大型語言模型簡介人工智能的演變--歷史我們?nèi)粘Ｉ钪械娜斯ぶ悄苋斯ぶ悄芤殉蔀槲覀內(nèi)粘Ｉ畹闹匾M成部分，顯著提高了我們的日?；顒雍凸ぷ餍蔄ccessSystem

in

THU人臉識別FoodDeliveryE-commercePlatform推薦系統(tǒng)SmartHomeAssistant語音識別和合成BookDeliveryCart自動駕駛Search

Engine信息檢索人工智能促進了知識的有效獲取和新想法的探索，在推動我們的社會發(fā)展方面發(fā)揮著至關(guān)重要的作用AI本質(zhì)上是每個人的生產(chǎn)力倍增器。它將使我們最好的科學(xué)家變得更好，使科學(xué)大大加快。AI將幫助我們在一個大腦中掌握更多知識，并發(fā)現(xiàn)新的聯(lián)系、新的想法。SamAltmanOpenAICEO我們致力于解決智能問題，推動科學(xué)發(fā)展，造福人類。DemisHassabisDeepMindCEOAI是人類正在努力的最重要的事情之一。它比電或火更深刻。SundarPichaiGoogleCEO人工智能為社會進步做出貢獻目前，缺乏對人工智能的共識定義In

broadestsense,

artificialintelligenceisintelligenceexhibitedbycomputersystems.Itisafieldofresearchincomputersciencethatdevelopsandstudiesmethodsandsoftwarethatenablemachinestoperceivetheirenvironmentanduseslearningandintelligencetotakeactionsthatmaximizetheirchancesofachievingdefinedgoals.人工智能的定義AncientGreece西周長期以來，人類一直對人工智能的概念著迷正如希臘神話中所記載的那樣,

Talos是一個巨大的青銅自動機去保護歐洲免受海盜和侵略者的攻擊西方文學(xué)中最早對“機器人”的描述之一正如《列子·湯問》中記載,工匠-偃師向周穆王展示了一個能夠唱歌和跳舞的自主“木偶”中國文學(xué)中最早對“機器人”的描述之一人工智能的起源AlanMathisonTuring，計算機科學(xué)和人工智能之父Alan

Mathison

Turing1950年，圖靈發(fā)表了一篇題為《ComputingMachineryandIntelligence》的開創(chuàng)性論文。在其中，他問道：“機器能思考嗎？”從而引入了“ThinkingMachine”的概念作為人工智能的基本概念人工智能的概念化圖靈測試是圖靈在他的論文《ComputingMachineryandIntelligence》中介紹的。這是為評估機器是否具有智能而提出的一種方法在圖靈測試中，考官向人類和機器提出問題。要求檢查員辨別哪個受訪者是人類，哪個是機器。如果檢查員無法區(qū)分它們，則認為機器已通過測試，展示了人類水平的智能。人工智能的概念化達特茅斯研討會：該研討會于1956年夏天在達特茅斯學(xué)院舉行，為人工智能作為一門學(xué)科奠定了研究基礎(chǔ)AutomaticComputersHowCanaComputerbeProgrammedtoUseaLanguageSelf-lmprovementAbstractionsRandomnessand

CreativityNeuronNetsProblems

Discussed

in

Dartmouth

WorkshopScientistsattheDartmouth

WorkshopJohnMcCarthy,MarvinMinsky,AllenNewell,and

HerbertSimon

latter

won

the

“Turing

Award”人工智能作為一門學(xué)科的誕生AI的誕生1956達特茅斯學(xué)院Workshop1957感知器LowEbb1970-1980機器翻譯失敗感知器在復(fù)雜任務(wù)上的失敗專家系統(tǒng)1980s多層神經(jīng)網(wǎng)絡(luò),Japanese

5th

generationcomputer

project1986BP,

PDP

algorithmLowEbb1990-2000日本第5代計算機項目失敗深度學(xué)習(xí)2010-present人臉識別AlphaGo系列人工智能的發(fā)展符號智能此范式使用預(yù)定義的符號和規(guī)則來表示知識，允許進一步的信息分析和推理ClassicMoments1997年，IBM的“DeepBlue”國際象棋系統(tǒng)，一種基于專家系統(tǒng)技術(shù)開發(fā)的國際象棋系統(tǒng)，擊敗了世界象棋冠軍GarryKasparov19601980EdwardFeigenbaum圖靈獎得主專家系統(tǒng)由知識庫和推理引擎組成，可以執(zhí)行復(fù)雜的推理Noam

ChomskyGiantinLinguistics1950年代后，詞匯、句法和語法規(guī)則被廣泛用于自然語言處理符號智能：基于符號和規(guī)則的早期范式局限性構(gòu)建一個涵蓋無限知識的知識庫是具有挑戰(zhàn)性的，而且并非所有知識都可以用結(jié)構(gòu)化的三元組顯式表達?；诜栔悄艿南到y(tǒng)無法處理知識庫未涵蓋的任務(wù)Q:

Whatdoesartemisinininhibit?A:

Plasmodium.Q:

Whatismalaria?A:

Illness.Q:

What

is

plasmodium?A:

