All articles
Tsinghua's SEED: self-evolving distillation teaches AI agents
A Tsinghua team introduced SEED — self-evolving on-policy distillation with hindsight supervision. On ALFWorld it improves success by 45.9 points over GRPO and matches it using 60% of the data.
PolicyTrim speeds up VLA robots up to 5.83× without retraining
A Sichuan University team introduced PolicyTrim — a two-stage RL framework that cuts VLA robot task time by up to 5.83× in simulation and 1.86× on a real robot, with no training from scratch.
Google ships three new Gemini models — but no 3.5 Pro
Google released Gemini 3.6 Flash, 3.5 Flash-Lite and the specialized 3.5 Flash Cyber. The flagship 3.5 Pro is missing — Bloomberg reports it fell short of internal benchmark targets. Gemini 4 pre-training is already underway.
GPT-Red: OpenAI trained an AI to break other AI models
OpenAI trained GPT-Red — an AI model specialized in breaking other models via prompt injection. GPT-Red achieved 84% attack success versus 13% for human red-teamers and was used to harden GPT-5.6 Sol to a 0.05% attack success rate.
China's free AI models split Trump's AI camp
Free, open Chinese models — chiefly Moonshot's Kimi — have triggered an open rift inside the Trump administration's tech circle over whether to back open source or curb it on national-security grounds.
Moonshot AI's Kimi K3: 2.8T parameters and open-weight release by end of July
Moonshot AI announced Kimi K3 — a 2.8 trillion parameter model that self-reported benchmarks place above Claude Opus 4.8 and GPT-5.5. The largest Chinese open-weight model to date, with weights promised by July 27, 2026.
Google AI Mode gains app integrations: Instacart, Canva and YouTube at launch
Google AI Mode expanded with direct app integrations — Instacart, Canva, and YouTube at launch. Users can now complete tasks without leaving the AI Mode interface. Available in the US.
Google renames NotebookLM to Gemini Notebook
Google has renamed NotebookLM to Gemini Notebook, deepening its integration with the Gemini ecosystem and Search. Every notebook will also gain a secure cloud computer for running code natively — currently available to AI Ultra and Workspace customers.
The AI agent evaluation gap: half of enterprises shipped agents that failed customers
VentureBeat Research survey of 157 enterprises reveals: 50% have shipped an AI agent that passed evals and failed in production. Only 5% fully trust automated evaluation, while 66% are moving toward zero-human-in-the-loop deployments.
Phantom Twist: a drone nearly invisible to the human eye
Northwestern University researchers presented Phantom Twist at RSS 2026 — it spins at 15–25 Hz, making it over 10x harder to spot than a standard quadrotor. Shape algorithmically optimized using the LPIPS perceptual metric.
Emergent from Bengaluru turns unicorn: $130M and 200K paying customers
Indian AI coding startup Emergent closed a $130 million Series C at a $1.5 billion valuation — five times higher than six months ago. The company targets small businesses as an engineering team in a box and grew 70% in four months.
Vint Cerf is designing AI agent identity — a DNS-based standard
Internet architect Vint Cerf has joined Innovation Labs to help build a standard for identifying AI agents across the open internet. The DNSid proposal links each agent to a domain name using cryptographic proofs.