Google’s New AI Can Now Think at Superhuman Level (Scary Fast)
AI Summary
Summary of Google’s New AI Developments
- Mixture of Depths (M-MoD)
- M-MoD is a method that dynamically allocates computational resources in transformer-based language models.
- It identifies key tokens that require more attention and allows less important tokens to bypass expensive computations.
- A per-token router assigns weights to tokens, and only the top K tokens undergo full processing.
- This approach reduces the number of floating-point operations (FLOPs) and maintains a static computation graph.
- M-MoD can reduce memory usage and speed up post-training sampling by up to 50%.
- An auxiliary predictor is used for autoregressive sampling to approximate top K decisions causally.
- When combined with Mixture of Experts (MoE), M-MoD can skip blocks for some tokens and route them among different MLP experts, improving efficiency.
- The model can outperform baselines with fewer FLOPs and can be more efficient with aggressive capacity lowering, as long as full capacity layers are interleaved.
- VO2 AI Video Generation Model
- Google’s VO2 AI video generation model is priced at 30 per minute or $1,800 per hour.
- It is targeted at professionals and businesses for creating sophisticated short video content.
- Freepic offers early public access to VO2, with the first 10,000 users getting two free generations, then charging in credits (1,000 credits for a 5-second video).
- AI Co-Scientist System
- The AI co-scientist system is designed to assist research scientists in forming hypotheses, proposing experiments, and producing research overviews.
- It uses a multi-agent architecture with various specialized agents and a supervisor agent for coordination.
- The system employs an ELO rating process to rank hypotheses, improving over time.
- It has demonstrated the ability to solve complex scientific problems, such as antibiotic resistance in bacteria, in a short time frame.
- The AI co-scientist system has shown higher novelty and impact ratings compared to other AI approaches.
- Google’s AI Video Creation Features Testing
- Google is testing new AI video creation features within its app, codenamed “Robin.”
- The features may allow users to generate videos from text prompts.
- Speculation exists about integrating these features into consumer-friendly formats and Google’s business-centric tools.
Conclusion
Google’s latest AI breakthroughs include advancements in efficient language processing with M-MoD, a premium-priced AI video generation model VO2, an AI co-scientist system that accelerates scientific research, and potential new consumer-friendly AI video creation features under testing.