• Starter AI
  • Posts
  • Enhancing models’ legibility, DeepMind’s PEER architecture, Introducing: Instruct-MusicGen

Enhancing models’ legibility, DeepMind’s PEER architecture, Introducing: Instruct-MusicGen

AI models learn to explain themselves better.

In partnership with

Hello, Starters!

One of the most interesting parts of being in the midst of AI development is the abundance of research being done. Each of them is a step further into advancing this technology, and we're here watching it happen from the front row!

Here’s what you’ll find today:

  • OpenAI shares its study on “Prover-Verified Games”

  • Google DeepMind introduces the PEER architecture

  • Instruct-MusicGen: An AI music editor

  • Salesforce presents its Einstein Service Agent

  • DeepL unveils a new LLM

  • And more.

OpenAI has recently shared a paper where, in the form of a "game," it uncovers a new algorithm to help models generate understandable outputs. Legibility is the main goal of the "prover-verifier" game, where an advanced model creates text that is easy to validate by a weaker one.

As the research explains, accuracy is important, but legibility also plays a significant role in the development of effective AI applications. If a model is only optimised to give correct answers, it may reach a point where it becomes difficult to understand, obstructing the work of evaluators.

Researchers at Google DeepMind are developing a new AI architecture that might dethrone transformers. They've introduced what they call the "Parameter Efficient Expert Retrieval" (PEER) technique, which draws inspiration from the well-known "Mixture of Experts", just that in this case, PEER focuses on more than a million small "experts" to enhance the performance and scalability of language models.

Results of the testing process are surprising, as the PEER technique was able to outperform most conventional transformer models, and those with the MoE approach, while using the same computing power.

By modifying the architecture of Meta's MusicGen AI models, researchers from Queen Mary University and Sony AI have developed a new prototype called "Instruct-MusicGen," which is capable of changing a song completely through the use of text instructions.

This happens thanks to the integration of text and audio fusion modules, which allows the model to work simultaneously with audio input and prompts, so users can easily add, remove, or separate music tracks, resulting in a valuable tool with the potential to transform the music production field.

Learn AI in 5 Minutes a Day

AI Tool Report is one of the fastest-growing and most respected newsletters in the world, with over 550,000 readers from companies like OpenAI, Nvidia, Meta, Microsoft, and more.

Our research team spends hundreds of hours a week summarizing the latest news, and finding you the best opportunities to save time and earn more using AI.

🛒Agents have been gaining traction in the AI field for a while now, and Salesforce is joining the club with its new Einstein Service Agent, a system based on its Einstein platform that focuses on customer service. The agent, besides answering customer queries, can also take action on their behalf, processing refunds or returns.

🌐DeepL, one of the leading translation platforms, is taking its efforts on AI to the next level by launching an LLM optimised for translation and editing. What sets this language model apart is that it has been trained on DeepL's proprietary data, and its fine-tuning process makes it surpass competitors such as Google and GPT-4 in translation quality.

What did you think of today's newsletter?

Login or Subscribe to participate in polls.

Thank you for reading!