Insane AI News You Missed This Week (Microsoft, OpenAI, Nvidia, Robots, Adobe, and more)

AI News Roundup

In this section, we cover the latest news in the field of AI from various companies and research papers.

OpenAI's Thoughts on Governance of Super Intelligence

• OpenAI team released a blog post discussing governance of super intelligence.

• They believe that within the next 10 years, AI systems will exceed expert skill level in most domains and carry out as much productive activity as one of today's largest companies.

• They lay out three ways they think governance for artificial intelligence should manifest: coordination between leading development efforts with AI, an international watchdog for everything AI-related, and technical capability to keep superintelligence safe through alignment with human goals.

• Some critics believe that OpenAI is engaging in regulatory capture by setting up hoops and red tape to go through with government. However, OpenAI made it clear that they want to allow companies and open-source projects to develop models below a significant capability threshold without burdensome mechanisms like licenses or audits.

Microsoft's Announcements

• Microsoft announced integrating AI into various products such as Chachi PT and Bing-powered browsing.

• Co-pilot is being brought to Windows which brings Chachi PT into the deepest layers of the Windows environment.

• Common plugins across all chat CPT platforms are now available which means Office 365 Windows chat GPT itself can be built once and work across all these different platforms.

Adobe's Latest Research Paper

• Adobe published a research paper on "GPT Understands, Too" where they explore how GPT models can be used for understanding natural language text beyond generating coherent responses.

• The paper proposes a new task called "Winograd Schemas" which requires commonsense reasoning about pronouns in ambiguous sentences.

Anthropomorphic Robotics Companies' Developments

• Anthropomorphic robotics companies are developing robots that can perform tasks such as cooking and cleaning.

• These robots use AI to learn from human demonstrations and improve their performance over time.

Research Papers on Prompts

• There have been several research papers published recently that explore the use of prompts in AI models.

• One paper proposes a new method for generating diverse responses using a small set of prompts, while another paper explores how to generate more informative responses by conditioning on additional information.

Introduction

The speaker introduces the topics that will be covered in the video.

Topics Covered

• The announcement of Microsoft Azure's new AI studio

• Adobe Photoshop's generative art tool, Firefly

• Drag Gan, an interactive point-based manipulation on the generative image manifold

• Nvidia and Intel's announcements regarding their HPC and AI products

• Two research papers published this week: Tree of Thoughts and Lima

Microsoft Azure AI Studio

The speaker discusses the announcement of Microsoft Azure's new AI studio.

Key Points

• Developers can now ground powerful conversational AI models such as OpenAI's chat GPT and GBT4 on their own data.

• Fine-tuning is going to be easily done in Azure.

Adobe Photoshop Generative Art Tool - Firefly

The speaker talks about Adobe Photoshop's generative art tool, Firefly.

Key Points

• Firefly is coming out of beta and it's getting put directly in Photoshop.

• Users can add generative art to any image or remove objects they don't want.

• A quick video showcases some of the things users can do with Firefly.

Drag Gan - Interactive Point-Based Manipulation on Generative Image Manifold

The speaker discusses Drag Gan, an interactive point-based manipulation on the generative image manifold.

Key Points

• With just a few clicks, users can completely manipulate what an image looks like.

• Users can change an image's angle or what objects are present in it.

• Code base for this is not released yet but it will be soon.

Nvidia and Intel Announcements Regarding HPC and AI Products

The speaker talks about Nvidia and Intel's announcements regarding their HPC and AI products.

Key Points

• Nvidia's stock price skyrocketed for a company whose stock price is already up tremendously this year.

• Intel called out Nvidia directly, stating that their broad portfolio of HPC and AI products provides competitive performance with Intel data center GPU Max series 1550 showing an average speed up of 30 percent over Nvidia's h100 on a wide range of scientific workloads.

• Intel announced plans to release generative AI models specifically for science.

Research Papers Published This Week

The speaker discusses two research papers published this week: Tree of Thoughts and Lima.

Key Points

• Tree of Thoughts is a model that gives large language models the ability to do planning, look forward, backtrack when they make mistakes, and test different possibilities for the same prompt. The tree of thoughts technique allows large language models to lay out all of the options for a potential solution, go through them one by one, rank them, and go down the tree into branches trying to figure out what is the best possible path to a solution. They were able to achieve a 900 Improvement on logic and reasoning problems.

• Lima is a paper that takes the 65 billion parameter llama model open-source model and trains it on only 1,000 fine-tuned instructions prompts and responses to achieve results on GPT3.

Anthropics Raises Funds for AI Language Model

In this section, the speaker discusses the recent funding raised by Anthropics, a company behind Claude, a large language model that is hosted and has a 100,000 token context length.

Anthropics Funding Details

• Anthropics recently raised funds from Spark Capital, Google, Salesforce, Sound Ventures, Zoom Ventures and others.

• The speaker believes that competition in the AI space is beneficial whether it's host or open source.