📆 26th July 2024 📆 

This is one to bookmark with lots of new model & service announcements across chat, video gen & the next stage in AI infrastructure. For those of you who like a quick read I’m sorry but this week jammed so much in we just have to go a little long form.

• Nvidia announce support to easily to train and deploy fine tuned LLMs. Useful for those in regulated industries of with high levels of security.

• A new study shows that re-training on generated content can lead to reduced quality of new LLMs - potentially leading to model collapse.

• Musk hits the button on an impressively large computing cluster to train the next big model. Touts December as the point were training will be complete

• Four major model announcements from Meta, Google, Mistral & StabilityAI

• As always, too much to put in headlines. Check out the Other News Rollup section for a quickfire review of the other top stories in AI.

Dive in for all the details!

🆘 Quick Tips 🆘 

Stop Googling and spending precious time trawling through search results and page links. Instead, use PerplexityAI.

Perplexity touts itself as an answer engine (as recently mention on the Lex Friedman podcast) and that is exactly what it gives you. By using AI to understand your question and collate the results into a single output you get a direct answer while still seeing all the links to the content that drove it. It then uses AI to suggest a set of likely follow-up questions, leading to more focused discovery beyond your original question.

And it’s free to use today!!!

Have an AI idea and want to learn how feasible it is? Contact Me for a free discovery call, no matter your businesses size I can help you unlock AI to propel your business.

💫Help Support my Work💫

I really enjoy researching and writing about AI & appreciate your support. If you enjoy this content please use any of the options below to help

• Buy Me a Coffee ☕️- AI is really fuelled by coffee - keep the tank filled

• Spread the Word 🗞️- Share on social or directly with your friends

• Follow me on X - Always happy to chat AI and software dev

📰 News 📰 

Elon Musk Launches Ambitious AI Training Project

Elon Musk has initiated training for what he claims is 'the world's most powerful AI' at xAI's new supercomputer facility in Memphis. The facility, featuring 100,000 liquid-cooled GPUs, aims to achieve significant advancements in AI by December. However, concerns about its substantial power requirements, potentially consuming up to 150 megawatts, have arisen, prompting discussions with local utility providers. To mitigate power issues, Musk is installing natural gas generators while seeking a contract with the Tennessee Valley Authority. The project represents a historic investment for Memphis, but its impact on local infrastructure remains uncertain.

Read More At: Ars Technica

Nvidia Launches AI Foundry for Custom Models

Nvidia has introduced its AI Foundry service, enabling businesses to create tailored large language models. This initiative aims to enhance enterprise AI adoption by providing access to pre-trained models and high-performance computing resources. The service allows companies to customize open-source models, improving accuracy significantly. Additionally, Nvidia's NIM (Nvidia Inference Microservices) facilitates easy deployment of these models. This strategic move positions Nvidia as a comprehensive AI solutions provider, addressing the growing demand for customized AI while navigating concerns about safety and governance in the industry.

Read More At: VentureBeat

AI Models Risk Collapse from Overuse

A study from the University of Oxford warns that AI models may collapse due to recursive training on AI-generated content. Researchers found that models trained on their own outputs can produce nonsensical results, a phenomenon termed 'model collapse.' This issue arises when models fail to capture less common data, leading to a cycle of over-representation of certain outputs. The study emphasizes the need for effective data filtering and community coordination to prevent this collapse, ensuring AI-generated content remains meaningful and diverse.

Read More At: The Register

🧱 Model Releases 🧱 

Stream of New AI Models Released This Week

I don’t know if there was a week with this many concurrent model announcements. This may be a longer read but it’s worth putting it out there. Here’s a look at the latest announcements:

Meta Unveils Largest Open AI Model Yet Meta has launched Llama 3.1 405B, their most advanced large language model to date. Trained on over 15 trillion tokens with 16,000 Nvidia GPUs, this model features 405 billion parameters and a 128,000 token context window. Meta emphasizes its open AI development approach, encouraging developers to utilize Llama’s outputs to enhance other models. The model is available for download, with Meta seeking community feedback for future improvements. Read More At: The Register

