Power of Llama 2 with 32k Context Length 🤯
PLUS: Understanding Matrices in 3D, More Customized and Extensible AI Agents
Today’s top AI Highlights:
Text-intensive Image Analysis with Microsoft’s Model
PyTorch’s Visualization Tool for Matrix Multiplication
Abacus AI releases Giraffe 70B with 32k Context Length
LoopGPT - A Modular Auto-GPT Framework
& so much more!
Read time: 3 mins
Latest Developments 🌍
Understanding Text-Rich Image 📜
Microsoft has introduced Kosmos-2.5, a multimodal literate model designed for machine reading of text-intensive images. It excels in two primary tasks: generating spatially-aware text blocks and producing structured text output in markdown format.
Key Highlights:
Unlike traditional LLMs, Kosmos-2.5 combines visual and textual information in a single model. It achieves its multimodal capabilities through a shared Transformer architecture, task-specific prompts, and flexible text representations.
The architecture comprises a pre-trained vision encoder based on the Vision Transformer (ViT) and a language decoder connected with a resampler module.
Kosmos-2.5 demonstrates strong performance in various text-intensive image understanding tasks, including few-shot and zero-shot learning scenarios.
3D Visualization of Matrix Multiplication 📦
PyTorch has introduced a visualization tool called "mm" that helps visualize matrix multiplication expressions (matmuls), a basic building block in machine learning, and compositions of matmuls in 3D. It's designed to provide a more intuitive understanding of matmuls.
Key Highlights:
The "mm" tool is fully interactive and can be run in a browser or notebook iframe. It retains its complete state in the URL, making it easy to share sessions with others.
This concept is illustrated by wrapping the matmul around a cube, which helps in understanding the relationships between argument shapes, result shape, and shared dimensions.
The tool extends its capabilities to compound expressions, allowing users to visualize and understand intricate compositions of matmuls within neural networks.
Giraffe 70B Model with 32K Context Window 🦒
Abacus AI has released 70 billion parameter Giraffe model, succeeding the 13B model. Giraffe is part of a model family that is fine-tuned from base Llama 2 with an extended context length from 4k to 32k.
Key Highlights:
The 70B Giraffe model achieves a remarkable 61% accuracy compared to the 18% accuracy of the 13B model on the AltQA dataset at the longest context windows. It also outperforms the LongChat-32k model at all context lengths.
The model was also evaluated on the MT-Bench evaluation set, examining LLM performance in multi-turn settings across various categories, where it emerged as a top performer in writing, coding, and math.
To scale LLMs to longer contexts, Abacus AI introduces the concept of context length extrapolation, unlocking the potential for complex retrievals, sustained conversations, and large-scale coding tasks without the need for extensive training.
AI Agents with More Modularity and Compatibility 🔌
Introducing LoopGPT, a re-implementation of the popular Auto-GPT project, designed as a Python package with modularity and extensibility in mind.
LoopGPT offers an extensible and modular framework that is easy to work with, allowing you to add new features, integrations, and custom agent capabilities directly from Python code, without the need for complex config files.
LoopGPT is designed to work effectively with GPT-3.5, providing excellent results for those who don't have access to GPT-4.
It enables the ability to "course correct" agents through human feedback, ensuring more accurate and reliable responses.
Tools of the Trade ⚒️
YouAgent by You.com: An AI agent with code execution capabilities, allowing it to answer STEM questions accurately by writing and running code in a computing environment.
Peslac: AI/ML Insurance Platform that simplifies insurance management through streamlined onboarding, real-time identity verification, fraud detection, risk assessment.
Magic Loops: Simplifies automation by combining LLMs and code to create customizable workflows for repetitive tasks, offering full control and various triggering options.
Silimate: AI copilot semiconductor chip designers that automates workflows, reducing designers' time spent on debugging and scripting.
Eightfold AI: AI-driven talent solutions for recruiting and talent management, offering skill-based insights and workforce flexibility.
😍 Enjoying so far, TWEET NOW to share with your friends!
Hot Takes 🔥
no real advancement has been in theoretical physics since 1997’s maldacena duality / holography principle ~ roon
Generative AI models are the worst they will ever be. ~ Emad Mostaque
A 10X engineer is not just someone who can do other engineer’s job 10X better. A 10X engineer is someone who makes 10 other engineers hate their job. ~ Bojan Tunguz
Meme of the Day 🤡
That’s all for today!
See you tomorrow with more such AI-filled content. Don’t forget to subscribe and give your feedback below 👇
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!