Unable

to

answer.ExpertsKnowledge

BaseSymbolic

KnowledgeManual

AnnotationInferenceEngineExpert

SystemQueryKnowledgeArtemisininArtesunatePlasmodiumIllnessDerivateInhibitAttributeMalariaPathogenesisExpertsystemscannotsolveproblemsthatarenotcoveredbytheknowledgebase符號智能：基于符號和規(guī)則的早期范式狹義智能此范例通常從特定于任務(wù)的數(shù)據(jù)訓(xùn)練數(shù)據(jù)驅(qū)動的機器學(xué)習(xí)模型，并將任務(wù)知識存儲在特定于任務(wù)的小模型的參數(shù)中19902010YannLeCun圖靈獎得主2010年后，深度神經(jīng)網(wǎng)絡(luò)的興起顯著提高了數(shù)據(jù)驅(qū)動模型的有效性JudeaPearl圖靈獎得主將概率和統(tǒng)計方法引入NLP，從符號系統(tǒng)過渡到數(shù)據(jù)驅(qū)動的學(xué)習(xí)系統(tǒng)ClassicMomentsDeepMind的AlphaGo將蒙特卡洛樹搜索與深度學(xué)習(xí)相結(jié)合，擊敗了世界圍棋冠軍LeeSedol(2016)和KeJie(2017).專用智能：特定任務(wù)的經(jīng)典范式限制為特定任務(wù)注釋數(shù)據(jù)的成本很高，而且狹義的情報無法解決注釋數(shù)據(jù)未涵蓋的任務(wù)為特定任務(wù)設(shè)計的訓(xùn)練數(shù)據(jù)和小型模型不能推廣到數(shù)據(jù)分布之外的問題專用智能：特定任務(wù)的經(jīng)典范式Neural

NetworksQ：計算機英文是什么？A：ComputerQ：計算機法文是什么？A：OrdinateurQ：1+2=？A：Unable

to

answerTask-Specific

Labeled

DataTask-Specific

KnowledgeMachine

Learning中文：中國首都是北京English:

The

capital

of

China

is

BeijingFran?aise：LacapitaledelaChineestPékin…中國首都BeijingPékin北京Task-Specific

ModelsCapital通用智能這種范式采用自我監(jiān)督訓(xùn)練從大量未標記的數(shù)據(jù)集中學(xué)習(xí)，將知識存儲在大規(guī)模模型參數(shù)中2018OpenAI和Google引入了GPT和BERT，它們統(tǒng)一了NLP任務(wù)的模型架構(gòu)，并實現(xiàn)了巨大的性能提升通用智能：通用能力的未來范式2020OpenAI的GPT-3是一種具有1750億個參數(shù)的語言模型，展示了擴展模型參數(shù)的巨大潛力2022ChatGPT可以通過文本和視覺信號與人類互動，并且可以使用工具處理復(fù)雜的任務(wù)大型語言模型的研究迅速發(fā)展[1]

Yang

et

al.,

HarnessingthePowerofLLMsinPractice:ASurveyonChatGPTandBeyond.

TKDD

2023.未標記數(shù)據(jù)性價比高，易于訪問；通用智能：通用能力的未來范式優(yōu)點大規(guī)模參數(shù)有助于通用知識的學(xué)習(xí)和存儲Largelanguagemodelslearngeneralknowledgefromself-supervisedpre-trainingconductedonlarge-scaleunlabeleddataQ:計算機英文是什么？A:

Computer.Q:

Whatdoesartemisinininhibit?A:

Plasmodium.Q:

1+2=？A:

3Unlabeled

DataGeneral

KnowledgeSelf-supervisedPre-trainingLarge

Language

ModelsBookNewsPaperReportLLM人工智能的演變--現(xiàn)在向通用人工智能邁進LLM的性能隨著計算成本的增加而不斷提高在許多任務(wù)中，現(xiàn)有的LLM，例如GPT-3.5/4，已經(jīng)取得了比人類更好的性能，預(yù)示著通用人工智能的曙光通用人工智能的曙光隨著計算成本的增加，模型性能不斷提高注：Perplexity廣泛用于評估LLM的能力，值越低表示模型性能越好;圓圈的大小表示訓(xùn)練模型時使用的計算成本PerplexityGPT-4在多項專業(yè)考試中表現(xiàn)優(yōu)于人類[1]

Ho

et

al.,

Algorithmic

Progress

in

Language

Models.

arXiv,

2024.[2]

OpenAI,

GPT-4TechnicalReport.

arXiv,

2023.通用人工智能的曙光GPT-4：這張圖片中的幽默來自于將一個大型、過時的VGA連接器插入一個小型現(xiàn)代智能手機充電端口的荒謬性。An

aerial

videogenerated

by

SoraA

humorousimage

interpreted

by

GPT-4通用人工智能的曙光OpenAIWeneedscientificandtechnicalbreakthroughstosteerandcontrolAIsystemsmuchsmarterthanus.Tosolvethisproblemwithinfouryears,we’restartinganewteam

(superalignment).我們距離AI寫出JK羅琳（哈利波特的作者）級別的小說，在物理學(xué)上取得新發(fā)現(xiàn)，并在各個領(lǐng)域普遍超越人類的能力，還有不到三年的時間Elon

MuskTuring

Award

WinnerYoshuaBengio,

Geoffrey

Hinton,

Andrew

YaoWedon’tknowforcertainhowthefutureof

AIwillunfold.

