AINoon Discussion

Thanks

  • To the host for the great venue!
  • To our sponsors

Administrivia

  • Fire escapes
  • Toilets
  • Cleaning up after ourselves
  • Wi-Fi

AINoon Discussion

  1. Recap the key points from previous lessons
  2. Questions and deeper discussion

GenAI for Business

  • AI Terminology and History
    • Artificial Intelligence
      • Machine Learning
        • Deep Learning
          • Generative AI
  • Chatbot Use Cases
    • Drafting, Brainstorming
    • Summarising, Extracting, Transforming
  • The Golden Rule of AI
    • Don’t trust the output of an AI unless you can verify it
    • E.g. Inconsistent results summarising the hotel reviews

Demystifying AI

  • Simple maths, but at a scale that makes it inscrutable
    • A lot of “why” questions can’t be answered definitively
  • Models trained on general sources and human feedback
    • Training is expensive and slow - any personalisation is from adding to the prompt
  • “Everything is a hallucination”
    • It predicts what sounds right, it isn’t “thinking” with logic
  • Even the same model can return different responses to the same prompt

Common AI Patterns

  • AI App = LLM + Clever inputs + Clever use of outputs
  • Retrieval Augmented Generation (RAG)
    • Easily equip a chatbot to answer questions from your documents
    • Separates knowledge searching from answer generation
  • Tools
    • Equip the LLM with tools that it decides when to use
    • E.g. Fetching data and taking action
  • Agents
    • “An LLM agent runs tools in a loop to achieve a goal”
    • Easy to build with just a prompt and some tools
    • E.g. Copilots in Zapier didn’t work reliably

Risks & Challenges

  • Inaccurate Outputs
    • Consider failure modes; human in/on the loop; testing; guardrails
  • Unwanted Bias
    • From training sources; provider intervention; LLM tendencies
  • Privacy & Security
    • Public chatbots training on your data; lethal trifecta
  • Copyright
    • Does AI training infringe copyright? Who owns the outputs?
  • Vendor Lock-in
    • Build modularly (LLM gateways) and with standards (MCP) to retain freedom to pick the best provider
  • Other risks
    • Environmental impact; identifying AI outputs; impact on jobs, AGI

Vibe Coding

  • Describing an app for AI to build and not reading the code
  • Build prototypes and low-stakes apps
  • Automate time-consuming tasks specific to your work
  • When security or correctness are important - verify the code

Discussion Starters

  • Do you have any questions on topics covered in the course?
  • What surprised you most about AI from the course?
  • Is there anything you’re interested in that we didn’t cover?
  • How are you using (or planning to use) AI?
  • What would you consider when weighing up if/how to use AI for a use case?
    • E.g. Describe a process where a human could follow each step to check that you aren’t expecting the AI to do “magic” that even a human couldn’t even do