Pair Programming with AI

Transcript

1. Pair Programming with AI Rebecca Vickery, Lead Data Scientist, EDF
2. Data Models Computational power AI is the ability for computers to think and learn
3. Generative AI Images Video/Audio Text
4. Worlds First Generative AI • ELIZA – early natural language processing program created between 1964 and 1966. • Created using pattern matching and substitution.
5. Data growth Model sophistication Greater computational power A new era of generative AI
6. How Does Generative AI Work? Large volumes of relevant data e.g., text, code, images, video, audio Training – neural networks. Learns patterns in the data to create algorithms Algorithms are used to generate responses e.g., text, code, images
7. Foundational Models Large model trained on massive amount of data e.g., GPT, BERT, DALL-E Github co-pilot, codewhisperer Duolingo Max, GrammarlyGo Niche tasks Further training on task-specific data General purpose tasks
8. GPT (Generative Pre-trained Transformer) Around 500 billion “tokens”. Text from books, articles, documents, the internet Neural networks + randomness + reinforcement learning Algorithms are used to generate text-based responses Large language model (LLM)
9. Neural networks + randomness + reinforcement learning platform.openai.com/tokenizer
10. Use Cases Data science Team Writing
11. Pair Programming with AI - Github Copilot • Generates code from comments, provides suggestions, auto-complete. • Works on most popular programming languages. • Integrates with popular IDE’s e.g., VSCode, PyCharm.
12. Github Co-pilot GPT foundational model Fine tuning with open source code including public Github repositories Algorithms are used to generate code
13. Github Copilot – Write a Function
14. Github Copilot – Generate the Doc String
15. Github Copilot – Write Tests
16. CodeWhisperer • Trained on Amazon specific code. • Supports most popular programming languages. • Available as part of the AWS Toolkit and can be used in most popular IDE’s. • Particularly good for AWS based code.
17. chatGPT • Not trained specifically but still makes a good coding assistant. • Good for understanding code bases. • Re-writing functions and modules into other programming languages.
18. Pair Programming with AI - Drawbacks • Bias. • Reinforcement learning means models are reluctant to say, “I don’t know”. Generates responses that look correct but are wrong. • Be careful entering sensitive information into these tools.
19. The Future of Generative AI • Growth in the use of foundational models. Expect to see more niche code assistants. • An important future skill will be prompt engineering. • Generative AI will become more sophisticated at an exponential rate. • Companies will go through an AI transformation like the digital transformation in recent years. • Increased demand for tech skills.
20. Thank you
21. Bias - Examples