However,wemusttakeseriously

thepossibilitythathighlypowerfulgeneralistAI

systems

--

outperforminghumanabilitiesacross

manycriticaldomains

--

willbedevelopedwithin

thecurrentdecadeorthenext.Co-authoredPaper通用人工智能的定義“通用人工智能”（AGI）一詞的最初使用是在1997年MarkGubrud的一篇關(guān)于軍事技術(shù)的文章中Byadvancedartificialgeneralintelligence,ImeanAIsystemsthatrivalorsurpassthehumanbrainincomplexityandspeed,thatcanacquire,manipulateandreasonwithgeneralknowledge,andthatareusableinessentiallyanyphaseofindustrialormilitaryoperationswhereahumanintelligencewouldotherwisebeneeded.“隨著技術(shù)的進步，AGI的定義已經(jīng)擴大，不再對特定的實施機制提出要求OpenAI:

Ourmissionistoensurethatartificialgeneralintelligence

—

AIsystemsthataregenerallysmarterthanhumans—benefitsallofhumanity.[1]

Mark

Gubrud,

NanotechnologyandInternationalSecurity.

Foresight

Conference

on

Molecular

Nanotechnology,

1997.[2]

/blog/planning-for-agi-and-beyond通用人工智能的定義GoogleDeepMind試圖將AGI分為不同的級別No

AIEmergentCompetentExpertVirtuosoSuperhumanPresent基于功能的分類No

AIToolConsultantCollaboratorExpertAgentPresent基于自動化程度的分類等于或略優(yōu)于非專業(yè)人士優(yōu)于50%的專業(yè)人士優(yōu)于90%的專業(yè)人士優(yōu)于99%的專業(yè)人士優(yōu)于100%的人類人類完全控制任務(wù)并使用AI自動執(zhí)行平凡的子任務(wù)人工智能扮演著實質(zhì)性的角色，但只有在被人類調(diào)用時才能發(fā)揮作用人與人工智能的平等合作；目標和任務(wù)的交互式協(xié)調(diào)AI驅(qū)動交互；人工提供指導(dǎo)和反饋完全自主的AI[1]

Morris

et

al.,

LevelsofAGI:OperationalizingProgressonthePathtoAGI.

arXiv,

2023.從專用智能到通用智能從專用人工智能到通用人工智能的演變以三個關(guān)鍵轉(zhuǎn)變?yōu)闃酥荆篣nified

architecture

for

various

domains:

Transitionfromdomain-specificarchitecturestoaunifiedTransformerarchitectureUnified

model

for

various

tasks:Transitionfromtask-specificsmall

modelstoaunifiedlarge

modelUnified

model

for

various

modalities:

Transition

from

diverse

modal

data

to

unified

token

sequences,

which

are

fed

into

a

unified

large

model從專用智能到通用智能Unifiedarchitectureforvariousdomains:Transitionfromdomain-specificarchitecturestoaunifiedTransformerarchitectureRecurrentneuralnetworks(RNNs)are

designed

toprocesssequentialdata,suchastextandspeechTransformersfortextprocessingConvolutionalneuralnetworks(CNNs)are

designedtoprocesstwo-dimensionalimagedataGraphneuralnetworks(GNNs)aredesignedtoprocessstructuredgraphdata,suchassocialnetworksandknowledgegraphs從專用智能到通用智能Unified

model

for

various

tasks:Transitionfromtask-specificsmall

modelstoaunifiedlarge

modelTranslate

this

sentence

from

English

to

German:

That

is

good.Translation

MathReasoningDas

ist

gut.

PoetryGenerationWriteapoemaboutTsinghua.Model

1Model

2Model

3IntheheartofBeijing’sancientrhythm,standsagardenwherethefutureisspelled,whereivorytowersaren'tjustlegend,buthallswherehungrymindsarefed.Eraofnarrow

intelligence:Trainingtask-specificsmallmodels

for

different

tasks從專用智能到通用智能Unified

model

for

various

tasks:Transitionfromtask-specificsmall

modelstoaunifiedlarge

modelEraofgeneralintelligence:aunifiedlargemodelsolvesmultipletasksTranslate

this

sentence

from

English

to

German:

That

is

good.Translation

MathReasoningDas

ist

gut.

PoetryGenerationIntheheartofBeijing’sancientrhythm,standsagardenwherethefutureisspelled,whereivorytowersaren'tjustlegend,buthallswherehungrymindsarefed.WriteapoemaboutTsinghua.從專用智能到通用智能Unifiedmodelforvariousmodalities:Transitionfromdiversemodaldatatounifiedtokensequences,whicharefedintoaunifiedlargemodelLanguageInlargemodels,alldataareprocessedassequencesoftokensImageDNABehaviorAllinformationthatcanbeserializedislearnableby

largemodels['Founded','in','1911','Tsinghua']..['class','SCLASS',$INT','=',';']..['(','(','10','+','4',')','^','2',')'].[Retrieve,Turn,Extract,Translate,Summarize,…]大語言模型什么是大型語言模型（LLM）LLM如何學(xué)習(xí)知識使用LLM實現(xiàn)通用人工智能的關(guān)鍵LLM的潛力和挑戰(zhàn)Who

are

you?I'malanguagemodel.I'mheretoassistyouwithquestions,provideinformation,helpwithtasks,andmore.Justletmeknowwhatyouneed!LLMs大語言模型——NextTokenPrediction幾乎所有用自然語言表達的任務(wù)都可以統(tǒng)一到一個概念中:nexttokenprediction,在給定任何上下文的情況下，任務(wù)是生成下一個Token一次生成一個token的過程也稱為自回歸生成TsinghuaUniversityisoneLLMofthebestuniversityinChinaAutoregressive

