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
AI Company
Search
Sungjoo Ha
May 24, 2022
Technology
0
940
AI Company
Google cloud academy 발표 자료
Sungjoo Ha
May 24, 2022
Tweet
Share
More Decks by Sungjoo Ha
See All by Sungjoo Ha
AI in Social Discovery -- Blending Research and Production
shurain
2
1k
Bridging the Gap: AI Research and Real-World Deployment in AI Companies
shurain
0
580
AI Company
shurain
0
220
제품에 기여하는 머신러닝 -- 연구에서 고객까지
shurain
0
230
Research in Production Environment
shurain
0
2.2k
Zen of NumPy
shurain
1
990
Other Decks in Technology
See All in Technology
Amazon Bedrock Multi-Agent Collaboration Workshop の紹介 - ワークショップでAIエージェントを学ぼう
nasuvitz
4
370
2000年てづくりキーボードの旅
tagomoris
1
170
開発者向けツールを魔改造してセキュリティ診断ツールを作っている話 - 第1回 セキュリティ若手の会 LT
pizzacat83
0
410
10分で学ぶKubernetesコンテナセキュリティ/10min-k8s-container-sec
mochizuki875
1
120
AI時代のデータセンターネットワーク
lycorptech_jp
PRO
1
200
WernerVogelsのKeynoteで語られた6つの教訓とOps
hatahata021
1
210
Explain EXPLAIN
keiko713
10
2.9k
KubeCon NA 2024 Recap / Running WebAssembly (Wasm) Workloads Side-by-Side with Container Workloads
z63d
1
130
GitHub Actions의 다양한 기능 활용하기 - GitHub Universe '24 Recap
outsider
0
570
B10-ひと目でわかる(といいなぁ)Microsoft Purview
seafay
PRO
0
660
Tailwind CSSとAtomic Designで実現する効率的な Web 開発の事例
toranoana
1
280
アップデート紹介:AWS Data Transfer Terminal
stknohg
PRO
0
110
Featured
See All Featured
The Invisible Side of Design
smashingmag
298
50k
[RailsConf 2023] Rails as a piece of cake
palkan
53
5k
Designing Experiences People Love
moore
138
23k
Gamification - CAS2011
davidbonilla
80
5.1k
The Psychology of Web Performance [Beyond Tellerrand 2023]
tammyeverts
45
2.2k
Responsive Adventures: Dirty Tricks From The Dark Corners of Front-End
smashingmag
251
21k
Put a Button on it: Removing Barriers to Going Fast.
