'EMO': Cutting MoE Experts to One-Eighth Barely Changes Performance

A research team from the Allen Institute for AI (Ai2) and UC Berkeley has taken a major step forward in making Mixture-of-Experts (MoE) models more efficient. Their newly announced 'EMO' is a novel architecture that specializes experts by content domain rather than by token type.

In conventional MoE models, the decision of which experts to activate for each token is made based on the token's own characteristics. EMO switches this to a 'domain-based' approach, assigning experts to content areas such as science, coding, and everyday conversation. With this design, even after removing three-quarters of the experts, the performance drop is reportedly only about one point.

As the first practical approach to actually running MoE models in memory-constrained environments, EMO is expected to enable applications on edge devices and on-device inference. The details of the paper still await peer review, but as a research direction, this looks like a highly promising result.

The '10,000 Millionaires' the AI Boom Created in Silicon Valley — and the Alienation of Everyone Else

According to Menlo Ventures partner Deedy Das, the AI boom has made roughly 10,000 people in Silicon Valley worth $20 million or more. The primary beneficiaries are employees and early investors of Anthropic, OpenAI, xAI, Meta, and Nvidia.

Behind the scenes, however, the vast majority reportedly feel left behind. The middle-management layer has hollowed out, and even among the winners, some voice struggles with a 'deep lack of purpose.' Wealth concentration driven by AI appears to be increasingly recognized as a new social problem, even inside the tech industry itself.

Frontier AI Has Broken the Premise of Security CTF Competitions

Frontier AI models are bringing fundamental change to CTF (Capture The Flag), the flagship competition of security education, and the debate is heating up. An article that gathered 88 points on HN argues that state-of-the-art AI models can now automatically solve CTF challenges, and that the traditional premise of 'competing on human skills' is collapsing.

For years, CTF has played an important role in training and evaluating security talent. But once AI can solve nearly every challenge, the very meaning of the competition comes into question. We may now need to rethink how security education itself is designed.

An Experiment Passing Off a Real Monet as 'AI-Generated' to Collect Critiques

According to PetaPixel, someone posted a genuine Claude Monet painting to an online community, claiming it was 'generated with AI,' and asked for critiques. The experiment highlights the strange situations that excessive suspicion of AI-generated content can create.

The posted work is a genuine Impressionist masterpiece, but once it was labeled as 'AI-made,' many people reportedly set out to find its flaws. It makes for a fascinating case study in the limits of human ability to judge whether something looks 'AI-generated.' It is also a sobering reminder of the accuracy of AI-detection tools and of how preconceptions shape our evaluation of content.

Qwen3.6-35B-A3B Beats Gemini 2.5 Pro on Terminal-Bench 2.0

Qwen3.6-35B-A3B scored 24.6% on the public benchmark Terminal-Bench 2.0, surpassing Gemini 2.5 Pro's 19.6% and Qwen3-Coder-480B's 23.9%. Despite its small size, combined with a scaffold (little-coder), it delivers results that outperform much larger models.

Meanwhile, Qwen3.5-9B posted a modest 9.2%, but the community sees real significance in the fact that a local model under 10B recorded a measurable score on such a demanding agentic benchmark at all. It is a sign that open-source local models are rapidly gaining practicality.