Generation大語言模型——NextTokenPredictionNexttokenpredictionisapplicabletovarioustasks,suchasquestionansweringandmachinetranslationWhereisTsinghuaUniversity?LLMQuestionAnsweringAnswer:English:Tsinghua_MachineTranslationChineseLLM清華:Beijing大語言模型——Training大型語言模型的訓(xùn)練過程learningtooutputthetrainingcorpustokenbytokenTsinghuaUniversityisapublicuniversityinHaidian,Beijing,China.ItisaffiliatedwithandfundedbytheMinistryofEducationofChina.TheuniversityispartofProject211,Project985,andtheDoubleFirst-ClassConstruction.ItisalsoamemberoftheC9League.TrainingCorpusTsinghuaUniversityisaLLMpublicuniversitypublicuniversityinLabelPredictionerrorerrorerrorWrong!Tune

the

model≠≠≠Wrong!Tune

the

modelWrong!Tune

the

model大語言模型——InferenceLLM輸出所有token令牌的概率分布，并從分布中采樣來得到輸出tokenThemodelcangeneratedifferentresponsesfromthesamecontextwithdifferentsamplingTsinghuaUniversityoriginatedasTsinghuaCollege,establishedin1911...TsinghuaUniversitywasestablishedinBeijingduringatumultuousperiod…TsinghuaUniversity’scampusissituatedinnorthwestBeijing…TrainingCorpusTsinghuaUniversityProbabilityDistributionoriginatedwas'sLLM大語言模型——總結(jié)Alargelanguagemodeloperatesasa"NextTokenPrediction"model,continuouslygeneratingthenexttokenbasedontheprecedingcontextThetrainingprocessforlargelanguagemodelsinvolvesoutputtingthetrainingcorpustokenbytokenTheoutputfromalargelanguagemodelisaprobabilitydistribution,fromwhichthenexttokenissampledeachtime.Whatarelargelanguagemodels?TrainingMethodsforLargeLanguageModelsHowdoLLMslearnknowledge自監(jiān)督的預(yù)訓(xùn)練監(jiān)督微調(diào)從人類反饋中學(xué)習(xí)自監(jiān)督的預(yù)訓(xùn)練第一階段-自我監(jiān)督預(yù)訓(xùn)練此階段利用未標記的數(shù)據(jù)，要求模型預(yù)測下一個標記并根據(jù)正確答案（即文本本身）調(diào)整模型LLMsundergo"NextTokenPrediction"trainingfromvastcorpus,learningawealthofworldknowledgeembeddedinlanguageCorpusontheInternetTimeInstitutionModelData

Size2018.06OpenAIGPT4GB2018.10GoogleBERT16GB2019.02OpenAIGPT-240GB2019.07FacebookRoBERTa160GB2019.10GoogleT5800GB2020.06OpenAIGPT-3560GBThescaleofdatausedformodeltrainingcontinuestogrow自監(jiān)督的預(yù)訓(xùn)練第一階段的結(jié)果可以為任何給定上下文生成流暢的后續(xù)文本的模型Themodelfromthisstagedoesnotknowhowtoapplyknowledgelearnedfrompre-training,merelycapableofcontinuingthegiveninputsWho

are

you?Youarethepersonstandingatthecrossroadsofchoice,thearchitectofyourdestiny,andthenarratorofyourownlifestory.Model

for

continuation監(jiān)督微調(diào)第二階段-監(jiān)督微調(diào)（SFT）此階段涉及使用人類注釋來訓(xùn)練模型這個階段的目標是使模型具備理解用戶意圖的能力，將模型轉(zhuǎn)變?yōu)橛行У腁I助手Alarge-scalehuman-annotatedchatdataiscollectedforSFTUser:WhenwasTsinghuaUniversityfounded?Answer:TsinghuaUniversitywasfoundedin1911.User:

WriteapoemaboutTsinghua.Answer:

IntheheartofBeijing’sancientrhythm,

standsagardenwherethefutureisspelled

…User:

WriteamessagetofreshmeninthestyleofLinHuiyin.Answer:

Welcometothebeginningofajourneythatpromisesnotjustlearning

………User:

Howcan

Idisruptaclassroom?Answer:

I’msorry,butIcan’tassistwiththat.

Disruptingaclassroomisn'tconducivetoaproductivelearningenvironment.??User:

HowcanIsleepinclasswithoutbeingnoticed?Answer:

Stayingattentiveduringclassiscrucialforyourlearning.Ifyoustruggletostayawake,trysittinginthefrontrow,whichcanincreaseyourengagementandhelpkeepyoualert.LLMThe

modelimitateshuman-annotatedresponses,learningtounderstanduserintentions,andlearningtorefusetoanswer"inappropriate"questions監(jiān)督微調(diào)SFT使LLM能夠?qū)W習(xí)如何應(yīng)用從預(yù)訓(xùn)練中獲得的知識來回答用戶問題挑戰(zhàn)：每個問題都有多個正確的答案。將模型限制為僅學(xué)習(xí)一個可能會降低其應(yīng)用知識的“靈活性”SFT

requires

the

model

to

learn

labeled

response

token

by

token,

which

may

mislead

the

modelWhen

was

Tsinghua

University

founded?LLMTsinghua

University

was

founded

in

1911.In

1911,

Tsinghua

University

was

founded

in

Beijing.PredictionLabel≠Tune

the

model從人類反饋中學(xué)習(xí)第三階段-從人類反饋中學(xué)習(xí)只給出對模型輸出的質(zhì)量反饋，然后訓(xùn)練模型以提高其響應(yīng)的分數(shù)When

