- NATURAL 20
- Posts
- Grok-1's Open Release
Grok-1's Open Release
PLUS: FTC Probes Reddit Data Sale, Groq Revolutionizes AI Chips and more.
Wes Roth
March 18, 2024
Today:
Grok-1's Open Release
YouTube Music rolling out song search, like Google Play Music
Nvidia in talks to acquire AI infrastructure platform Run:ai in massive deal
Reddit’s Sale of User Data for AI Training Draws FTC Inquiry
AI chip startup Groq grabs the spotlight
Elon Musk STUNNING Reveal of Grok | NOT aligned, MUCH Bigger, Open Source. There is no doubt left...
Elon Musk delivered on his promise to open-source Grok, announced by XAI on March 11th. The long-awaited move generated much speculation until the surprise release on Sunday. Grok, with 314 billion parameters and a mixture of eight experts, is now available under the Apache 2.0 license. Despite potential regulatory hurdles, Musk's commitment to open-source AI counters monopolistic control, benefiting the global community.
Despite concerns from some factions, including discussions of potential jail sentences for similar actions, Musk's move aligns with the push for open-source AI to counterbalance corporate control. Musk's commitment to further open sourcing reflects a broader trend in the tech industry, shaping the future landscape of AI development.
YouTube Music rolling out song search, like Google Play Music
YouTube Music is adding a song search feature, mimicking Google Play Music's old tool. Available on Android, users can search by playing, singing, or humming. iOS users might have it already. Unlike Spotify, it doesn't save findings to a playlist. It's accessible through the search bar or a dedicated button.
Google Assistant, Search, and Pixel's Now Playing offer similar functionalities. Users hope for improvements in core functions and better audio quality, comparing it unfavorably to Spotify. Despite demand for lossless audio, the focus remains elsewhere.
Nvidia in talks to acquire AI infrastructure platform Run:ai in massive deal
Nvidia is in talks to acquire Israeli AI infrastructure platform Run:ai for potentially over $1 billion. Run:ai, founded in 2018, offers software tailored for AI tasks on GPUs, with $118 million raised in funding. Co-founders Omri Geller and Ronen Dar lead the company. If sealed, it would mark Nvidia's first Israeli acquisition since Mellanox in 2019.
Nvidia's soaring market value, now over $2 trillion, is driven by AI chip dominance, fueling Wall Street records. The deal aligns with Nvidia's AI focus amid market shifts. Neither company commented on the potential acquisition.
Reddit’s Sale of User Data for AI Training Draws FTC Inquiry
Reddit faces an FTC inquiry over selling user data for AI training, potentially making $203 million. The FTC questions the fairness and privacy risks. Other platforms like Stack Overflow and Tumblr have similar deals. Reddit claims innocence but acknowledges the costly inquiry.
Google's deal to access Reddit's data draws attention, seen as a competitive threat to Reddit and publishers. FTC's scrutiny extends to other big tech partnerships. Reddit's move to monetize data for AI training raises concerns about ownership and privacy. The saga reflects broader debates on data use and AI ethics.
AI chip startup Groq grabs the spotlight
Groq, an AI chip startup in Mountain View, plans to raise new funds after securing recent backing. With chips claimed to be 10 times faster and one-tenth the cost of traditional ones, Groq is central to recent shake-ups at Social Capital.
Founded in 2016 by Jonathan Ross, who helped create Google's Tensor Processing Unit, Groq has deployed 4,500 chips and aims for 1.5 million by 2025. It targets both cloud developers and data center operators, recently acquiring Definitive Intelligence. Groq's focus on inference aligns with growing demand for AI chips, positioning itself as an environmentally friendly option.
🧠RESEARCH
"StreamMultiDiffusion" tackles the challenge of combining fast image synthesis with precise user control. By merging speed enhancements and a novel model architecture, it achieves tenfold faster panorama generation and 1.57 frames per second for region-based image synthesis on a single GPU. This paves the way for real-time interactive image creation.
