The AI Gazette: The Latest from ChatGPT, Meta and Google

After a two-week break, Paul and Martin return with a heap of AI news from the last two weeks. Strap in!

• ChatGPT user growth may have stalled, with data suggesting fewer users in June compared to May.
• OpenAI provided all API users access to GPT 4, a significant development for developers using the API.
• ChatGPT's internet access was revoked due to users bypassing paywalls and accessing gated content on news sites.
• OpenAI's team is building a human-level automated alignment researcher to scale efforts and align superintelligence more effectively.
• OpenAI made Code Interpreter available to everyone, enabling various functionalities within the chat, data analysis, visualization, predictive modeling, and more.
• OpenAI seeks new data sources, striking deals with Shutterstock and the Associated Press.
• ChatGPT's performance has changed significantly, with some tasks showing deterioration over time.
• ChatGPT increases message limits, allowing some users up to 100 messages every 3 hours.
• ChatGPT introduces custom instructions, letting users provide information about themselves and how they want ChatGPT to respond.
• Claude 2 is released, nearly as good as ChatGPT 4 on various benchmark tests with a 100k context window.
• Meta releases commercial version of its open source language model, LLaMA 2, available in three model sizes with some usage restrictions.
• Google updates to Bard, allowing image uploads, voice responses, and modifications to response tones.
• Google launches NotebookLM, an AI-powered notes app for select users in the US, designed to gain insights from language models and existing content.
• Google is testing an AI tool called "Genesis" for writing news articles, acting as a co-pilot for journalists.
• Hubspot's chatspot learns your writing style for blogs and emails, but the tool is limited to 5000 characters.
• Microsoft plans to charge $30 per user for Microsoft Office AI features, potentially leading to significant ROI for enterprises.
• Inflection AI develops an $880M AI supercomputer despite scarce GPUs, raising $1.5 billion at a $4 billion valuation.
• ClipDrop introduces Stable Doodle, allowing users to create sketches for image generation prompts.
• Midjourney launches the Panning feature, enabling image extension in various directions.
• Meta's new image generation model, CM3Leon, creates images from text and vice versa, with the ability to edit existing images.