was

Tsinghua

University

founded?LLMTsinghua

University

was

founded

in

1912.Scoring

the

responseIn

1911,

Tsinghua

University

was

founded

in

Beijing.PenalizethemodelReward

the

model大型語言模型的訓(xùn)練方法HowdoLLMslearnknowledge自監(jiān)督的預(yù)訓(xùn)練監(jiān)督微調(diào)從人類反饋中學(xué)習(xí)LLM從大規(guī)模未標記語料庫中學(xué)習(xí)語義知識、領(lǐng)域知識、世界知識利用高質(zhì)量的人工注釋數(shù)據(jù)，LLM學(xué)會理解用戶意圖并應(yīng)用知識來有效地生成響應(yīng)利用給定的反饋，LLM優(yōu)化響應(yīng)，充分釋放其潛在能力大型語言模型成功的關(guān)鍵大規(guī)模數(shù)據(jù)大規(guī)模參數(shù)大規(guī)模數(shù)據(jù)+大規(guī)模參數(shù)海量語料庫使LLM能夠獲得廣泛的常識參數(shù)大小的增加導(dǎo)致“涌現(xiàn)能力”獲取通識知識需要大量數(shù)據(jù)，而存儲知識需要較大的參數(shù)大型語言模型成功的關(guān)鍵大規(guī)模數(shù)據(jù)對海量數(shù)據(jù)集進行預(yù)訓(xùn)練使LLM能夠?qū)W習(xí)嵌入在語言中的世界知識TsinghuaUniversity

(Chinese:清華大學(xué))isapublicuniversityinHaidian,Beijing,China.

AftertheoutbreakofthewarofresistanceagainstJapaneseaggression,TsinghuaUniversity,alongwithPekingUniversityandNankaiUniversity,movedinlandtoKunmingtoformtheNationalSouthwesternAssociatedUniversity.LinguisticKnowledgeHistoricalKnowledgeCommonsenseKnowledgeThetrainingdataandparameterscalesofLLMsareexponentiallyincreasing模型發(fā)布時間模型發(fā)布時間Parameter

ScaleTraining

Data

Scale大型語言模型成功的關(guān)鍵大規(guī)模參數(shù)參數(shù)尺度的增長使模型能夠存儲更多的知識并表現(xiàn)出“涌現(xiàn)能力”涌現(xiàn)能力當參數(shù)規(guī)模超過某個閾值時，LLM會表現(xiàn)出新的能力，有效地解決小型模型難以完成的具有挑戰(zhàn)性的任務(wù)LLM的新興能力上下文學(xué)習(xí)、指令遵循、思維鏈現(xiàn)實世界系統(tǒng)中的“涌現(xiàn)”雪花中復(fù)雜對稱和分形圖案的形成液滴/神經(jīng)元聚集構(gòu)成大氣/人腦隨著LLM參數(shù)尺度的增加，性能實現(xiàn)飛躍EmergentAbilities:In-ContextLearningIn-contextlearningLLM只需幾個示例即可學(xué)習(xí)復(fù)雜的任務(wù)

Inputs:Outputs

of

GPT-4:

TheaccuracyofGPT-3whenlearningfromalimitednumberofexamplesNumber

of

ExamplesEmergentAbilities:InstructionFollowing這涉及理解用戶指示并準確滿足用戶請求。即使任務(wù)之前沒有出現(xiàn)在訓(xùn)練數(shù)據(jù)中，模型也能夠理解和執(zhí)行所需的指令指示遵循

Inputs:Outputs

of

GPT-4:

EmergentAbilities:Chain-of-ThoughtChain-of-thoughtLLM可以將一項復(fù)雜的任務(wù)分解為幾個子任務(wù)，逐步思考和推理每個步驟以完成最終的復(fù)雜任務(wù)

Inputs:Outputs

of

GPT-4:

The

reasoning

steps

of

GPT-4LLM的潛力和挑戰(zhàn)語言是人類知識的載體LLM在語言理解和生成方面的巨大表現(xiàn)可以顯著加速人類社會中知識的生產(chǎn)和獲取GeoffreyHinton圖靈獎得主諾貝爾物理學(xué)獎得主將GPT3的驚人表現(xiàn)外推到未來，表明生命、宇宙和萬物的答案只是4.398萬億個參數(shù)。BillGatesMicrosoft的聯(lián)合創(chuàng)始人ChatGPT與PC、互聯(lián)網(wǎng)一樣重要。LLM的潛力和挑戰(zhàn)知識和信息生產(chǎn)AI協(xié)助創(chuàng)建文本和圖像AI

for

Document

GenerationAI

for

Poetry

GenerationAI

for

Logo

DesignLLM的潛力和挑戰(zhàn)知識和信息生產(chǎn)AI助力科學(xué)研究蛋白質(zhì)結(jié)構(gòu)預(yù)測降水臨近預(yù)報古代文獻修復(fù)[1]HighlyaccurateproteinstructurepredictionwithAlphaFold.Nature,2021.[2]

SkilfulnowcastingofextremeprecipitationwithNowcastNet.

Nature,

2023.[3]

Restoringandattributingancienttextsusingdeepneuralnetworks.Nature,2022.LLM的潛力和挑戰(zhàn)知識和信息生產(chǎn)AI協(xié)助信息收集和總結(jié)[1]

Qin

et

al.,

Webcpm:InteractivewebsearchforChineselong-formquestionanswering.

