Upgrade to Pro — share decks privately, control downloads, hide ads and more …

What if...? 처음부터 다시 LLM 어플리케이션을 개발한다면

Hoon Heo
July 25, 2024
Technology
0
2.2k

What if...? 처음부터 다시 LLM 어플리케이션을 개발한다면

Liner LLM 어플리케이션 개발 배움: Orchestration, Cost Operation, Fine-tuning

Hoon Heo

July 25, 2024
Tweet

More Decks by Hoon Heo

See All by Hoon Heo
신뢰할 수 있는 AI 검색 엔진을 만들기 위한 Liner의 여정
huffon
0
610
​고군분투 LLM 프로덕트 적용기: Blind Prompting 부터 Agent까지
huffon
3
1.8k
Autonomous Agent in Production
huffon
2
1.1k
Generative UX in LLM Application
huffon
1
990

Other Decks in Technology

See All in Technology
組織成長を加速させるオンボーディングの取り組み
sudoakiy
2
190
生成AIが変えるデータ分析の全体像
ishikawa_satoru
0
170
[CV勉強会@関東 ECCV2024 読み会] オンラインマッピング x トラッキング MapTracker: Tracking with Strided Memory Fusion for Consistent Vector HD Mapping (Chen+, ECCV24)
abemii
0
220
オープンソースAIとは何か？ --「オープンソースAIの定義 v1.0」詳細解説
shujisado
9
1.1k
複雑なState管理からの脱却
sansantech
PRO
1
150
20241120_JAWS_東京_ランチタイムLT#17_AWS認定全冠の先へ
tsumita
2
300
強いチームと開発生産性
onk
PRO
35
11k
誰も全体を知らない ～ ロールの垣根を超えて引き上げる開発生産性 / Boosting Development Productivity Across Roles
kakehashi
1
230
Evangelismo técnico: ¿qué, cómo y por qué?
trishagee
0
360
障害対応指揮の意思決定と情報共有における価値観 / Waroom Meetup #2
arthur1
5
480
Incident Response Practices: Waroom's Features and Future Challenges
rrreeeyyy
0
160
100 名超が参加した日経グループ横断の競技型 AWS 学習イベント「Nikkei Group AWS GameDay」の紹介/mediajaws202411
nikkei_engineer_recruiting
1
170

Featured

See All Featured
We Have a Design System, Now What?
morganepeng
50
7.2k
Refactoring Trust on Your Teams (GOTO; Chicago 2020)
rmw
31
2.7k
Testing 201, or: Great Expectations
jmmastey
38
7.1k
Become a Pro
speakerdeck
PRO
25
5k
Building Better People: How to give real-time feedback that sticks.
wjessup
364
19k
A Modern Web Designer's Workflow
chriscoyier
693
190k
GitHub's CSS Performance
jonrohan
1030
460k
Visualizing Your Data: Incorporating Mongo into Loggly Infrastructure
mongodb
42
9.2k
Embracing the Ebb and Flow
colly
84
4.5k
Building Your Own Lightsaber
phodgson
103
6.1k
Visualization
eitanlees
145
15k
Building Applications with DynamoDB
mza
90
6.1k

Transcript

  1. ,BSUFS)FP!-JOFS What if…? ୊਺ࠗఠ ׮द LLM য೒ܻா੉࣌ਸ ѐߊೠ׮ݶ : ૑դ

    1֙ р੄ ߓ਑ ӒܻҊ ࢿࣼ೧оח ী੉੹౟ ѐߊ ࢤక҅

  2. ݾର 0SDIFTUSBUJPO $PTU0QFSBUJPO 'JOFUVOJOH

  3. Orchestration

  4. Orchestration ݽ؛ࢶఖदੌ߈੸ੋҊ۰ࢎ೦ ↟$POUFYU-FOHUI ↟1FSGPSNBODF ↟-BUFODZ ↟$PTU ↟3BUFMJNJU ↟4BGFUZ'JMUFS

  5. Orchestration ഑਷ࢎਊ੗ীѱݽ؛ࢶఖӂਸઁҕ೧ঠೡࣻ੓਺

  6. Orchestration ׮নೞѱߊࢤೡࣻ੓ח࠙ӝ߂'BMMCBDL۽૒ ↟૑ա஖ѱӟਃ୒੉ٜয৳ਸٸ ੉ܳযڌѱ୊ܻೡѪੋо  ↟઱۱ݽ؛੉3BUFMJNJUীѦܻѱغ঻ਸٸ ੉ܳযڌѱ୊ܻೡѪੋо  ↟ࢿמࠁ׮ࣘبա࠺ਊ੉઺ਃೠ҃਋ ੉ܳযڌѱ୊ܻೡѪੋо

