New MosaicLeaks Framework Tests AI Research Agents' Secrecy
ServiceNow's MosaicLeaks, published on Hugging Face, probes whether research agents can keep confidential information secret.
AI & ML
Frontier models, research results, agents, and developer tooling.
Import AI 463: Robots, China's 10k GPUs, and the Human Era
The latest Import AI brief ties together self-improving robots, a large Chinese GPU cluster, and an essay about the end of the human era.
CopilotKit CLiF Turns Agent Traces into Continuous Learning
The new platform converts real agent traces and user actions into skills, evals, and fine-tuning data to help teams deploy reliable AI agents.
Nvidia expands open Nemotron family with Super VL and Bolt edge models
Nvidia's open-weight Nemotron 3.5 Super VL and Bolt edge models push agentic AI efficiency from data centers to local hardware, as US open-weight models gain ground.
Google Antigravity accelerates search for antimicrobial peptides
An agentic science workbench from Google DeepMind compresses fragmented database searches into one session and surfaces a wet-lab candidate.
Modal hits $1B ARR, raises $750M at $15.75B valuation
The AI infrastructure startup has tripled usage since its Series C and now counts sandboxes as a $300M run-rate product.
Datalab's OmniParseBench Audits OCR Parsers Across 92 Languages
A new open benchmark gives developers a standard way to measure OCR parsers using pass/fail tests on multilingual pages.
Grok Voice 3.0 Tops Speech-to-Speech Index with 86.9 Score
SpaceXAI's new speech-to-speech model leads the Artificial Analysis ranking, aiming at developers with tool-calling and a commercial endpoint.
Why Specialization, Not Generalization, May Define AI’s Next Phase
A Hugging Face blog post argues that AI’s future lies in specialised models rather than ever-broader general-purpose systems.
OpenAI Paper: Agents Enable Longer, More Complex Work
New research from OpenAI suggests AI agents are shifting work from short, isolated tasks to longer, more complex assignments.
Open Model Releases from Zyphra, Cohere, and Poolside Broaden AI Ecosystem
An assessment of recent open artifacts shows how these three firms are expanding the range of open models and why they choose to release them.
Gemini 3.5 Flash Gains Computer Use Abilities
DeepMind's new model can interact with screens, moving AI toward practical agentic tasks.
Why Reasoning Helps LLMs Recall What They Know
Google Research shows that chain-of-thought prompting can surface parametric knowledge that direct questioning misses.
Google Research Proposes Linear Elastic Caching for Cloud Cost Savings
A new algorithm treats cache as an elastic resource to reduce cloud spending.
Google's Frozen Multi-Token Prediction Speeds Up Gemini Nano on Pixel
A new technique called frozen multi-token prediction lets Gemini Nano models generate multiple tokens at once on Pixel devices, cutting latency without retraining.
OpenAI Signals Data Shows Growing Global ChatGPT Adoption
New data from OpenAI Signals indicates that ChatGPT users are engaging more deeply while adoption expands across regions and languages.
openKylin Pushes Linux Toward Intent-Based AI Management
openKylin's new AI runtime and desktop features signal a shift from managing resources to understanding user intent.
Hugging Face Adds One-Command vLLM Server for HF Jobs
A new integration lets developers launch a vLLM inference server directly inside a Hugging Face Job, cutting deployment friction.