01Tencent Releases Hy3: 295B MoE Model Rivals Flagship Open-Source Models [AI Models & Research]Tencent's Hy3 is a 295B-parameter Mixture-of-Experts model with 21B active parameters, Apache 2.0 licensed, rivaling open-source flagships with 2-5x the parameters. → source
02LLM-as-a-Verifier: Verification as a New Scaling Axis [AI Models & Research]Stanford researchers propose verification as a new scaling axis for LLMs, with verifiers trained alongside generators yielding compounding gains. → source
03Weak-to-Strong Generalization via Direct On-Policy Distillation [AI Models & Research]ByteDance and Tsinghua show that distilling weak model on-policy rollouts into stronger models cuts post-training cost while preserving reasoning gains. → source
04Harness Engineering for Self-Improvement [AI Models & Research]Lilian Weng maps design patterns for recursive self-improvement through harness engineering, reframing harnesses as the OS layer for AI. → source
05LeRobot v0.6.0: Imagine, Evaluate, Improve [AI Tools & Ecosystem]HuggingFace's LeRobot v0.6 brings simulation-first training, new benchmarks including Robometer-4B, and NVIDIA GR00T support. → source
06Google Now Trains AI on Your Media by Default [AI Tools & Ecosystem]A quiet June update opts users into AI training on images, audio, and video across Google Search, Maps, Translate, and Lens. → source
07Vercel's Guillermo Rauch on Why Models and Agents Must Split [AI Tools & Ecosystem]Vercel's CEO argues models should become infrastructure while agents become the product, with auditing and tool-call trails as the real hard problems. → source
08The Complete Guide to Tool Selection in AI Agents [AI Tools & Ecosystem]Six practical techniques for keeping tool selection accurate at scale as agent tool catalogs grow, without needing a bigger model. → source
09"First" AI-Run Ransomware Attack Still Needed a Human [AI Applications & Industry]Sysdig's JadePuffer case shows AI handled ransomware execution, but a human chose the target, set up infra, and provided stolen credentials. → source
10Microsoft Cuts ~4,800 Jobs Across Xbox and Commercial Sales [AI Applications & Industry]Microsoft laid off 2.1% of its workforce, citing transformation in how technology is built, deployed, and used. → source
11Your Family's $300 Stake in OpenAI [AI Applications & Industry]MIT Tech Review analyzes Sam Altman's equity-for-all proposal, finding each household would get about $300 worth at current valuations. → source
12Reddit Uses LLMs to Fight the Spam That LLMs Created [AI Applications & Industry]Reddit blocks 23M spam views/day using LLMs to catch the same technology's output, claiming 20% reduction in spam exposure. → source
Get the digest delivered
AI intelligence, curated daily by autonomous agents. Free, no spam, unsubscribe anytime.