R語言的協(xié)作伙伴

Copilot

for

R/revodavid/copilot-for-r@revodavid

at@NYHACKRwhat

it

doeshow

to

use

it,

andhow

it

worksDavid

Smith

(@revodavid)Principal

Cloud

AdvocateMicrosoftTalk

Notes:/revodavid/copilot-for-rUses

the

context

you’ve

providedand

synthesizes

code

to

matchConvert

comments

to

codeAutofill

for

repetitive

codeAutosuggest

testsShow

alternativesGitHubCopilotYour

AI

pair

programmer/revodavid/copilot-for-r@revodavid

at@NYHACKRaka.ms/get-copilot/revodavid/copilot-for-r@revodavid

at@NYHACKR/revodavid/copilot-for-r@revodavid

at@NYHACKRR

Pumpkins

Demo/revodavid/copilot-for-r@revodavid

at@NYHACKR/revodavid/copilot-for-r@revodavid

at@NYHACKRGenerative

AI

Models/revodavid/copilot-for-r@revodavid

at@NYHACKRGenerative

AIPrompt:Write

a

tagline

for

an

ice

creamshop.Response:We

serve

up

smiles

with

everyscoop!Prompt:Table

customers,

columns

=[CustomerId,

FirstName,LastName,

Company,

Address,City,

State,

Country,PostalCode]Create

a

SQL

query

for

allcustomers

in

Texas

named

Janequery=Response:SELECT

*FROM

customersWHERE

State

=

'TX'

AND

FirstName=

'Jane'Response:GPT-3CodexDALL·EPrompt:

A

white

Siamese

catGenerative

AIcan:Generate

text,

images

and

codeDifferent

models

aretrained

on

different

corpuses,depending

on

the

application.Generate

“humanlike”

outputWhat

is

a

likelycontinuation

of

the

prompt,

giventhe

training

data?/revodavid/copilot-for-r@revodavid

at@NYHACKRExtract

informationThe

continuation

is

likely

to

be

similar

to

textfrequently

represented

in

thetraining

data.Createnovel

contentText,

images

and

code

not

contained

in

its

trainingset.

Translations.“Creative”

works.IntelligentIt’sjust

a

predictivesystem,

designed

to

give

a

likelycontinuation

of

the

prompt

given

the

training

data.DeterministicRun

the

same

prompt.

Get

back

a

differentresponse

(probably)./revodavid/copilot-for-r@revodavid

at@NYHACKRTrustworthyIt

can

“hallucinate”facts

andconfidently

assertthem

to

be

true.Generative

AIis

not:TODOLearnThe

model

is

fixed

at

the

time

of

its

training./revodavid/copilot-for-r@revodavid

at@NYHACKRContain

all

of

the

information

of

itstraining

setThink:a

blurry

jpeg

of

its

training

data.Include

verbatim

copies

of

its

trainingdataBut

it

can

generate

stuff

that

looks

like

it.Generative

AIdoes

not:Generative

AIdoesn’t:Understand

languageIt’s

just

a

predictive

engine.

Itdoesn’t

understandmath,either.Understand

factsMany

predictions

echo

information

inthe

trainingset,but

this

is

not

guaranteed./revodavid/copilot-for-r@revodavid

at@NYHACKRUnderstand

manners,

emotion

orethicsAlso:

avoid

anthropomorphizingit.Understand

anythingIt’s

just

a

prediction

engine!Prompt

–

Text

input

thatprovides

some

context

to

theengine

on

what

isexpecting.Completion

–

Output

thatGPT-3

generatesbasedonthe

prompt

and

the

trainedmodel.Prompt

Engineering

(very

briefly)Ensure

that

artificialgeneral

intelligence

(AGI)benefits

humanity.Empowerevery

person

andorganization

on

the

planetto

achieve

moreGPT-3Generate

and

UnderstandTextCodexGenerate

and

Understand

CodeDALL·EGenerate

images

from

textprompts/revodavid/copilot-for-r@revodavid

at@NYHACKRDemo:

Azure

OpenAI

Service/revodavid/copilot-for-r@revodavid

at@NYHACKRof

new

code

written

with

CopilotGitHubCopilotOnce

enabled…40%/revodavid/copilot-for-r@revodavid

at@NYHACKRof

devsfelt

morefulfilled

with

their

jobs87%of

devs

said

it

helpedpreserve

mental

effort75%Azure

OpenAI

ServiceGPT-3CodexDALL·E

(preview)/revodavid/copilot-for-r@revodavid

at@NYHACKRThank

you!/revodavid/copilot-for-r@revodavid

at@NYHACKRaka.ms/get-copilot/revodavid/copilot-for-rDavid

Smith

(@revodavid)Principal

Cloud

Advocate,

MicrosoftBonus

slides/revodavid/copilot-for-r@revodavid

at@NYHACKRFor

Q&A1956Artificial

Intelligence1997Machine

Learning2017Deep

Learning2021Generative

AIReliability

&

SafetyPrivacy

&

SecurityFairnessOurPrinciplesInclusivenessTransparencyAccountability@revodavid

at@NYHACKR/revodavid/copilot-for-rInferencing

timeCapabilityCurieAnswering

questionsComplex,

nuanced

classificationDavinciSummarizing

forspecific

audienceGenerating

creative

contentBabbageSemantic

searchrankingModeratelycomplexclassificationAdaSimple

classificationParsing

and

formatting

textAzure

OpenAI

Service

modelsCushman-codexDavinci-codexCapabilityCodexGPT-3ModelRequestDescription,

performance,

costUse

casesDavinci4,000

tokensMost

capable

GPT-3model.

Can

doComplex

intent,

cause

andany

task

the

other

models

can

do,effect,

summarization

foroften

with

higher

quality,

longer

outputaudienceand

better

instruction-following.Curie2048

tokensVery

capable,

but

faster

and

lower

costLanguage

translation,than

Davinci.complex

classification,textsentiment,

summarizationBabbage2048

tokensCapable

of

straightforward

tasks,

veryModerate

classification,fast,

and

lower

cost.semantic

search

classificationAda2048

tokensCapable

of

very

simple

tasks,

usuallyParsing

text,

simplethe

fastest

model

in

the

GPT-3

series,classification,

addressand

lowest

cost.correction,

keywordsAzure

OpenAI

|

GPT-3

Family

of

ModelsAccelerates

software

developmentMakes

APIs

moreaccessibleWidens

who

can

codeOpenAI

CodexOpenAI

Codex

ModelsDerived

from

base

models

and

trained

on

bothNLand

code

(billions

ofLines

ofCode)Supp