Pondhouse Data AI - Tips & Tutorials for Data & AI 17

Enterprise AI Agent Case Studies | Custom LLMs Training Guide | AI-Ready Web Crawling | Architecture Diagrams with AI | Meta's Training Controversy

Author

Andreas Nigg
14 Jan

Hey there,

This week, we're featuring an excellent tutorial on training small LLMs, an innovative web crawling solution, great case studies on AI agent implementation, and much more. Plus, our handy tip of the week that could save you hours of development time!

Enjoy!

Cheers, Andreas & Sascha

In todays edition:

📚 Tutorial of the Week: Aligning and Training Small LLMs: A Practical Guide from Hugging Face

🛠️ Tool Spotlight: Crawl4AI: Smart Web Scraping for AI Applications

💡 Tips: Generate Architecture Diagrams Automatically with Swark

💼 Case Studies: Early Adopters of AI Agents

📰 Top News & Analysis: Meta's AI Training Controversy & Sam Altman's 2025 AI Workforce Prediction

Let's get started!

Find this Newsletter helpful?
Please forward it to your colleagues and friends - it helps us tremendously.

Tutorial of the week

Aligning and train small LLMs for your specific use case

This week's tutorial highlights a fantastic new offering from Hugging Face: "a smol course" (the name is kinda silly, but bare with us ;-) ).

While large language models have shown impressive capabilities, they often require significant computational resources and can be overkill for focused applications. Small language models offer several advantages for domain-specific applications:

  • Efficiency: Require significantly less computational resources to train and deploy

  • Customization: Easier to fine-tune and adapt to specific domains

  • Control: Better understanding and control of model behavior

  • Cost: Lower operational costs for training and inference

  • Privacy: Can be run locally without sending data to external APIs

  • Green Technology: Advocates efficient usage of resources with reduced carbon footprint

  • Easier Academic Research Development: Provides an easy starter for academic research with cutting-edge LLMs with less logistical constraints

Building on these advantages, Hugging Face's "a smol course" provides a focused, hands-on introduction to aligning these smaller language models without requiring massive resources or paid services. What makes this course particularly appealing is its accessibility. You can run everything on a standard local machine with minimal GPU requirements. It focuses on the SmolLM2 series but emphasizes that the skills learned are transferable to other models. Plus, it's completely free and open, encouraging community participation through a unique peer-review process.

The course structure is very practical. You're invited to fork the repository, work through the materials, perform exercises, and even add your own examples. By submitting your work as a pull request to the december_2024 branch, you contribute to the course and receive feedback. This is a great way to learn while contributing to a growing resource.

Here's a peek at the modules you'll find:

  • Instruction Tuning: Master supervised fine-tuning, chat templating, and the fundamentals of instruction following.

  • Preference Alignment: Dive into DPO (Direct Preference Optimization) and ORPO (Odds Ratio Preference Optimization) techniques for aligning models with human preferences.

  • Parameter-efficient Fine-tuning: Learn techniques like LoRA (Low-Rank Adaptation) and prompt tuning for efficient model adaptation.

  • Evaluation: Discover how to use automatic benchmarks and even create custom evaluations tailored to your specific domain.

  • Vision-language Models: Explore the exciting world of adapting multimodal models for tasks that involve both images and text.

  • Synthetic Datasets: Learn the art of creating and validating synthetic data to boost your training efforts.

We are aware that that this course is quite extensive. But if you ever want to run your own models - specialized and optimized for your specific use case - there is nothing better to get you going.

Ready to get started? You can find the course repository here.

Tool of the week

Crawl4AI - An LLM-Ready Web Scraping Solution

If you've ever tried to build AI applications that rely on up-to-date information from the web, you know the pain of dealing with modern, JavaScript-heavy websites. This week's featured tool, Crawl4AI, is designed to alleviate that headache.

Crawl4AI is an open-source web crawler specifically built with AI applications in mind. It addresses the challenge of dynamically loaded content by functioning like a real browser, executing JavaScript to capture the data you need. This is a significant step up from basic web scrapers that often miss crucial information loaded after the initial page load.

Two key features stand out: asynchronous processing for handling numerous requests quickly, crucial for large-scale data collection, and its seamless integration with LLMs. Instead of wrestling with messy HTML, Crawl4AI can deliver cleaner, structured data ready for your AI models.

Key Features:

  • Asynchronous processing for high-performance crawling

  • JavaScript rendering support

  • Direct LLM integration for intelligent data extraction

  • Clean output formats (HTML, Markdown, structured metadata)

  • Docker support for easy deployment

  • Built-in anti-detection features

Why It's Special:
What sets Crawl4AI apart is its AI-first approach. Instead of just scraping web pages, it can use LLMs (via providers like OpenAI or Azure) to intelligently extract specific information from crawled content. This makes it particularly valuable for applications like competitive analysis, research assistants, or e-commerce monitoring.

To get started, follow our free guide here.

Top News of the week

Case Studies: How Early Adopters Are Implementing Automated AI Agents

The business case for AI agents is becoming clearer as major corporations report their first successful deployments. These autonomous systems are now handling tasks that previously required significant human involvement, offering a glimpse of how companies might reshape their workflows in the coming years.

