Unveiling the Latest AI Models and Developments: Must-Watch AI News!

Table of Contents

1. Introduction
2. Wired commits to buying $551 million worth of AI chips from Open AI
3. Changes in the Open AI board
4. New release: Open GPT Vision
5. Introduction to Seamless models by Meta AI
6. Microsoft's paper on short-talking Avatar generation
7. NTU Singapore and Salesforce review of open source models
8. Scalable extraction of training data from production language models
9. Introduction to Meditron: a fine-tuned model for medical use
10. New release: Stable Diffusion XEL Turbo for image generation

Introduction

Welcome to One Little Coders' weekly AI News update! In this video, we have a lot of exciting updates on the latest AI models and developments. Let's dive right in!

Wired commits to buying $551 million worth of AI chips from Open AI

In a surprising turn of events, Wired has announced that they will be purchasing $551 million worth of AI chips from a startup backed by Open AI's CEO, Sam Altman. This news has caused some controversy due to potential conflicts of interest. Sam Altman previously stated that he has no equity in Open AI and is purely motivated by his passion for the field. However, with this sizable investment, it raises questions about the true intentions behind the partnership. To learn more about this topic, check out the article linked in the YouTube description.

Changes in the Open AI board

In another update related to Open AI, there have been significant changes in the initial board of the company. The CEO has returned, and new members have joined the board, including some well-known names like Brett Tor. However, what caught our attention is the removal of Ilia S. Waker, the chief scientist of Open AI. While the CEO, Sam Altman, expressed his admiration for Ilia in his letter, stating that he is a guiding light in the field, it is unclear why Ilia will no longer serve on the board. Despite this, discussions are ongoing about how Ilia can continue his work at Open AI. The situation raises questions about the dynamics and future direction of the organization.

New release: Open GPT Vision

Technium, one of the leading AI research groups, has released a groundbreaking new model called Open GPT Vision. This multimodal model, built upon 220,000 data points from GPT-3 and trained on various vision datasets, showcases impressive capabilities in vision language understanding and function calling. Initial reviews highlight its excellent performance in tasks such as Vision Plus function calling. As AI enthusiasts, we are eagerly looking forward to trying out this model and seeing how it compares to existing computer vision models like LAVANet.

Introduction to Seamless models by Meta AI

Meta AI has recently launched a suite of models known as Seamless models. These models focus on audio, speech, and related tasks. The lineup includes Seamless EXP, an expressive model designed to preserve the expressions and intricacies of speech during translation. Another model, Seamless Streaming, allows for real-time Text-to-Speech and Speech-to-Text capabilities. Furthermore, Seamless M4T V2 is a multilingual and multitask model, while the Seamless model combines all these capabilities into a single comprehensive model. The demos of these models showcase their impressive performance, and we are excited to see further reviews and evaluations of their capabilities.

Microsoft's paper on short-talking Avatar generation

Microsoft has recently published a research paper titled "GAAZero: Short-Talking Avatar Generation." The paper details a unique approach to create video-driven talking avatars that can mimic the speech and expressions of reference videos. This technology has far-reaching implications, blurring the boundaries between reality and virtual content. While its applications are numerous, it raises concerns about the potential misuse of such capabilities. Nonetheless, it is an impressive development that showcases the rapid progress in the field of AI-generated content.

NTU Singapore and Salesforce review of open source models

A joint effort by NTU Singapore and Salesforce has resulted in a comprehensive review of open source models that aim to catch up with ChatGPT. The paper provides valuable insights into various models' strengths and limitations. It is worth noting that comparing ChatGPT with other models may not yield accurate results, as each model serves specific purposes. The paper highlights that different models excel in specific areas, such as logical reasoning or agent-level benchmarks. For those seeking the most suitable model for a particular use case, this paper serves as a valuable resource.

Scalable extraction of training data from production language models

A newly published paper discusses the extraction of training data from Large Language Models. Often, these models rely on compressed knowledge bases to generate responses. Researchers have explored the possibility of accessing and extracting the training information stored within these models. This raises concerns about the security and privacy of such data. The paper presents a unique approach that leverages the limitations of large language models to extract sensitive information. The implications of this research are significant and highlight the need for robust privacy measures in the development and deployment of AI models.

Introduction to Meditron: a fine-tuned model for medical use

Meditron, a fine-tuned model based on Llama, has been released. This model is specifically tailored for medical applications, offering a promising solution for improving medical knowledge and accuracy. Google's MedPALM, another medical model, has proven to be useful in the field. However, the low level of trust in medical professionals in some regions calls for alternative sources of information. Meditron presents an opportunity to empower individuals with accurate medical knowledge, beyond traditional MBBS education. The open sourcing of the data used to fine-tune the model further enhances transparency and collaboration in the medical AI field.

New release: Stable Diffusion XEL Turbo for image generation

In the realm of image generation, a remarkable model has emerged: Stable Diffusion XEL Turbo. Unlike previous approaches that required iterating steps, this model can generate high-quality images in just a single step. With only four steps, it showcases impressive performance and realism. The combination of Stable Diffusion XEL Turbo with latent consistency models like LCM unlocks the potential for real-time image and video generation. The ease of running these models on consumer hardware adds to their accessibility and practicality. Although some may have reservations about the ethical implications of image generation, this advancement undoubtedly marks an exciting milestone in AI technology.

Highlights

• Wired's significant investment in AI chips from a startup backed by Open AI raises questions about potential conflicts of interest.
• The changes in the Open AI board introduce new members but also lead to the removal of the chief scientist, Ilia S. Waker, creating uncertainty about the organization's direction.
• Technium's release of Open GPT Vision, a multimodal model, promises enhanced capabilities in vision language understanding and function calling tasks.
• Meta AI's Seamless models offer innovative solutions for audio, speech, and translation tasks, showcasing impressive results.
• Microsoft's GAAZero paper introduces an impressive technology for creating short-talking avatars, blurring the lines between reality and virtual content.
• NTU Singapore and Salesforce's review of open source models provides valuable insights into their strengths and limitations, helping users choose the most suitable model for specific use cases.
• A paper on scalable extraction of training data from language models raises concerns about the security and privacy of sensitive information contained within these models.
• Meditron, a fine-tuned medical model, presents an opportunity to improve medical knowledge and empower individuals in the field.
• Stable Diffusion XEL Turbo revolutionizes image generation, allowing for high-quality images in a single step and real-time video generation.

FAQs

Q: What are Seamless models by Meta AI? A: Seamless models are a suite of models developed by Meta AI that focus on audio, speech, and translation tasks. They provide solutions for preserving expressions during translation, real-time text-to-speech and speech-to-text capabilities, and a comprehensive multitask model combining these functionalities.

Q: What is the significance of the changes in the Open AI board? A: The changes in the Open AI board, including the return of the CEO and the removal of the chief scientist, Ilia S. Waker, raise questions about the organization's dynamics and future direction.

Q: How does Stable Diffusion XEL Turbo revolutionize image generation? A: Stable Diffusion XEL Turbo enables high-quality image generation in a single step, unlike previous approaches that required multiple iterations. This advancement, combined with latent consistency models, opens up possibilities for real-time image and video generation.