OpenAI “Strawberry” Model, Microsoft “SpreadsheetLLM”, Reuters Momentum AI conference

🌐 AI News

• 🔜Up Next:
  - July 26-28, 20246th International Conference on Artificial Intelligence and Computer Science (AICS 2024) Wuhan, China. Register Now
  - July 30-31, 2024 Fortune Brainstorm AI Singapore Register Now
  

OpenAI’s “Strawberry” Model: Stage 2 Of 5-Level AI Development?

OpenAI is developing a new AI model named “Strawberry” to advance AI tools towards human-level intelligence through enhanced reasoning capabilities. Building on the previous Q* project, Strawberry aims to autonomously scan the internet and perform deep research.

Capabilities and Features

• Advanced Reasoning: Enhancing AI’s human-like thinking and reasoning.

• Autonomous Research: Performing complex tasks like scientific discoveries and software development.

• Post-Training Analysis: Refining models to improve human-like responses.

Development of Strawberry
An internal document from May 2024 reveals that Strawberry will enhance the reasoning ability of OpenAI’s models to handle complex scientific and mathematical problems, allowing models to plan ahead and perform in-depth research. The Q* project, which led to CEO Sam Altman’s temporary dismissal due to AI safety concerns, laid the groundwork for Strawberry.

OpenAI’s Roadmap for Future AI Development
OpenAI outlines a five-level AI development roadmap:

1. Chatbots

2. Reasoners

3. Agents

4. Innovators

5. Organizers

Current models are at the first level, with expectations to reach the second level, reasoning, soon.

The Future of Strawberry
OpenAI aims to achieve doctoral-level intelligence in specific tasks within a year to a year and a half. Improving the ability of large language models to perform long-time tasks (Long-Horizon Tasks, LHT) is a key focus. If successful, Strawberry could redefine AI’s ability to make scientific discoveries, develop software, and perform complex tasks independently, bringing humanity closer to AGI.

Read the full version Here!

Microsoft’s “SpreadsheetLLM”: AI in Spreadsheets

Microsoft’s SpreadsheetLLM is a groundbreaking AI framework designed to optimize large language models (LLMs) for understanding and analyzing complex spreadsheet data. Traditional LLMs struggle with spreadsheets' structured and diverse formats, but SpreadsheetLLM addresses this with an innovative encoding method called SheetCompressor, which simplifies and compresses spreadsheet content for efficient processing.

SheetCompressor consists of three main modules: structural-anchor-based compression, inverse index translation, and data-format-aware aggregation. These modules reduce token usage by up to 96%, significantly improving performance in tasks such as table detection and question-answering. This advancement allows SpreadsheetLLM to outperform existing methods and opens up new possibilities for AI-assisted data analysis and decision-making in the enterprise.

By making spreadsheet data more accessible and understandable through natural language processing, SpreadsheetLLM enhances productivity and democratizes data access. Users can query and manipulate data using plain English, automating tedious tasks and enabling more individuals to make data-driven decisions. As SpreadsheetLLM moves from research to real-world applications, it promises to revolutionize tools like Microsoft Excel, transforming how businesses interact with and derive insights from spreadsheet data.

Read the full version Here!

AMD's AI head explains how it's tackling Nvidia's 'lock-in' and the GPU shortage with an open-source approach

At the Reuters' Momentum AI conference, Ramine Roane, AMD's corporate vice president of data center, cloud, and AI, discussed the company's strategy to tackle Nvidia's dominance and the ongoing GPU shortage. Roane emphasized AMD's commitment to an open-source approach with its ROCm software for GPU programming, contrasting Nvidia's closed CUDA platform. This open-source model aims to provide flexibility and lower costs for developers and businesses.

AMD is addressing the high demand for GPUs, driven by the growth of generative AI, by maximizing production. Roane noted that major companies like Meta and Microsoft are starting to shift from Nvidia to AMD, leveraging AMD's ROCm and HIP programming language to convert existing CUDA programs for broader compatibility.

Analysts like Alvin Nguyen from Forrester believe AMD's open-source strategy could help it gain market share by reducing adoption barriers and costs. However, Nguyen also pointed out that maintaining compatibility with CUDA will require ongoing support from AMD to avoid potential frustrations from users switching from Nvidia's ecosystem.

