🚀 Microsoft Unveils Phi-3-vision

PLUS: Mapping Claude's Mind

Welcome back!

Microsoft unveiled an ambitious AI strategy at its annual Build 2024 Conference, with announcements in infrastructure, models, tools, and applications. Highlights include the introduction of Phi-3-vision, an open-source multimodal model, and several other developments and partnerships. Let's unpack this...

Today’s Summary:

  • Microsoft reveals new Phi AI models

  • Mapping AI concepts in Claude Sonnet’s mind

  • PwC's AI Jobs Barometer Report

  • Seoul AI safety summit

  • Scale AI closes $1B financing

  • 3 new tools

TOP STORY

Microsoft Unveils Ambitious AI Plans at Build 2024

The Summary: At its annual Build conference, Microsoft presented a comprehensive AI strategy spanning infrastructure, models, tools, and applications. The new AI models include Phi-3-vision, an open-source multimodal model combining language and vision capabilities, and Phi-3-small and medium language models on Azure, offering cost-effective yet powerful options for developers. Additionally, Microsoft partnered with Khan Academy to provide AI-powered teaching assistants.

Source: Microsoft

Key details:

  • Phi-3-vision is a 4.2B parameter multimodal model for language and vision tasks

  • Phi-3-small (7B) and Phi-3-medium (14B) are available on Azure for generative AI applications

  • Phi-3-Silica (3.3B) included in Copilot+ PCs for on-device tasks

  • Phi-3 models outperform many larger models despite their smaller size

  • Azure infrastructure updates, including AMD MI300X

  • Free access to Khan Academy's Khanmigo for Teachers on Azure

Why it matters: By offering cutting-edge models, robust AI infrastructure, and deep product integrations, Microsoft is positioning itself as a leading AI platform company. The new open-source AI models help developers with practical solutions bridging the gap between powerful capabilities and resource constraints.

RESEARCH

Inside the Mind of an AI: Mapping Millions of Concepts in Claude Sonnet

The Summary: Researchers have achieved a fascinating breakthrough by mapping millions of concepts inside Claude Sonnet, one of the models powering claude.ai. They found features for cities, people, or science concepts. By changing the features, researchers were able to alter Claude's behavior. 

Source: Anthropic

Key details:

  • Researchers found millions of features representing concepts

  • These include both concrete things and abstract ideas like biases and security vulnerabilities

  • Feature mapping shows related concepts close together (eg Golden Gate Bridge is near San Francisco)

  • Features are multimodal, activating for both text and images of the same concepts, and across different languages

  • Changing features can change model behavior (eg creating scams)

  • First detailed look inside a modern LLM in production

Why it matters: This work provides new ways to understand large language models, and suggests new ideas to make them safer and more reliable. As AI systems become more powerful, these insights will be crucial for maintaining high standards and ensuring responsible use.

INSIGHTS

PwC Publishes "AI Jobs Barometer" Report

Image: DALL·E

The Summary: PwC has published the AI Jobs Barometer report, analyzing over 500 million job ads to study AI's impact on the workforce. The data reveals a productivity revolution underway in AI-exposed sectors like IT and finance, transforming the skills required for various occupations. While job growth in these roles is slowing, overall job numbers are still increasing. Jobs that need specialized AI skills are growing faster and offer a wage premium.

Key details:

  • 4.8x higher growth in productivity in AI-exposed sectors where AI can be used for some tasks

  • Job growth is 27% slower in AI-exposed occupations, though jobs are still growing overall

  • Skills changed 25% faster in AI-exposed roles

  • 3.5x faster job growth for roles requiring specialist AI/ML skills with up to 25% wage premium

  • 69% of CEOs expect that AI will require the workforce to develop new skills

Why it matters: The PwC report provides evidence of AI's effects on the job market, causing productivity gains and changing job skill requirements. Workers who upskill to leverage AI can access emerging opportunities, boost their value to employers, and potentially earn higher wages.

QUICK NEWS

Quick news

  • Global leaders and AI companies have agreed on international collaboration and safety measures at the AI summit in Seoul

  • Scale AI closes $1B financing for its “data foundry”

TOOLS

🥇 New tools

  • Findr - Write documents from your workplace apps

  • Canvid - Next generation AI powered screen recorder

  • Nowadays - AI assistant for corporate event planning

That’s all for today!

If you liked the newsletter, share it with your friends and colleagues by sending them this link: https://thesummary.ai/