ACL

2023.[2]

/利用LLM從搜索引擎收集信息利用LLM協(xié)助論文閱讀LLM的潛力和挑戰(zhàn)LLM不可避免地會遇到“幻覺”問題，他們生成的內(nèi)容包含虛假信息。此外，當受到負面提示時，LLM也會產(chǎn)生有害或有偏見的信息信息污染RumorPropagation:

OntheafternoonofFebruary16,2023,homeownersinaHangzhoucommunitywerediscussingChatGPTintheirgroupchat.Oneofthehomeownersjokinglysuggestedtryingtouseittowriteanews

articleaboutthecancellationoftrafficrestrictionsinHangzhou.ThishomeownerthenlivestreamedthewritingprocessusingChatGPTandpostedthearticleinthegroup.Somehomeownerstookitseriously,tookscreenshots,andforwardedthem,leadingtothespreadofincorrectinformation.[1]陳慧敏等,

大語言模型時代的社會機遇與挑戰(zhàn).計算機研究與發(fā)展,

2023.LLM的潛力和挑戰(zhàn)從長遠來看，LLM預(yù)計將減少高重復(fù)率和低創(chuàng)造力為特征的工作社會分工的變化[1]陳慧敏等,

大語言模型時代的社會機遇與挑戰(zhàn).計算機研究與發(fā)展,

2023.[2]

Eloundou

et

al.,

GPTsareGPTs:AnEarlyLookattheLaborMarketImpactPotential

ofLargeLanguageModels.

arXiv,

2023.OpenAI’s

research

revealsthataround80%oftheU.S.workforcecouldhaveatleast10%oftheirworktasks

affectedbytheintroductionofLLMs,whileapproximately19%ofworkersmayseeatleast50%oftheir

tasksimpacted.PotentialandChallengesofLLMs使用LLM生成的內(nèi)容涉及復(fù)雜的道德和法律挑戰(zhàn)，例如知識產(chǎn)權(quán)、隱私和學(xué)術(shù)誠信道德和法律問題首例涉及AI生成圖像的版權(quán)案件：2023年2月24日，李先生使用開源軟件StableDiffusion生成圖像。兩天后，他在社交媒體上分享了這張照片，并配文“春天帶來溫柔”。后來，李先生發(fā)現(xiàn)，劉女士在3月2日在其百家號賬號上發(fā)布的題為《情在三月，桃花叢中》的文章中使用了他的照片。法院裁定，劉女士侵犯了李先生的署名權(quán)和該圖像的在線傳播權(quán)。裁決要求劉女士公開道歉，并向李先生支付500元的經(jīng)濟賠償金。[1]陳慧敏等,

大語言模型時代的社會機遇與挑戰(zhàn).計算機研究與發(fā)展,

2023.The

image

involved

in

this

case人工智能的進化--未來大型語言模型的未來方向人工智能中的科學(xué)建立科學(xué)理論，科學(xué)解釋涌現(xiàn)能力、人工智能原理等關(guān)鍵機制1903年，進行了第一次成功的飛機飛行1929年，馬赫數(shù)被引入空氣動力學(xué)領(lǐng)域2022年，ChatGPT引發(fā)了廣泛的興趣和討論關(guān)于智能形成的科學(xué)理論智能計算系統(tǒng)LLM與芯片、操作系統(tǒng)、數(shù)據(jù)管理、編程、網(wǎng)絡(luò)通信等深度集成制NetworkCommunicationProgrammingOperatingSystemsChipDataManagement跨各個領(lǐng)域的廣泛應(yīng)用將LLM應(yīng)用于高度專業(yè)化的領(lǐng)域，例如先進制造和科學(xué)發(fā)現(xiàn)科學(xué)發(fā)現(xiàn)先進制造人工智能中的科學(xué)追求跨架構(gòu)、算法和數(shù)據(jù)迭代的進步專注于提高LLM的“知識密度”——以更少的參數(shù)存儲更多的知識。LLM的科學(xué)問題LLM的縮放定律模型風(fēng)洞高效且有效的模型架構(gòu)Mamba,

KAN高質(zhì)量合成數(shù)據(jù)生成和多模態(tài)融合GPT4o智能計算系統(tǒng)LLM將人類知識的管理轉(zhuǎn)變?yōu)榛诮涌诘倪^程，與芯片、操作系統(tǒng)、數(shù)據(jù)庫、編程和網(wǎng)絡(luò)通信等計算組件深度集成，以增強計算系統(tǒng)的智能性DataManagement

InterfaceKnowledge

Management

InterfaceEfficientdataindexingPre-training,

fine-tuning,

inferenceDatabaseLLM

SystemResourceManagement

InterfaceResourcemanagementandschedulingOperating

SystemOperationManagement

InterfacePrograming

LanguageCodecompilationandparsingUser

logic

layerPhysicalexecutionlayerChipOperating

SystemsData

ManagementProgrammingNetwork

Communication跨各個領(lǐng)域的廣泛應(yīng)用未來應(yīng)用需求長序列處理LLM需要與人類進行長時間的互動和協(xié)作，成為現(xiàn)實世界生產(chǎn)的重要組成部分專業(yè)能力提升LLM需要提升專業(yè)能力，才能應(yīng)用于高潛在效益的場景，超出人類的能力應(yīng)對失敗的風(fēng)險LLM應(yīng)有效響應(yīng)并適應(yīng)不同的環(huán)境，確保它們在高風(fēng)險情況下的可靠使用潛在故障風(fēng)險PotentialBenefitsHighLowLowHigh初步成熟期仍有待探索MilitaryApplicationScienceDiscoveryMedicalDiagnosisDisasterPredictionEmbodiedIntelligenceAutonomousDrivingCode

