Phi-3: Microsoft unveils its most compact AI model to date

Welcome to AI Disruptor! If you want to join our growing community of readers, click the button below.

TODAY’S TOP STORIES:

🧠 Phi-3: Microsoft unveils its most compact AI model to date
🔍 UK launches investigation into Amazon and Microsoft's AI partnerships
😎 Ray-Ban and Meta expand smart glasses lineup
🚀 Tines raises $50M to extend workflow automation beyond security
🛡️ OpenAI proposes instruction hierarchy to mitigate LLM vulnerabilities

🧠 Phi-3: Microsoft unveils its most compact AI model to date

Microsoft has introduced the latest iteration of its lightweight AI model, Phi-3 Mini, the first of three compact models in the pipeline. With just 3.8 billion parameters, Phi-3 Mini is trained on a smaller dataset compared to large language models like GPT-4.

What to know:

• Phi-3 Mini is now accessible on Azure, Hugging Face, and Ollama.

• Microsoft is set to launch Phi-3 Small (7B parameters) and Phi-3 Medium (14B parameters) in the near future.

• Phi-3 outperforms its predecessor and can generate responses on par with models 10 times its size.

• Developers trained Phi-3 using a "curriculum" drawing inspiration from children's learning through bedtime stories and simplified books.

• Compact models like Phi-3 are often better suited for companies' custom applications, given their smaller internal datasets and reduced computing power needs.

The introduction of Phi-3 Mini underscores the rising significance of smaller, more efficient AI models that can excel at specific tasks while requiring fewer resources. As businesses strive to incorporate AI into their offerings, lightweight models like Phi-3 are likely to gain popularity for tailored applications.

🔍 UK launches investigation into Amazon and Microsoft's AI partnerships

The U.K.'s Competition and Markets Authority (CMA) is initiating preliminary inquiries to determine whether the close-knit collaborations and hiring practices involving Microsoft, Amazon, and three AI startups fall under its merger regulations and could potentially impact competition within the U.K. market.

What to know:

• The CMA's action is part of a growing trend of scrutinizing Big Tech's M&A strategies in the AI sector, where critics claim "quasi-mergers" are employed to circumvent regulatory oversight.

• The regulator is evaluating whether these partnerships are equivalent to mergers from a regulatory perspective and if they could affect competition in the U.K.'s rapidly expanding AI industry.

• Microsoft has invested in and collaborated with OpenAI, Inflection AI, and Mistral AI, while Amazon recently concluded a $4 billion investment in Anthropic.

• The CMA's initial call for comments is open until May 9, potentially triggering a formal "phase 1" review directly involving Microsoft and Amazon.

The CMA's probe into Big Tech's AI partnerships and investments may establish a precedent for how regulators approach these "quasi-mergers" moving forward. The findings will shape how tech giants interact with emerging AI companies and influence the competitive dynamics of the AI industry in the U.K. and beyond.

Some other quick developments you should know…

😎 Ray-Ban and Meta expand smart glasses lineup, introduce video calling and AI-powered vision

Ray-Ban and Meta are broadening their smart glasses portfolio with fresh designs, incorporating video calling via WhatsApp and Messenger, and introducing Meta AI with Vision for hands-free access to information about the user's surroundings.

 🚀 Tines raises $50M to extend workflow automation beyond security

Tines, a Dublin-based startup specializing in workflow automation, has secured $50 million to expand its platform into infrastructure, engineering, and product applications, leveraging the rising trend of automation in the enterprise sector.

🛡️ OpenAI proposes instruction hierarchy to mitigate AI language model vulnerabilities

OpenAI researchers have introduced an instruction hierarchy method to decrease the susceptibility of large language models to prompt injection attacks and jailbreaks. By assigning varying priority levels to instructions from developers, users, and third-party tools, the hierarchy has led to notable improvements in robustness when applied to GPT-3.5.