AI in Marketing Research: Expert Panel Discussion
Résumé
TLDRIl webinar ha esplorato l'uso dell'intelligenza artificiale nella ricerca di marketing, con esperti che hanno condiviso le loro esperienze e applicazioni pratiche. Sono stati discussi temi come la qualità dei dati, l'analisi qualitativa e l'uso di modelli di linguaggio per migliorare le indagini. I relatori hanno anche parlato di un prossimo summit sull'analisi e le intuizioni, dove verranno presentati ulteriori approfondimenti sull'argomento. Le preoccupazioni etiche e le sfide legate all'uso dell'AI sono state evidenziate, insieme all'importanza della progettazione dei prompt per ottenere risultati significativi.
A retenir
- 🤖 L'AI sta trasformando la ricerca di marketing.
- 📊 L'analisi qualitativa è più efficiente con l'AI.
- 🛠️ Strumenti come Voxpopme sono utilizzati per l'analisi dei dati.
- 📅 Il summit sull'analisi e le intuizioni si svolgerà dal 29 aprile al 3 maggio.
- ⚖️ Le questioni etiche sono fondamentali nell'uso dell'AI.
- 📝 La progettazione dei prompt è cruciale per risultati accurati.
- 📈 L'AI può migliorare la qualità dei dati nei sondaggi.
- 💡 L'AI può aiutare a identificare temi nei dati qualitativi.
- 🔍 L'AI può potenzialmente sostituire i sondaggi tradizionali.
- 💬 Le allucinazioni dell'AI possono essere mitigate con prompt ben progettati.
Chronologie
- 00:00:00 - 00:05:00
Benvenuti al webinar di Satu Software su AI e ricerca di marketing, con quattro ospiti che condivideranno le loro esperienze sull'uso dell'intelligenza artificiale nella ricerca. Il webinar ha registrato un alto numero di iscrizioni, evidenziando l'interesse per il tema. Ci sarà una sessione di domande e risposte alla fine, e una registrazione sarà inviata a tutti gli iscritti.
- 00:05:00 - 00:10:00
Brian Orm, presidente di Satu Software, introduce i relatori e anticipa che ci saranno 11 presentazioni all'imminente Analytics and Insights Summit. I relatori condivideranno le loro scoperte sull'AI e ci sarà un'opzione per accedere ai contenuti in differita per chi non può partecipare di persona.
- 00:10:00 - 00:15:00
Kevin Cardi, esperto di analytics, discute le sfide della qualità dei dati nelle indagini di mercato, evidenziando problemi come il calo dei tassi di partecipazione e l'aumento delle risposte non valide. Propone di rendere i sondaggi più coinvolgenti e umani attraverso l'uso dell'AI, come l'implementazione di domande dinamiche e l'analisi automatizzata.
- 00:15:00 - 00:20:00
Kevin condivide la sua esperienza con l'uso di AI per umanizzare i sondaggi, menzionando l'uso di ChatGPT e i modelli di linguaggio di grandi dimensioni. Sottolinea l'importanza di migliorare l'esperienza degli utenti nei sondaggi per ottenere dati di qualità superiore.
- 00:20:00 - 00:25:00
Mangela Budia, direttrice associata di Ipsos, presenta uno studio sull'uso dei modelli di linguaggio di grandi dimensioni per replicare i risultati di studi di congiunzione. Discute le sfide e le opportunità nell'uso di AI per analizzare le preferenze dei consumatori e l'importanza di testare diversi parametri nei modelli AI.
- 00:25:00 - 00:30:00
Mangela esplora l'impatto delle impostazioni di temperatura nei modelli di linguaggio e come queste influenzano le risposte. Sottolinea la necessità di comprendere come i modelli AI possano gestire design complessi e la questione del bias posizionale nelle risposte.
- 00:30:00 - 00:35:00
Dan Penny di Microsoft discute l'adozione dell'AI in vari casi d'uso di ricerca, inclusa l'analisi dei dati aperti e l'uso di dati sintetici. Sottolinea l'importanza di testare diversi scenari per determinare l'efficacia dell'AI nella ricerca di mercato.
- 00:35:00 - 00:40:00
Dan condivide che Microsoft sta sperimentando con dati sintetici e l'uso di AI per migliorare l'efficienza nella ricerca. Sottolinea che ci sono ancora molte opportunità da esplorare e che l'AI è presente in molte aree della ricerca.
- 00:40:00 - 00:45:00
Jeffrey Doson, professore di marketing, presenta un progetto sull'uso dell'AI generativa per creare stimoli sperimentali. Discute le sfide etiche legate all'uso di AI e alla proprietà intellettuale, evidenziando l'importanza di considerare il valore dello stile artistico.