kastner
59
3.6k
Learning to Love Humans: Emotional Interface Design
aarron
273
40k
Adopting Sorbet at Scale
ufuk
73
9.1k
Let's Do A Bunch of Simple Stuff to Make Websites Faster
chriscoyier
507
140k
RailsConf & Balkan Ruby 2019: The Past, Present, and Future of Rails at GitHub
eileencodes
132
33k
Art, The Web, and Tiny UX
lynnandtonic
298
20k
Transcript
AI Company Google Cloud Academy Hyperconnect Director of AI Sungjoo
Ha May 24th, 2022 Sungjoo Ha 1
AI Company • য়ט ઁ: AI company • ੋఠ֔ ӝࣿਸ
ੜ ഝਊೠ ӝস internet companyۄ ࠛ۷Ҋ ח ݽ߄ੌ दө য • Amazon, Alphabet, Facebook, Alibaba, Tencent, etc. • AI द AI companyח ޖ ܳө? • ݽܰ݅ ѐੋਵ۽ ࢤп೮؍ ݻ о ನੋী ೠ ࣗѐ Sungjoo Ha 2
ࣳೝށ + ਢࢎ ≠ Internet Company Sungjoo Ha 3
Jeff Bezos in 1997 In the book space, there are
more than three million different books worldwide active and in print at any given time across all languages, so when you have that many items, you can literally build a store online that couldn't exist any other way. Sungjoo Ha 4
Aggregators • Zero marginal cost • ٣ణ ইమਸ ౸ݒೞחؘী ୶о
࠺ਊ ٜ ঋ • Distributionী ࠺ਊ ٜ ঋ • Transactionী ࠺ਊ ٜ ঋ • ܳ ӓೠਵ۽ ഝਊೠ Ѫ ҳӖա ಕझ࠘җ э super-aggregator • ੋఠ֔ী ઓೞח Ѫ value proposition ইש • ੋఠ֔ ࢜۽ ח оܳ ߉ইٜҊ ੋఠ֔ ইפݶ ࢿ݀ೡ ࣻ হח ࠺ૉפझܳ ݅ٝ Sungjoo Ha 5
Internet-Enabled Technology • Internet द ӝࣿ • ੋఠ֔ दী ਢಕח
݅ ੋఠ֔ਵ۽ ੋ೧ оמ೧ ӝࣿਸ ഝਊೞח ӝস ઁೠ • ਬ ೯زী ೠ ࣻਸ ߄ఔਵ۽ ࢎਊܳ ೧ೞח Ѫ • ೠ ߊ աইоࢲ A/B test ೞח Ѫ • 1֙ী ೠ ف ߣ ߓನೞ؍ Ѫ • Continuous integration, continuous deployment ١ਵ۽ ݒੌݒੌ ߓನ • ӓױਵ۽ ૣ ఠۨ࣌ ࢎਸ ٜ݅যղҊ product-market-fitਸ ӝ ਤ೧ ࡅܰѱ ח Ѫ оמ೧ • ܳ оמாೞח ઑ ҳઑ Sungjoo Ha 6
Learnings From The Past • Internet company द۽ࠗఠ ࢤп೧ࠁח AI
company • AIо ইפݶ ࢿ݀غ ঋח ࠺ૉפझ • ӝઓী ഃ ࠛоמೞ؍ Ѫ оמ೧ח Ѫ • ઑӘ ؊ ևѱ ೧ࢳ೧ࠁݶ AI दী оמ೧ ࢜۽ ӝࣿਸ ੜ ഝਊೞ ח ഥࢎ Sungjoo Ha 7
Modern ML • അ ࠗ࠙ "AI"ۄ ܻࠛח Ѫ بणী Ӕрਸ
