Hello, Starters!

Once again, keeping track of the most recent models becomes a bit difficult as the AI frenzy grows, yet here we take our time to explain the most relevant ones to you. Just sit down and keep reading!

Here’s what you’ll find today:

• Prometheus 2: A model for evaluation

• Microsoft AI is reportedly working on MAI-1

• Nvidia introduces DrEureka to improve robot training

• OpenAI & Stack Overflow team up

• Apple’s highway to AI with the “ACDC” project

• And more.

🔥 Prometheus 2: A model for evaluation (1 min)

Developed by KAIST AI, Prometheus 2 is an open-source language model trained on a dataset of over 1,000 evaluation criteria. It seamlessly optimises model evaluation while promoting accessibility in AI development.

Many developers and researchers evaluate the performance of their models using GPT-4, which often leads to biases and lacks affordability. With Prometheus 2, both code and data are available on GitHub, allowing users to test their models through various methods such as direct evaluation, scoring scales, pairwise comparison, and user-defined criteria.

🔎 Microsoft AI is reportedly working on MAI-1 (1 min)

Sources close to Microsoft have announced that the company is currently working on a large language model called MAI-1, which is being developed under the guidance of Mustafa Suleyman, CEO of Microsoft AI. This showcases a new approach for Microsoft, which, until now, has relied on OpenAI models for most of its work.

Although there's not much detail on how the model will be used, it's known that it boasts 500 million parameters. Depending on its progress, it may be announced at Microsoft's Build developer conference this month.

🤖 Nvidia introduces DrEureka to improve robot training (3 min)

A joint effort between Nvidia, the University of Pennsylvania, and UT Austin has resulted in DrEureka, a groundbreaking method that automates the transfer of robot skills learned in simulation to reality using language models.

This process, known as sim-to-real transfer, typically requires a considerable amount of manual work. However, DrEureka only utilises physics simulation for the target task to generate appropriate functions and configurations, allowing robots to adapt and perform well in real-world situations (domain randomization). It represents a significant advancement in robotics and AI.

🤝OpenAI and the programming forum for developers, Stack Overflow, are collaborating to improve their AI offerings. From OpenAI's side, they'll use the data from the forum to enhance its models' programming-related responses, and from Stack Overflow, they'll use OpenAI's expertise for new AI integrations on its platform.

🛣️Continuing with their hinting of their work on AI, reports announce that Apple is secretly working on a project to develop their own AI chips for data centres, codenamed "Project ACDC." Apple is not unfamiliar with chip-making, so these "Apple Silicon" chips are likely set to optimise most of its AI applications.

⚡️Quick links

Google’s AI plans now include cybersecurity

U.S. had most newly funded AI startups in the past decade

Amazon’s AWS to double down on Singapore with additional $9 billion investment

Gen AI innovation race is leading to security gaps, according to IBM and AWS

AI engineers report burnout and rushed rollouts as “rat race” to stay competitive hits tech industry

Thank you for reading!