• Starter AI
  • Posts
  • Algorithms for maths, Runway introduces Gen-3 Alpha, Apple’s Core ML

Algorithms for maths, Runway introduces Gen-3 Alpha, Apple’s Core ML

Apple goes open source!

Hello, Starters!

Slow but steady efforts lead to great developments, and that has been Apple's approach to most of its work for a while. AI will not make a difference! We are eager to see what they are planning for us.

Here’s what you’ll find today:

  • Improving models' maths skills

  • Runway releases Gen-3 Alpha

  • Apple’s new Core ML models

  • Researchers introduce “LiveBench”

  • DeepMind’s video-to-audio efforts

  • And more.

In the current AI scene, many researchers aim to improve models' maths skills, a task even advanced LLMs like GPT-4 still struggle with. A study from the Shanghai Artificial Intelligence Laboratory proposes an interesting approach to this challenge by integrating the Monte Carlo Tree Search (MCTS) algorithm with LLMs, already showing positive results.

The MCTS algorithm is a decision-making tool often used in AI for strategic planning, such as in games or complex problem-solving, and it has been employed in systems like AlphaGo and AlphaZero. Combining the algorithm with LLMs could help improve how models address logical and strategic tasks.

Competitiveness is strengthening in video generation, and lately, many AI companies have shown the results of their developments. Runway is included in this list with the unveiling of Gen-3 Alpha, its new AI model that signals a significant improvement from the former Gen-2, surpassing its performance and enhancing details, consistency, and motion representation.

Besides text-to-video and image-to-video functions, Gen-3 Alpha allows users to create unique content with advanced tools such as camera controls, a motion brush, and even a director mode. Runway stated that they intend to release more features soon.

A lot has been said about Apple's approach to AI, yet there's an incredible effort the company has been making that may not be noticeable if you're not in touch with sites like Hugging Face. Apple is focusing on supporting developers in the open-source community by uploading 20 new Core ML models and 4 datasets to the AI platform.

These models are optimised for on-device performance, leveraging Apple Silicon, and reducing memory and power consumption. This allows developers to craft AI applications with models mostly focused on text and images, which aligns with the company's approach to "practical" AI.

🏅A group that includes experts from Abacus.AI, New York University, Nvidia, the University of Southern California, the University of Maryland, and AI pioneer Yann LeCun, has introduced LiveBench, a groundbreaking benchmark for LLMs. LiveBench defies test data contamination by constantly updating its questions, using only recent material, scoring answers automatically, and presenting challenging tasks.

🎥Google DeepMind is also making efforts in video generation. Recently, they presented advancements in their video-to-audio (V2A) technology, which synchronises video output with soundtracks that align with given text prompts, something regular video generation models do not yet do. As a result, videos created with tools like their Veo model can include realistic sound effects that match the footage, giving users more creative power.

What did you think of today's newsletter?

Login or Subscribe to participate in polls.

Thank you for reading!