tempo ai - designing ai interfaces

20
Designing AI Interfaces Raj Singh CEO Tempo AI, Inc. 510-282-4229 (M) [email protected] @mobileraj

Upload: raj-singh

Post on 23-Jan-2018

430 views

Category:

Technology


8 download

TRANSCRIPT

Page 1: Tempo AI - Designing AI Interfaces

DesigningAIInterfaces

RajSinghCEOTempoAI,Inc.

510-282-4229(M)[email protected]@mobileraj

Page 2: Tempo AI - Designing AI Interfaces

“Itisthescienceandengineeringofmakingintelligentmachines,especiallyintelligentcomputerprograms. Itisrelatedtothesimilartaskofusingcomputerstounderstandhumanintelligence,butAIdoesnothavetoconfineitselftomethodsthatarebiologicallyobservable.”

“Itpasses theTuringtest.”

“Thestudyanddevelopmentofintelligentagents.”

“Softwarethatlearnsandcompletestasksforyou.”

WhatIsAI?

Page 3: Tempo AI - Designing AI Interfaces

3

Raj:“WhatisAI?”Nephew:“It’srobotsandshit…”

AndthenthereisHollywood’sdefinition:

Page 4: Tempo AI - Designing AI Interfaces

4RecommendationEngines- ImplicitandExplicitLearning

AIHasBeenOmnipresent

Page 5: Tempo AI - Designing AI Interfaces

5ButthepuristssayWatsonisdumb!

ManyhavesaidWatsonwasthefirstmainstreamdemonstrationofAI

Page 6: Tempo AI - Designing AI Interfaces

Anticipatory:“Predictyournextwantoraction”

Smart:“GivemeonlytheinformationIneed”

Assistant:“Completetasksforme”

Ihaveasmartphonebutit’snotsmart!

ButWhatIsMobileAI

Page 7: Tempo AI - Designing AI Interfaces

7

DesigninganAI-AssistedUXIsNotEasy

Designing95%UIYoudon’tknowifyourrecommendations arerightorwrong

Whatifthesuggestionwaswrong?Certainappsarealotmoreforgiving

Howdoestheusertrainthesystem?Mostsurveythattheywilltrainbutfewactuallydo.

Howdoyougetthedata?AIsuffersfromfalsestarts.

Allthataside,anticipatoryUIdesignisthenextfrontier!

Page 8: Tempo AI - Designing AI Interfaces

SomeExamples(RecommendationEngines)

Pandora FourSquare LinkedIn

Page 9: Tempo AI - Designing AI Interfaces

Doesthisbotheryou?

AnotherExample(RecommendationEngine)

Itdoesbecausetheserecommendationshaverightandwronganswers.

Page 10: Tempo AI - Designing AI Interfaces

Flipboard vs Zite

ExtensiveML

LittletoNoML

Page 11: Tempo AI - Designing AI Interfaces

• “Suggestions”– indicatesintelligence• “Recommendations”– slightlylowerbarthan“suggestions”• “Searched/Results”– lessintelligent

RecommendationLanguageMatters

TempoSearchedMeetingLocations

Page 12: Tempo AI - Designing AI Interfaces

• Howdoyoudeterminethebalance?‒ Usertestingdoesn’talwayswork

• Constrainthedomain‒ Segment theusersviacohortanalysis‒ Whatistherightnumber ofsegments?

• The“More”buttoncanbeyourbestfriend‒ Infinitescroll

• Trainthesystem‒ Tempousertestingindicatedlessthan3%wouldtrain

‒ Thumbsup/down, ratingsvs implicitlearning?

BalanceofNoisevs Precision

PrismaticNewsTraining

Page 13: Tempo AI - Designing AI Interfaces

13

Siriisentertainingbutbeingunconstrainedkilledtheirengagement

Page 14: Tempo AI - Designing AI Interfaces

• Itwouldannoyyouless• Youknowwhatyoucanaskanddo• You“settheexpectations”

ExamplesAutomatedsupportsystemsTellMe /Free411SalesforceVoiceAccess

ConstrainingCanSetExpectations

Lexee AppVoiceCommandsforSalesforce

Page 15: Tempo AI - Designing AI Interfaces

• Animationsworkmosteffectively‒ Butifittakestoolong, ithinders theUX

• Speedofapplicationdirectlycorrelatedtoretentionrate‒ 15searchresultsvs 10searchresults(Google SearchResults)

• InTempo,wenumberedtheresults• Searchenginesusedtonumbertheirresultsaswell

IndicatingIntelligence

NumberedResultstoIndicateSmarts

Page 16: Tempo AI - Designing AI Interfaces

• Falsestartsareverycommon‒ Introducerecommendations andanticipatoryactionsthrough use

• Keepon-boardingaslight-weightaspossible‒ Toomuchtimebetweenon-boardingandfirst-usewillcauseproblems

‒ Canyoucollectdataasyougoalong

• First3-Dayusagewillbeheavyexperimentationtoseewhatthesystemdoes‒ SiriusersexperimentbyaskingalotofQs

‒ Tempouserscreate10sofmtgs inthefirstfewdays

AINeedsData

Sosh Setup

Page 17: Tempo AI - Designing AI Interfaces

BeingAnticipatory

• Anticipatewithoutthenoise‒ Pushnotificationsdriverepeatusagebutifnoisyresultinbounced users

• NotifyingyouwhentoleaveinTempo‒ Wewantedtobeveryanticipatorybutwe’renot95%yet

‒ Falsenotifications resultinangstandalostuser

Page 18: Tempo AI - Designing AI Interfaces

18

It’sJustBeginning!

FourSquare GoogleNow Prisimatic

Bigdatatosuggestplacestoexplore

Searchhistorytobecomemoreanticipatory

Machinelearningtosuggestrelevantnews

Page 19: Tempo AI - Designing AI Interfaces

• Recommendationsworkbestwhentheusercan’ttellwhat’srightorwrong.

• Bespecificwithyourlanguagebecauseithelpssetexpectations• Bettertoundersellandover-deliver• Userswillwanttotrainthesystembutfewwilldoit.Beconsciousthattrainingmaycreateanaversereaction

• Cold-startsarecommon;needtohaveacompellingcaseandintegratetheuserdataover-time

• Useclustering/segmentationtoimprovethecold-start(egchooseyourinterests)

• Besensitiveaboutnotificationsandtrackengagementtomachinelearnonyournotifications

• UnderstandthatwhatyoumaythinkofasAI,theuserthinksisdumb(andvice-versa);incorporateanimationsorothertoindicateAI

Summarizing

Page 20: Tempo AI - Designing AI Interfaces

Whatdoessmartmeanineverycoreapp?

RajSinghCEOTempoAI,Inc.

510-282-4229(M)[email protected]@mobileraj