Today’s top AI Highlights:

1. ChartGemma - A small powerful opensource model for chart understanding

2. Go global with search - Jina AI’s multilingual retriever

3. Humanoid robots for homes are coming earlier than we expect

4. Train FLUX AI model with LoRA for just $5 in a few minutes

5. Build and share interactive web apps with no-code

& so much more!

Read time: 3 mins

Latest Developments

Chart Understanding through Visual Instruction-tuning 📈

ChartGemma, a new opensource multimodal model, excels at understanding and reasoning about charts in real-world applications. Instead of relying on the data tables behind charts, ChartGemma learns directly from the chart images themselves, capturing visual trends and patterns more effectively. Built on Google's small vision language model PaliGemma, it utilizes a large alignment dataset and a compact model size to achieve impressive results.

Key Highlights:

1. Image-based learning - ChartGemma is trained using instructions generated from chart images, enabling it to grasp visual information better than methods that depend on data tables. This allows it to analyze charts even when the underlying data is unavailable.

2. Stronger foundation - It benefits from a strong connection between its visual and language processing components. This allows it to interpret a wide range of charts used in real-world scenarios and understand complex relationships between visual elements and text.

3. Performance - Despite its smaller size compared to existing chart understanding models, ChartGemma achieves state-of-the-art results in summarization, question answering, and fact-checking. The code, model checkpoints, dataset, and demos are available here.

Building a Better Search Engine with Jina ColBERT 🕸️

Jina AI has released Jina ColBERT v2, a new retrieval model built upon the ColBERT architecture. This model provides improved search accuracy, supports numerous languages, and customizable output settings, making it a valuable resource for creating search applications that work across the globe. It achieves this high performance while remaining as efficient as traditional retrieval methods. You can access Jina ColBERT v2 through the Jina API, Hugging Face, Qdrant, and other platforms.

Key Highlights:

1. Improved Search Accuracy - Jina ColBERT v2 shows a 6.5% improvement in search accuracy compared to the original ColBERT v2, based on tests using 14 standard English benchmarks.

2. Wide Language Support - The model supports 89 languages, including commonly spoken languages like Arabic, Chinese, Japanese, Russian, and Spanish, as well as programming languages.

3. Customizable Output - Jina ColBERT v2 uses a technique called Matryoshka representation learning, allowing users to choose between different output sizes (128, 96, or 64 dimensions). This lets you balance the need for precise results with the need for fast and efficient processing.

4. Handles Longer Text - Jina ColBERT v2 can process documents up to 8192 tokens long, a significant upgrade from the original ColBERT v2's limit of 512 tokens. This means it can handle much larger and more complex pieces of text.

Quick Bites

1. AI robotics startup 1X revealed its latest humanoid robot Neo Beta, a prototype of its bipedal humanoid designed for home use. 1X aims Neo to be safe to work among people and be fully capable of performing a wide range of tasks in diverse environments.

2. Cohere has released upgraded versions of its Command R and Command R+ models built for enterprise RAG. These versions bring significant improvements in efficiency, affordability, and performance, with enhancements in coding, math, reasoning, and latency.

3. Fal.ai has reduced the price of its FLUX LoRA trainer. You can now train FLUX text-to-image model to generate your images for just credits worth $5 in a few minutes.

4. Amazon’s new Alexa will be powered by Anthropic’s Claude AI due to struggles with its own in-house AI models. The improved version, called “Remarkable Alexa,” is expected to launch in mid-October with new features and may require a subscription.

Tools of the Trade

1. V0: Build and share interactive web applications or "Blocks" using React, HTML, Markdown, and runnable code. You can also share these Blocks, fork existing ones, and get insights into Block popularity with view counts.

2. RagBuilder: A toolkit to create optimal production-ready RAG setup for your data automatically. It achieves this by tuning parameters like chunking and embedding models against a test dataset.

3. CodeViz: A VS Code extension that creates an interactive map of codebases - from high-level system architecture to function calls to give you deep contextual knowledge of the codebase.

4. Awesome LLM Apps: Build awesome LLM apps using RAG to interact with data sources like GitHub, Gmail, PDFs, and YouTube videos through simple text. These apps will let you retrieve information, engage in chat, and extract insights directly from content on these platforms.

Hot Takes

1. all the people working on agents are actually just doing neurosymbolic AI, but they wouldn't admit it in a million years because of its association with g*ry m*rcus ~
   James Campbell
   

2. If California had passed an Internet safety bill like SB 1047 in 1994, none of the Internet giants would be there and Silicon Valley would be a backwater. ~
   Pedro Domingos
   

3. prediction: microsoft will acquire cursor. ~
   Santiago

Meme of the Day

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!

Unwind AI - Twitter | LinkedIn | Instagram | Facebook

Awesome LLM Apps | Sponsor Us

PS: We curate this AI newsletter every day for FREE, your support is what keeps us going. If you find value in what you read, share it with at least one (or 20) of your friends!