GPT / LLMs
• Internet sized change
• Change in capability
• Change in how we build and use AI
Slide 3
Slide 3 text
No content
Slide 4
Slide 4 text
No content
Slide 5
Slide 5 text
No content
Slide 6
Slide 6 text
Level 1: GPTs are incredible!
Level 2: GPTs make things up and aren’t trustworthy.
Level 3: GPTs can be incredible when used right
Slide 7
Slide 7 text
See them as engineering components
Separate out aspects accidentally bundled
Slide 8
Slide 8 text
What is GPT?
Slide 9
Slide 9 text
Training objective: token prediction
Slide 10
Slide 10 text
Training objective: token prediction
Slide 11
Slide 11 text
No content
Slide 12
Slide 12 text
No content
Slide 13
Slide 13 text
• A sequence model
• That uses ‘attention’
• Gradient descent
Slide 14
Slide 14 text
• A sequence model
• That uses ‘attention’
• Gradient descent
Slide 15
Slide 15 text
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’
Slide 16
Slide 16 text
Model: Database + Reasoning Engine
• The reasoning engine is key
• Often, the database is a liability
Slide 17
Slide 17 text
Reasoning capabilities
Slide 18
Slide 18 text
No content
Slide 19
Slide 19 text
No content
Slide 20
Slide 20 text
Model:
‘Interpolative’ vs
‘Extrapolative’
tasks
Slide 21
Slide 21 text
No content
Slide 22
Slide 22 text
• Less reliable at extrapolation
• Favour interpolation
• Perform a task, given a context
• ‘Retrieval Augmented Generation’
Slide 23
Slide 23 text
Model: Human intuition
Ask a human to answer a historical question
vs
Give them a history book and ask them the question
Slide 24
Slide 24 text
Note: Context window limited
• Thousands of words
• Can’t put a whole KB, or context, in it
• Synergizes well with Vector Search
Slide 25
Slide 25 text
How we build with GPTs
Slide 26
Slide 26 text
No content
Slide 27
Slide 27 text
No content
Slide 28
Slide 28 text
No content
Slide 29
Slide 29 text
No content
Slide 30
Slide 30 text
30 November 2022:
ChatGPT
Slide 31
Slide 31 text
First features we built
• Summarisation
• Edit tone of voice
• Expand from shorthand
Slide 32
Slide 32 text
No content
Slide 33
Slide 33 text
No content
Slide 34
Slide 34 text
• 5th Dec: Rolling
• 20th Dec: Internal use
• ~13th Jan: Customer beta
• 31st Jan: Launch with testimonials
Timeline
Slide 35
Slide 35 text
Model: Easy vs Hard
AI features
Slide 36
Slide 36 text
• ‘Easy’:
• Out-of-box accuracy high
• Cost of error low
• E.g. ‘Draft me a summary’
Slide 37
Slide 37 text
• ‘Hard’:
• Out-of-box accuracy low
• Cost of error high
Slide 38
Slide 38 text
Development Tactics
Slide 39
Slide 39 text
• 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
Slide 40
Slide 40 text
Hard feature: Fin
• GPT-powered
question answering Bot
Slide 41
Slide 41 text
• An LLM can seem inert
• However, can easily be turned into an agent
Slide 42
Slide 42 text
My key points
• Internet sized change
• Good model: DB+Reasoning
• Changes how we build ML
• Feature dif
fi
culty varies
Slide 43
Slide 43 text
Guessing what’s next
Slide 44
Slide 44 text
• 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