     ↟ࢎਊ੗оਗೞחݽ؛ਸ૒੽ࢶఖೠ҃਋ ੉ܳযڌѱ୊ܻೡѪੋо  ↟ਗ஖ঋח࢚ടীࢲ೙ఠ݂੉Ѧܽ҃਋ ੉ܳযڌѱ୊ܻೡѪੋо  ↟j

  7. Orchestration .PEFM'BNJMZ߹۽׹߸ࢤࢿ۽૒ਸ߹ѐҳഅ೧઱যঠೣ

  8. Orchestration ੉ਬחݽفة੗੸ੋ4%,ܳഝਊೞҊ੓ӝٸޙ 0QFO"*4%, (PPHMF(FNJOJ4%,

  9. Orchestration ೞա੄4%,ীࢲࢲ۽׮ܲী۞ী؀ೠ'BMMCBDLਸೠٲೠٲҳഅ೧ঠೣ $POUFYU-FOHUI 4BGFUZ 3BUF-JNJU

  10. Orchestration ী੉੹౟ѐߊਸҳਗ೧઴-JUF--.

  11. Orchestration ాੌػੋఠಕ੉झ۽ࢲ۽׮ܲݽ؛ഝਊҗ'BMMCBDL؀਽੉оמ

  12. Orchestration ాੌػੋఠಕ੉झ۽ࢲ۽׮ܲݽ؛ഝਊҗ'BMMCBDL؀਽੉оמ

  13. Orchestration ాੌػੋఠಕ੉झ۽ࢲ۽׮ܲݽ؛ഝਊҗ'BMMCBDL؀਽੉оמ

  14. Orchestration ઑӘ؊਄बਸղࠄ׮ݶ3PVUF--.ѐ֛ө૑੸ਊ

  15. Orchestration ׮নೞѱߊࢤೡࣻ੓ח࠙ӝ߂'BMMCBDL۽૒ ↟૑ա஖ѱӟਃ୒੉ٜয৳ਸٸ ੉ܳযڌѱ୊ܻೡѪੋо  ↟઱۱ݽ؛੉3BUFMJNJUীѦܻѱغ঻ਸٸ ੉ܳযڌѱ୊ܻೡѪੋо  ↟ࢿמࠁ׮ࣘبա࠺ਊ੉઺ਃೠ҃਋ ੉ܳযڌѱ୊ܻೡѪੋо

     ↟ࢎਊ੗оਗೞחݽ؛ਸ૒੽ࢶఖೠ҃਋ ੉ܳযڌѱ୊ܻೡѪੋо  ↟ਗ஖ঋח࢚ടীࢲ೙ఠ݂੉Ѧܽ҃਋ ੉ܳযڌѱ୊ܻೡѪੋо  ↟j

  16. Orchestration 3PVUF--.਷औѱ݈೧௪ܻܳۄ਋ఠীకਕ੸׼ೠݽ؛۽ࠁղחѐ֛

  17. Orchestration 3PVUF--.ѐ֛੄੸ਊਸా೧ࣘب৬࠺ਊ୭੸ചоמ

  18. Cost Operation

  19. Cost Operation ݆਷ӝস੉--.࠺ਊҙܻীয۰਑ਸѻҊ੓਺

  20. Cost Operation ৔ܻӝস਷ѾҴઁಿਸా೧੉ਮਸٜ݅যյࣻ੓যঠೣ

  21. Cost Operation ࠺ऴоѺഋࢿਵ۽߸ز࠺ࢿѺীоӰѱ࠙ܨغח--.࠺ਊ

  22. Cost Operation ѾҴ--.য೒ܻா੉࣌਷--.࠺ਊ୭੸ചীٮۄӝস੉ਮী৔ೱਸ઱ѱؽ

  23. Cost Operation ޷द੸ҙ੼ীࢲஹನք౟ױਤ࠺ਊ࠺઺ী؀ೠхਸղ੤ച೧ঠೣ

  24. Cost Operation xyz ࡺ݅ইפۄੌ੗߹۽যځೠন࢚ਸࠁ੉ח૑ ౠ੉੼਷হ঻ח૑١بҙஏ

  25. Cost Operation Ѣद੸ҙ੼ীࢲੌр࠺ਊ૑ࠛ੉যڌѱഋࢿغח૑ب੗زചా೧ҙஏ

  26. Cost Operation ب؀୓੉۞ೠࣻ஖חযڌѱࢸ੿ೞѱغחѦө

  27. Cost Operation ೠݺ੄ࢎਊ੗ীѱ঴݃ܳߥযٜੌࣻ੓ח૑҅࢑

  28. Cost Operation ೠݺ੄ࢎਊ੗ীѱੌрदр࠙рಣӐ঴݃ܳై੗ೡࣻ੓ח૑҅࢑

  29. Cost Operation ੸੿ై੗оמࣻ஖ܳઁಿী੸ਊ

  30. Cost Operation ୭Ӕ--.оѺ࠺Үبҳب݆੉աয়Ҋ੓਺

  31. Cost Operation ↟࠺ਊী੓য--.ীై੗оמೠࣻળী؀ೠхਸ૑פҊ੓যঠೣ ↟੉ীӝ߈ೞৈ#VEHFUJOH 3PVUJOH 5ISFTIPME١੸ਊ ↟૑ࣘ੸ਵ۽ઑ૒ղ࠺ਊхпਸ੊൧ࣻ੓ب۾ݽפఠ݂ दпചъച ↟--.ਸ׸׼ೞחઑ૒਷ࢿמ ࠺ਊ

