Microsoft's New Massive AI MAI-1: The Final Answer to OpenAI's GPT-4 (500B Parameters)

Microsoft is about to launch its own new AI language model and it's going to be a big one. They're calling it Mai1 and it's being developed entirely in house by Microsoft. So first let's look at who's behind this new AI model, because it's quite an interesting connection. The development of Mai1 is being led by Mustafa Suleiman, who used to work at Google AI and was also the CEO of an AI startup called Inflection. Now, Microsoft recently made a big move by acquiring Inflection's intellectual property for a whopping $650 million and brought on board much of their team, including Suleiman himself. While MAI one might use some of the training data and technologies from Inflection, it's important to note that this is a brand new creation from Microsoft, not just a rebranded version of something Inflection was working on. With Suleiman's experience at Google AI and Inflection, he brings a lot of expertise to the table for MI1's development. Alright, now let's talk about how this model compares to other major players in the AI field. Microsoft has ambitious plans for Mai1, aiming to make it significantly larger than their earlier smaller open source models. They're stepping it up to about 500 billion parameters, which marks a substantial advancement for them. For a bit of context, OpenAI's GPT4, one of the most advanced AI language models currently available, boasts over 1 trillion parameters. Meanwhile, Meta's Llama 2 models come with up to 70 billion parameters. So while Mai 1 might not match GPT4's size, it's. It's definitely shaping up to be a strong competitor in the AI language model field. And Microsoft has a lot of resources at their disposal to make Mai1 a success. They have vast amounts of data and computing power, including large clusters of servers with Nvidia's graphic processing units. All of this will play a crucial role in training and supporting the development of Mai1. It seems like Microsoft is investing heavily in this model because they want to stay competitive in the rapidly evolving world of AI. They've already invested $10 billion in OpenAI and acquired Inflection's intellectual property and staff. This suggests that Microsoft is making a strategic shift towards building their own in house AI capabilities. While Microsoft chose not to comment to Reuters on the story, Microsoft CTO Kevin Scott shared his thoughts on LinkedIn regarding the company's strategy. He explained, we build big supercomputers to train AI models. Our partner OpenAI, uses these supercomputers to train cutting edge models, then we both offer these models in our products and services, making them accessible to many people. We're quite happy with this setup. Additionally, Scott highlighted that for many years Microsoft has been actively developing AI models, both in Microsoft Research and within their product groups. He noted that working on AI models is fascinating and that their researchers have done excellent work in studying and building them. Recently, emails released as part of the US Justice Department's antitrust case against Google revealed that Microsoft's investment in OpenAI was partly driven by concerns over Google's advancements in AI, according to a report by Business Insider. This came to light through an internal email involving Microsoft co founder Bill Gates CEO Satya Nadella and Scott In a June 2019 email titled Thoughts on OpenAI, the discussion focused on the investment opportunities in the AI organization and highlighted areas where Google was notably ahead of Microsoft in AI research and models. Scott initially was somewhat dismissive of OpenAI and Google DeepMind's achievements, which he referred to as game playing stunts, possibly alluding to AlphaGo Zero's demonstrations. However, he grew concerned when he recognized the advanced capabilities of competitors natural language processing models. Now, if you are curious about what exactly Mai1 will be used for and what it's capable of like I am. Unfortunately, those details haven't been made public yet. We'll have to wait until the product is released to really evaluate its performance and capabilities. However, given the scale and ambition behind the project, it's clear that Microsoft is aiming for the model to compete with the most advanced AI language models out there. So how does Mai1 compare to GPT4, one of the most well known AI language models out there? Well, from what we know so far, one key difference is the size of the models. As we mentioned earlier, MAI1 is expected to have around 500 billion parameters, which is a lot, but still less than the over 1 trillion parameters in GPT4. This might give GPT4 a bit of an advantage when it comes to raw processing power and handling complex tasks. However, Microsoft has a lot of data and computing resources at their disposal, which could help close that gap. Both models are designed to be really good at a wide range of tasks, from processing natural language to generating code. GPT4 has already shown some impressive results on academic and professional tests like the Uniform Bar Exam. We don't know the full extent of what MAI one will be capable of yet, but with Microsoft's investment and the leadership of Mustafa Suleiman, it's likely to be a highly capable model. Another thing to consider is how these models approach safety and ethics. GPT4's training helps it avoid giving harmful responses and refuse requests for content that's not allowed. We don't know exactly how MAI one will handle these issues yet, but given Microsoft's commitment to responsible AI development, it's likely to be a priority for them as well. Alright now. In other news, Apple is reportedly developing its own AI chips tailored for data centers. These AI chips could potentially give Apple a crucial advantage in the competitive AI industry, according to sources familiar with the matter. Apple has been closely collaborating with its chip manufacturing partner, TSMC to design and initially produce these chips. However, it's still unclear if the final version of the chips has been manufactured yet. It's speculated that Apple's server chips may focus on executing AI models, particularly in AI inference, using trained AI models to make predictions or decisions, rather than AI training the process of teaching AI models Where Nvidia's chips currently dominate the market over the past decade, Apple has gradually become a significant player in chip design for products like the iPhone, iPad, Apple Watch and Mac computers. This latest project, internally known as Project acdc, short for Apple Chips and Data center, will integrate Apple's chip design capabilities into the operation of servers used by Apple's clients. According to sources, the project has been ongoing for several years, although the timeline for launching these server chips remains uncertain. Apple is expected to unveil more new AI products and AI related updates at its Worldwide Developers Conference in June. An Apple spokesperson declined to comment on these reported developments. Now one more interesting update from YouTube. The platform is rolling out a new AI powered Jump Ahead feature for premium subscribers. This feature makes it easier to skip directly to the most popular or interesting parts of videos, enhancing the viewing experience by letting users quickly access the highlights. The feature uses AI to analyze patterns in how viewers watch and interact with videos. When you double tap on a video, it will recommend skipping ahead to the next section that is commonly watched or considered the best part based on viewer behavior data. Previously, double tapping would just skip ahead a set amount of time or jump between chapters if the creator had set them up. But this could be clunky, especially on mobile devices. The new AI Jump Ahead feature aims to make skipping to the good parts more seamless. For now, the Jump Ahead feature is rolling out as an experiment for YouTube Premium subscribers using the Android app. Premium is YouTube's paid membership that removes ads and allows offline viewing. When using an eligible video, double tapping will bring up a jump Ahead button that will skip the video to the next popular segment, as determined by YouTube's AI system analyzing viewer patterns. The feature currently only works for English videos, but YouTube may expand it further if the experiment is successful. Creators will also be able to use Jump Ahead when watching their own videos, even without a premium subscription. Alright, don't forget to hit that subscribe button for more updates. Thanks for tuning in and we'll catch you in the next one.