Microsoft AI’s Skeleton Key, AutoML with AutoGluon, MultiOn AI's Retrieve API, Narrative BI’s Hybrid AI, Python's Duck Typing, Gibbs Diffusion

Subscribe to our Data Pro newsletter for the latest insights. Don't miss out – sign up today!

👋 Hello,

Happy Friday! 

Welcome to DataPro#101—Your Essential Data Science & ML Update! 🚀  

This week, we’ve curated the latest techniques in data extraction, transforming unstructured data into structured formats, best practices for prompt engineering in NL2SQL, and much more. Consider this your all-in-one guide to staying informed in the ever-evolving world of data science and machine learning. Now, dive in and explore these exciting new ideas! 

⚡ Tech Highlights: Stay Updated! 

Prompt Engineering with Claude 3: Learn hands-on techniques on Amazon Bedrock. 

Accelerated PyTorch: Boost models with torch.compile on AWS Graviton. 

BigQuery Data Canvas: Perfect your prompts. 

Skeleton Key AI: New AI jailbreak method. 

GraphRAG: Complex data discovery tool on GitHub. 

📚 New from Packt Library 

Data Science for Web3 - Guide to blockchain data analysis and ML. 

🔍 Latest in LLMs & GPTs 

NASA-IBM's INDUS Models: Advanced science LLMs. 

EvoAgent: Evolutionary multi-agent systems. 

Kyutai's Moshi: Real-time AI model. 

MultiOn AI's Retrieve API: Accurate web search. 

Gibbs Diffusion (GDiff): Bayesian image denoising. 

Narrative BI’s Hybrid AI: Business data analysis. 

WildGuard: Safe LLM interactions. 

ProgressGym: Ethical AI alignment. 

OmniParse: Structuring unstructured data for GenAI. 

✨ What's Fresh 

Claude 3.5 Sonnet Use Cases: Future AI capabilities. 

Explainability in ML: Make models understandable. 

Group-By Aggregation: Powerful EDA tool. 

OpenAI and PandasAI: Series operations. 

AutoML with AutoGluon: ML in four lines of code. 

Python's Duck Typing: Flexible coding concept. 

🔰 GitHub Finds: Add These Repos 

fal/AuraSR 

arcee-ai/Arcee-Spark-GGUF 

pprp/Pruner-Zero 

ruiyiw/patient-psi 

hrishioa/rakis 

ragapp/ragapp 

Doriandarko/claude-engineer 

hao-ai-lab/MuxServe 

DataPro Newsletter is not just a publication; it’s a complete toolkit for anyone serious about mastering the ever-changing landscape of data and AI. Grab your copy and start transforming your data expertise today! 

📥 Feedback on the Weekly Edition

Take our weekly survey and get a free PDF copy of our best-selling book, "Interactive Data Visualization with Python - Second Edition."

We appreciate your input and hope you enjoy the book!

Share your Feedback!

Cheers,
Merlyn Shelley
Editor-in-Chief, Packt

Sign Up | Advertise | Archives

🔰 Data Science Tool Kit 

➔ ️ fal/AuraSR: AuraSR, a GAN-based super-resolution model for upscaling images. Implemented in PyTorch, it's inspired by the GigaGAN paper, enhancing image quality significantly. 

➔ arcee-ai/Arcee-Spark-GGUF: Arcee Spark, a 7B model from Qwen2, excels with fine-tuning and DPO, outperforming GPT-3.5 on tasks, ideal for efficient AI deployment. 

➔ pprp/Pruner-Zero: Pruner-Zero automates symbolic pruning metric discovery for Large Language Models, surpassing current methods in language modeling and zero-shot tasks. 

➔ ruiyiw/patient-psi: Patient-Ψ uses Large Language Models to simulate patient interactions for training mental health professionals, emphasizing cognitive modeling and practical deployment. 

➔ hrishioa/rakis: Rakis is a browser-based permissionless AI inference network enabling decentralized consensus without servers, emphasizing open-source and educational use. 

➔ ragapp/ragapp: RAGapp simplifies enterprise use of Agentic RAG models, configurable like OpenAI's custom GPTs, deployable via Docker on cloud infrastructure. 

➔ Doriandarko/claude-engineer: Claude Engineer, powered by Anthropic's Claude-3.5-Sonnet, aids software development through an interactive CLI blending AI model capabilities with file operations and web search. 

➔ hao-ai-lab/MuxServe: MuxServe efficiently serves multiple LLMs using spatial-temporal multiplexing, optimizing memory and computation resources based on LLM popularity and characteristics. 

📚 Expert Insights from Packt Community

Data Science for Web3: A comprehensive guide to decoding blockchain data with data analysis basics and machine learning cases 

By Gabriela Castillo Areco 

Understanding the blockchain ingredients 

If you have a background in blockchain development, you may skip this section. Web3 represents a new generation of the World Wide Web that is based on decentralized databases, permissionless and trustless interactions, and native payments. This new concept of the internet opens up various business possibilities, some of which are still in their early stages. 