Mistral Large 2: A New AI Contender Mistral AI has released Mistral Large 2, a 123-billion-parameter language model that competes with top-tier models from OpenAI and Meta. It supports a 128,000 token context window and multiple languages, achieving an impressive 84% on the MMLU test. Mistral highlights ML2's efficiency, offering faster responses and reduced operational costs, though it requires a commercial license for business use. This model aims to enhance accuracy and reduce hallucinations in AI responses. Read More At: The Register

Gemini's Major Upgrade: Enhanced Features Unveiled Google's Gemini has undergone a major upgrade with the introduction of 1.5 Flash, delivering faster responses in over 40 languages. This update improves reasoning and image understanding, expanding access for teens and mobile users. New features include related content links for fact-checking and the ability to upload files for analysis. Gemini is now available in more regions, showcasing Google’s commitment to responsible AI development and user safety. Read More At: The Official Google Blog

Stability AI Launches Innovative Stable Video 4D Stability AI has unveiled Stable Video 4D, a pioneering generative AI model that creates videos from multiple perspectives. This technology synthesizes videos by generating eight novel-view videos from given input, enhancing applications in movie production, gaming, and AR/VR. Utilizing a unique network for video generation and novel view synthesis, it ensures better coherence and smoothness. Stable Video 4D is currently available for research on Hugging Face, with commercial options pending. Read More At: VentureBeat

🔨 AI Powered Tools 🔨 

• Claude - developed by Anthropic, is an advanced AI assistant designed to enhance productivity and collaboration. It leverages natural language processing to offer conversational support, helping users manage tasks, seek information, and streamline workflows securely and accurately. Claude is currently in open beta and is aimed at both individual users and teams to improve their work efficiency.

• Pictory - Transforms text and long-form content into engaging videos, targeting content marketers, social media managers, and educators. Key features include script-to-video conversion, blog-to-video transformation, and text-based video editing.

• Runway - Primarily targeting the media and entertainment industry, Runway ML supports high-fidelity video generation, text-to-image, and image-to-image synthesis. Its Gen-3 Alpha tool enables creators to produce videos from text prompts, while Gen-2 and Gen-1 tools offer image and video manipulation capabilities. These tools are designed to expand creative possibilities and streamline the production process for artists and filmmakers.

🗞️ Other News Rollup 🗞️ 

SearchGPT Launch - OpenAI introduces SearchGPT, an AI search engine aimed at competing with Google and enhancing user experience.

AI Math Breakthrough - DeepMind's AI systems achieved silver medal-level performance at the International Mathematical Olympiad, showcasing advanced mathematical reasoning capabilities.

Bing's New AI Feature - Microsoft introduces generative search in Bing, providing AI-generated summaries alongside traditional search results.

AI Video Scraping Controversy - Runway's AI video generator allegedly scrapes YouTube videos, raising copyright concerns and violating platform policies.

Google AI Cost Management - Google keeps AI costs flat while boosting engagement and revenue, despite mixed investor reactions.

Gaming Jobs - Generative AI is threatening job security in the video game industry, prompting fears and calls for unionization among workers.

AI for Government - HelloGov AI simplifies government document applications, reducing processing times and enhancing user experience through advanced AI technology.

Salesforce & Workday AI Alliance - Salesforce and Workday's partnership introduces AI solutions, enhancing enterprise software and workforce management capabilities.

AI Safety Innovations - OpenAI's new Rules-Based Rewards enhance AI safety by automating compliance checks, reducing human subjectivity.

On-Demand GPUs - Lambda offers on-demand access to Nvidia H100 GPUs, ideal for companies needing flexible, short-term model training.

Reddit Blocks AI Access - Reddit restricts AI search engines, impacting Bing and DuckDuckGo, while maintaining a deal with Google.

Deepfake Policy Update - Meta's Oversight Board calls for clearer rules on sexually explicit deepfakes amid rising concerns over online harassment.

🎶 Prompts 🎶 

Prompts used to generate some of this issues images. Unless otherwise stated all images are using Dall-E and ChatGPT Plus.

Thank You for Subscribing