- 00:45:00 - 00:50:00
Il panel discute l'importanza di testare l'AI in vari contesti e la necessità di un approccio etico nell'uso dell'AI nella ricerca. Viene sottolineato che l'AI può migliorare l'efficienza, ma ci sono anche rischi e limitazioni da considerare.
- 00:50:00 - 00:59:44
Il webinar si conclude con una sessione di domande e risposte, dove i relatori rispondono a domande su come l'AI sta influenzando la ricerca di mercato e quali strumenti stanno utilizzando per analizzare i dati qualitativi.
Carte mentale
Vidéo Q&R
Qual è l'argomento principale del webinar?
Il webinar si concentra sull'uso dell'intelligenza artificiale nella ricerca di marketing.
Chi sono i relatori del webinar?
I relatori sono esperti del settore che condividono le loro esperienze con l'AI nella ricerca.
Quando si svolgerà il summit sull'analisi e le intuizioni?
Il summit si svolgerà dal 29 aprile al 3 maggio.
Quali sono alcune delle applicazioni dell'AI nella ricerca di mercato?
L'AI viene utilizzata per migliorare la qualità dei dati, analizzare risposte aperte e progettare sondaggi.
Ci sono preoccupazioni etiche riguardo all'uso dell'AI?
Sì, ci sono molte discussioni sulle questioni etiche legate all'uso dell'AI nella ricerca.
Come viene utilizzata l'AI per analizzare i dati qualitativi?
L'AI viene utilizzata per riassumere e identificare temi nei dati qualitativi.
Quali strumenti vengono utilizzati per l'analisi qualitativa?
Strumenti come Voxpopme vengono utilizzati per analizzare dati qualitativi.
L'AI può sostituire i sondaggi tradizionali?
Potenzialmente, ma ci sono domande su come e quando dovrebbero essere utilizzati.
Qual è l'importanza della progettazione dei prompt nell'AI?
La progettazione dei prompt è cruciale per ottenere risultati accurati e significativi dall'AI.
Come si affrontano le allucinazioni dell'AI?
Si possono mitigare attraverso una progettazione attenta dei prompt.
Voir plus de résumés vidéo
INTRODUZIONE AL CEMENTO ARMATO MEDIANTE MODELLINI DI TRAVI INFLESSE - (770)
Ha ancora senso mettere le persone in carcere?
Per cosa si vota ai referendum su cittadinanza e lavoro?
La peste del 600 ne I promessi sposi di Manzoni | La peste manzoniana del 1630 a Milano
Fatti smart! Tutela la tua privacy su smartphone e tablet
App-rova di privacy. I suggerimenti del Garante privacy per un uso consapevole delle app
- 00:00:04okay let's get started we want to
- 00:00:05welcome everyone to this satu software
- 00:00:08webinar entitled Ai and marketing
- 00:00:10research uh we have four amazing guests
- 00:00:13on our webinar today and we'll they'll
- 00:00:15share each of them will share with us
- 00:00:17their experience with artificial
- 00:00:19intelligence and how they are applying
- 00:00:21it to research um each of them will also
- 00:00:24be speaking a little later this month uh
- 00:00:27at the analytics and in it Summit uh
- 00:00:31that we're having so we're excited to
- 00:00:32hear them now and uh in the conference a
- 00:00:35little bit
- 00:00:36later uh AI is a very interesting topic
- 00:00:39uh actually this webinar has had the
- 00:00:41second uh second highest number of
- 00:00:44signups that we've ever had in a webinar
- 00:00:46so lots of interest um we're excited
- 00:00:49that uh you'll be able to uh hear what
- 00:00:52they have to say and interact with our
- 00:00:54guests I want to mention a few things
- 00:00:56before we get
- 00:00:58started
- 00:01:00uh just a few details about the webinar
- 00:01:02um we have a Q&A section here at the
- 00:01:05bottom of Zoom that you can use to ask
- 00:01:08questions throughout uh and then at the
- 00:01:10end of the the end of the webinar we'll
- 00:01:14have a Q&A uh session uh where we'll
- 00:01:17we'll try to answer as many questions as
- 00:01:19we can in the time allotted um also a
- 00:01:21recording of This webinar will be sent
- 00:01:23out to everybody who
- 00:01:26registered uh little plug here for saw
- 00:01:28to software uh we have an amazing survey
- 00:01:31platform that we're building we're
- 00:01:33really good at conjurant analysis and
- 00:01:35Max sff easy to use we have incredible
- 00:01:39support we're really friendly and you
- 00:01:41can try Us for free at
- 00:01:43discover.
