Learn more
Join Dexa
Published May 23, 2024

📅 ThursdAI - May 23 - OpenAI troubles, Microsoft Build, Phi-3 small/large, new Mistral & more AI news

Delve into the latest AI headlines as Alex Volkov unpacks the legal and ethical challenges of voice cloning, advancements in AI interpretability and safety, and cutting-edge developments from Microsoft's Build conference, including breakthrough AI models that promise to revolutionize language and vision tech.
Episode Highlights
ThursdAI logo

Topics covered

Popular Clips

Episode Highlights

  • Microsoft AI

    Microsoft's Build conference showcased groundbreaking AI integrations, particularly with their new Copilot hardware. highlights Microsoft's move into hardware with Copilot+ PCs, which feature NPUs designed to enhance AI processing efficiency 1. These devices promise improved battery life and performance by offloading AI tasks to NPUs, making them ideal for AI-driven applications.

    Microsoft Copilot is now the main one. And GitHub Copilot, the copilot we all know love as developers, is the secondary one.

    ---

    Additionally, the integration of AI into everyday tech was demonstrated through interactive demos, such as an AI assistant in Minecraft, highlighting the seamless blend of AI with consumer technology 2.

       

    Open Source

    Open source AI models are gaining momentum, with significant contributions from companies like Hugging Face. notes Hugging Face's $10 million commitment to Zero GPU, enabling developers to run AI applications on high-end GPUs without cost 3. This initiative democratizes access to powerful AI tools, fostering innovation in the open-source community.

    Hugging Face introduced a new functionality that adds the ability to run local models with one click.

    ---

    Moreover, Hugging Face's integration with local models like LM Studio enhances model accessibility, allowing users to easily deploy AI models on their devices 4.

       

    Model Performance

    Recent AI model releases demonstrate impressive performance improvements, particularly in language processing and vision tasks. discusses Microsoft's Phi models, which are competitive on benchmarks despite using fewer training tokens than competitors 5. These models, including the Phi-3 Mini and Vision, excel in reasoning tasks and are optimized for local device use.

    Phi three, for example, is trained for reasoning test specifically, and it's notably improved when used on hard harder prompts.

    ---

    Additionally, the introduction of new vision models like OpenBNB's Mini CPM highlights advancements in AI's ability to handle complex visual data, offering better performance and lower latency 6.