AI news for 2 September 2026
Wednesday 2 September 2026 · Latentown
On Wednesday 2 September 2026, Latentown tracked 2 posts from 1 AI lab, 3 research papers. Google DeepMind was the busiest lab with 2 posts. The highest-ranked story was "Introducing Gemini 3.8 Flash and 3.8 Flash Cyber".
Top stories
- Introducing Gemini 3.8 Flash and 3.8 Flash Cyber
Two Flash models aimed at agentic workflows, including a Cyber variant built for security work.
- Proactive cyber defense for governments and enterprises
The Fairwind Program gives governments and trusted partners limited access to Google DeepMind cyber defense tools.
- StudentSim: Training LLM-based Student Simulators
Top paper on Hugging Face that day. Trains language models to simulate students, so AI tutors can be tested against realistic strengths and weaknesses.
- Qwen-Drive-1.0: An Initial Step towards a Vision-Language Foundation Model for Autonomous Driving
A vision-language foundation model for autonomous driving from the Qwen team, described as a first step.
- SMELT: Scaling Laws for Compute-Matched MoE Looped Transformers
Compares looped transformers, which reuse a shared block of layers, against standard models at matched compute rather than matched size.
From the AI labs
Google DeepMind
- Introducing Gemini 3.8 Flash and 3.8 Flash Cyber
Two Flash models aimed at agentic workflows, including a Cyber variant built for security work.
- Proactive cyber defense for governments and enterprises
The Fairwind Program gives governments and trusted partners limited access to Google DeepMind cyber defense tools.
Research papers
- StudentSim: Training LLM-based Student Simulators
Top paper on Hugging Face that day. Trains language models to simulate students, so AI tutors can be tested against realistic strengths and weaknesses.
- Qwen-Drive-1.0: An Initial Step towards a Vision-Language Foundation Model for Autonomous Driving
A vision-language foundation model for autonomous driving from the Qwen team, described as a first step.
- SMELT: Scaling Laws for Compute-Matched MoE Looped Transformers
Compares looped transformers, which reuse a shared block of layers, against standard models at matched compute rather than matched size.
Every story links to its original source. Ranking is by automatic rules, not an editor. Something wrong? Tell us.