NVIDIA set to release a next-gen flagship GeForce RTX 5090D GPU in China

NVIDIA plans to release a next-gen GeForce RTX 5090D GPU in China in response to U.S. export restrictions on high-performance GPUs for AI. These restrictions have already led to the GeForce RTX 4090 being replaced with the 4090D in China. The RTX 5090D is expected to launch in January 2025 and will likely differ from the standard RTX 5090 by reducing AI capabilities while maintaining strong gaming performance.

The upcoming RTX 5090 is rumored to be significantly more powerful than the RTX 4090, though the exact performance increase is still speculative. NVIDIA might implement a custom BIOS for the RTX 5090D to limit its AI capabilities without affecting gaming performance. The January 2025 release aligns with broader expectations for the RTX 50 Series, which may see a global launch around the same time.

Overall, NVIDIA's strategy includes adapting its GPU offerings to comply with export restrictions while continuing to cater to the gaming market in China. The specifics of the RTX 5090D and the potential for an RTX 5080D remain to be seen.

Researchers Recreate Human Episodic Memory to Give LLMs Infinite Context
Researchers from Huawei and University College London have developed EM-LLM, a model that integrates human-like episodic memory into large language models (LLMs) to provide infinite context. This approach uses cognitive science principles, organizing information into episodic events and employing a two-stage retrieval process. EM-LLM significantly improves performance on long-context tasks and maintains efficiency without increased computational demands. It outperforms similar models, showing notable improvements in context handling.

Jensen Huang, Mark Zuckerberg to Discuss Future of Graphics and Virtual Worlds at SIGGRAPH 2024
NVIDIA CEO Jensen Huang and Meta CEO Mark Zuckerberg will discuss the future of graphics and virtual worlds at SIGGRAPH 2024. Their conversation will explore AI and simulation's potential impact. The event, held in Denver from July 28 to August 1, will feature nearly 100 exhibitors and highlight advancements in computer graphics, including generative AI. Notable sessions will include a fireside chat with Huang and Lauren Goode from WIRED, and the introduction of the Stephen Parker Award.


Intel venture arm’s China tech stakes raises alarm in Washington
Intel's venture capital arm has raised concerns in Washington due to its investments in Chinese tech companies, highlighting tensions over U.S. firms' involvement in China's tech sector. These investments, part of Intel Capital's strategy, are seen as potentially conflicting with U.S. national security interests, especially amid increasing scrutiny of tech ties with China. The situation underscores the delicate balance between fostering business growth and adhering to geopolitical and security considerations.


Meta to drop Llama 3 400b next week — here’s why you should care
Meta will release Llama 3, a new language model with 400 billion parameters, next week. This model offers advanced reasoning, coding capabilities, and multimodal functions, and is significantly more powerful than its predecessor. Initially, smaller versions will be available, with the larger model to follow. Llama 3, trained on 15 trillion tokens, might not be open-sourced as Meta integrates it into consumer products. The model will be accessible on platforms like Amazon SageMaker, Google Cloud, and Microsoft Azure, highlighting Meta's focus on AI in its product ecosystem


Android devices get Anthropic’s Claude app after iOS launch
Anthropic has released its AI chatbot, Claude, on Android devices following its iOS launch. The app allows users to sync chat history across devices and perform real-time image analysis. A new Team plan, designed for higher usage and processing longer documents, is available at $30 per user per month. This expansion aims to increase accessibility and competitiveness against AI tools like ChatGPT and Google's Gemini

Mistral releases Codestral Mamba for faster, longer code generation
Mistral has introduced Codestral Mamba, an advanced AI model for faster and longer code generation, handling inputs up to 256,000 tokens. It outperforms models like CodeLlama and DeepSeek. Open-sourced under Apache 2.0, it's available on GitHub and HuggingFace. Mistral also released Mathstral, aimed at mathematical and scientific tasks, expanding their open-source AI offerings

👥 Connect & Feedback!

👉 Join Us:

📧 Advertise In Weekly AI News:

📧 Contact directly at [email protected]

😍 Share with your friends!