中文 EN

Bubble's Brain - 2025-12-08

Daily Brief

AI News 2025-12-08

AI Daily Brief

Summary

Google Gemini 3 can now generate real-time interactive 3D particle systems from plain text, and its CEO predicts AGI may arrive before 2030.
Google also introduced the Titans architecture for ultra-long context, making progress on long-term memory and continual learning.
Separately, the Qwen team published a new view on why RL training for LLMs can be unstable and proposed methods to stabilize training, especially for MoE models.

Today’s AI News

  1. Gemini 3 can generate interactive 3D particle systems from text, no coding required: Google Gemini 3 now supports generating real-time interactive 3D particle systems from simple text prompts, without programming. Users can even control particle effects with hand gestures captured by a camera. The article contrasts two Gemini tools: Gemini Canvas, an integrated real-time rendering environment aimed at quickly producing interactive deliverables for non-developers; and Google AI Studio, positioned as a developer “arsenal,” supporting up to 2M tokens of context and system-level instructions for building and debugging applications.

  2. DeepMind CEO Demis Hassabis: AGI could arrive before 2030; Google unveils “Titans” architecture: Demis Hassabis predicted AGI may be achieved before 2030, but said it requires one or two “Transformer-level” breakthroughs. Meanwhile, at NeurIPS 2025, Google introduced Titans, a new AI architecture seen as a potential successor to Transformers. Titans combines the fast response of RNNs with the power of Transformers and leverages the MIRAS theoretical framework to handle >2M tokens of ultra-long context, with key advances in long-term memory and continual learning.

  3. Microsoft open-sources VibeVoice, a voice AI project: Microsoft open-sourced a cutting-edge voice AI project called VibeVoice on GitHub, which has already gained 12,000+ stars.

  4. ai-engineering-hub: deep tutorials on LLMs, RAG, and real-world AI agents: ai-engineering-hub is a GitHub project providing in-depth tutorials on LLMs, RAG, and real-world AI agent applications, currently with 21,762 stars.

  5. claude-quickstarts: Anthropic’s quickstart repository collection for Claude API: claude-quickstarts is Anthropic’s set of quickstart repositories to help developers get started with the Claude API and build deployable apps, currently with 11,142 stars.

  6. Qwen team explains why RL for LLMs can be unstable and proposes stabilization techniques: The Qwen team published a paper offering a new “first-order approximation” view of instability in LLM reinforcement learning (RL). The work argues that token-level objectives can be seen as a first-order approximation of expected sequence-level rewards, and that the approximation’s validity depends on numerical differences between training and inference and on the magnitude of policy updates. This perspective unifies why techniques like importance sampling and clipping stabilize training. For Mixture-of-Experts (MoE) models, the team proposed Routing Replay to freeze expert routing and improve stability. Extensive experiments on a 30B MoE model found that on-policy training is most stable with importance-sampling correction; off-policy training requires both clipping and routing replay to avoid collapse. The paper also reports that once training is stable, different cold-start methods converge to similar final performance—suggesting future work should focus more on RL methods than on cold-start details.

Comments