Llama-3 with 1 Million Context
PLUS: Cluade 3 for teams and iOS, Generative AI meets Jira and Confluence
Today’s top AI Highlights:
Anthropic releases a new Team Plan for businesses and an iOS app for Claude 3 on-the-go
Sanctuary AI and Microsoft team up to build smarter general-purpose robots
Llama 3’s context window expanded from 8K to 1 Million tokens!
Atlassian releases generative AI assistant for businesses to search for information, learn, and create AI agents
Call for Google Gemini’s help by typing “@Gemini” in Chrome
& so much more!
Read time: 3 mins
Latest Developments 🌍
Claude 3 Goes Mobile and Expands for Businesses 🤳
Anthropic has released a “team plan” making available its Claude-3 models to cater to the growing demand for AI solutions in enterprises. The company has also released the Claude iOS app, available on the App Store today. For $30 a month, the Team plan offers advanced features and controls specifically designed for businesses, while the iOS app brings the power of Claude 3 directly to your fingertips.
Key Highlights:
All models available: The Team plan will give access to all three Claude-3 models - Haiku for speed and cost efficiency, Sonnet for a balanced approach, and Opus for maximum intelligence. You can choose the model that suits your requirements the most.
Increased usage: Provides greater usage per user, priority access during high-traffic periods, and early access to new features, compared to Claude’s Pro plan which costs $20 a month. The usage limit depends on the message length, conversation length, and Claude’s current capacity.
Context window benefit: The teams can benefit from Claude 3’s huge context window of 200K tokens. This is important for businesses where documents can be long and complex, and long conversations are necessary to maintain.
Other features: It also includes administrative tools to conveniently control user and billing management for easier onboarding.
Upcoming features: Features like citations for AI-generated content, integration with data repositories, and a collaborative workspace for team projects will be rolled out soon.
Next-Gen Robots Will Learn from the Real World 🏕️
The future of AI is embodied AI agents, essentially general-purpose humanoid robots, that can navigate the complex physical world just like we do and can effectively reason their actions. Onto this path is Sanctuary AI, a Canada-based robotics company. They have partnered with Microsoft to speed up the development of Phoenix, Sanctuary AI’s robot, utilizing Microsoft Azure cloud resources for AI workloads.
Key Highlights:
Large Behavior Models (LBMs): Moving beyond language-based AI, Sanctuary AI is developing LBMs that enable robots to understand and learn from physical interactions and experiences.
Microsoft Azure: Access to Microsoft’s robust cloud infrastructure will provide Sanctuary AI with the computational power and resources needed to train and refine its advanced AI models efficiently.
The Next-Gen Phoenix: Sanctuary AI has also recently unveiled the latest iteration of the Phoenix robot with major improvements, including increased uptime, enhanced visual and tactile sensing, and a wider range of motion, allowing for more complex task execution and faster task automation.
Llama 3 on Par with Gemini Pro and GPT-4 🧠
Llama 3 models are certainly impressive but their limited context window of 8k tokens makes them inapplicable for tasks that have longer pieces of text like research papers or reports.
However, two separate teams have tackled this challenge head-on, significantly expanding the context window of Llama 3 models. First, Gradient AI has increased the context window of Llama 3 8B to a staggering 1 Million tokens, making it the second LLM after Gemini Pro to have such a huge capacity.
Second, Abacus AI has released 128K long-context support for Llama 3 70B, making it head-to-head with GPT-4. Both the models are available on Hugging Face.
Gradient AI’s Llama 3 8B Gradient Instruct 1048k:
Token Milestone: Gradient AI has taken the Llama 3 8B model to a whole new level by extending its context window to over 1 million tokens! This is a staggering 128x increase from the original, opening doors for processing massive amounts of information.
Training: Their approach involved NTK-aware interpolation and progressive training on increasingly longer contexts. This allowed them to achieve these remarkable results with minimal additional training data.
