Lytle讲座：大语言模型 2023-2024.pdfVIP

LytleLecture2023-2024

田NEWYORKUNIVERSITYCX)MetaAl

Objective-DrivenAl

TowardsAlsystemsthatcanlearn,

remember,reason,plan,

havecommonsense,

yetaresteerableandsafe

YannLeCun

NewYorkUniversity

Meta-FundamentalAlResearch

UniversityofWashington

LytleLecture

2024-01-24

．

冒，，

MachineLearningsucks!(comparedtohumansand1:1..,_,

►Supervisedlearning(SL)requireslargenumbersoflabeledsamples.

►Reinforcementlearning(RL)requiresinsaneamountsoftrials.

►Self-SupervisedLearning(SSL)worksgreatbut…

►Generativepredictiononlyworksfortextandotherdiscretemodalities

►Animalsandhumans:

►Canlearnnewtasksveryquickly.

►Understandhowtheworldworks

►Canreasonanplan

►Humansandanimalshavecommonsense

►Therebehaviorisdrivenbyobjectives(drives)

WeNeedHuman-LevelAlforIntelligentAssistant

►Smartglasses

►Communicatesthroughvoice,vision,display,

electro-myograminterfaces(EMG)

►IntelligentAsistant

►Cananswerallofourquestions

►Helpsusinourdailylives

Her

►Knowsourpreferencesandinterests(2013)

►Forthis,weneedmachineswithcommonsense

►Machinesthatunderstandhowtheworldworks

►Machinesthatcanremember,reason,plan.

FutureAlAssistantsneedHuman-LevelAl

►Alassistantswillrequire(super-)human-levelintelligence

►Likehavingastaffofsma