LLM Resources

Webinars

• Navigating LLM Threats: Detecting Prompt Injections and Jailbreaks
  Good information on how to detect some common scenarios, they are pushing this library: https://github.com/whylabs/langkit and here is a jupyter notebook about some simple examples: https://colab.research.google.com/drive/1RMjiJK9Nd-tP7kBXo8h9A0vtCCdY1ikS?usp=sharing#scrollTo=7fm-oNxDJNGI

Courses

• Courses - DeepLearning.AI
  This is the main page of the courses offered by Deep Learning, following are a selection of courses I’ve taken and seems appropiate to our work
  
  

• Functions, Tools and Agents with LangChain
  Functions, Tools and Agents with LangChain - DeepLearning.AI
  Good resource to learn a bit more on langchain and also how the functions works in OpenAI
  
  

• LLM Evaluation
  courses/prompt_evaluations at master · anthropics/courses
  How to grade LLM results

Email newsletter

• AlphaSignal
  AlphaSignal | The Most Read Technical Newsletter in AI
  Daily newsletter with news, papers recommendations, new projects, selection of github repositories and some code examples

• DeepLearning.AI The Batch
  The Batch | DeepLearning.AI | AI News & Insights
  News, AI Concepts, explanation of papers, non tech stuff (ethics, impact of the AI) and announcements of new courses

Papers

• What is the Role of Small Models in the LLM Era: A Survey
  https://arxiv.org/pdf/2409.06857
  
  Examples of where Small Models (LSM) can be used, special interest:

  • 2.1.3 Efficient Inference
    Use Small Models to route input into proper LLM model since each model has strong and weak points so the Small Model can decide which to use, this strategy has been tested in some projects, but doing mostly threshold checks (so running LLM and testing the output to decide if to go into another LLM)
    
    

  • 2.1.6 Retrieval Augmented Generation
    Instead of using “generic” LLM for embeds, use a small model in specialized topics similar to what is used, we are actually using this in cogcache
    
    

  • 2.1.8 Deficiency Repair
    Using Small Model to handle things like mitigate hallucinations or safeguarding user privacy
    
    

  • 2.2.2 Data Synthesis
    Generate data from LLM to train smaller LLM, kind of segmenting a generic LLM into a specialized LSM