AI for Software Developers

Kursmaterialet är på engelska, med detta innehåll:

Day 1

AI in the software development industry

• Impact of AI on tech industry
• Reaction of the big players
• AI won't take your job (yet)
• From simple questions to working code
• AI improves software development
• AI will transform software as we know it
• Connection between AI and software developers

The AI landscape

• Welcome to the AI Revolution
• Demo - Github Copilot
  • Your AI Pair Programmer What can it do?
  • Writing Unit Tests
  • Writing Regular Expressions
• Demo - Amazon CodeWhisperer
  • AI Code Generator
• Tabnine – Tabnine
  • Code completion
  • Tabnine chat
• Exercise: Cloud VM Intro
• Exercise: OpenAI DALL-E 2 Introduction
  • Generating images with simple prompts
  • Calling DALL-E from the command prompt
• Demo - Stable Diffusion XL

Developer Plugins

• Demo - Advanced Data Analytics
  • How to enable Advanced Data Analytics?
  • Math examples
  • Upload and download files with data analysis
  • Create and interpret code

Further important AI concepts

• Fine-tuning and embedding

Prompt Engineering

• Fundamentals summarized
  • The 5 rules for prompting
• Intermediate techniques summarized
• Advanced techniques summarized
• Advanced techniques - SudoLang: pseudocode for LLMs
  • Introducing SudoLang
  • SudoLang examples
  • Behind the tech
  • A SudoLang program
  • Integrating an LLM
  • Integrating an LLM with SudoLang
• Exercise: Translate a prompt to SudoLang
• Tree of Thoughts - Advanced techniques
  • How does it work?
  • How can you experiment with this?
• Exercise: Tree of Thoughts with LangChain framework
• Limitations in development
  • Limitations of ChatGPT
  • Example - hallucinated answer
  • Example - the correct function

Copilot ecosystem

• GitHub Copilot
  • Inside GitHub Copilot
  • Prompt Engineering inside Copilot
  • Fill-In-the-Middle (FIM)
  • Life of a Completion
  • Copilot over the network
  • Improved semantic understanding with embedding
  • Features and use-cases
  • Productivity increase
  • Boosted coding speed
• GitHub Copilot – Autocomplete
  • Most trivial feature: autocomplete
  • Basic usage
• Exercise: Setting up GitHub Copilot
• Exercise: GitHub Copilot autocomplete
• GitHub Copilot - Prompt Engineering
  • Code generation from comments
• Exercise: Prompt Engineering in Copilot

Day 2

Copilot ecosystem

• GitHub Copilot – Testing
  • Writing unit tests
• Exercise: Testing with Github Copilot
  • First tests for an Express app
• GitHub Copilot - Unfamiliar territory and boilerplate
  • Boilerplate code generation
  • Unfamiliar framework or libraries
• Exercise: Writing boilerplate code
  • Write boilerplate code
  • Express boilerplate with Copilot
• Exercise: Unfamiliar framework or libraries
  • Calling an API
• Exercise: ChatGPT in your IDE
  • CodeGPT
• Copilot Limitations
• Amazon CodeWhisperer
  • Code with confidence
  • Enhance code security
  • Language and IDE compatibility
• Tabnine
  • Code privacy
  • Local machine mode
  • Tabnine for teams
  • Writing unit tests
• Comparison
  • Choose your fighter wisely

Building an AI integrated application

• Fresh Cart
  • Introducing Fresh Cart
  • The vision for Fresh Cart 2.0
  • How will we get there?
• Exercise: Fresh Cart Setup
  • Setting up the exercise
  • The perfect output
  • Fill the database with 100 ingredients
• Fresh Cart - Integrating an LLM chatbot
  • OpenAI API introduction
  • Context crafting and streaming
• Exercise: Fresh Cart chatbot
• Exercise: Recipe chatbot
• Fresh Cart - Prompt routing
  • Using multiple contexts
  • From routing to orchestration
• Exercise: Prompt routing
  • Prompt orchestration summary
• Exercise: Output formatting
• Fresh Cart - Basic Semantic Search
  • What is semantic search?
  • Semantic search vs. classical search?
• Exercise: Semantic search
• Fresh Cart - State mutation
  • Mutate state based on semantics
• Fresh Cart - Next steps
  • Improvements
  • Upcoming features
• Operating an AiLLM
  • Logging
  • Self correction
• Attacks on AI systems - Prompt injection
  • Prompt injection
  • Impact
  • Examples
  • Indirect prompt injection
  • From prompt injection to phishing
• Exercise: Prompt injection exercise
  • Prompt injection exercise Get the password!

