Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
Fergal Reid - Building products in the age of Ai
Search
Turing Fest
PRO
July 05, 2023
Technology
0
110
Fergal Reid - Building products in the age of Ai
Turing Fest
PRO
July 05, 2023
Tweet
Share
More Decks by Turing Fest
See All by Turing Fest
Andy Budd: The Growth Equation: 7 Essential Steps to Finding Product Market Fit
turingfest
PRO
0
7
Andrey Vinitsky: Babe Are You OK? You've Barely Touched The Dashboard You Claimed Was Mission Critical
turingfest
PRO
0
9
Finbarr Taylor:From Scotland to Silicon Valley: Lessons Learned Raising $100m & Building a Global SaaS Business
turingfest
PRO
0
6
Megan Caywood: A Product Playbook to Building a Unicorn
turingfest
PRO
0
11
Jason Miller: Branding in the Age of AI
turingfest
PRO
0
6
Petra Wille: Lessons on Storytelling for Product Builders
turingfest
PRO
0
14
Meri Williams: Career Vectors: Navigating Modern Careers
turingfest
PRO
0
3
Todd Olson: How AI Supercharges Product-led Growth
turingfest
PRO
0
4
Rand Fishkin: Zero-Click Marketing
turingfest
PRO
0
13
Other Decks in Technology
See All in Technology
AOAI Dev Day - Opening Session
yoshidashingo
2
470
AI研修【MIXI 24新卒技術研修】
mixi_engineers
PRO
0
130
ゆめみのアクセシビリティの現在地と今後
ryokatsuse
3
290
ABEMAにおけるLLMを用いたコンテンツベース推薦システム導入と効果検証
cyberagentdevelopers
PRO
1
750
推薦システムを本番導入する上で一番優先すべきだったこと~NewsPicks記事推薦機能の改善事例を元に~
morinota
0
130
コンテナ・K8s研修 - 後半 Kubernetes 基礎&ハンズオン【MIXI 24新卒技術研修】
mixi_engineers
PRO
1
120
20240724_cm_odyssey_hibiyatech
hiashisan
0
110
技術負債による事業の失敗はなぜ起こるのか / Why do business failures due to technical debt occur?
i35_267
0
190
E2Eテスト自動化プラットフォームにおけるAIの活用
shift_evolve
0
190
「単なる OAuth 2.0 を認証に使うと、車が通れるほどのどでかいセキュリティー・ホールができる」のか検証してみた
terara
0
380
ここがすごいよ! AWS Systems Manager!
saichan11
0
1.8k
OSSコミットしてZennの課題を解決した話
dyoshikawa1993
0
150
Featured
See All Featured
Clear Off the Table
cherdarchuk
89
320k
How To Stay Up To Date on Web Technology
chriscoyier
784
250k
A better future with KSS
kneath
231
17k
Learning to Love Humans: Emotional Interface Design
aarron
269
39k
Templates, Plugins, & Blocks: Oh My! Creating the theme that thinks of everything
marktimemedia
23
1.9k
GraphQLとの向き合い方2022年版
quramy
36
13k
Bash Introduction
62gerente
607
210k
GraphQLの誤解/rethinking-graphql
sonatard
59
9.6k
Distributed Sagas: A Protocol for Coordinating Microservices
caitiem20
325
21k
The Straight Up "How To Draw Better" Workshop
denniskardys
229
130k
Debugging Ruby Performance
tmm1
71
11k
Design by the Numbers
sachag
277
18k
Transcript
Building products in the Age of AI @fergal_reid
GPT / LLMs • Internet sized change • Change in
capability • Change in how we build and use AI
None
None
None
Level 1: GPTs are incredible! Level 2: GPTs make things
up and aren’t trustworthy. Level 3: GPTs can be incredible when used right
See them as engineering components Separate out aspects accidentally bundled
What is GPT?
Training objective: token prediction
Training objective: token prediction
None
None
• A sequence model • That uses ‘attention’ • Gradient
descent
• A sequence model • That uses ‘attention’ • Gradient
descent
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’
Model: Database + Reasoning Engine • The reasoning engine is
key • Often, the database is a liability
Reasoning capabilities
None
None
Model: ‘Interpolative’ vs ‘Extrapolative’ tasks
None
• Less reliable at extrapolation • Favour interpolation • Perform
a task, given a context • ‘Retrieval Augmented Generation’
Model: Human intuition Ask a human to answer a historical
question vs Give them a history book and ask them the question
Note: Context window limited • Thousands of words • Can’t
put a whole KB, or context, in it • Synergizes well with Vector Search
How we build with GPTs
None
None
None
None
30 November 2022: ChatGPT
First features we built • Summarisation • Edit tone of
voice • Expand from shorthand
None
None
• 5th Dec: Rolling • 20th Dec: Internal use •
~13th Jan: Customer beta • 31st Jan: Launch with testimonials Timeline
Model: Easy vs Hard AI features
• ‘Easy’: • Out-of-box accuracy high • Cost of error
low • E.g. ‘Draft me a summary’
• ‘Hard’: • Out-of-box accuracy low • Cost of error
high
Development Tactics
• 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
Hard feature: Fin • GPT-powered question answering Bot
• An LLM can seem inert • However, can easily
be turned into an agent
My key points • Internet sized change • Good model:
DB+Reasoning • Changes how we build ML • Feature dif fi culty varies
Guessing what’s next
• 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
• Breakneck progress • Smaller models, open? • Exciting but
overhyped today • Productisation • Larger models
Thank you! @fergal_reid