Key Implementation Strategies:

  • Start Small: Deutsche Telekom began with an internal "ask me anything" system for employee queries about policies and benefits, reaching 10,000 weekly users

  • Define Clear Boundaries: Cosentino treats AI agents as "digital workers" with specific skills, training requirements, and strict operational procedures

  • Multiple Agent Approach: Moody's deploys 35 different agents, each specialized for specific tasks, working under supervisor agents in a "multi-agent system"

  • Human Oversight: Johnson & Johnson maintains employee review of AI outputs while developing more systematic oversight processes

ROI Indicators:

  • Cosentino reports that AI agents have fully automated work previously requiring 3-4 full-time employees

  • eBay is seeing efficiency gains in code writing and marketing campaign creation

  • Pure Storage suggests executive-level support can now be extended to all employees through AI assistance

Risk Considerations:
According to Gartner, while 15% of business decisions will be AI-driven by 2028, companies need to prepare for increased security risks, with predictions of 25% of enterprise breaches being tied to agent abuse.

For businesses considering AI agent implementation, these early case studies suggest starting with well-defined, repetitive tasks while maintaining clear oversight protocols. The technology appears most successful when treated as an augmentation of existing workflows rather than a complete replacement of human workers.

Source: WSJ

Opinion: Small, Specialized LLMs - The Future of Practical AI

The AI industry's obsession with large language models like GPT-4 has created an unrealistic expectation of what AI can deliver. While these models are impressive in their knowledge, they often produce generic, inconsistent results - even with careful prompt engineering, testing, automated prompting etc. Their probabilistic nature means you can never fully rely on their output without human verification.

However, this shouldn't discourage us from using LLMs' - quite the contrary. LLMs are enormously helpful and useful. While there are certainly also use-cases for these huge LLMs, the real breakthrough in corporate AI usage lies in creating smaller, highly specialized models focused on specific use cases. Think of a model trained specifically to analyze customer support tickets and route them to the right department, or one that specializes in converting draft documents into accessible knowledge.

For instance, a specialized LLM could process incoming invoices, automatically matching them with purchase orders, flagging discrepancies, and routing them through the appropriate approval chain - replacing manual invoice processing workflows. Or consider a model trained to monitor customer communication channels, automatically categorizing and prioritizing urgent issues while routing routine requests to self-service solutions - streamlining traditional customer service triage processes.

The advantages of smaller, specialized models extend beyond just better results. They can run on local hardware, eliminating the reliability issues that plague large cloud-based models. Anyone who's dealt with GPT-4 in production knows the frustration of slow response times and service outages. With small, specialized models, you can deploy them close to your actual processes - even having dedicated models for each specific task. This not only improves reliability but also reduces latency significantly.

Furthermore, these smaller models are easier to fine-tune and align with specific business needs. You can train them on your domain-specific data and implement strict guardrails to ensure they operate within well-defined boundaries. This controlled environment is more than just a simple advantage for business applications where consistency and reliability are non-negotiable.

It's time to move beyond the one-size-fits-all approach and embrace the power of specialized AI.

To get started, we highly recommend the HuggingFace guides related to model fine-tuning and alignment. One is provided as part of this newsletter at the very top.

Also in the news

Zuckerberg Accused of Approving Pirated Books for AI Training

To no one's surprise, Meta's Mark Zuckerberg finds himself at the center of another controversy, this time accused of knowingly approving the use of pirated books to train the company's AI systems. Internal communications revealed in a lawsuit filed by authors Sarah Silverman and Ta-Nehisi Coates show that Zuckerberg greenlit the use of LibGen, a notorious pirated book database, despite warnings from his own AI team about potential regulatory backlash. The revelation is just the latest chapter in Big Tech's ongoing struggle with copyright laws in the AI era.

Read more about this story here.

Sam Altman predicts that AI Agents will “join the workforce” in 2025

OpenAI CEO Sam Altman has predicted that AI agents will begin "joining the workforce" as early as 2025, potentially transforming how businesses operate. The forecast came in a lengthy blog post reflecting on ChatGPT's second anniversary, where Altman also discussed the company's progress toward artificial general intelligence (AGI) and his vision of eventual superintelligent AI. The announcement follows a turbulent year for OpenAI, which saw Altman briefly ousted as CEO before being reinstated amid widespread employee support.

Find Altmans full blog post here.

Tip of the week

Automatically create architecture diagrams from your code.

Diving into a new codebase, especially a complex one, can feel like trying to navigate a maze blindfolded. It would be great to have a quick way to get at least a basic understanding of the architecture. Our this week's tip introduces Swark, a free and open-source VS Code extension that aims to do just that.

Swark leverages the power of large language models, specifically GitHub Copilot, to automatically generate architecture diagrams from your code. Think of it as asking an AI to draw you a map of your project. What's particularly appealing is its simplicity: if you're already using GitHub Copilot (which now has a free tier!), you're pretty much good to go. No extra authentication or API keys are needed.

Because Swark relies on the "intelligence" of the LLM, it boasts universal language support, a significant advantage over traditional code visualization tools. It generates diagrams in the popular Mermaid.js format, which you can then easily edit and refine.

Why might this be useful?

  • Quickly understanding unfamiliar code: Onboarding onto a new project? Swark can give you a high-level view of its structure in minutes.

  • Visualizing AI-generated code: As AI tools churn out more code, Swark can help you ensure it's architecturally sound.

  • Improving documentation: Keep your project documentation current with automatically generated diagrams.

  • Tackling legacy code: Unraveling the mysteries of older codebases becomes a bit less daunting with a visual guide.

  • Spotting potential design issues: Seeing the dependencies laid out can help identify areas for optimization or unwanted connections.

You can download the tool here - for free.

GitHub - swark-io/swark: Create architecture diagrams from code automatically using LLMs

github.com/swark-io/swark?

We hope you liked our newsletter and you stay tuned for the next edition. If you need help with your AI tasks and implementations - let us know. We are happy to help