- 00:01:46software.com also as I mentioned this
- 00:01:48analytics and insights Summit uh will be
- 00:01:50having April 29th through May 3rd so
- 00:01:53coming right up uh down in San Antonio
- 00:01:56Texas it's a great venue uh we have
- 00:01:59amazing speed ERS amazing
- 00:02:02um uh schedule and everything planned
- 00:02:05there come down and interact with uh
- 00:02:07some of the brightest researchers in the
- 00:02:09industry uh if you can't make it down
- 00:02:11there uh you also can get this delayed
- 00:02:14virtual access uh that that we'll we'll
- 00:02:18provide for you you can see more
- 00:02:19information at
- 00:02:22software.com
- 00:02:24conference with that I'm going to turn
- 00:02:27uh the time over to Brian orm president
- 00:02:30of satu
- 00:02:32software super thank you Justin so we
- 00:02:36put this together uh we have 11
- 00:02:38presentations coming up at the analytics
- 00:02:40and insight Summit on AI and our four
- 00:02:44guests are four of those 11 speakers who
- 00:02:46are going to give us a taste of some of
- 00:02:49their findings for the AI Summit if you
- 00:02:52want to see the rest of the story please
- 00:02:55sign up it's not too late to sign up and
- 00:02:57come and also if you can't can't come in
- 00:03:00person you can sign up for the delayed
- 00:03:02option which gives you the video in the
- 00:03:04slides about one week after the
- 00:03:07conference is over for you to be able to
- 00:03:09enjoy at your own schedule and your own
- 00:03:11Leisure so we've lined up these four
- 00:03:13speakers really happy that they've
- 00:03:15prepared I'm GNA introduce each of them
- 00:03:18one at a time I'm going to give them
- 00:03:20just one follow-up question uh after
- 00:03:23their five minute presentations after
- 00:03:25that we're going to allow the panelists
- 00:03:28to kind of talk to one another ask
- 00:03:30questions of one another and have about
- 00:03:32a 10 or 15 minute panel discussion
- 00:03:34amongst themselves and then the last 15
- 00:03:37minutes or so of this webinar we're
- 00:03:39going to open it up to the questions
- 00:03:41that you've put in the Q&A I don't think
- 00:03:43we'll be able to get to all of them but
- 00:03:45we'll pick some of those and ask them of
- 00:03:47the panelists and see where we get from
- 00:03:49there so I'm really looking forward to
- 00:03:51this and we're going to start out with
- 00:03:53Kevin cardi Kevin has been in analytics
- 00:03:57for over 20 years with 15 years in
- 00:03:59marketing research and eight years in
- 00:04:01new product development he co-founded in
- 00:04:03tufi to humanize surveys using
- 00:04:06technology and AI while Sim
- 00:04:08simultaneously enhancing data quality
- 00:04:10and depth of insights prior to launching
- 00:04:12in tufi he led a technology incubator in
- 00:04:16a major real estate finance company and
- 00:04:18ran analytics and Innovation for aenova
- 00:04:20before its acquisition by neelen he
- 00:04:23earned his PhD in quantitative methods
- 00:04:25and political science from MIT and has
- 00:04:28published papers and P patents on
- 00:04:30various research methods with that go
- 00:04:33ahead and take the
- 00:04:36floor hi Brian thank you it's a pleasure
- 00:04:39to be here um good to meet know you guys
- 00:04:41um we're covering as everyone here is a
- 00:04:43lot of stuff um at the upcoming confence
- 00:04:45so we're just give you a quick taste of
- 00:04:47some of the the things that we'll be
- 00:04:49talking about and I'm going to share
- 00:04:50screens real quick um hopefully you guys
- 00:04:55can all see my screen here if not let me
- 00:04:57know real quick before I uh before I get
- 00:04:59into it so these are some of the things
- 00:05:01that we're going to be discussing things
- 00:05:02like impact and Aon survey cheating
- 00:05:04Dynamic questions coding open end some
- 00:05:06automated insights some other things we
- 00:05:08won't be talking about at least we won't
- 00:05:10because other folks here on this call
- 00:05:12the conference can be covering them in
- 00:05:13great detail with some fantastic papers
- 00:05:16and I'm certainly looking forward to it
- 00:05:18to give you a little taste of the types
- 00:05:19of things that we're going to be
- 00:05:20covering just take a quick mental
- 00:05:22Journey Back to the year 20122 feels
- 00:05:24like an eternity ago remember the
- 00:05:26Ukraine war just started Queen Elizabeth
- 00:05:28had died inflation was running a mock in
- 00:05:31the market research industry um the big
- 00:05:34issue was Data quality back then and it
- 00:05:35still is today um in fact it's probably
