7
Confidential
“Scaling Laws” for Neural Language Models
言語モデルのパラメーター数(=モデルのサイズ)や、データセットのサイズ、
トレーニングに使用される計算(Compute)量が増えるにつれて、損失
(Loss、誤差)が「べき乗則」に従って減少する
出典:Jared Kaplan et al., "Scaling Laws for Neural Language Models" , arXiv:2001.08361
Slide 8
Slide 8 text
8
Confidential
“Emergent” Phenomena in Large Language Models
特定の転換点で、精度が劇的に向上した、という研究もある
出典:Jason Wei et al., "Emergent Abilities of Large Language Models", arXiv:2206.07682
Slide 9
Slide 9 text
9
Confidential
GPT3含む、様々なLLMの勃興
直近では、さらにその勢いが加速している
出典:Wayne Xin Zhao et al., "A Survey of Large Language Models", arXiv:2303.18223
Slide 10
Slide 10 text
10
Confidential
現時点でも、色んなジャンルで多数のサービスが登場
出典:KONSTANTINE BUHLER, "Generative AI Is Exploding. These Are The Most Important Trends To Know", APRIL 11, 2023
Slide 11
Slide 11 text
11
Confidential
ベーシックな使い方だけでも様々な活用方法
● Text
○ Marketing
○ Sales
○ Support
○ Knowledge
○ General writing
● Video
○ Video editing / generation
○ Personalized videos
● Image
○ Image generation
○ Consumer / Social
○ Media / Advertising
○ Design
● Code
○ Code generation
○ Documentation
○ Web app builders
○ Text to SQL
● 3D
○ 3D models / scenes
● Speech
○ Voice synthesis
● Others
○ Gaming
○ Music
○ Audio
○ RPA
○ Biology / Chemistry
Slide 12
Slide 12 text
12
Confidential
AIの急激な進化への対応
Slide 13
Slide 13 text
13
Confidential
大きな変化に備えて、必要な変革をすべき
Direction Investment Execution
Slide 14
Slide 14 text
14
Confidential
Direction
“AI Driven”
● AI Driven For UX
● AI Driven For Engineering
● AI Driven For Employee
Direction
Slide 15
Slide 15 text
15
Confidential
Direction: AI Driven For UX
● AIの進化をみた上で、1,3,5年後のメルカリの姿にどうAIが統合されていくと思う
か?
● 生成AI/LLM時代の、AIを活かしたプロダクト体験はどんなものであるべきか?
AIを活かした
お客さまへのプロダクト体験
Slide 16
Slide 16 text
16
Confidential
Direction: AI Driven For Engineering
● AI活用をしやすいデータ・API設計はどういうものか?そのために必要なアクショ
ンは何か?
● AIによる自動化やパーソナライズのための必要情報は十分reachableになっ
ているか?必要な情報量・幅はあるか?
AIを活かすための基盤投資
Slide 17
Slide 17 text
17
Confidential
Direction: AI Driven For Employee
● 全ての領域・職種で、AIと協働してパフォーマンスを発揮できているか?
● AIの最新の進歩と可能性を継続的に学習し、業務に適用する仕組みを構築でき
ているか?
AI活用を前提とした
業務設計・組織設計
Grand Vision
● Chat that understands what the user wants, and
provides what our service can.
● Search the app and data, not just items
● “I want to buy Pixel 7” -> Display Pixel 7 listings
● “I want to list my iPhone 13” -> List with pre-filled
information
● “How much did I earn in Mercari?” -> “You earned
XX yen total.”
Slide 32
Slide 32 text
Demo time!
@allan.conda
Recording
Slide 33
Slide 33 text
Technical Details
Slide 34
Slide 34 text
Implementation Plan
1. Provide GPT a list of Mercari APIs and deeplinks as context,
with a simple description
2. GPT should be able to give instructions to call the correct API
or deeplinks to satisfy the request.
3. Chat: Send the APIs response to GPT to find the information
4. Present GPT’s final response as Chat to user.
Slide 35
Slide 35 text
Initial Prompt for Search
''''''
Possible deeplinks in the app:
Item Detail Screen xxxxx://xxxx/xxxxxxxx?id={item_id}
Listing Screen xxxxx://xxxxx/xxxxxx
Search Results Screen mercari://xxxxx/xxxxxxx?param={param}
xxxxx://xxxx
Bought Items /xxxxxxxx
Coupon List /xxxxxxx
Draft Listings /xxxxx
…
''''''
Display the deeplink first, followed by a newline, then a human readable action label
{deeplink}
{button action label}
''''''
Slide 36
Slide 36 text
Initial Prompt for Chat
'''’''
Mercari API:
Description: {endpoint_path} [{query_name: Type}]
Liked Items History: xxx/xxxxx [xxxxx: Int]
My User Profile: xxxxx/xxxxx?_user_format="xxxx"
''''''
If it is plain user prompt, create only API call instruction tokens in order to retrieve the information by the Mercari user. Strict
completion format for tokenization.
Do not return anything else but the instruction in this case.
example format: GET: xxx/xxxx?pageSize=10
If the prompt is json response data, consider it as the success response of the API call and use it to retrieve the data
desired by the user and describe to the user.
''''''
Slide 37
Slide 37 text
Mercari comment assistant
powered by ChatGPT
kenmaz
@merpay iOS team
Slide 38
Slide 38 text
My LINE message log
I don’t like typing long
sentences on iPhone