Currently, we are in the Web2 stage, where centralized companies store significant amounts of data sourced from our interactions with apps. The promise of Web3 is that we will interact with Decentralized Apps (dApps) that store only the relevant information on the blockchain, accessible to everyone. 

As of the time of writing, Web3 has some limitations recognized by the Ethereum organization: 

Velocity: The speed at which the blockchain is updated poses a scalability challenge. Multiple initiatives are being tested to try to solve this issue. 

Intuition: Interacting with Web3 is still difficult to understand. The logic and user experience are not as intuitive as in Web2 and a lot of education will be necessary before users can start utilizing it on a massive scale. 

Cost: Recording an entire business process on the chain is expensive. Having multiple smart contracts as part of a dApp costs a lot for the developer and the user. 

Blockchain technology is a foundational technology that underpins Web3. It is based on Distributed Ledger Technology (DLT), which stores information once it is cryptographically verified. Once reflected on the ledger, each transaction cannot be modified and multiple parties have a complete copy of it. 

Two structural characteristics of the technology are the following: 

It is structured as a set of blocks, where each block contains information (cryptographically hashed – we will learn more about this in this chapter) about the previous block, making it impossible to alter it at a later stage. Each block is chained to the previous one by this cryptographic sharing mechanism. 

It is decentralized. The copy of the entire ledger is distributed among several servers, which we will call nodes. Each node has a complete copy of the ledger and verifies consistency every time it adds a new block on top of the blockchain. 

This structure provides the solution to double spending, enabling for the first time the decentralized transfer of value through the internet. This is why Web3 is known as the internet of value. 

Since the complete version of the ledger is distributed among all the participants of the blockchain, any new transaction that contradicts previously stored information will not be successfully processed (there will be no consensus to add it). This characteristic facilitates transactions among parties that do not know each other without the need for an intermediary acting as a guarantor between them, which is why this technology is known as trustless. 

The decentralized storage also takes control away from each server and, thus, there is no sole authority with sufficient power to change any data point once the transaction is added to the blockchain. Since taking down one node will not affect the network, if a hacker wants to attack the database, they would require such high computing power that the attempt would be economically unfeasible. This adds a security level that centralized servers do not have. 

This excerpt is from the latest book, "Data Science for Web3: A comprehensive guide to decoding blockchain data with data analysis basics and machine learning cases” written by Gabriela Castillo Areco. Unlock access to the full book and a wealth of other titles with a 7-day free trial in the Packt Library. Start exploring today!   

Read Here!

⚡ Tech Tidbits: Stay Wired to the Latest Industry Buzz! 

AWS  

➤ Prompt engineering techniques and best practices: Learn by doing with Anthropic’s Claude 3 on Amazon Bedrock. In this blog post, the focus is on crafting effective prompts for generative AI models to achieve desired outputs. It emphasizes the importance of well-constructed prompts in guiding models like Claude 3 Haiku on Amazon Bedrock to produce accurate and relevant responses, showcasing examples of prompt variations and their impact. 

➤ Accelerated PyTorch inference with torch.compile on AWS Graviton processors. In this blog post, AWS optimized PyTorch's torch.compile feature for AWS Graviton3 processors, significantly enhancing performance for Hugging Face and TorchBench model inference compared to the default eager mode. These optimizations, available from PyTorch 2.3.1, aim to streamline model execution on Graviton3-based Amazon EC2 instances. 

Google

➤ How to write prompts for BigQuery data canvas?  This blog post focuses on leveraging generative AI, specifically Gemini in BigQuery, to perform data tasks via natural language queries (NL2SQL and NL2Chart). It highlights how refining NL prompts can enhance query accuracy, promoting collaboration and efficiency among data professionals using BigQuery's data canvas tool. 

Microsoft

➤ Microsoft AI Unveils Skeleton Key: A Novel Generative AI Jailbreak Method. This blog post discusses a newly discovered type of attack in generative AI called Skeleton Key, also known as Master Key. It explores how this attack bypasses AI guardrails, allowing models to generate unauthorized content, and outlines Microsoft's mitigation strategies using Prompt Shields in Azure AI. 

➤ GraphRAG: New tool for complex data discovery now on GitHub. The update introduces GraphRAG, a graph-based approach to retrieval-augmented generation (RAG), now available on GitHub. It enhances information retrieval and response generation by automating knowledge graph extraction from text datasets, offering structured insights for global queries. An Azure-hosted API facilitates easy deployment without coding. 

Email Forwarded? Join DataPro Here!

🔍 From Bits to BERT: Keeping Up with LLMs & GPTs 

🔸 NASA-IBM Collaboration Develops INDUS Large Language Models for Advanced Science Research. The blog explores NASA's collaboration with IBM to develop INDUS, a suite of specialized language models (LLMs) tailored for scientific domains. INDUS enhances data analysis, retrieval, and curation across Earth science, heliophysics, and more, advancing research capabilities in diverse scientific disciplines. 

🔸 EvoAgent: Expanding Expert Agents to Multi-Agent Systems with Evolutionary Algorithms. EvoAgent automates the extension of expert agents to multi-agent systems using evolutionary algorithms, applicable to any LLM-based agent framework. It enhances agent diversity and performance across tasks, exemplified in debates by generating varied opinions and improving content quality dynamically. 

