Hello Tuners,

In the last week, I faced a dilemma: should we stop AI companies from using IP or enjoy the latest Metro Boomin Song? Peruse the legal landscape, trying to understand the severity of unlawful use of Intellectual Property theft (or would you call it borrowing) in the form of a potential landmark case from the world’s top Music Labels to ensure that their songs are not being used to make the next diss track in another Rap Feud!

Join us as we explore OpenAI's strategic move into data analytics with its acquisition of Rockset, Sohu's emergence as a potential frontrunner in AI chip technology, and the latest updates on the European Union's AI regulatory framework.

• OpenAI Acquires Rockset to Transform Data Utilization

• Sony, Universal, and Warner Records Sue AI Startups

• Sohu, Fastest AI Chip of our Time?

• European AI Office: EU AI Act takes a New Turn

OpenAI Acquires Rockset to Transform Data Utilization

OpenAI recently made headlines by acquiring Rockset, a leading real-time analytics company, to strategically enhance its data processing, analytics, and retrieval capabilities. This move is set to strengthen OpenAI's position in the enterprise sector, providing a competitive advantage through improved data handling and real-time insights.

Rockset, founded in 2016 by Venkat Venkataramani and Dhruba Borthakur, specializes in providing a serverless search and analytics engine that enables developers to build applications without extensive data pipelines. Venkataramani, the CEO, has a background with Facebook and Oracle, while Borthakur, the CTO, was instrumental in developing the high-performance key-value store RocksDB at Facebook. Rockset leverages the principles of RocksDB to deliver scalable, real-time indexing and SQL-based query capabilities, facilitating complex queries on live data streams.

Only time will tell how this acquisition plus Multi Acquisition (though not confirmed by OpenAI yet) is to shape the company’s future direction.

Sony, Universal, and Warner Records Sue AI Startups

In a potentially landmark case, the world's biggest record labels, including Sony Music, Universal Music Group, and Warner Records, are suing two artificial intelligence (AI) start-ups, Suno and Udio, for alleged copyright violations. The labels claim that these start-ups' software steals music to "spit out" similar works, demanding compensation of $150,000 (£118,200) per work.

The lawsuits, announced on Monday by the Recording Industry Association of America, are part of a broader wave of legal challenges from authors, news organizations, and other groups against AI firms using their work without permission. Suno, based in Massachusetts, launched its first product last year, attracting over 10 million users. The company, which has a partnership with Microsoft and has raised $125 million from investors, charges a monthly fee for its service. Despite the allegations, Suno did not respond to requests for comment.

New York-based Udio, known as Uncharted Labs, is backed by high-profile venture capital firms such as Andreessen Horowitz. Released in April, Udio's app quickly gained fame as the tool behind "BBL Drizzy," a parody track related to the feud between artists Kendrick Lamar and Drake. In a blog post on Tuesday, Udio stated it was "completely uninterested in reproducing content," denying the allegations of copyright infringement.

Sohu, Fastest AI Chip of our Time?

Etched, a chip company that recently raised $120 million to compete with Nvidia's AI chips, has announced a groundbreaking new chip named Sohu. Claimed to be the fastest AI chip of all time, Sohu can process over 500,000 tokens per second while running Llama 70B, making it a revolutionary tool for building products that are impossible on traditional GPUs. One 8xSohu server can replace the power of 160 H100 GPUs.

Sohu is the first specialized chip (ASIC) designed specifically for transformer models, offering unparalleled performance by focusing exclusively on this type of AI model. Unlike general-purpose chips, Sohu cannot run CNNs, LSTMs, SSMs, or other AI models. Given that major AI products like ChatGPT, Claude, Gemini, and Sora are powered by transformers, Sohu is poised to become the cornerstone of future AI infrastructure. Within a few years, it is expected that every large AI model will run on custom chips like Sohu.

European AI Office: EU AI Act takes a New Turn

Amid AI uncertainties, the European Union is opening the European AI Office to study AI's societal and economic benefits while mitigating risks. This office will be the central hub of AI expertise across the EU, providing advice on best practices, regulation, compliance, safety, and innovation.

The EU aims to position itself as a global leader in AI governance. Despite the potential of AI, only 8% of European companies with over 10 employees used AI last year, mainly in Denmark, Finland, and Luxembourg.

Margrethe Vestager, Executive Vice President of the European Commission, stated that the office will ensure a coherent implementation of the AI Act, which will enter into force at the end of June. The European AI Office will guide the AI ecosystem, aiming to make it innovative, competitive, and compliant with EU values.

LLM Of The Week

DeepSeek v2 Coder is a Mixture-of-Experts (MoE) code LLM that approaches GPT4-Turbo in code-specific tasks. It comes in two sizes, 16B and 236B, and was trained on 6T tokens and over 300 programming languages, leveraging the advances from DeepSeek V2 MoE.

By utilizing 236B parameters with 160 experts and 16B active parameters, DeepSeek v2 Coder sets new state-of-the-art results in HumanEval, MBPP+, and LiveCodeBench for open models. The lite version, with 16B parameters and 2.4B active parameters, is perfect for on-device use and achieves 81.1% on HumanEval. Additionally, it enhances mathematical reasoning, achieving 75.7% on the MATH benchmark.

DeepSeek v2 Coder supports a 128K context length and was initiated from DeepSeek V2, further trained on 6T tokens. This makes it a robust choice for coding tasks, with the lite version being especially strong for on-device coding assistants.

Weekly Research Spotlight 🔍

TextGrad: Automatic “Differentiation” Via Text

TextGrad introduces a novel approach to prompt optimization in module LLM programming, leveraging a method akin to the chain rule. In the forward pass, each module receives a (query, prompt) pair and generates an answer. During the backward pass, an LLM refines the prompt based on the (query, answer) pair and suggests improvements for the query based on the (prompt, answer) pair. This end-to-end bootstrapping significantly enhances prompt optimization.

The performance of TextGrad is impressive, matching or surpassing DSPy's random search method. This innovative approach not only makes prompt optimization more systematic but also opens exciting possibilities for applications such as code generation. The paper is a must-read for those interested in the future of LLM programming.

Best Prompt of the Week 🎨

Today's Goal: Try new things 🧪

Acting as a Fitness Journey Planner

Prompt: I want you to act as a fitness journey planner. You will create a personalized fitness plan based on the client’s current health condition and fitness objectives, emphasizing weight loss, healthy living, and overall fitness improvement. You will identify a target client profile, develop essential strategies and action plans, select the tools and resources for effective exercise routines and nutrition, and outline any additional activities needed to ensure long-term success. My first suggestion request is: "I need help designing a fitness plan for an overweight person aiming to lose weight, maintain a healthy lifestyle, and get fit."

This Week's Must-Read Gem 💎