Daily AI Newsletter: Microsoft developing AI model with 500B parameters 🤯

In today’s email:

🏢 Apple working on AI chips for data centers
💃 This year’s Met Gala theme is AI deepfakes
📚 How LLMs Work, Explained Without Math
🧰 8 new AI-powered tools and resources. Make sure to check the online version for the full list of tools.

Microsoft reportedly developing MAI-1 AI model with 500B parameters

Microsoft Corp. is currently developing a large language model known as MAI-1, which has approximately 500 billion parameters, according to a report by The Information. This model is positioned between the capacities of OpenAI's GPT-3 and GPT-4, aiming to deliver high accuracy with lower power consumption and inference costs. The development is led by Mustafa Suleyman, who recently joined Microsoft through the acquisition of Inflection AI Inc. for $625 million.

MAI-1's training data includes text from GPT-4, web content, and is being developed using Nvidia's GPUs on a large server cluster. Although it's unclear how Microsoft will use MAI-1, it's likely to be deployed in data centers to enhance services like Bing and Azure due to its complexity and inability to run on consumer devices.

Microsoft may reveal MAI-1 at the upcoming Build developer conference if it meets development milestones. This news comes shortly after Microsoft open-sourced the Pi-3 Mini, a smaller model with 3.8 billion parameters that reportedly outperforms models ten times its size, part of a series that includes two larger neural networks.

Stack Overflow signs deal with OpenAI to supply data to its models

OpenAI has formed a partnership with Stack Overflow, the widely-used Q&A platform for software developers, to enhance the capabilities of its generative AI models for programming tasks. This collaboration announced on Monday, aims to improve ChatGPT and other OpenAI models in handling programming queries. Stack Overflow, reversing its initial stance of banning ChatGPT responses due to concerns about spam, stands to gain from OpenAI's expertise by incorporating new AI features into its platform. The first set of these features is expected to be available by the end of June.

Stack Overflow has been integrating AI into its services since last April, starting with a conversational search tool that leverages its vast repository of over 58 million questions and answers. Despite some backlash from its community over the accuracy and privacy implications of AI-generated content, the company continues to explore AI tools. A recent Stack Overflow poll revealed that a significant portion of developers are now using or planning to use AI in their programming work, highlighting a shift in the development landscape.

The collaboration with OpenAI is part of a broader strategy by Stack Overflow to rejuvenate its platform and offset declining traffic, which has been impacted by the availability of advanced generative AI models. In addition to its non-exclusive deal with Google to enhance Google’s Gemini models, Stack Overflow is looking to establish more licensing agreements with AI providers. Prashanth Chandrasekar, CEO of Stack Overflow, emphasized the company's commitment to redefining the developer experience and promoting socially responsible AI through accurate, trusted data with the launch of OverflowAPI.

This year’s Met Gala theme is AI deepfakes

The 2024 Met Gala featured an intriguing twist with a surge of fake images of celebrities, notably Katy Perry and Rihanna, stirring up the internet. Katy Perry was depicted in a whimsical, woodland-like gown that clashed with the actual carpet colors of the event, hinting at the image's falsity, despite it garnering massive online attention. Another image showed her in a bronze corset with a floral skirt, which also seemed artificially altered due to the lighting and texture mismatches, leading to speculation since no fashion outlets reported her attendance.

Rihanna, who is traditionally a highlight of the Met Gala, was portrayed in a regal, garden-themed dress complete with a sculptural halo. However, this too was a fabrication, mismatched with her newly dyed pink hair, and confirmed by reports of her absence due to illness. These AI-generated images, while visually captivating, lacked the genuine creativity that real Met Gala outfits usually embody, such as Cynthia Erivo's suit or Lana Del Rey's woodsy attire.

This year's Met Gala also saw Zendaya return spectacularly with two different outfits, including a fairy-tale villain-esque ensemble. Her presence, amid the fake visuals, reminded attendees and viewers alike to scrutinize the details in images—such as carpet patterns and background paparazzi—to discern reality from fabrication. The prevalence of AI in generating deceptive celebrity visuals has turned a usual glamorous event into a more cautious examination of what's real and what's digitally conjured, emphasizing the broader implications of AI in our perception of media.

Apple working on AI chips for data centers

Apple is reportedly developing its own artificial intelligence chip for use in data centers, according to the Wall Street Journal. Named Project ACDC, the initiative aims to harness Apple’s expertise in chip design specifically for its server infrastructure. This development comes as Apple seeks to catch up with its competitors in the AI field, with CEO Tim Cook expressing optimism about the company's prospects in generative AI.

The project, which has been in progress for several years, focuses on inference tasks in AI, which involve applying trained models to new data, rather than on the training of AI models themselves. However, details such as the launch date of Project ACDC remain undisclosed.

Apple is collaborating with Taiwan Semiconductor Manufacturing Co to design and possibly start production of these AI chips. Despite these efforts, it is still unclear if the project has achieved any definitive outcomes. Apple has not yet responded to inquiries for comments on the matter.

Other stuff

'It would be within its natural right to harm us to protect itself'
How LLMs Work, Explained Without Math
Meta’s new multi-token prediction makes AI models up to 3X faster
Hallucination-Free RAG: Making LLMs Safe for Healthcare
Quora CEO Adam D’Angelo talks about AI, chatbot platform Poe and why OpenAI is not a competitor
Google's medical AI destroys GPT's benchmark and outperforms doctors
YouTube rolling out JumpAhead - It uses AI to let you jump to parts of a video where most viewers typically skip ahead.