- 00:05:37gotten worse things like dropping
- 00:05:39participation rates Rising cheater rates
- 00:05:41um professional survey takers with
- 00:05:43individuals taking 50 plus surveys a day
- 00:05:46um challenges with empty answers and
- 00:05:48quality of content and one of the major
- 00:05:51reasons that we were facing these kinds
- 00:05:53of problems was survey design um if
- 00:05:55you've used surveys you kind of are very
- 00:05:58familiar with those types of issues
- 00:05:59people people have had tremendous
- 00:06:01progress in the experience with
- 00:06:02computers in general um and what they
- 00:06:05expect from the experiences they engage
- 00:06:07in but the experiences on surveys have
- 00:06:08remained kind of a little bit behind the
- 00:06:10times in fact this is an actual survey
- 00:06:12from 1997 if anyone know is around back
- 00:06:15in the Greenfield online days and then
- 00:06:1620 thou you know 2008 and 2022 over here
- 00:06:19you know the progress is pretty limited
- 00:06:21compared to other areas of technology
- 00:06:23and so we had set off on this journey
- 00:06:26asking the question how do we make
- 00:06:27surveys you know dramatically more
- 00:06:29engaging much better so that we get much
- 00:06:32better data and we can kind of address
- 00:06:33some of these issues and part of that
- 00:06:36was really making surveys more human um
- 00:06:39in a lot of different ways Graphics but
- 00:06:41you know looking at AI we asked the
- 00:06:42question can AI make surveys and other
- 00:06:44types of research more human not less
- 00:06:46human but more human kind of a little
- 00:06:48different from some of the simulated
- 00:06:50response um approaches that are out
- 00:06:51there that will be discussed as well and
- 00:06:54so when we first heard about chat GPT
- 00:06:57back in um you know in 200 this is my
- 00:07:00initial response right when I first
- 00:07:01heard about chat GPT this is me after
- 00:07:03using chat GPT for 30 minutes um and
- 00:07:06immediately started seeing a lot of the
- 00:07:08applications that you know we could
- 00:07:10start experimenting with um and you know
- 00:07:15and my skepticism was kind of born out
- 00:07:17of this um this issue if you've been in
- 00:07:19modeling you know for a period of time
- 00:07:21you kind of noticed that you know we've
- 00:07:23had models get bigger and bigger and
- 00:07:25lots of heavy duty promises um for a
- 00:07:27long period of time by the way this is
- 00:07:29you know here on the bottom and this is
- 00:07:30the size of the model parameter in in um
- 00:07:33in the literature the largest models
- 00:07:35that were being published in the
- 00:07:36literature this is a log scale so you
- 00:07:38know every every Point increase is like
- 00:07:4010x and you see the steady increase in
- 00:07:42the size of the models um over time and
- 00:07:45then this you know inflection in around
- 00:07:482019 with generative Ai and then llms
- 00:07:50large language models come in and you
- 00:07:51see this massive jump in the size of the
- 00:07:54models and I was skeptical because you
- 00:07:56know big models does not necessarily
- 00:07:57translate into better outcomes um and
- 00:07:59gp4 by the way is up here that's how big
- 00:08:02it is um literally trillions of
- 00:08:04parameters and this is gpt3 compared to
- 00:08:06gp4 and so you know there these huge
- 00:08:09differences in models but did they make
- 00:08:11an impact um and so you know one of the
- 00:08:14you know the applications we looked at
- 00:08:16in terms of humanizing surveys was
- 00:08:17things like conversational voice um sort
- 00:08:20of using AIS to actually almost converse
- 00:08:22with people um as part of a survey and
- 00:08:25survey questions and you know I won't
- 00:08:27kind of go into details now we certainly
- 00:08:28don't have time much I hope you guys get
- 00:08:30to join the the final survey but I will
- 00:08:32give you a problem that emerged out of
- 00:08:34you know these Solutions so Solutions
- 00:08:36create their own problems and it was
- 00:08:37this we did a presentation um last year
- 00:08:40um at at quirks event with Pepsi and you
- 00:08:42know we had a a voice survey an
- 00:08:44interactive voice survey and we had
- 00:08:4620,000 voice responses with 240 hours of
- 00:08:50content and so the next question that AI
- 00:08:53creates is okay we can we can humanize
- 00:08:56surveys what do we do with all of that
- 00:08:57and that's another challenge that
- 00:08:59convert that that llms and other
- 00:09:01existing AI models um are helping us to
- 00:09:05um to deal with and we'll talk about
