Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Fergal Reid - Building products in the age of Ai

Fergal Reid - Building products in the age of Ai

Turing Fest

July 05, 2023
Tweet

More Decks by Turing Fest

Other Decks in Technology

Transcript

  1. Building products


    in the Age of AI
    @fergal_reid

    View full-size slide

  2. GPT / LLMs
    • Internet sized change


    • Change in capability


    • Change in how we build and use AI

    View full-size slide

  3. Level 1: GPTs are incredible!
    Level 2: GPTs make things up and aren’t trustworthy.
    Level 3: GPTs can be incredible when used right

    View full-size slide

  4. See them as engineering components


    Separate out aspects accidentally bundled

    View full-size slide

  5. What is GPT?

    View full-size slide

  6. Training objective: token prediction

    View full-size slide

  7. Training objective: token prediction

    View full-size slide

  8. • A sequence model


    • That uses ‘attention’


    • Gradient descent

    View full-size slide

  9. • A sequence model


    • That uses ‘attention’


    • Gradient descent

    View full-size slide

  10. Not a useful model
    • Human = genes and evolution ?


    • Distrust:

    ‘It de
    fi
    nitely can’t do X because its just trained to predict
    the next word’

    View full-size slide

  11. Model: Database + Reasoning Engine
    • The reasoning engine is key


    • Often, the database is a liability

    View full-size slide

  12. Reasoning capabilities

    View full-size slide

  13. Model:


    ‘Interpolative’ vs


    ‘Extrapolative’


    tasks

    View full-size slide

  14. • Less reliable at extrapolation


    • Favour interpolation


    • Perform a task, given a context


    • ‘Retrieval Augmented Generation’

    View full-size slide

  15. Model: Human intuition
    Ask a human to answer a historical question


    vs


    Give them a history book and ask them the question

    View full-size slide

  16. Note: Context window limited
    • Thousands of words


    • Can’t put a whole KB, or context, in it


    • Synergizes well with Vector Search

    View full-size slide

  17. How we build with GPTs

    View full-size slide

  18. 30 November 2022:

    ChatGPT

    View full-size slide

  19. First features we built
    • Summarisation


    • Edit tone of voice


    • Expand from shorthand

    View full-size slide

  20. • 5th Dec: Rolling


    • 20th Dec: Internal use


    • ~13th Jan: Customer beta


    • 31st Jan: Launch with testimonials
    Timeline

    View full-size slide

  21. Model: Easy vs Hard

    AI features

    View full-size slide

  22. • ‘Easy’:


    • Out-of-box accuracy high


    • Cost of error low


    • E.g. ‘Draft me a summary’

    View full-size slide

  23. • ‘Hard’:


    • Out-of-box accuracy low


    • Cost of error high

    View full-size slide

  24. Development Tactics

    View full-size slide

  25. • Fast customer contact


    • Assume you can build v1 of most ML with powerful LLM


    • Make cheap later


    • “LLMs aren’t all of AI”


    • How we build software has changed

    View full-size slide

  26. Hard feature: Fin
    • GPT-powered

    question answering Bot

    View full-size slide

  27. • An LLM can seem inert


    • However, can easily be turned into an agent

    View full-size slide

  28. My key points
    • Internet sized change


    • Good model: DB+Reasoning


    • Changes how we build ML


    • Feature dif
    fi
    culty varies

    View full-size slide

  29. Guessing what’s next

    View full-size slide

  30. • V1: text tools, working around clunky interfaces


    • V2: features reasoning can enhance


    • V?: End to end problems where intelligence can help


    • Don’t underestimate the reasoning capability, very
    sophisticated

    View full-size slide

  31. • Breakneck progress


    • Smaller models, open?


    • Exciting but overhyped today


    • Productisation


    • Larger models

    View full-size slide

  32. Thank you!
    @fergal_reid

    View full-size slide