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
ChatGPT活用サービスの スケール時の落とし穴と対策 - 生成AIにおけるpiconのこれまでのチャレンジと今後 -
Search
Shibuya Yukito
July 05, 2023
0
140
ChatGPT活用サービスの スケール時の落とし穴と対策 - 生成AIにおけるpiconのこれまでのチャレンジと今後 -
Shibuya Yukito
July 05, 2023
Tweet
Share
Featured
See All Featured
Save Time (by Creating Custom Rails Generators)
garrettdimon
PRO
1
130
Intergalactic Javascript Robots from Outer Space
tanoku
266
26k
A better future with KSS
kneath
231
16k
I Don’t Have Time: Getting Over the Fear to Launch Your Podcast
jcasabona
23
1.7k
The Invisible Customer
myddelton
114
12k
Docker and Python
trallard
35
2.7k
Optimising Largest Contentful Paint
csswizardry
13
2.4k
What's new in Ruby 2.0
geeforr
338
31k
VelocityConf: Rendering Performance Case Studies
addyosmani
321
23k
Atom: Resistance is Futile
akmur
260
25k
実際に使うSQLの書き方 徹底解説 / pgcon21j-tutorial
soudai
123
39k
The Straight Up "How To Draw Better" Workshop
denniskardys
228
130k
Transcript
גࣜձࣾpicon COO ौ୩ਓ ChatGPT׆༻αʔϏεͷ εέʔϧ࣌ͷམͱ݀͠ͱରࡦ - ੜAIʹ͓͚Δpiconͷ͜Ε·ͰͷνϟϨϯδͱࠓޙ -
౦ژେֶதୀޙɺגࣜձࣾQJDPOڞಉۀɻ "*νϟοτ͘Μͷ1.։ൃ13Ϗζσϒ ͳͲɺ෯͘୲ɻ ͖αφɺΫϥϑτϏʔϧɺφνϡϥϧϫ ΠϯɺʢαʔϑΟϯʣ 5XJUUFS!ZVLJUP@TIJCVZB ौ୩ਓ4IJCVZB:VLJUP
None
None
None
None
None
None
None
None
None
8IJTQFS"*Λ׆༻
None
3/2 • ىচ: ChatGPT APIͷϦϦʔεϝʔϧ͕ಧ͍ͯΔ • ޕલத: ݩʑ͋ͬͨΞϓϦͷΞϓσ࡞ۀʢ࣭ͱͰֵ໋Λײ͡Δʣ • ޕޙ:
ࣗͰɺεϚϗͰ͏ͳΒLINEͩͳͱࢥ͍ϓϩτλΠϓ࡞Δ • ༦ํ: දͷshosemaruʹϦϦʔε͍͍͔ͯ͠ฉ͘ -> OK͕ग़Δ • : ʮAIνϟοτ͘ΜʯͷϦϦʔεπΠʔτ ίʔυͷ8ׂ͙Β͍ChatGPTʹॻ͍ͯΒͬͨ ։ൃ·ͰͷܦҢ
• Server: Cloud Functionsʢnode.jsʣ • DB: Firestore • ͍ͬͯΔAPI: ChatGPT
/ LINE API ॳͷߏ Φʔτεέʔϧ / αʔόʔϨεͰ؆୯ϥΫϥΫʢͱࢥ͍ͬͯͨʣ
None
• 7/4࣌ ొऀ12ສ͑ • ࣗલPCϞσϧͷηοτΞοϓෆཁ • ຊޠରԠ • image 2
imageʹରԠ AIΠϥετ͘Μͷಛ ຊޠରԠʂLINEͰStable Di ff usionΛ͑ΔαʔϏε
ChatGPT׆༻αʔϏεͷ εέʔϧ࣌ͷམͱ݀͠ͱରࡦ
