Evaluate an Open LLM with the EleutherAI Eval Harness: A Hands-On Guide
Evaluate open LLMs with the EleutherAI lm-evaluation-harness: run standard benchmarks, compare models fairly, and read the results like a practitioner.
Read Full News
Cisco Puts Claude and OpenAI Agents Inside Webex Meetings
Cisco is putting AI agents from both Anthropic and OpenAI inside Webex — so the agents come to your meeting instead of making you go to them. Plus RoomOS 27 with Google Meet support in November and a customer-experience harness called Dialog.
Nvidia China AI Chip Market Share Falls to 34% as Huawei Closes the Gap, IDC Finds
Nvidia shipped 678,000 AI accelerators in China in the first half of 2026, down 44% from a year earlier, as Huawei and a wave of domestic chipmakers filled the space left by US export controls. New IDC numbers seen by Reuters show the world’s second-largest AI chip market being remade.
“Don’t Drop Dropout”: The Case for Layer Sparsity in Efficient LLM Training
A Cerebras paper argues layer dropout belongs in state-of-the-art LLM training: up to 25% training FLOP savings and 1.5× faster inference, 2,400 experiments.
OpenAI Dots Agent Explained: The Always-On Assistant That Works While You Sleep
OpenAI’s Dots is an always-on AI agent that keeps working after you close the laptop — and it’s racing Meta’s Muse in a new privacy battle. Here’s what Dots actually does, what it costs, and whether it’s worth your attention.
Anthropic’s Threat Report: How State-Backed Hackers Turned Claude Into a Cyber Weapon
Anthropic’s Sept 2026 threat report: 190M stolen model exchanges, self-rebuilding malware, and state hackers wielding Claude as a cyber weapon.
Gemma 4: Google’s Open Models Get Smaller, Faster, and Multimodal
Google’s Gemma 4 family spans 2.3B to 31B parameters, with an encoder-free 12B that reads raw audio — all Apache 2.0 licensed and phone-friendly.
Trump’s ‘Super Intelligence Force’: The New White House AI Power Center, Explained
Trump’s Super Intelligence Force, announced Oct 4 and chaired by DNI Jay Clayton, coordinates federal AI policy. Here’s what it is — and isn’t.
The ‘Threat-Model Coverage Gap’: Why Automated Red-Teaming Isn’t Enough
A new paper argues automated red-teaming beating humans on cost and count doesn’t retire humans — benchmarks measure search thoroughness, not harm coverage.
Ornith-1.5: The Open-Weights Model That Trains Itself
DeepReinforce’s Ornith-1.5 ships MIT-licensed open weights in 9B, 35B, and 397B sizes — trained in a loop where the model generates its own tasks.