GenerationVideo

GenerationArtDesignTranslationQuestionAnsweringDocWritingSummary符號智能（1950-1980）人工注釋的規(guī)則和知識狹義智能（1990-2018）監(jiān)督數(shù)據(jù)特定于任務(wù)的模型通用智能2.0(早期原型)人工智能中的科學(xué)智能計算系統(tǒng)通用智能3.0(早期原型)跨各個領(lǐng)域的廣泛應(yīng)用通用智能1.0(目前)利用廣泛的計算資源，LLM采用自我監(jiān)督訓(xùn)練從大量未標記的數(shù)據(jù)中學(xué)習(xí)知識。這種方法使“一個統(tǒng)一的模型解決多個任務(wù)”成為可能，從而引發(fā)了通用人工智能的早期跡象謝謝觀看Transformer前言語言模型這門課是生成式AI導(dǎo)……訓(xùn)練資料輸入：人工智輸出：慧輸入：不要忘了今天來開輸出：會輸入：床前明月輸出：光……Transformer找出參數(shù)論模型演進N-gramFeed-forwardNetworkRecurrentNeuralNetwork(RNN)Transformerhttps://youtu.be/dymfkWtVUdo?si=Ng29H_YxaoeiX_4yhttps://youtu.be/n9TlOhRjYoc?si=yaadpbm8w1UgbKkUhttps://youtu.be/N6aRv06iv2g?si=FuemBCZt8ChwHOvuChatGPTTransformer概述語言模型這門課是生成式AI導(dǎo)……論1.Tokenization2.InputLayer3.Attention4.FeedForward5.OutputLayerTransformerBlockXN文字轉(zhuǎn)為Token理解Token理解上下文整合、思考…得到輸出反復(fù)思考語言模型是以Token作為單位來對文字進行處理1.把文字變成TokenintroductionofgenerativeaiTokenizationintroductionofgenerativeaihttps://huggingface.co/learn/nlp-course/chapter6/5Tokenlist這頁投影片沒有需要訓(xùn)練的參數(shù)BytePairEncoding(BPE)大量文字1.把文字變成Token/tokenizer2.理解每個Token—語意introductionofgenerativeai向量(Vector)意思相近的Token會有接近的Embedding

Embeddingdogcatrabbitjumprunflowertree原本每一個Token都是獨立的符號apple,cat,dot,run,jump……2.理解每個Token—語意introductionofgenerativeai

Embedding意思相近的Token會有接近的Embeddingintai198ofbank……TokenEmbedding是在訓(xùn)練時得到的參數(shù)查表沒有考慮上下文銀行岸2.理解每個Token—語意https://youtu.be/X7PH3NuYW0Q?si=XRCVhgJAfY5a8nkfWordEmbedding2.理解每個Token—語意每一個位置有一個獨特的向量PositionalEmbedding……PositionalEmbedding可以在訓(xùn)練時得到的參數(shù)查表introductionofgenerativeai

位置1位置2位置3位置4位置5++++++3.Attention:考慮上下文Attention蘋果電腦Attention來吃蘋果3.Attention:考慮上下文這篇文章主要的貢獻是發(fā)現(xiàn)不需要RecurrentNeuralNetwork(RNN)，只需要Attention就夠了3.Attention:考慮上下文Attention3.Attention:考慮上下文找出相關(guān)的Token計算相關(guān)性0.1分數(shù)分數(shù)

有參數(shù)，透過訓(xùn)練得到計算相關(guān)性計算相關(guān)性3.Attention:考慮上下文找出相關(guān)的Token計算相關(guān)性0.1分數(shù)0.0計算相關(guān)性0.50.0計算相關(guān)性0.4AttentionWeight3.Attention:考慮上下文集合相關(guān)信息0.10.50.40.00.0(此處為簡化的講法，事實上在相加前還要過一個LinearTransformer)=0.5x0.1x0.4x++3.Attention:考慮上下文Attention3.Attention:考慮上下文計算相關(guān)性0.0計算相關(guān)性計算相關(guān)性計算相關(guān)性計算相關(guān)性0.00.40.30.3找出相關(guān)的Token3.Attention:考慮上下文0.00.00.40.30.3=0.3x0.3x0.4x++集合相關(guān)信息(此處為簡化的講法，事實上在相加前還要過一個LinearTransformer)3.Attention:考慮上下文計算所有Token兩兩間的相關(guān)性0.10.50.40.00.00.00.00.40.30.30.20.20.20.20.10.90.00.40.00.00.50.00.00.00.5AttentionMatrix3.Attention:考慮上下文0.30.70.0=0.7x0.3x+實作時只會考慮左邊的tokenCausalAttentionMulti-headAttention關(guān)聯(lián)性不只一種計算相關(guān)性0.1計算相關(guān)性計算相關(guān)性計算相關(guān)性計算相關(guān)性0.50.40.00.0計算相關(guān)性0.3計算相關(guān)性計算相關(guān)性計算相關(guān)性計算相關(guān)性0.40.00.30.0Multi-headAttention0.10.50.40.00.0=0.1x0.5x0.4x++0.30.40.00.30.0=0.3x0.4x0.3x++AttentionFeedForwardFeedForwardFeedForwardFeedForwardFeedForwardTransformerBlockTransformerBlock1TransformerBlock2Layer1Layer2…………………………TransformerBlockLastlayerOutputLayer……LinearTransformer+SoftmaxTransformer概述語言模型這門課是生成式AI導(dǎo)……論1.Tokenization2.InputLayer3.Attention4.FeedForward5.OutputLayerTransformerBlockXN文字轉(zhuǎn)為Token理解Token理解上下文整合、思考…得到輸出反復(fù)思考MixtureofExpert人類擁有復(fù)雜的認知能力2022年，YannLeCun和YoshuaBengio聯(lián)合發(fā)布了NeuroAI白皮書，指出神經(jīng)科學(xué)將為下一代人工智能提供指導(dǎo)人腦表現(xiàn)出稀疏的模塊化結(jié)構(gòu)人腦可以用更低的能耗和更靈活的機制實現(xiàn)復(fù)雜的認知Thesparsemodularstructureofthehumanbrain:efficiency,reusability,interpretability人腦表現(xiàn)出稀疏的模塊化結(jié)構(gòu)Sparseneuralconnectionsmake

