The AI Weekly Breakthrough | Issue 6 | April 17, 2024
Welcome to The AI Weekly Breakthrough, a roundup of the news, technologies, and companies changing the way we work and live
Meta Breaks Silence, Announces Llama 3
Meta, relatively quiet as of late, breaks its silence this week to make two AI-related announcements. First, Meta announces that Llama 3, its new suite of next-gen foundation models, will be released within the next month. Llama 3 is expected to have about 140 billion parameters, twice as many as are available in the largest Llama 2 model. Users should expect a release similar to Anthopic’s Claude 3, which included three models that differed in size and cost. Second, Meta announced the next generation of their Meta Training and Inference Accelerator (MTIA), their family of custom-made chips designed for Meta’s AI workloads. This is something we’re seeing all the big players do – improve and elevate their AI infrastructure in order to make bigger and better performing AI models.
March RAGness Results Are In
Tonic.ai has published the results of their March RAG leaderboard that looks surprisingly similar to another March bracket with an acronym. In this competition, Tonic tested the top RAG systems against each other. Using their Tonic Validate answer similarity score, they evaluated the response quality of each RAG system on a set of benchmark Q&As. Cohere placed first with a score of 4.62 out of 5, followed by Haystack in second place at 4.40, and Langchain landing in the third spot with a 4.0 score. At the bottom of the list were Amazon Bedrock at 3.16 and OpenAI’s November 2023 API version at a mere 2.42.
Cloudless Skies at Google Cloud Next
This year’s Cloud Next Conference, hosted by Google in Las Vegas, wasn’t so much about the cloud as it was about generative AI. The event showcased a suite of generative AI-powered tools across Google’s ecosystem, including Gemini Code Assist, Google’s answer to enhancing coding efficiency and quality, rivaling Github’s Copilot. Google also unveiled the ability to use generative text to create live images with Imagen 2, and its Vertex AI Agent Builder, which empowers developers to craft multimodal AI agents.Topping off the roster, Google announced that Gemini 1.5 Pro is in public preview on Vertex AI, accessible in over 180 countries and boasting an industry-leading one million token context window.
Rerank 3 Boosts Relevance, Lowers Costs
Cohere, a leader in AI models specifically designed for enterprise generative AI projects, has released Rerank 3. Working with a RAG pipeline, Rerank 3 helps assess the relevancy of retrieved documents before the LLM processes the retrieved data. This improves the overall relevance and accuracy of the LLM’s response and, since the LLM only has to process a smaller set of relevant data, it brings the cost down of the entire generative AI system. Improvements in Rerank 3 include a larger 4K token context window length; the ability to search over multi-aspect and semi-structured data like emails, codes and tables; and multilingual coverage for 100+ countries.
X Releases Grok 1.5V
X announced the release of Grok 1.5V, its first-generation multimodal AI model. In addition to Grok’s text capabilities, it can now process visual information like documents, diagrams, screenshots, and photographs. According to X, Grok performs comparably with other frontier models like GPT-4V, Gemini 1.5 Pro and Claude 3 Opus, and outperforms them all in their newly defined “RealWorldQA” benchmark that measures real-world spatial understanding.
Augment Yourself 🤖
- Find out how you can use AI for predictive analytics to make smarter business decisions
- Explore the benefits, applications and limitations of Reinforcement Learning from Human Feedback (RLHF)
- Read these practical ETL strategies to refine AI models and decision making
- Discover why Meta’s researchers think AI needs to be embodied to become truly intelligent
🔥 For more AI News brought to you via email subscribe to our newsletter here.
👀 Want to know more about Shelf’s suite of AI solutions? Check out our website here.
