Good morning! Intel recently announced a major deal with Microsoft to manufacture a custom AI accelerator chip for Azure, advancing Intel's foundry capabilities. However, Google paused features in its new AI image generator, Gemini, after users highlighted issues with how it depicted race, showing ethical challenges remain with rapidly advancing AI systems. Looking behind the scenes, building music streaming services requires special infrastructure to store large volumes of songs and scale up to handle spikes in demand from millions of users.

— Forrest Knight & Meghanadh Vasireddy

Intel Lands Microsoft to Make Custom AI Chip

Intel recently shared the major announcement that Microsoft has signed on to have Intel manufacture a custom AI accelerator chip, likely to enhance Microsoft's Azure cloud computing platform. Landing a high-profile customer like Microsoft signals meaningful progress for Intel’s foundry efforts as they invest heavily to rebuild their chip fabrication capabilities and compete against TSMC to become the industry’s leading processor manufacturer.

What Intel Announced:

• The undisclosed Microsoft chip will be produced using Intel’s new 18A process technology. Some key innovations with 18A:

  • Uses extreme ultraviolet lithography to enable denser transistor packing

  • Incorporates advanced FinFET and ribbonFET transistor architectures

  • Utilizes compound semiconductors like SiGe

• Together these 18A advances promise substantial gains in efficiency and performance – ideal for AI workloads that demand fast, powerful chips.

This Microsoft foundry deal offers tangible proof that Intel’s hefty investments to re-establish the dominance of their US-based chip manufacturing operations appears to be progressing as planned. Customers are seeking resilient supply chains and increasingly turning to advanced AI chips – two areas Intel is hoping to leverage. 

However, while the secured partnership with Microsoft signals meaningful headway, Intel still faces tremendous financial and engineering challenges in fully achieving their target of overtaking rival TSMC by 2025 and cementing their long-term foundry superiority through technology like their planned 14A node.

Read More Here

Google Pauses Gemini AI Image Making After Race Problems

Google recently launched Gemini, a new AI system that can generate realistic images from text prompts. It uses deep learning to produce striking, photo-realistic images.

The Issue: This week, users highlighted problems with Gemini depicting people of incorrect or unlikely races when prompted to generate images of public figures. For example, generating images of medieval European kings as non-white.

Google's Response: Google has now paused Gemini's ability to generate images of people in response to the feedback. They acknowledge issues with built-in racial and gender bias reflecting real-world discrimination. These biases originate from patterns in the training data used to develop AI systems.

Google says it is working quickly to improve Gemini's understanding of race to address the biases. The system will be retrained before image generation of people is reinstated. Fixing issues around bias remains an ongoing challenge as AI creation tools rapidly advance.

Read More Here

Designing Scalable Music Streaming Services

As music streaming services like Spotify surge in popularity, building robust infrastructure to handle growth is crucial behind the scenes. Initial requirements may be supporting 500,000 users and a library of 30 million songs.

Storage needs can start at around 90TB estimated to store the actual song files, plus additional smaller databases to hold metadata.

The high-level design would include:

• A mobile app for the user interface

• Load-balanced web servers providing APIs for the backend

• Blob storage (like Amazon S3) to store unstructured song data

• A SQL database to hold structured metadata

When scaling up to handle 50 million users and 200 million songs, some key principles apply:

• Separate storage for large song files versus structured metadata

• Distribute loads with content delivery networks (CDNs) that cache popular songs geographically closer to users

• Expand storage and optimize database replication for more reads

Read More Here

🔥 More Notes

• Google releases Gemma, a new AI model designed with AI researchers in mind

• iMessage with PQ3: The new state of the art in quantum-secure messaging at scale

• Python packages caught using DLL sideloading to bypass security

🎥 Youtube Spotlight

Can AI help us predict extreme weather?

AI is playing a crucial role in helping predict extreme weather events, with models developed by tech companies like Google, Huawei, and Nvidia rivaling traditional forecasting methods. These models utilize vast historical weather data to make accurate predictions, and can produce forecasts in a matter of seconds. While AI forecasting models still have limitations, such as predicting the intensity of weather events, they offer potential for improving the accuracy and understanding of uncertainty in weather predictions.

Was this forwarded to you? Sign Up Here