TLDR AI 2024-05-03

If you missed Netlify's Compose Web + AI virtual event, now's your chance to watch every session on-demand, including:

• Netlify CEO Mathias Biilmann's unpopular opinion about AI's impact on web development
• How human and AI developers will work together. Featuring Scott Wu of Cognition Labs (creator of AI software engineer Devin)
• Why embracing AI is key to workplace productivity, featuring Vip Gupta, Product Lead at Google
• Netlify's AI roadmap for the Composable Web Platform

Find out why Netlify has become a leading provider of composable web infrastructure for AI — powering chatbots, agents, genUI, semantic stores, image generation, and more.

▶️ Watch the videos now (no registration required!) or read the guide for building AI experiences on Netlify

Eight prominent U.S. newspapers owned by investment giant Alden Global Capital are suing OpenAI and Microsoft for copyright infringement in a complaint filed Tuesday in the Southern District of New York. On top of a similar case filed by the New York Times against both companies, the new suits add heft to publishers' claims. Until now, the Times was the only major newspaper to take legal action against AI firms for copyright infringement.

Mustafa Suleyman, co-founder of DeepMind, has been named the chief executive of Microsoft AI. He will contribute to Microsoft's expansion into AI consumer products, while his former colleague and DeepMind's other co-founder, Demis Hassabis, will lead AI research at Google. The journey of these two influential figures reflects the personal and competitive undercurrents driving the race to develop the next major computing platform.

Apple CEO Tim Cook is anticipated to tease new AI features during the "Let Loose" event next week, with further details expected at the WWDC in June. The new iPad Pro models may receive the M4 chip, hinting at potential advanced AI capabilities in Apple's upcoming products.

Full-Frequency Dynamic Convolution (FFDConv) is a new method that enhances 2D convolution for sound event detection. By generating unique frequency kernels for each band, FFDConv improves the accuracy of detecting sound events, especially in terms of their frequency characteristics.

This study discusses how combining algorithmic adjustments with tailored hardware can enhance ViTs' efficiency, particularly through model quantization.

Spikformer V2 combines the self-attention mechanism with the biological efficiency of Spiking Neural Networks (SNNs). This innovative model uses a Spiking Self-Attention mechanism and a Convolutional Stem, enhancing its ability to process visual features while being energy-efficient.

The Segment Anything Model (SAM) from Meta AI, a notable foundation model in computer vision, excels at image segmentation but struggles in certain specific areas. This project presents ASAM, an advancement on SAM that uses adversarial tuning to boost its performance.

This project introduces SUNDAE, a new approach that enhances memory efficiency through spectral pruning and neural compensation.

InstructDr is a model designed to excel in various visual document understanding tasks like question answering and information extraction. InstructDr can adapt to new tasks and datasets by combining document images with large language models, outperforming existing models.

Domino Data Lab is running a series of free in-person events for data science and IT leaders to network, innovate, and share perspectives. Hear from industry leaders from EY, Moderna, Bridgewater, Moody's Analytics, & more — register now to shape the discussion on responsible AI practices that can accelerate impact.

Join now: New York | Philadelphia | London

This study reveals a method to significantly extend the context length of language models up to 128K, emphasizing the importance of both the amount and variety of training data.

NYC's ascendancy in AI highlights the city's robust talent pool and growth as an AI hub. The NYC tech scene has drawn AI unicorns and tech workers. It is fuelled by resources like elite universities and a $400 million AI Research Consortium fund.

In recent years, the majority of AI applications have adopted traditional subscription-based pricing models, with a focus on per-user charges, reflecting their role as digital assistants rather than replacements for human workers. Innovative pricing strategies, such as outcome-based models, are emerging among newer AI companies, potentially enhancing customer adoption and revenue by charging only for successful results.

StreamMultiDiffusion is a framework that enables real-time region-based text-to-image generation.

Learn how to build a virtual banking assistant with access to real time data using Amazon Bedrock and Confluent Cloud. Watch the tech session

The $800 million investment aims to create over 100 jobs and is expected to be operational by the end of 2026.

Microsoft updated Azure OpenAI Service's terms to prohibit U.S. police from using its generative AI for facial recognition, clarifying restrictions on law enforcement applications globally.