- AI Tangle
- Posts
- ☕️ Microsoft's MAI-1 to Rival The World
☕️ Microsoft's MAI-1 to Rival The World
From Microsoft's 500B parameter state-of-the-art LLM, MAI-1, to Wayve's U.K. record-shattering AI fundraiser, the world of AI starts the week off with excitement. Meanwhile, coding Q&A platform Stack Overflow pens a deal with OpenAI to help train models, Google brews a top-of-the-line cybersecurity system out of Gemini, Mandiant, and VirusTotal, and YouTube gives U.S. Premium users some AI features to play with. Join us at AI Tangle as we untangle this week's happenings in AI.
THE BIG AI STORY
Microsoft is reportedly developing a large-scale AI language model, dubbed MAI-1 and led by former Google AI leader Mustafa Suleyman. This marks Microsoft's first in-house development of such a model in a long while, aiming to compete with state-of-the-art models like Google's Gemini, Anthropic's Claude family line, and even OpenAI's GPT-4, which Microsoft has closely collaborated with in the past, remaining one of its biggest backers.
So what's MAI-1 made of?
MAI-1, with approximately 500 billion parameters, is significantly larger than Microsoft's previous models like Phi-3, requiring more computing power and training data. Focusing on both smaller models run locally and larger cloud-powered models, this dual approach to AI mirrors strategies reportedly being explored by Apple as of late. The exact purpose of MAI-1 is unknown as of now, and its use will depend on its performance, though the world might catch a glimpse of it at the Build developer conference later this month in a likely preview by Microsoft.
6 QUICK HITS
Known as the de facto Q&A forum for software developers, Stack Overflow recently penned a deal with OpenAI to allow the latter to boost its model's programming-related knowledge using the platform's massive amounts of developer talk. In return, OpenAI will introduce a few sets of generative AI features to Stack Overflow's platform, the first of which will go live at the end of June. However, many active Stack Overflow members aren't on board with the idea as they worry about the validity of information generated by AI, information overload, and data privacy.
Google is developing a new cybersecurity product, Google Threat Intelligence, which uses the company's Gemini 1.5 Pro large language model to reverse engineer malware attacks and summarize threat reports. The product aims to help companies assess potential attacks and better monitor threats before it's too late. Google also plans to make use of Mandiant, a cybersecurity company it acquired in 2022, to assess security vulnerabilities around AI projects along with VirusTotal threat intelligence.
U.S. Air Force Secretary Frank Kendall recently flew in an AI-piloted F-16 fighter jet, marking a significant advance in military aviation. The AI-controlled jet, named X-62A Vista, performed maneuvers alongside a human-piloted F-16. Kendall expressed that he was confident in the AI technology's role in future air combat, stating it's a security risk to lack it. This event serves as a public statement of confidence in AI's role in air combat, with plans to use the technology to operate a pilotless fleet of 1,000 aircraft.
Following the release of YouTube's "Ask" feature, which allowed users to prompt a generative AI to answer questions and comment on a video, the company has recently unveiled "Jump Ahead." As the name implies, YouTube's recent "Jump Ahead" button allows users to skip to the most-watched sections of a video by using AI and viewer data to identify points of interest in videos, currently only available to YouTube Premium users in the United States.
Wayve, a startup that's developing self-learning autonomous driving systems, recently came out with an announcement touting a $1.05 billion win during its Series C round of funding. Backed by the likes of SoftBank Group, Nvidia, and Microsoft, this makes Wayve the single most fundraised AI startup in the United Kingdom. Much of that capital will be going to developing its product for "eyes on" assisted driving and "yes off" fully automated driving and other AI-assisted automotive applications, with further plans to expand operations globally.
Co-founded by Sharon Zhou and Greg Diamos in Palo Alto, the startup aimed at helping enterprises deploy generative AI tech, Lamini, recently came out with $25 million in Series A funding, backed by AMD Ventures, Amplify Partners, and First Round Capital. The two co-founders claim that Lamini does what general-purpose AI platforms don't, which are solutions and infrastructure geared to meet the needs of corporations.
4 AI TOOLS
Turbolearn - Instant notes, flashcards, quizzes, and more - upload any audio, video, website, or PDF and get beautifully formatted notes with the help of Turbolearn.
Synthesia - Turn text into studio-quality videos in minutes with Synthesia's AI-powered platform bundled with AI avatars and voiceovers in 130+ languages.
RoeAI - Roe is a data warehouse for unstructured data that leverages AI-powered SQL to do data extraction, classification, and RAG on documents, web pages, videos, images, and audio.
Visme - Visme is an all-in-one marketing and AI-enhanced design platform made to stand out from the crowd with beautiful and engaging on-brand content.
AI READ & WATCH
The Rise of AI Scams (3-min read)
Praised as one of America's most successful investors, Warren Buffett recently chimed in on AI and the potential rise of AI-piloted scams as a "growth industry." Likening it to that of the atomic bomb, Buffett sure hopes that whoever lets the AI genie out of the bottle knows what they're doing.
AI Conquers Gravity (16-min watch)
DrEureka might signal the start of a transition, from humans training robots, to machines teaching machines. Nvidia has demonstrated how LLMs can have immense impacts, even with their flaws. This video, presented by AI Explained, is about one paper, one concept - a genius one at that.