Confoo 2024 Gettings started with OpenAI and data science
Download as PPTX, PDF0 likes272 views
The document provides steps for getting started with using AI to assist with data science work. It outlines 5 steps: 1) Find a large language model (LLM) that can be installed in your integrated development environment (IDE), 2) Install the extension or library for the LLM in your IDE, 3) Try a simple "hello world" type command to test it, 4) Evaluate the code suggested by the AI, and 5) Continue practicing and exploring with the help of the AI. The document uses setting up Jupyter notebooks with the OpenAI LLM as a case study and example to walk through the steps.
Confoo 2024 Gettings started with OpenAI and data science
- 1. Getting started with OpenAI and Data science SUSAN IBACH | HOCKEYGEEKGIRL SUSAN.IBACH@LIVE.COM
- 2. You can't go anywhere these days without hearing about Generative AI
- 3. AI won't replace you, but someone with your skills + AI might
- 4. Coders are more productive when they use AI to help them code Over 80% of coders say they are more productive when they use a code helper such as GitHub Copilot 74% say it enables them to focus on more satisfying work 96% say they are faster completing repetitive tasks When studying two control groups, the group using a built in AI to help with coding completed their tasks 50% faster
- 5. Okay I get it Susan this AI thing looks useful, how do I get started using it for data science?
- 6. You could just open up ChatGPT ask it to write code for you then copy & paste
- 7. But the real win is doing it inside your IDE! This Photo by Unknown Author is licensed under CC BY
- 8. Step 1 Find a Large Language Model (LLM) you can install inside your IDE
- 9. This takes a bit of research OpenAI – Owned by Microsoft Codeium – VS Code, Vim, Jupyter Notebook, Eclipse GitHub Copilot – comes as an extension for VS Code, Visual Studio, JetBrains Obsidian Integration, heroml, Superpower extension, llmops.space, cursor.so, ChatGPT, CometLLM, Cohere
- 10. I use Jupyter notebooks so I'm going with Jupyter AI
- 11. Jupyter AI is vendor neutral and can connect to different LLMs AI21 Anthropic AWS Cohere HuggingFace Hub OpenAI
- 12. I chose OpenAI because I had played with it a bit already
- 13. Step 2 Install the extension or library in your IDE
- 14. If you want to use Jupyter AI with OpenAI in a Jupyter Notebook Software versions required Requires Jupyter Lab 4 Python 3.8 – 3.11 (I installed Python 3.11.6 64 bit) Accounts required (you can start with the free version) OpenAI
- 15. If you want to use Jupyter AI with OpenAI in a Jupyter Notebook Install the openai library pip install openai Create an environment variable and set it to the API key for your OpenAI account OPENAI_API_KEY=xxxxxxxxxxxxxxxxxxxxxxxxxxxx Each LLM supported has a specific environment variable name Install the jupyter_ai extension load_ext jupyter_ai
- 16. Not all OpenAI models are created equal Version GPT-3.5 Turbo GPT-4.0 Speed Faster Slower Database size 10X size of ChatGPT 3.5 and can handle images Quality of output 40% more likely to produce factual responses than 3.5, better at dialects $ Input / 1000 tokens $0.00005 $0.03 $ Output / 1000 tokens $0.0015 $0.06 You can find more information on pricing at openai.com
- 17. So what is a token anyway? You can think of tokens as pieces of words Wayne Gretzky’s quote "You miss 100% of the shots you don't take" contains 11 tokens 1 token is about 4 characters in English 1 English word is typically 1.3 tokens 1 French word is typically 2 tokens Punctuation marks are counted as one token Special characters are one to three tokens Emojis are between two to three tokens
- 18. Step 3 Try a hello world type command
- 19. Ask the AI to create "Hello World" %%ai chatgpt --format code display a message that says hello world Possible successful outputs include print("Hello World") System.out.println("Hello World"); console.log("Hello World"); echo "Hello World";
- 20. Step 4 Evaluate the suggested code
- 21. AI does not replace programmers. Programmers with AI replace programmers There is more than one way to write code to complete a task LLMs make an educated guess based on code it has seen in the past The coder provides the knowledge to evaluate the suggestion from the AI and make modifications to the prompt as needed (referred to as prompt engineering)
- 23. How much did that cost?
- 24. How many tokens and calls was it?
- 25. Step 5 Now we can play!
- 26. Maybe I need a dataframe with some sample data
- 27. Maybe I forgot the syntax for returning entries that start with a particular letter
- 28. Let's read a .csv file and then do some linear regression
- 29. Let's read a .csv file and then do some linear regression
- 30. ValueError: Input y contains NaN
- 31. AI does not replace programmers. Programmers with AI replace programmers
- 32. What would a coder do? We'd get rid of the rows with Nulls and try again!
- 33. Victory! I have successfully produced a plot but if you don't know how to read it this isn't going to help you
- 34. AI does not replace data scientists. Data scientists with AI replace data scientists
- 35. Until today, I have never done a live code demo - with this much code - in a session this short - without having to look up method names and parameters - without spending time in the session having the audience help me find my typing mistakes
- 36. AI doesn't replace presenters. Presenters with AI replace presenters
- 37. References ChatGPT Open AI Project Jupyter | Installing Jupyter Generative AI in Jupyter. Jupyter AI, a new open source project… | by Jason Weill | Jupyter Blog GitHub - jupyterlab/jupyter-ai: A generative AI extension for JupyterLab What are tokens and how to count them OpenAI Pricing
- 38. Questions? SUSAN IBACH | HOCKEYGEEKGIRL SUSAN.IBACH@LIVE.COM
- 39. Thank you!