فҊ • ীࢲ ܳ Ҋ ܳ ח Ѫ • ҃ઁ ࠗооܳ о ݆ ٜ݅য ղח Ѫ ب//࠺ب/ъച ण ࣽࢲ • п ޙઁب ѾҴ بण ޙઁ۽ ജೞח ߑߨٜ ߊѼغҊ ਵݴ ী ٮܲ ࢿҗо աয়ח Sungjoo Ha 8
Supervised Learning Examples • male/female • ҃/ߓ҃ •
ೠҴয য • োয ࢸݺ • োয ࢸݺ • ҟҊ, ਬ ܼ • ݂ ܼ Sungjoo Ha 9
Algorithm + Compute + Data • ML ޙઁীࢲ ݏׯڰܻח ف
о ޙઁ • ݽ؛ ࢿמ णࣇীࢲ ծ ҃ • underfitting: ݽ؛ ࣻਊ۱ ૐо • ݽ؛ ࢿמ పझࣇীࢲ ծ ҃ • overfitting: ؘఠ ୶о • അ ML ӝࣿ ݽ؛ ࣻਊ۱(capacity)ਸ ޖೠ טܾ ࣻ • ܳ جܻӝ ਤೠ ҅ ਗҗ • ࠙ೠ ন ؘఠ݅ ਵݶ ҅ࣘ ࢿמ જই • 2010֙ ୡ߈ ীח ۠ ѐ֛ হਵݴ ୭Ӕ ٜয ۠ ܴ ؊ ъചغҊ Sungjoo Ha 10
Solvable Problems • അ AI ӝࣿ۽ ಽ ࣻ ח ޙઁ
• ಣߧೠ ࢎۈ 1ୡ ب दрਸ ٜৈࢲ ಽ ࣻ ח ޙઁ • द௫झܳ ஏೞח ޙઁ • ҳઑܳ צ ݽٚ ޙઁח زചؼ Ѫ Sungjoo Ha 11
AI Strategy • ܲ ۨযٜ যڃ ࢤпਸ ೞҊ ਸө? AI
ۚ ަө? • ٜ ৈ۞ оמࢿਸ ࠁݶࢲ ߬ೞҊ • Google: AI 3ਃࣗܳ ೠ ۽ ݽف ы୶Ҋ ח ݻ উغח ӝস • OpenAI: Ѣ ݽ؛ী ೠ ъೠ ਸ ాೠ ߬ • Tesla: അઓೞח ӝࣿਸ ӓೠਵ۽ ഝਊೞח ߡ౭ஸ ా Sungjoo Ha 12
ӝઓഥࢎ + AI/ML/DL ≠ AI Company Sungjoo Ha 13
Zero Marginal Content • AIо ইפݶ ࢿ݀ೡ ࣻ হח ࠺ૉפझח
ޖ ੌө? • ݻ о ൦ח • Zero marginal cost content creation • GPT-3 (copilot), DALL·E 2 • Super-human decision making • AlphaGo, AlphaFold Sungjoo Ha 14
AI-Enabled Technology • ӝসٜঠ নೠ ߬ਸ ೡ ࣻ ݅
ഥࢎח যڌѱ? • AI दীח ੋఠ֔ द ܴҗ ࠺तೞѱ AI ݽ؛ ٜ ࢎਊೡ Ѫ • ৈӝীࢲ بػ 2ରੋ ѐ֛/ӝࣿ/ޙചܳ ݃ա ੜ ഝਊೞջب ਃೠ ನੋо ؼ Ѫ • ݃ A/B పझܳ ೞח ഥࢎ৬ ইצ ഥࢎо աҊ • CI/CDܳ ഝਊೞח ഥࢎ৬ ইפפ ഥࢎо աח Ѫۢ Sungjoo Ha 15
Learned Business Logic • ࠺ૉפझ ۽ -> ݽ؛ Үജ •
࠺ૉפझ ۽: A ݶ B • ۽Ӓېݠٜ ٜ݅যղח ٘ ݆ ࠗ࠙ ࠺ૉפझ ۽ • ѱ ޤо ؘܲ? ۽Ӓېݠо ೞܖݶ ٜ݅؍Ѧ ҡ ML ॄࢲ ೠ׳ Ѧ۰ࢲ ݅٘ח Ѣ ইפঠ? • ࢎۈࠁ ੜ ೡ ࣻ ח ҃о • Aݶ Bীࢲ A ઑѤ ցޖցޖ ࠂ೮ݶ ࢎۈ ޅೣ • Software 2.0 Sungjoo Ha 16
Software Rot • ࠺ૉפझ ۽ب ࣗਝযب ॅ • ߸ ജ҃
߄Շӝ ٸޙ • ࢜۽ ӝמ ࢤӝҊ, ઁಿਸ Ցয оҊ ೞח ߑೱ ߄ՇҊ, ࢎਊо ߄ՇҊ... Ӓېࢲ ॅ • ࠁా Ѧ যڌѱ ೞջ? ࣗਝয ূפযо ٘ܳ ࣻ • Aݶ B ೮חؘ ઁח Aݶ C ೧ঠ ೠ • Ӕؘ ѱ ݽ؛ݶ • ࣘਵ۽ ؘఠ ࣻೞҊ ܳ ߄ఔਵ۽ ݽ؛ ঌইࢲ ജ҃ী • ؘఠ ऺݶ ࢿמب ؊ જই ࣻ Sungjoo Ha 17
Example • ോܻझ౮ vs ݽ؛ • द ࢎۈ ݅ٚ ോܻझ౮
જওਵա ࣘਵ۽ ࢿמ ೞۅ • ݽ؛ द աࡍਵա ࣘਵ۽ ࢿמ ࢚ थ Sungjoo Ha 18
Ideal • ݽٚ ࢎ Ѿ ਃೠ Ҕী ݽ؛ਸ ഝਊ •
