces a general–specialized evaluation paradigm that fuses multi-source leaderboard data via PCA-based information integration, enabling...

AI tools promise efficiency gains in documentation, imaging, and workflows for stretched care teams, yet the same capabilities can drive up spending...

• Committee Setup: Three LLM agents manage data analysis, risk assessment, and trade decisions via Chain-of-Thought prompting
• Performance Edge:...

CellTok converts continuous single-cell transcriptomes into compact token sequences that join a pretrained LLM's vocabulary, enabling cells, text...

Colibrì lets enthusiasts run the 744-billion-parameter GLM-5.2 model on modest CPUs with only 25 GB RAM by dynamically loading MoE experts per token,...

A new open-source hardware accelerator performs on-the-fly spherical-to-Cartesian conversion of VLP-16 LiDAR packets directly inside a bufferless...

PrismML achieved the largest AI model yet runnable on an iPhone, supporting complex chat, reasoning, fully autonomous agents, and software coding. This signals a major step toward practical edge AI deployment for on-device intelligence.

• H.R. 2152 mandates a U.S. strategy against AI-driven financial crimes and misinformation to protect economic security.
• Led by Treasury, Homeland...

Task-agnostic LLMs and VLMs like Vicuna and Otter can generate natural language descriptions of geospatial changes in a zero-shot setting, bypassing...

An Expert-AI hybrid approach—pairing human architectural design with LLM-driven prototyping, diagnosis, and refactoring—cut feature implementation...

A new IEEE survey fills a critical gap by reviewing LLM-based insider threat detection methods.

• Core approaches: BERT-, GPT-, LLaMA-based, and...

General-purpose LLMs like Llama 3.1 70B can generate valid ABB RAPID code for robotic automation tasks using only few-shot prompting, reaching ~99%...