🔸 Kyutai Releases Moshi: A Real-Time AI Model that Understands and Speaks. Kyutai introduces Moshi, a real-time native multimodal foundation model surpassing GPT-4o functionalities. Moshi understands emotions, speaks with accents like French, and handles dual audio streams, enabled by joint pre-training on text and audio. It supports open-source transparency and runs efficiently on consumer hardware. 

🔸 MultiOn AI's Retrieve API Boosts Web Search with Real-Time Accuracy for Advanced Applications. MultiOn AI has launched the Retrieve API, a cutting-edge tool for autonomous web information retrieval. It enhances data extraction from web pages with real-time processing, catering to diverse applications such as personalized shopping assistants, automated lead generation, and content creation tools, setting new standards in web data extraction technology. 

🔸 Gibbs Diffusion (GDiff): A Bayesian Blind Denoising Method for Images and Cosmology. The study introduces Gibbs Diffusion (GDiff) as an innovative method for blind denoising with deep generative models. It enables simultaneous sampling of signal and noise parameters, improving Bayesian inference for scenarios like natural image denoising and cosmological data analysis, enhancing accuracy in noise characterization and signal recovery. 

🔸 Narrative BI Introduces Hybrid AI Approach for Business Data Analysis: The research explores hybrid approaches in business data analysis, combining rule-based systems' precision with Large Language Models' (LLMs) pattern recognition. This integration aims to generate actionable insights from complex datasets, improving efficiency and accuracy in decision-making processes for businesses. 

🔸 WildGuard: A Lightweight Moderation Tool for User Safety in LLM Interactions. The paper introduces WildGuard, an open and lightweight moderation tool for enhancing safety in Large Language Models (LLMs). It focuses on identifying malicious intent in user prompts, detecting safety risks in model responses, and evaluating model refusal rates. WildGuard achieves state-of-the-art performance across these tasks, addressing critical gaps in existing moderation tools.  

🔸 ProgressGym: ML Framework for Ethical Alignment in Frontier AI. This research addresses the influence of AI systems, particularly large language models (LLMs), on human epistemology and societal values. It introduces progress alignment as a technical solution to prevent AI reinforcement of problematic moral beliefs. ProgressGym, an experimental framework, facilitates learning from historical data to advance real-world moral decision-making challenges. 

🔸 OmniParse: AI Platform for Structuring Unstructured Data for GenAI Applications. OmniParse tackles the challenge of managing diverse unstructured data types—documents, images, audio, video, and web content—by converting them into structured formats optimized for AI applications. It integrates various tools like Surya OCR and Florence-2 for accurate data extraction, enhancing workflow efficiency and data usability across platforms. 

✨ On the Radar: Catch Up on What's Fresh

🔹 10 Use Cases of Claude 3.5 Sonnet: Unveiling the Future of Artificial Intelligence AI with Revolutionary Capabilities. Claude 3.5 Sonnet by Anthropic AI marks a leap forward in AI capabilities, showcasing versatility across diverse domains. It excels in generating n-body particle animations, interactive learning dashboards, escape room experiences, virtual psychiatry, interactive poster designs, educational visual demonstrations, customizable calendar applications, real-time object detection, financial tools, and advanced physics simulations. 

🔹 Explainability, Interpretability and Observability in Machine Learning: The article explores the nuances of machine learning (ML) transparency through concepts like explainability, interpretability, and observability. It discusses their definitions, distinctions, and importance in fostering trust, accountability, and effective deployment of ML models across various industries and applications. 

🔹 A Powerful EDA Tool: Group-By Aggregation. The article dives into Exploratory Data Analysis (EDA) techniques, focusing on group-by aggregation in Pandas. Using the Metro Interstate Traffic dataset as an example, it demonstrates how to derive insights such as monthly traffic progression, daily traffic profiles, hourly traffic patterns by weekday versus weekend, and identifying top weather conditions associated with congestion rates. 

🔹 Using OpenAI and PandasAI for Series Operations: This article explores PandasAI, leveraging AI models like OpenAI to enhance Pandas data manipulation tasks. It covers querying Series values, creating new Series, conditional value setting, and reshaping data using natural language commands. Examples include summarizing statistics, conditional operations, and reshaping COVID-19 and NLS youth study datasets efficiently. 

🔹 AutoML with AutoGluon: ML workflow with Just Four Lines of Code. The article explores AutoGluon, an automated machine-learning framework developed by Amazon Web Services (AWS). It discusses how AutoGluon simplifies the entire machine-learning process—from data preprocessing to model selection and hyperparameter tuning—making it accessible and efficient for users across various data types like tabular, text, and image data. 

🔹 Understanding Python's Duck Typing: The article explores the concept of duck typing in Python, emphasizing behavior over type. It allows objects to be used based on their methods rather than explicit types, promoting flexibility and polymorphism. Duck typing simplifies code but requires careful handling to avoid runtime errors. 

See you next time!