ݽٚ Ѫ زച • ҳӖ द • ݂ झாેਸ زਵ۽ ೞѱ ೞৈ ؘఠࣃఠ պߑ࠺ܳ 40 ಌࣃ х • ౠ ࠺ૉפझ/۽؋ ਃ ޙઁܳ AI ޙઁ۽ ജೡ ࣻ ਵݶ ߬झ • Ӓ۞ݶ ҅ࣘ೧ࢲ ࣘਵ۽ ઁಿ ѐࢶغח ҃ਸ ೡ ࣻ ਸ Ѫ Sungjoo Ha 19
Easy? • औ֎? No • ML/ঌҊ્ܻ, ҅ਗ, ؘఠ ыঠ
ೣ • ݅ ॶݽӝ ਤ೧ ೦࢚ ఠޖפ হ ݆ ҅ਗҗ ؘఠо ਃೞח ঋ • ೡ ࣻח ח ب • Ӓېب ݆ਸࣻ۾ ٙ • ӒܻҊ അ ӝࣿ۽ب оמ • ҙ۲ػ ঠӝ ઑӘ ؊ ೧ࠁݶ... Sungjoo Ha 20
ML Model • ೠػ ؘఠ۽ જ Ѿҗܳ যղ۰ݶ ݽ؛ ஏݶ
Ҋ ݆ ਃ • AI ӝࣿ ߊ ࣘبо • ী ղܽ Ѿۿ ࢜۽ ӝࣿ աয়ݶࢲ Ҋغযঠ ೞח ҃ب ݆ • ୭न ӝࣿ زೱਸ ੜ ಝঠ ೣ • ݽ؛۞ٜ ࣗਝয ূפয݂ب ੜೞח Ѫ ൨ٝ • ݽ؍ೠ ஶపց ӝ߈ ূפয݂ী ೠ Ҋ ਃ • ӝࣿ Բ݅ ߄Շযࢲ ࡅܰѱ ٮۄоࠁݶ ߸҃ ݆ • ഈস ҳઑী ೠ Ҋ ਃ Sungjoo Ha 21
Compute • ݽ؛ਸ ٜ݅ӝ ਤ೧ ҅ ਗ ਃ •
҅ ਗਸ ؘఠ۽ ߄Բח ҃ب ݆ • AlphaGo • Data augmentation • ݽ؛ ࢲࡂب য۰ Sungjoo Ha 22
MLOps • ؘఠ ۄੋ • ࣘਵ۽ ݽ؛ णೞח ۄੋ •
ݽ؛ ࢿמ ز ಣо • ز ݽ؛ ߓನ • Shadow mode Sungjoo Ha 23
Data • ઁಿਸ ా೧ ؘఠܳ ദٙೡ ࣻ ח ؘఠ ۚ
ਃ • যڃ ؘఠܳ যڌѱ ࣻೡө? • যڌѱ ਸ ਸө? • Negative sampling ١ • human-in-the-loop दझమਸ যڌѱ ੜ ೡө? • ݽ؛ ੜ ೞҊ ח যڌѱ ࣘਵ۽ ݽפఠ݂ೞ? • ా ؘఠ ਝযೞझ Sungjoo Ha 24
AI PM • ۽؋ ਃ ޙઁܳ AI ޙઁ۽ যڌѱ ജೡ
ࣻ ਸө? • ܳ ా೧ ࢎਊٜ ݅بо য়ܰҊ ؊ ݆ ؘఠܳ ٜ݅যղҊ ؊ ա ઁಿਵ۽ যח • ਸ ӝ ए ҃ب Ҋ য۰ ҃ب Ҋ • AI ӝࣿਸ ೧ೞҊ ী ݏ ઁಿਸ э ҳࢿ೧ঠ ೣ • ӝദী ೠ ࢎ ࣗాਸ ਤೠ ࢜۽ য/بҳ/۽ࣁझ ਃ Sungjoo Ha 25
TikTok • ߑೠ ন బஎ • ন ਸ ࠁೠח Ѻೞী
ӏݽ ݃ாਸ ాೠ ݽٚ פ ൚ࣻ • ࣚऔѱ బஎܳ ٜ݅ ࣻ ب۾ ب৬ח নೠ ઁ بҳ ઁҕ • ݺߔೠ ਸ ਵ۽ ਸ ࣻ ח ߑधਵ۽ UIܳ ҳࢿ • ೠ ߣী ೞա ࠺٣য়݅ ࠁݴ ࢎਊ swipe ৈࠗ۽ ஂೱਸ ݺߔೞѱ ഛੋ • implicit feedback -> explicit feedback • ஏ ӝ߈ ࢎਊ ݽ؛݂ • ױࣽ ੋध(recognition)ীࢲ Ӓח Ѫ ইפۄ ࢎਊо যڌѱ ೯زೞח ݺߔೠ ೖ٘ߔਸ ߉ب۾ ҳࢿ Sungjoo Ha 26
Culture • ੌೞח ߑध, ޙചب ਃ • AI ӝദۄחѱ ই
ઓೞח ঋ • Ӓ۞ࠁפ э যڌѱ ੌೡী ೧ࢲب Ҋ೧ࠊঠ ೣ • ݽٚ ઑ/ҳࢿਗ ী ೠ Ӓܿਸ ݠ݁ࣘী ыҊ ӝৈ೧ঠ ೣ • ࢎ Ѿӂٜب ח ࣻળਵ۽ AIী ೧ ೧೧ঠ • ઑҳઑب ۞ೠ ղਊਸ ߈೧ঠ • Conway ߨ Sungjoo Ha 27
Conclusion • AI ӝࣿਸ ా೧ ী ೡ ࣻ হ؍ Ѫਸ
ೡ ࣻ ח ഥࢎٜ ࢤӡ Ѫ • AI ӝࣿ ݅ ੜ ഝਊೞח Ѫਵ۽ ର߹ചח оמ • ӝࣿী ೠ ೧৬ ઑ ҳઑ ߂ সޖ ߑध ߂ ޙച ١ নೠ ஏݶীࢲ Ҋ۰೧ঠ AI ӝࣿਸ ੜ ഝਊೡ ࣻ • ࢎ Ѿ ܖযח ݽٚ Ҕীࢲ ݽ؛ ഝਊؼ ৈо • ۽؋ ਃೠ ޙઁܳ AI ޙઁ۽ ജೞݶ ࢚ • ܳ ߄ఔਵ۽ ML ݽ؛ ѐࢶ ۽؋ ѐࢶਵ۽ যҊ ח ؊ ݆ ؘఠ۽ য ח ࢶࣽജ ҳઑܳ ٜ݅যյ ࣻ Sungjoo Ha 28