News Archive
Pick a date to read that day's stories.
Days with a blue dot have stories.
- 01 of 10POLICY
UN Scientific Panel Says AI Safeguards Are Losing the Race to Capability
A new UN Independent International Scientific Panel report published today warns that AI capabilities are outpacing global safeguards, with agentic systems already operating faster than any oversight regime member states have in place.
OUR TAKEDiplomats print another PDF while the models ship another version.
Read more → - 02 of 10REGULATION
Vietnam Publishes Official List of 46 High-Risk AI Systems
Vietnam's Ministry of Science and Technology released its formal registry of 46 AI systems classified as high-risk, covering biometrics, credit scoring, hiring, healthcare triage, and public-sector decision tools that will face mandatory disclosure and audit.
OUR TAKEThe EU AI Act finally has a fan in the room.
Read more → - 03 of 10ROBOTICS
Realbotix Puts a Humanoid Robot in Front of a New York Classroom
Realbotix has begun a pilot at Salamanca City Central School District using its M-Series humanoid and the district-trained Optio AI tutor, with plans to reach roughly 500 students this fall in what it calls the first full humanoid teaching-assistant deployment in a US school.
OUR TAKESubstitute teacher shortage, meet uncanny valley.
Read more → - 04 of 10OPEN SOURCE
Meituan Drops LongCat-2.0, a 1.6-Trillion Parameter Open MoE for Agent Coding
Chinese super-app Meituan open-sourced LongCat-2.0, a 1.6T total parameter Mixture-of-Experts model activating about 48B per token, with a native 1M-token context and a sparse attention scheme aimed squarely at long-horizon agent coding.
OUR TAKEAnother weekend, another Chinese lab casually posts a trillion-parameter weight file.
Read more → - 05 of 10NVIDIA
NVIDIA Open-Weights Nemotron TwoTower, a Diffusion LM Bolted Onto a Frozen Nemotron Backbone
NVIDIA released Nemotron-Labs-TwoTower, an open-weight diffusion language model that keeps a pretrained Nemotron-3-Nano-30B-A3B autoregressive backbone frozen and trains a parallel diffusion decoder on top, aiming for faster non-sequential generation without retraining the base.
OUR TAKEIf your base model still works, don't throw it out, bolt a diffusion head on it.
Read more → - 06 of 10AGENTS
Nous Research Ships Hermes MoA 2.0, Claims It Beats GPT-5.5 and Opus 4.8
Nous Research launched Mixture of Agents 2.0 inside Hermes Agent, letting users stack GPT, Claude and DeepSeek models as a single selectable virtual model; internal HermesBench numbers show the presets beating Claude Opus 4.8 by 8% and GPT-5.5 by 11%.
OUR TAKETurns out the shortest path to a frontier model is a group chat of other people's models.
Read more → - 07 of 10DEVTOOLS
Kimi K2.7 Code Becomes the First Open-Weight Model in the GitHub Copilot Picker
GitHub made Moonshot's Kimi K2.7 Code generally available inside Copilot, hosted on Microsoft Azure, the first open-weight model users can select directly in the Copilot model picker, billed at provider list pricing.
OUR TAKEMicrosoft quietly puts a Chinese open-weight model inside its flagship dev tool. Nobody blinks.
Read more → - 08 of 10SECURITY
Meta Freezes Mercor Contract After Breach Exposes AI Training Playbooks
Meta has paused its work with $10B AI data startup Mercor while investigators probe a supply-chain attack, reportedly a poisoned LiteLLM build, that exposed not just personal data but the training methodologies used by several leading labs to fine-tune frontier models.
OUR TAKEThe moat wasn't the model. It was a Python package nobody audited.
Read more → - 09 of 10ENTERPRISE
Ford Pulls Back on In-House AI, Emboldening Skeptics Across Detroit
Ford is scaling down several of its in-house generative AI programmes after cost overruns and thin ROI, a retreat that people inside GM and Stellantis are now citing internally as proof that automakers should buy AI capacity rather than build foundation models themselves.
OUR TAKEDetroit tried to be a hyperscaler for six months and remembered it sells cars.
Read more → - 10 of 10RESEARCH
Hugging Face's Pulpie Models Aim to Clean the Open Web More Cheaply Than Any Frontier LM
Hugging Face published Pulpie, a family of small Pareto-optimal classifiers built specifically to filter and deduplicate web crawl data, matching frontier-model quality scores on the FineWeb pipeline at a fraction of the compute cost per token.
OUR TAKEThe unglamorous part of AI, janitorial work on the training set, quietly gets its own model family.
Read more →