In 2024 I started learning AI after finishing my full stack training. These are my notes from that year, from deciding to build chatbots through to my first full stack AI project.
Narrowing my focus (March 2024)
For several months I had been trying to decide what to focus on. I had already considered eCommerce, learning platforms and a portfolio built with Django and Flask. AI chatbot development interested me too, along with creative coding.
A chatbot is software that simulates conversation, used for customer service, sales, marketing and entertainment. A chatbot developer designs, tests, debugs and maintains them. The tools I found included Botkit, Rasa and Dialogflow, Postman and Chatbase for testing, AI libraries such as spaCy, TensorFlow and NLTK, and hosting on Heroku, Firebase or AWS. I would need to understand natural language processing and machine learning. Before paying for any certified course I planned to start with YouTube.
The Microsoft Learn AI Skills Challenge
I joined the month long challenge late, with about three weeks left. I picked the Azure OpenAI track, which has 27 modules including prompt engineering, retrieval augmented generation, Azure AI Search and Copilot Studio. I aimed for three or four modules a day and no more, to avoid overwhelm.
Three ways to build a chatbot
My aim was a service: build an AI persona chatbot for a business or person from their own data. I had three options.
- Build from scratch with Python, Flask and an API. I get to code, but I have to find hosting and learn more cloud hosting.
- Use a low code or no code tool such as Bubble, stack-ai.com, Botpress or Voiceflow. It saves coding time, but I would pay for the work those platforms have already done.
- Learn Microsoft Azure, which has everything in one place and a trusted name. I still needed to compare its cost with a self build on Flask.
AI abbreviations
AI loves abbreviations. The one I learned that week was RAG, retrieval augmented generation: a model looks up the answer in a knowledge base of text and uses it to respond. Others I had met:
- ML: machine learning
- RL: reinforcement learning
- DL: deep learning
- NLP: natural language processing
- LLM: large language model
- GPT: generative pre-trained transformer
- DNN, GAN, RNN and LSTM: types of neural network
- LARF: logical consistency, accuracy, relevance and factual correctness (prompt engineering)
- SALT: style, audience, length and tone (prompt engineering)
DataCamp (April 2024)
I joined DataCamp because it had everything I thought I needed as a beginner in one place. I liked the constant prompts to keep going, the tests that show your level and the learning plan built around my goals. My AI track started with generative AI, what large language models are and the challenges of building language models, then moved on to how models are trained, the transformer and attention mechanisms.
A week or so later I finished the AI Fundamentals and ChatGPT Fundamentals courses, which cover machine learning, large language models, generative AI, prompt engineering, advanced data analysis and custom ChatGPT models. After that I turned back to improving my Python, because DataCamp's Python courses are extensive even though they lean towards data science. I felt I was starting from scratch with AI and chatbots, so I planned to begin with apps, where I was already trained.
Claude 3.5 crash course (October 2024)
I recommended Brandon Hancock's Claude 3.5 crash course. It covers Claude Projects, Artifacts and knowledge bases, which need a Pro account, and a three step process he uses to speed up coding: set a north star goal, stub out the files, then code them fully. The templates are free to download. I am in his Skool community and was using his approach to build an AI marketing platform with Next.js, React, Clerk and TypeScript, my first full stack AI project.