"BurstAttention" addresses the challenge of processing extremely long sequences efficiently in large language models. It introduces a distributed attention framework that optimizes memory access and communication operations across multiple devices. Compared to other solutions, BurstAttention reduces communication overhead by 40% and achieves a 2x speedup during training on 32K sequence length using 8 A100 GPUs.
"Griffon v2" tackles the challenge of limited image resolution in large vision language models, hindering their performance in complex scenarios. It introduces a high-resolution generalist model enabling precise object referencing through visual and textual cues. With lightweight down-sampling and visual-language co-referring capabilities, it achieves state-of-the-art performance in object perception tasks like object detection and counting.
"Follow-Your-Click" presents an open-domain regional image animation framework emphasizing controllability and local animation. It simplifies user input to a single click for object selection and a short motion prompt. With novel strategies like first-frame masking and flow-based motion control, it achieves superior animation quality and precision compared to existing methods.
The study examines language model scaling beyond conventional norms, focusing on over-training and downstream task performance. It analyzes 104 models with varying parameters and training token sizes. Findings reveal improved predictability of validation loss and downstream task error, aiding efficient model development.
🛠️TOP TOOLS
Framer - where teams design and publish stunning sites
Sudowrite - designed as a creative partner for writers, offering AI-powered tools to brainstorm, edit, and enhance writing
Giga - scans billions of discussions on reddit and other online communities to find the most useful posts and comments
Dart - AI project management tool
Hidden Biases - interactive experience that challenges users to recognize their own assumptions and biases in ambiguous writing
Lora Land - Fine-tuned LLMs that outperform GPT-4, served on a single GPU
Antworks - break down complex software development projects into micro tasks.
🗞️MORE NEWS
Jensen Huang is about to give AI believers a glimpse of the future
Nvidia's CEO, Jensen Huang, is set to host a major AI event dubbed the "Woodstock of AI." Expected revelations include details on the next-gen H100 chip, potentially named B100, promising significant performance boosts. With high anticipation, industry leaders gather, raising questions about AI's future challenges and opportunities amidst fierce competition. BUSINESS INSIDER
Behind The Scenes And Beyond Hype: What We Learned About AI At SXSW
SXSW 2024 showcased a shift in AI perception, moving beyond hype to address its challenges. Blockchain integration emerged as a solution to combat AI-related issues like fake content. Startups, notably Atomo, showcased AI's transformative potential, while crypto's bull run underscored the importance of Web3 investments. FORBES
How AI is clearing the waters in urban rivers
Researchers have developed a new machine learning system using Ant Colony Optimization and LSTM networks to enhance sewer-river system models, reducing calibration time and improving accuracy. This innovation aids in predicting urban water pollution, crucial for effective environmental management. The MLPS offers rapid adaptation to new data, facilitating targeted pollution control strategies. PHYS ORG
AI Scam Complaints Are Just the Tip of the Iceberg
Fraudsters are exploiting the hype around generative artificial intelligence, using it in their schemes and ads on social media platforms. Complaints to the US FTC about deceptive AI-related advertising have surged, especially on sites like Twitter, Facebook, and YouTube. Older and financially vulnerable individuals are often targeted, highlighting a growing issue in online misinformation. BLOOMBERG
Beyond ChatGPT: Four Types Of AI Your Marketing Team Should Use Now
The article discusses how AI advancements revolutionize marketing strategies for businesses, particularly in construction. It highlights four AI types: Large Language Models for automating tasks, Neural Radiance Fields for 3D scene representation, Recognition for brand protection, and Computer Vision for collaborative visibility. These technologies enhance marketing efficiency and reach. FORBES
These 61 robotics companies are hiring
61 robotics firms are hiring, offering over 60 varied job opportunities in fields like surgery and space exploration. Despite industry layoffs, Robotics and AI sectors flourish, doubling hiring listings. It's a promising market for job seekers and recent graduates seeking their next big career move. TECHCRUNCH
What'd you think of today's edition? |
What are MOST interested in learning about AI?What stories or resources will be most interesting for you to hear about? |
Reply