AI Weekly Update

Overview

• Tracking significant AI releases and news.
• Focus on practical applications usable by the public.

Moshi AI by Cute AI Labs

• Open-source real-time voice assistant.
• French-based, aims to compete with OpenAI's GPT-4 voice assistant.
• Features:
  • Low-latency web interface.
  • 7 billion parameters in base model.
  • Emotional awareness and tone modification (performance varies).
• Issues encountered during testing:
  • Interrupts user frequently.
  • Limited emotional detection and voice modulation capabilities.
• Expected to be integrated into many applications due to open-source nature.

Gen-3 Video Generator

• State-of-the-art AI video generator now widely available.
• Previously showcased by Andrej Karpathy.
• Examples and capabilities:
  • Various scenarios like a T-Rex surfing or a medieval village in the style of Hieronymus Bosch.
  • High-quality, immersive video generation.
• Cost concerns:
  • Uses credits; expensive ($1 per 10-second clip).
  • Requires multiple iterations for desired output, increasing cost.
• Real-World Use Case:
  • Motorola’s AI video tools in their ad campaign.

11 Labs Reader App

• iOS app (US only, future global rollout planned) for text-to-speech using AI voices.
• Features iconic voices like James Dean and Burt Reynolds.
• Introduced AI tool to isolate voices, enhancing audio quality.

Suna Mobile App

• iOS app (US only, future Android version and global rollout planned) for AI music generation.

Luma AI Key Frames

• New feature allowing smooth transitions in AI video from one object to another.
• Testing showed variable success; requires many iterations to avoid hard cuts.
• Potential limited by necessity for frequent regeneration to achieve desired effect.

Perplexity AI Pro Search

• New feature with multi-step reasoning and access to math, programming, and Wolfram Alpha.
• Aimed at enhancing agent-based search processes.

Other Interesting Developments

WebSIM AI

• Creates websites from scratch using Claude 3.5.
• Fun application example: recreating interdimensional TV from Rick and Morty.

AI Multimodal Model: Dolphin Vision 72B

• Largest multimodal, uncensored AI model to date.
• Requires significant computing power to run.

Figma AI Features

• New AI-powered tools for UI design, including prompt-to-UI and visual search.
• Recent controversy over copyright issues with the generated designs.
• Features disabled temporarily for ethical review.

Hugging Face Leaderboard Update

• Introduction of new benchmarks (MLU pro, GPT QA, MSU).
• Community voting system added for result validation.
• Quen 72B model leads current rankings.

Noteworthy Mentions

• Google’s AI-powered crossword game.
• New 11 Labs audio isolation feature.
• Motorola ad campaign using AI video tools.

Conclusion

• AI advancements are rapidly evolving, with practical applications becoming more accessible.
• Enthusiasts encouraged to explore and test new tools and features.
• Stay updated with weekly AI news for more insights and developments.