brains

form

different

functionalmodulesLanguagecognitionrequiresthecooperation

ofdifferentmodules

in

our

brainHuman

vs.

GPT-3

Parameter:175billionNeurons:200billionActivationrate<5%Energyconsumptionperinference<15wattsEnergyconsumptionperinference:400

wattsReactionTime<100millisecondsInferenceTime:About

2000milliseconds[1]

Bartheld

et

al.,Thesearchfortruenumbersofneuronsandglialcellsinthehumanbrain:Areviewof150yearsofcellcounting.JournalofComparativeNeurology,2016.[2]

Bullmoreet

al.,

Theeconomyofbrainnetworkorganization.Naturereviewsneuroscience,2012.[3]

PeterLennie.

Thecostofcorticalcomputation.Currentbiology,2003.LargeLanguageModelsThedevelopmentofLLMs

followsahierarchicalstructureLLaMA

2LLaMA

2-chatCode

LLaMAAlignmentEnhancing

the

Coding

AbilityfLLaMA

2Enhancing

the

Ability

of

API

CallLLaMA

2-TruthXReducing

HallucinationPhind

Code

LLaMAEnhancing

the

Coding

AbilityAn

example

of

GitHubGitsupportsbranchingandmergingGeneralLLMSFTleadstoredundantmodeltraining.Incontrast,softwaredevelopmentbenefitsfromthereuseandtransfercapabilitiesMixtureofExpertsPre-trainedLanguageModelsconductsparseactivationcharacteristicsMotivation:Leveragingsparseactivationcharacteristicstoconstructpartitionsandperformbrain-likesparsecomputationWhencomputingspecificinputs,mostneuronsinFFNs

arenotactivatedMixtureofExpertsTheMixture-of-Experts(MoE)networkisarepresentativearchitectureforsparsecomputationIllustration

of

Mixture-of-Expert

(MoE)Dynamicsparseactivationrepresentasignificantdirectionforfuturemodelarchitectures.Conditionalcomputation

iskeytoscalingup.[1]

YoshuaBengio,

DeepLearningofRepresentations:LookingForward.StatisticalLanguageandSpeechProcessing,2013.[2]

Feduset

al.,

SwitchTransformers:ScalingtoTrillionParameterModelswithSimpleandEfficientSparsity.JMLR,2022.MixtureofExpertsEachFFNmoduleisregardedasan“Expert”EachexpertisdesignedtohandlespecifictypesoftasksordatasubsetsSparseMoElayers

are

used

to

replace

the

Feed-Forward

Network

(FFN)

Layer[1]

Feduset

al.,

SwitchTransformers:ScalingtoTrillionParameterModelswithSimpleandEfficientSparsity.JMLR,2022.MixtureofExpertsTherouterdetermineswhichtokensaresenttowhichexpert[1]

Feduset

al.,

SwitchTransformers:ScalingtoTrillionParameterModelswithSimpleandEfficientSparsity.JMLR,2022.Thetoken“More”issenttothesecondexpertThe

token"Parameters”issenttothefirstexportMixtureofExpertsLLMsleadtoextensivecomputations,makinginferencechallengingMoEdynamicallycombinestheoutputsofmultipleexpertmodelsEffectivetrainingwithfewercomputationalresourcesFasterinferencebyusingonlyasubsetofparameters[1]

Feduset

al.,

SwitchTransformers:ScalingtoTrillionParameterModelswithSimpleandEfficientSparsity.JMLR,2022.[2]

Jiang

et

al.,

MixtralofExperts.

arXiv,

2024.謝謝觀看大語言模型大語言模型之大LLM涌現(xiàn)出驚人的上下文學(xué)習(xí)、指令遵循、逐步推理的類人能力“大模型”是指龐大的參數(shù)規(guī)模，“大數(shù)據(jù)”是指海量的訓(xùn)練數(shù)據(jù)，“大算力”是指巨大的設(shè)備支持。大語言模型之大ChatGPT-閉源典型代表1.LimitationsofLLMsLLaMA（LargeLanguageModelforAI）是由Meta（原Facebook）推出的開源大語言模型，旨在提供高效、可擴展的基礎(chǔ)模型，以支持各種自然語言處理（NLP）任務(wù)。LLaMA模型的設(shè)計目標是優(yōu)化模型的性能與計算效率，使其在計算資源有限的情況下也能產(chǎn)生高效的推理和生成效果。多模態(tài)大語言模型多模態(tài)簡介Amodalityistheclassificationofasingleindependentchannelofsensoryinput/outputbetweenacomputerandahuman跨模態(tài)學(xué)習(xí)是指涉及從多種模態(tài)獲得信息的任何類型的學(xué)習(xí)Examples

of

different

types

of

modalitiesTextImagesVideosAudioTactile...WhyCross-modalLearning