Abacus AI’s Llama-3-Giraffe-70B:
Extended Reach: They’ve boosted the context window of the Llama 3 70B model to a whopping 128k tokens, which is 16x longer than before, making it suitable for tasks like analyzing lengthy documents or generating more comprehensive and coherent text.
Training: Abacus AI employed a clever combination of Positional Skip-wise Training and dynamic-NTK interpolation to achieve this. They plan to push it further, aiming for a 1M token context window soon.
Search Information, Learn, and Create Virtual Teammates 👨👦👦
Finding the information you need from several data sources in your company can feel like searching for a needle in a haystack. Atlassian recognized this and has introduced Rovo, their AI-powered solution to turn information into action. While several companies offer AI-powered search solutions, Atlassian’s Rovo goes beyond simple search by acting as a central hub for knowledge discovery. Rovo not only connects data from both Atlassian and third-party tools for searching information, it also lets you build Rovo AI agents that act like your virtual teammates to automate routine tasks.
Key Highlights:
Unified Search: Rovo eliminates the need to jump between different apps by offering a centralized search function that includes Atlassian tools like Jira and Confluence, along with third-party platforms like Google Drive, Slack, and GitHub.
AI-Powered Insights and Learning: Beyond search, Rovo provides knowledge cards with contextual information about projects, goals, and teammates. It also features a chat interface for interactive learning and clarification of company-specific jargon.
AI Teammates: These AI agents can be readily integrated into existing workflows to automate tasks, generate content, and provide recommendations. Users can build custom agents using a simple natural language interface, requiring no coding expertise.
Security and Permissions: Rovo ensures that search results and agent actions adhere to established access permissions, guaranteeing data privacy and security.
😍 Enjoying so far, share it with your friends!
Building Semantic Search apps with RAG
Learn the quickest way to build and deploy semantic search solutions with minimal code in this FREE session on May 3, 2024. It is for anyone looking to enhance user experience, build smarter chatbots, and stay competitive in the rapidly evolving digital landscape.
Tools of the Trade ⚒️
Chat with Gemini shortcut: Google has added a shortcut on Chrome to chat with Gemini. Just type “@gemini” in the address bar and select Chat with Gemini. Put in your prompt and get your answer! This will let you get help from AI without navigating to a separate website.
Framedrop: Convert highlights from videos into reels, TikToks, clips, or anything else. Just paste the URL, and Framedrop will automatically identify and extract highlights from videos and lets you quickly create short-form content like TikToks, YouTube shorts, and Instagram reels.
Brainy Docs: Convert PDFs into videos, creating engaging videos and presentations in seconds. It uses AI to extract content from PDFs, including text and images, and generate summaries and detailed explainer videos that can be customized, downloaded, and shared.
Martin: A better Siri with an LLM brain. Through voice conversations, Martin builds a personal relationship with you. He can find information, brainstorm ideas, or just chat with you about your week! You can also ask Martin to send you notes or set reminders in any conversation.
Hot Takes 🔥
Originally in the process of saving up for buying a house,
But now, with AI and robots, possible Singularity, Abundance around the corner, or war?
…
I don’t know any more.
How will AGI/ASI affect real estate?
If abundance achieved, I’d assume house worth $0 ~
Annamy best guess:
gpt4 knowledge + q* search reasoning = gpt2
general knowledge seems near identical to gpt4, with much better reasoning and planning capabilities.
more expensive inference from tree of thought search would explain relatively slow inference and low rate limit. ~
Siqi Chen
Meme of the Day 🤡
How AI companies collect publicly available data for model training.
That’s all for today! See you tomorrow with more such AI-filled content.
Real-time AI Updates 🚨
⚡️ Follow me on Twitter @Saboo_Shubham for lightning-fast AI updates and never miss what’s trending!
PS: I curate this AI newsletter every day for FREE, your support is what keeps me going. If you find value in what you read, share it with your friends by clicking the share button below!