AI Updates July 2024 - Monthly Overview

RouteLLM: Achieves 90% GPT-4 Quality at 80% Lower Cost

RouteLLM, an innovative project from lmsys.org, aims to reduce the cost of running large language models (LLMs) by up to 80% while maintaining 95% of GPT-4’s quality. This open-source framework uses smaller, local models for most queries, routing only the most complex tasks to more expensive models like GPT-4. By leveraging a principled approach to LLM routing based on preference data and using augmented datasets, RouteLLM achieves significant cost reductions without compromising performance. Evaluations using benchmarks like MT Bench, MMLU, and GSM8K show that RouteLLM can maintain high performance while being over 40% cheaper than commercial systems. The framework's ability to generalize across different models and its emphasis on local computing make it a promising solution for efficient, cost-effective AI deployment.

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

OpenAI's new AI model, "Strawberry," aims to advance towards human-level intelligence through enhanced reasoning and autonomous research capabilities. Building on the Q* project, Strawberry is designed to tackle complex real-world problems by autonomously scanning the internet for deep research, improving scientific discovery, business intelligence, education, and software development. The model's development focuses on achieving advanced reasoning to perform long-term tasks, with the potential to reach doctoral-level intelligence in specific tasks within a year to a year and a half. This initiative marks a significant step towards artificial general intelligence (AGI), promising to revolutionize various fields by enabling AI to think and reason more like humans.

Mistral AI and NVIDIA Unveil Mistral NeMo 12B

Mistral AI and NVIDIA have collaborated to create the Mistral NeMo 12B, a state-of-the-art enterprise AI model designed for various applications including chatbots, multilingual tasks, coding, and summarization. This model, trained on NVIDIA's DGX Cloud AI platform using 3,072 H100 80GB Tensor Core GPUs, boasts unprecedented accuracy and efficiency with a 128K context length and 12 billion parameters. Released under the Apache 2.0 license, it encourages innovation and widespread adoption. Packaged as an NVIDIA NIM inference microservice, Mistral NeMo ensures easy deployment, high efficiency, and enhanced security, making it a powerful and reliable AI solution for enterprises.

Llama 3.1, With 405B Parameters Is Out!

Meta has announced Llama 3.1, its latest open-source AI model, available in three versions, including the advanced 405B model with 405 billion parameters. This release underscores Meta's investment in AI and collaboration with Nvidia, using their GPUs for training. Unlike commercial models, Meta's open-source strategy aims to attract top talent, reduce computing costs, and foster a community of developers. Llama 3.1, designed for complex tasks like long text understanding and coding, is accessible through cloud providers and Meta's platforms. Meta also emphasizes AI safety, collaborating with global organizations and providing tools like Llama Guard 3 to ensure secure AI applications.

Google Gemma 2: Lightweight, State-of-the-art Open Model Series

Google has launched Gemma 2, a state-of-the-art open AI model series available in 9 billion (9B) and 27 billion (27B) parameter sizes. Designed for efficiency and superior performance, the 27B model rivals larger proprietary models while running cost-effectively on a single GPU or TPU host. Gemma 2 supports major AI frameworks, offers rapid inference, and is optimized for NVIDIA hardware. It emphasizes responsible AI development with tools like text watermarking and an LLM Comparator. Accessible via Google AI Studio, Kaggle, and Hugging Face, Gemma 2 aims to revolutionize AI development with its robust, versatile design.

Mistral Large 2 Vs. Llama 3.1 Vs. GPT-4?

Mistral Large 2, the latest AI model from Mistral AI, offers unparalleled performance and cost efficiency with its 128k context window and 123 billion parameters.

Supporting multiple languages and coding languages, it excels in instruction-following, multilingual proficiency, and advanced reasoning, outperforming models like GPT-4, Claude 3 Opus, and Llama 3.1.

Trained for accuracy and reliability, it achieves an 84.0% accuracy on the MMLU benchmark. Released under the Mistral Research License, it is available for research and non-commercial use, with commercial licenses available.

Accessible via various cloud platforms, Mistral Large 2 is poised to drive the development of innovative AI applications.

Microsoft’s “SpreadsheetLLM”: AI in Spreadsheets

Microsoft's SpreadsheetLLM optimizes AI for understanding spreadsheets by using the SheetCompressor framework, which reduces token usage by up to 96%. This improves AI performance in table detection and question-answering tasks. SpreadsheetLLM enables natural language interaction with spreadsheets, democratizing data access and enhancing productivity. As a research project, it holds potential to revolutionize tools like Microsoft Excel, making data analysis more efficient and user-friendly.

OpenAI Announced GPT-4o Mini fine-tuning and SearchGPT

OpenAI has announced free fine-tuning for their GPT-4o mini model until September 23, enabling developers to enhance their AI applications, and introduced SearchGPT, a new AI-driven search engine that delivers conversational responses with real-time information from trusted sources. These innovations aim to make advanced AI more accessible and reshape online search, despite facing legal challenges over copyright use.

Google DeepMind’s JEST Method

DeepMind's JEST (Joint Example Selection and Trust) method significantly enhances AI training efficiency by requiring up to 13 times fewer iterations and 10 times less computational power than traditional methods. By selecting batches of data for training and starting with a highly curated initial dataset, JEST reduces energy consumption and environmental impact. This innovative approach promises to make AI development more accessible and sustainable, despite challenges for smaller developers in accessing high-quality initial data. As AI's power demands grow, JEST offers a critical solution for reducing costs and resource consumption in the competitive AI landscape.

Elon Musk’s xAI Unveils World’s Most Powerful AI Training Cluster

Nvidia: The IBM of the AI Era?

NVIDIA’s Leap from GPUs to Generative AI

Suleyman on AI: Grappling with ‘The Containment Problem’

👥 Connect & Feedback!

👉 Join Us:

📧 Advertise In Weekly AI News:

📧 Contact directly at [email protected]

😍 Share with your friends!