大模型和Copilot在微軟的產(chǎn)業(yè)應用

大模型和Copilot在微軟的產(chǎn)業(yè)應用祝曉鳳微軟(美國)高級應用科學家經(jīng)理目錄自然語言(NLP)處理以及日常應用大語言模型的前世今生ChatGPT在Microsoft的工程化落地狂潮背后隱藏的風險大模型的發(fā)展前景SpeechRecognitionSpamDetectionMachineTranslation案例簡介各個領域對大模型寄予厚望，各式應用尤其是Copilot層出不窮。在本次討論和分享中我們會了解自然語言處理技術的日常應用，從概率模型到網(wǎng)絡語義模型到大語言模型的迭代過程和背后技術提升的原理。微軟作為Copilot的發(fā)起者在各個產(chǎn)品比如Github,Office進行了大模型的深度集成。對于上線的Copilot產(chǎn)品，如何趨利避害有效使用用戶反饋是一個重要的課題。我們會一起深入淺出大語言模型的局限比如幻覺，信息不能及時更新，并且探討落地的解決方案比如使用檢索增強生成。語音識別

(Speech

Recognition)垃圾郵件檢測(Spam

Detection)機器翻譯(Machine

Translation

)檢索自動補全(Autocomplete

in

Search

Engine)情感分析

(Sentiment

Analysis)語法檢錯(Grammar

Checkers)聊天機器人(ChatBot)文本總結(Text

Summarization)圖片標題生成

(Image

Captioning

(multimodal)視覺問答(Visual

Question

Answering(multimodal)

)

…5/area/natural-language-processing你可以列舉出常見的自然語言應用嗎？NLP解決的問題是我們如何讓機器理解并且生成語言使用統(tǒng)計模型到深度學習模型將語言符號轉化成機器可以理解的格式Awordisan

index.Awordisa

vector.Asub-wordisacontextual

vector.6https://dugas.ch/artificial_curiosity/GPT_architecture.html語言模型真的是“大大益善”嗎(統(tǒng)計基于March

2024)Everyday,GPT-3generatestheequivalentofanentireUSpubliclibrary(80,000books)ofnew

content.GPT-3wasestimatedtoconsume500billiontokensandtakeatleast34dayson1024A100(80GB)GPUs~$5

millions!7訓練一個大模型需要燒多少錢？8https://lifearchitect.ai/models/Hey,

GPT!

請你猜猜我要說的下一個字是什么?請猜一猜下一個單詞是“capes”,“pants”,or

“socks”…的概率Test

onunseen

inputs9https://jalammar.github.io/how-gpt3-works-visualizations-animations/Prompts:

您的愿望就是我的命令10Theartofknittingorpaperquillingin

MidjourneyAskRightQuestions:Role+Background+

ObjectiveGPT-4

–

多任務和多模態(tài)學習多任務(Multi-task):

同時訓練多個任務,e.g.,translation,questionanswering,summarization,

etc.多模態(tài)(Multi-modal):多種類型的數(shù)據(jù),e.g.,text,image,audio,

etc.11/blog/2023/03/13/what-are-foundation-models/GitHub

Copilot–

程序也是一種語言!Presentation

Title/items?itemName=GitHub.copilot12您打算如何入圍

ChatGPT

Copilot?Presentation

Title13https:///en-us/microsoft-365/blog/2023/03/16/introducing-microsoft-365-copilot-a-whole-new-way-to-work/AI

還沒有達到準確解釋自己的程度https:///r/ChatGPT/comments/120w4zz/googles_bard_is_truly_something/14我們在使用過程中發(fā)現(xiàn)了哪些局限呢?15TrainedondatauntiltheendofApril2023

(GPT-4)Highcostper

queryNotopen-sourceyet(challengingto

reproduce)Noabilitytosearchtheinternet

directlyCannotcitesourcesaccurately(itcanbeconvincedtoinventreal-sounding

sources!)Verbose

answersInputandoutput

lengthsComparably

slowEthicalsafeguardsareeasily

defeatedNo

personalizationCanbeveryconfident(andright-seeming)evenwhenwrong-

hallucination12.…h(huán)ttps:///r/ChatGPT/comments/zls17p/according_to_chatgpt_1kg_iron_is_heavier_than_1kg/我們如何讓ChatGPT更可靠?16ChatGPT

煉金術–

獎勵還是懲罰?17/2023/01/16/what-is-rlhf/ProximalPolicyOptimization

(PPO)如何讓ChatGPT

接地氣–

檢索增強生成

(RAG)18/en-us/azure/search/retrieval-augmented-generation-overview終極選擇:Finetuning,Prompt

Engineering,

Retrieval-augmented

Generation19ScenariosPreferenceSmallandMedium

LLMsFinetuningSpecificDomainOptimization&

variousmodalitiesFinetuningLonger/shorterinput

lengthsFinetuningUptodate

informationRAGPrivate

dataRAGFactual

groundingRAGLower

costIt

variesSimple&quickintegrationand

testingPromptEngineering&

RAGReasoningLackthegeneralcapabilitiesoflarger

modelsUpdateLLMswithdomain-specificcontextAdaptLLMswithcustomizedinput

requirementsQueryup-to-dateinformationfromweband

databasesStaycompliantwithuserdata

boundaryEvaluateandinspectusingretrieval

resultsPrompt

engineering

&

RAG

does

not

need

model

fine-tuning

butlongpromptsarealso

expensiveAzure

OpenAI

on

Your

Data,

Azure

ML

PromptFlow,

LangChain/@bijit211987/when-to-apply-rag-vs-fine-tuning-90a34e7d6d25ChatGPT

公平性AI-

AzureAI

Content

Safety20/en-us/pricing/details/cognitive-services/content-safety/ExplainableRAI

MetricsHarmful(text)contentdetection(100+

languages)Jailbreaking,Hallucinations,Copyrightinfringements,

etc.Microsoft

產(chǎn)品中蓬勃發(fā)展的Copilot

體驗21/en-us/coilot/定制你的Copilots

-Microsoft

Copilot

Studio22/en-us/microsoft-copilot/microsoft-copilot-studio#modal-12根據(jù)項目任務找到最合適的基礎模型–

Azure

ML23/en-us/azure/machine-learning/concept-model-catalog?view=azureml-api-2總結

Summary24/blog/2023/03/16/introducing-microsoft-365-copilot-your-copilot-for-work/大語言模型的風口

(NLP

in

Daily

Life)大模型風起云涌(TheExplosion

of

LargeLanguage

Models)ChatGPT在微軟的成功產(chǎn)品案例

(ChatGPT

Use

Cases)安全性，可解釋性的緩解措施(Limitations

&

Resolution

Discussions)如火如荼的發(fā)展前景和規(guī)劃(Expectations

on

LLMs)期待您的反饋25感謝致敬Xiaofeng.Zhu@microsoft.co