    ࣘبࣁਃੋী؀೧ࢻࣁೞѱ੽Ӕ ↟ࣁਃੋ઺बਵ۽ઑ૒ݾ಴ܳࢸ੿ೞҊ ୭੸ചܳਤೠز۱ਸٜ݅যঠೣ

  32. Fine-tuning

  33. Fine-tuning: ࣁ ױ҅۽ ա־য ࠁח LLM য೒ܻா੉࣌ ߊ੹ ઱ӝ 1IBTF

    0QFO"* "OUISPQJD (FNJOJ١ 1SPWJEFSഝਊ೧য೒ܻா੉࣌ҳഅ 1&'5

  34. Fine-tuning: ࣁ ױ҅۽ ա־য ࠁח LLM য೒ܻா੉࣌ ߊ੹ ઱ӝ 1IBTF

    1IBTF 0QFO"* "OUISPQJD (FNJOJ١ 1SPWJEFSഝਊ೧য೒ܻா੉࣌ҳഅ 'JSFXPSLT"* 5PHFUIFS"*١ ݽ؛೟ण߂ࢲࡂ೒ۖಬഝਊ 1&'5 1&'5

  35. Fine-tuning: ࣁ ױ҅۽ ա־য ࠁח LLM য೒ܻா੉࣌ ߊ੹ ઱ӝ 1IBTF

    1IBTF 1IBTF 0QFO"* "OUISPQJD (FNJOJ١ 1SPWJEFSഝਊ೧য೒ܻா੉࣌ҳഅ 'JSFXPSLT"* 5PHFUIFS"*١ ݽ؛೟ण߂ࢲࡂ೒ۖಬഝਊ (16௿۞झఠഛࠁ೧ ૒੽ݽ؛೟ण߂ࢲࡂ୭੸ച 1&'5 1&'5 1&'5 'VMMQBSBNFUFSUVOJOH 1PTUUSBJOJOH

  36. Fine-tuning: ࣁ ױ҅۽ ա־য ࠁח LLM য೒ܻா੉࣌ ߊ੹ ઱ӝ 1IBTF

    1IBTF 1IBTF 0QFO"* "OUISPQJD (FNJOJ١ 1SPWJEFSഝਊ೧য೒ܻா੉࣌ҳഅ 'JSFXPSLT"* 5PHFUIFS"*١ ݽ؛೟ण߂ࢲࡂ೒ۖಬഝਊ (16௿۞झఠഛࠁ೧ ૒੽ݽ؛೟ण߂ࢲࡂ୭੸ച 1&'5 1&'5 1&'5 'VMMQBSBNFUFSUVOJOH 1PTUUSBJOJOH -JOFSח੉ױ҅ীਤ஖ ֙ਘ

  37. Fine-tuning 'JOFUVOJOHਵ۽ֈযоӝ੹'JOFUVOJOHਸ೧ঠ݅ೞח੉ਬী؀೧೤੄ܳݏ୶যঠೣ

  38. Fine-tuning 'JOFUVOJOHਸా೧--.࠺ਊ୭੸ചܳ੉ܖযյࣻ੓ਸ૑  ೐۽ં࣌೙ਃ

  39. Fine-tuning 'JOFUVOJOHਸా೧ఋࢎࠁ׮ڪযդஹನք౟ٜܳ݅যյࣻ੓ਸ૑  ࢿמૐݺ೙ਃ

  40. Fine-tuning 'JOFUVOJOHਸా೧ۨߡܻ૑ೡࣻ੓ח1SPQSJFUBSZؘ੉ఠо੓ח૑  झझ۽ঌҊ੓ਸѪ

  41. Fine-tuning খ੄җ੿ਸ૑աݶಽҊ੗ೞחޙઁী؀ೠ੿੄о೙ࣻ੸

  42. Fine-tuning ޙઁ੿੄о࢚ࣁ൤ ೤੄ػ଻۽غ૑ঋਵݶ೟ण੉়ڨೞѱൗ۞тࣻ੓਺

  43. Fine-tuning ੉റ೟णҗ੿ীࢲஏ੿ೡࣻ੓חݫ౟ܼী؀ೠ੿੄о೙ਃ

  44. Fine-tuning ੹ా੸ੋݫ౟ܼ੄੸ਊ੉য۵׮ݶ--.BTB+VEHFܳഝਊ೧ࠅࣻ੓਺

  45. Fine-tuning ੿੄ػޙઁ৬ݫ౟ܼ ӒܻҊపझ౟ؘ੉ఠࣇী؀ೠ"4*4ஏ੿

  46. Fine-tuning ೟णػݽ؛੄ࢿמҗ"4*4࠺Үೞݴࢿמ ࠺ਊ ࣘبୋب۽ߓನѾ੿

  47. We are hiring!

  48. Thank you 5IBOLTUP5ZMFS ,BOF ,BJ +VTUJO♥