Day 3

Working with models

• Foundation Models
  • What is a foundation model?
  • Foundation Models in practice
• Open-source models
  • Open-source models vs. closed-source models
  • Examples of open-source models
• Open-source models - Meta - Llama 2 Variants
  • Providers
  • Running a model locally
• Exercise: Running LLama 2 locally
  • LLama 2 at home
  • LLama 2 in the VM
  • Interact with a local model
• Open-source models - Stable Diffusion
  • Stability AI
  • Reminder: How diffusion models work
  • Stable Diffusion
  • Latent Diffusion model (LDM)
  • Advantages of using LDM
• Exercise: Local Stable Diffusion
  • Running a model using OnnxStream
• Open-source models - HuggingFace
  • Introduction
  • Models
  • Types of models
  • Model properties
  • Datasets
  • Spaces
  • The Open LLM Leaderboard
  • Benchmarks
  • AutoTrain
• Exercise: AutoTrain
  • Comparing fine-tuned models
  • A larger training set isn't always better
• Transformers - Hugging Face
  • What is Transformers?
  • Transformers Pipelines
• Exercise: Transformers Using the Transformers pipeline
  • Using the Transformers pipeline solution
• Inference Endpoints
  • The easiest way to deploy AI models
  • Inference Endpoints UI
  • The supported models of Inference Endpoints
• Exercise: Inference Endpoints
  • OpenChat with Inference Endpoints
  • Sending a request to the Endpoint
• Other models
  • BERT
  • Falcon family
  • OpenChat 3.5
  • Tons of other text-based models
  • CLIP - Connecting text and images
• Cloud providers - Microsoft Azure AI Tools
  • Azure AI services
  • Azure OpenAI Studio
  • Hosting a model
• Cloud providers - Google Cloud Platform
  • Google Vertex AI
  • PaLM 2
  • Vertex AI - Generative AI Studio
  • Vertex AI - Model Garden
• Cloud providers - Amazon AI Tools
  • Amazon Bedrock
  • Other Amazon tools for building generative AI applicatons
• Cloud providers - Replicate
  • Cloud platform for machine learning
  • Cog - Containers for machine learning
• Exercise: Cog containers Running a model with Cog

Machine Learning

• TensorFlow
  • Open-source machine learning platform
  • Keras
  • The main concepts of TensorFlow
  • Training loop in Keras
  • Machine learning on other devices
  • Kaggle Models / Model Garden
• Exercise: Training a model
  • Custom training
  • Saving and loading
  • Hyperparameter tuning
  • TensorFlow Serving with Docker

Advanced AI concepts

• Fine-tuning What is fine-tuning?
  • Parameters and hyperparameters
  • Behind the tech - How to fine-tune a model?
  • Customizing OpenAI models
  • Start a fine-tuning in OpenAI Playground
  • Fine-tuning with Vertex AI
  • Fine-tuning with Vertex AI - RLHF
  • Fine-tuning on AWS
• Exercise: Fine-tuning
  • Fine-tuning with OpenAI Playground
• Embedding
  • What is embedding?
  • Vector space
  • Euclidean distance (L2)
  • Euclidean distance in high-dimensions
  • The change of distance in high-dimensions
  • Cosine similarity
  • Dot product
  • Vector databases and indexes
  • Use cases
• Retrieval Augmented Generation
  • RAG basics
  • RAG in action
• Exercise: Consensus.app
  • What is Consensus.app?
  • Research like you never did before
• Exercise: Search with Perplexity AI
  • Research with Perplexity AI
• Exercise: Finding similar products in FreshCart
  • Explore the embedding flow
  • Embedding products
  • Query for similar products

Summary and closing - What did we learn?

• Coding assistants summarized
• AI-integrated applications summary
• Open-source models summary
• Cloud providers summary
• Machine learning summary
• Advanced concepts summarized