$0.002/1000 tokenɺ1000 token / 1requestͱԾఆ͢Δͱ 60000request/month…શવΓͳ͍ 1: usage limitͷॳظ͕খ͗͢͞Δ… ՝ۚޙͷusage
limit͕MAXͰ120υϧ/month
• piconͰ… • 2/4ʹAIνϟοτ͘ΜͷલͰGPT-3ΛνϟοτܗࣜͰ͑ΔΞϓϦ ʢFlutterʣΛϦϦʔεࡁΈͩͬͨ • ͦͷͨΊɺૣΊʹʹͿͪͨΓɺগ࣮ͣͭ͠Λ࡞Εͨ • Tips •
Quotaͷਃɺଟ͗͢Δ͔ΒϦδΣΫτͬͯ͜ͱͳ͍ͷͰଟΊ ͰOK • ਃ͔Βঝೝ·Ͱ͙Β͍ ରࡦ: ૣΊʹૣΊʹҾ্͖͛ਃ͢Δ 1: usage limitͷॳظ͕খ͗͢͞Δ…
• OpenAIͷμογϡϘʔυͰ͔͔͍ͬͯΔඅ༻Λݟ͍ͯͨ • ϦϦʔε͔ΒޙʹԿނ͔ίετ͕10ഒ͙Β͍ʹͳ͍ͬͯͨ • μογϡϘʔυͷόάͩͱࢥ͏ͷͰɺࠓ࣏ͬͯͦ͏͚ͩͲ… ରࡦ: ϦΫΤετ x total_tokenͰ༧ଌ͓͍ͯͨ͠ํ͕͍͍
2: μογϡϘʔυʹө͞Ε͍ͯͨՁ͕֨όάͬͯΔ
ରࡦ: 3: RateLimitͲ͏͠Α͏ͳ͍… OpenAI Azure OpenAI MAXϦΫΤετ/min 3,500 300 →
ഇࢭʹͳͬͨʁ MAXτʔΫϯ/min 90,000 120,000 200,000 500τʔΫϯ/req ͷͱ͖ͷmaxϦΫΤετ/min 180 240 400
ରࡦ: ΠϯελϯεΛෳཱͯͯෛՙࢄ 3: RateLimitͲ͏͠Α͏ͳ͍… • RateLimitͷҾ্͖͛ɺAzureOpenAI΄΅ແཧͬΆ͍ • Azure OpenAIͩͱɺϦʔδϣϯ͝ͱʹ2ΠϯελϯεཱͯΒΕΔ •
OpenAIͱAzure OpenAIͷซ༻͋Γ • ΠϯελϯεΛෳཱͯͯɺϦΫΤετΛࢄͤ͞Δ͜ͱͰճආ͢Δ ͔͠ͳ͍ • ࢀߟ: Azure OpenAI Serviceͷෛՙࢄ • https://logico-jp.io/2023/06/08/request-load-balancing-for-azure- openai-service/
• ʮAIνϟοτ͘ΜແྉͰ͔͢ʁʯͱ͔ΊͬͪΌฉ͘ • ͔͠͠ɺదͳ͕͑ฦ͖ͬͯͯ͠·͏ -> UXతʹ࠷ѱ • ར༻ن / ϓϥΠόγʔϙϦγʔ
/ ղಋઢͳͲ… ରࡦ: ༧ޠΛ࡞ͬͯɺఆܕจΛฦ͢Α͏ʹ͢Δ ͦͷଞ: ࣗࣗͷใΛΒͳ͍
piconͷࠓޙʹ͍ͭͯ
• ݱঢ়ɺڵຯຊҐͰ৮ͬͯΈ͚ͨͲɺৗʹਁಁ͢Δͱ͜Ζ·Ͱདྷ͍ͯͳ͍ →ͬͱ༷ʑͳར༻ͷํΛಧ͚͍͖͍ͯͨ • AIνϟοτ͘ΜYahoo!JAPANΛࢦ͍ͯ͘͠ʢChatGPTGoogleʣ →୭Ͱؾܰʹར༻Ͱ͖ͯɺ͠Έͷ͋ΔαʔϏεΛࢦ͍ͯ͘͠ • piconɺੜAI͕ίϯγϡʔϚʔͷੜ׆ʹͲ͏ͨ͠Βਁಁ͍ͯ͘͠ͷ͔ʁʹ ͖߹͍ͬͯ͘ →
ݱࡏɺ͢ͰʹاըதͷϓϩμΫτ2ͭ͋ΓɺϦϦʔεΛࢦ͢ → ·ͩਖ਼ղͷͳ͍ྖҬͰɺٕज़໘ɺUX໘ʹνϟϨϯδ͍ͯ͘͠ ChatGPT/ੜAIΛ·ͩ·ͩΈΜͳ͍͜ͳ͍ͤͯͳ͍ ݱঢ়ͷ՝
• ܦӦϝϯόʔͷ1ਓͯ͠ࣄۀܭըʹଇΓɺϓϩμΫτϩʔυϚοϓͷࡦఆͱ։ൃͷϚωδϝϯτ • ෳϓϩμΫτʹ͓͚Δऩӹੑͷ্ • ։ൃνʔϜͷϚωδϝϯτʢ࠾༻/ҭʣ CPOީิ: ϏδϣϯΛϓϩμΫτʹམͱ͠ࠐΈɺऩӹԽ·Ͱ͍͚࣋ͬͯΔํ ੜAIͷະདྷΛҰॹʹͭ͘ΔਓɺେืूதͰ͢ɻ CTOީิ:
ࣾ֎ͷνʔϜͱͱʹpiconΛٕज़ͰϦʔυͯ͘͠ΕΔํ • ܦӦϝϯόʔͷ1ਓͯ͠։ൃνʔϜͷ • ٕज़ઓུͷࡦఆͱ࣮ߦ • ֎෦ύʔτφʔͷϚωδϝϯτ C͚ͷྖҬͰɺϢʔβʔʹ͖߹ͬͨ։ൃΛ͢ΔจԽͰ͢ɻ
ͥͻ͜ͷ͋ͱ͓͠·͠ΐ͏ !ZVLJUP@TIJCVZB 5XJUUFS 'BDFCPPL ͝࿈བྷ͓͓ͪͯ͠Γ·͢ɻ