Daily AI Newsletter: Andrej Karpathy’s startup aims to apply AI to education

In today’s email:

☠️ Apple, Nvidia, Anthropic Used Thousands of Swiped YouTube Videos to Train AI
👀Mistral releases Codestral Mamba for faster, longer code generation
👨🏻‍⚖️ Microsoft faces UK antitrust probe after hiring Inflection AI founders and employees
🧰 9 new AI-powered tools and resources. Make sure to check the online version for the full list of tools.

Andrej Karpathy’s startup aims to apply AI assistants to education

Andrej Karpathy, a notable figure in AI known for his roles at Tesla and OpenAI, has launched Eureka Labs, an AI-driven education platform. Based in San Francisco and registered as an LLC in Delaware, Eureka Labs aims to integrate AI teaching assistants into classrooms to aid human teachers in guiding students through course materials. Though the platform's initial product will be an AI course, LLM101n, designed to help students train their own AI, details on the efficacy of these AI assistants and the specifics of their classroom integration remain sparse.

Karpathy's announcement on X and the Eureka Labs website hints at a future where AI assistants mimic real personalities, similar to Meta’s and Character AI’s chatbots. However, the startup's immediate focus appears to be on developing educational courses that combine digital and physical cohorts. The first course, according to a linked GitHub repository, will teach students how to build a "Storyteller AI Large Language Model," promising to cover everything from the basics to creating a functioning web app similar to ChatGPT.

Details about Eureka Labs' funding, business model, and team composition are currently unclear. Karpathy has not disclosed whether he has received investment backing or is working with other prominent AI leaders. His motivation for founding Eureka Labs stems from his long-standing passion for AI and education, reflected in his past roles and ongoing educational efforts, such as his Neural Networks: Zero to Hero online course and his YouTube lectures on AI.

Google is rolling out access to Vids

Vids is an AI-powered video creation app designed for work and deeply integrated with the Workspace suite that users interact with daily. This new medium unlocks powerful video storytelling capabilities, allowing users to be among the first to explore and utilize Google Vids.

With Google Vids, users can build their narratives using high-quality templates or leverage Gemini in Vids to expedite the creation of a first draft. The app empowers creativity with a royalty-free stock content library and enables users to deliver their messages effectively using the Vids recording studio, all within the confines of Workspace. For more information, users can visit the help center or learn how to use Google Vids through the provided resources.

Apple, Nvidia, Anthropic Used Thousands of Swiped YouTube Videos to Train AI

A recent investigation revealed that Apple, among other tech giants, trained AI models on YouTube videos without obtaining consent from the creators. This was done by utilizing subtitle files downloaded from over 170,000 videos by a third party, EleutherAI, a non-profit organization that compiles datasets for AI training. Notable creators affected include Marquees Brownlee (MKBHD), MrBeast, and PewDiePie. The dataset, known as the Pile, was used not only by small developers and academics but also by large companies like Apple, Nvidia, and Salesforce to train AI models, including Apple’s OpenELM.

The use of these subtitle files, which are essentially transcripts of the videos, contravenes YouTube’s policies against unauthorized data harvesting. The situation underscores a significant legal grey area in copyright law, which has struggled to keep pace with technological advancements since the 1970s. While traditional copyright laws are clear on the unauthorized broadcasting of content, applying these principles to AI training on vast datasets is complex and contentious. The principle of copyright extending to derivative works, such as movies based on novels, becomes murky when applied to AI models trained on millions of words from various sources.

Although Apple did not directly download the data, its use of the dataset provided by EleutherAI places it in a contentious position. The case illustrates the broader challenges and legal uncertainties surrounding the use of scraped web data for AI training. Instances of AI systems generating plagiarized content highlight the risks of using unapproved materials. Apple and other involved companies may have acted in good faith, but this incident highlights the urgent need for updated regulations to address the ethical and legal implications of AI development.

Mistral releases Codestral Mamba for faster, longer code generation

Mistral, a prominent French AI startup, has unveiled two new large language models (LLMs): Codestral Mamba 7B and Mathstral 7B. Codestral Mamba, based on the innovative Mamba architecture, offers faster response times and longer context windows, making it ideal for code productivity and local coding projects. The model, tested with inputs up to 256,000 tokens, outperformed competitors like CodeLlama 7B and DeepSeek in HumanEval tests. It is available for free under an Apache 2.0 license on GitHub and HuggingFace.

Mathstral 7B is designed for math-related reasoning and scientific discovery, developed in collaboration with Project Numina. With a 32K context window, Mathstral excels in benchmarks and is also available under an Apache 2.0 license. Mistral emphasized its commitment to creating purpose-built models, highlighting Mathstral's performance and speed tradeoffs as a key example of this philosophy. Both models can be accessed and fine-tuned through Mistral’s la Plateforme and HuggingFace.

Mistral's open-source approach and competitive edge against AI giants like OpenAI and Anthropic are bolstered by recent funding success, having raised $640 million in series B funding, reaching a valuation close to $6 billion. The company's strategic investments from Microsoft and IBM underscore its significant position in the AI industry, continuously driving advancements in AI model efficiency and specialization.