- 00:09:06different solutions that that we and
- 00:09:08other companies are using to approach um
- 00:09:11those problems so I certainly hope um
- 00:09:12everyone gets a chance to join us
- 00:09:14there's a lot to talk about um and
- 00:09:16there's just so many excellent
- 00:09:17presentations and the folks you know on
- 00:09:19this call in the presentation are
- 00:09:20definitely folks um that you know I'm
- 00:09:22I'm privileged to be with and and you'll
- 00:09:24definitely want to to connect
- 00:09:26with thanks so much for that intro
- 00:09:29and I got a question for you we're we're
- 00:09:32hearing a lot of interest in AI for the
- 00:09:34marketing research sector perhaps to the
- 00:09:36level of hype what tasks is AI most
- 00:09:40delivering on and where do you think
- 00:09:42it's falling
- 00:09:44flat yeah so that I mean there's a lot
- 00:09:46of I just got back from Insight
- 00:09:48association annual and it's been a topic
- 00:09:50of conversation like every conference
- 00:09:52this year and you know last year as well
- 00:09:54um so there's there's a lot of Focus
- 00:09:56right now on coding data which I think
- 00:09:59is certainly one of the you know the
- 00:10:01biggest opportunities so you know how do
- 00:10:02we de with large volumes of what we used
- 00:10:04to be dark data and then now that we can
- 00:10:06deal with that how does that change the
- 00:10:08types of data that we want to to connect
- 00:10:11with there's um a lot of use cases on
- 00:10:14creating um sort of better content on
- 00:10:17improving efficiency in the research
- 00:10:19process I me that's everything from like
- 00:10:21writing rfps to you know summarizing
- 00:10:24content there have been a lot of um or
- 00:10:27has been a lot of investment in
- 00:10:29Improvement in um areas around
- 00:10:33qualitative analysis um in particular
- 00:10:36you know if you have 20 Idis um you know
- 00:10:39you can just out of the box use existing
- 00:10:42technology um to summarize that content
- 00:10:45and then uh and then you know reduce
- 00:10:47that so really speeding up and and
- 00:10:49improving the operational efficiency of
- 00:10:50qualitative research so those those have
- 00:10:52been some areas that I think are widely
- 00:10:54used right now I think we're still we're
- 00:10:57still scratching the surface of the
- 00:11:00things we can do with it um and you know
- 00:11:04I know we're experimenting with ways of
- 00:11:07just radically changing surveys um
- 00:11:09things like brand tracking with
- 00:11:11open-ended analysis um or or emotional
- 00:11:14initation so what what is enabled by the
- 00:11:17technology out there is you know is are
- 00:11:21tools and processes that then give us
- 00:11:23abilities to collect data in different
- 00:11:26ways and to analyze that data in
- 00:11:27different ways I think that's kind of
- 00:11:28the the long poll um but in the short
- 00:11:31term the operational efficiency
- 00:11:32particularly in the coding and the open
- 00:11:33data is huge and constructing surveys is
- 00:11:36huge um in just writing surveys so you
- 00:11:38know people were discussing um use cases
- 00:11:41where they take an existing survey and
- 00:11:43say hey I have this new topic take my
- 00:11:46prior survey and just adapt it for this
- 00:11:49new domain this new vertical and just
- 00:11:51like you know save them a couple hours
- 00:11:53worth of work so I think those are a lot
- 00:11:55of the use cases people are discussing
- 00:11:56these
- 00:11:58days
- 00:12:00very
- 00:12:01interesting thank you so much for that
- 00:12:04we're going to go ahead and move on to
- 00:12:06mangela budia she is an associate
- 00:12:09director in the advanced analytics team
- 00:12:11at ipsos her area of expertise lies in
- 00:12:14advanced analytics across a wide range
- 00:12:17of multivariate techniques she
- 00:12:18specializes in Choice models such as
- 00:12:20conjoint and Max diff mangela has over
- 00:12:2420 years in research experience working
- 00:12:26with leading clients within a range of
- 00:12:28Industries and is respons responsible
- 00:12:29for Designing the advanced methodologies
- 00:12:31to ensure that all objectives are met
- 00:12:34and actionable insights can be taken
- 00:12:36from the
- 00:12:37findings go ahead mangela tell us a
- 00:12:39little bit about what you're going to be
- 00:12:41talking about at the upcoming SAU
- 00:12:43software analytics and insight
- 00:12:45Summit thank you very much Brian so I'm
- 00:12:48just gonna um share my screen to give
- 00:12:51you an idea of the things that we we've
- 00:12:53been going to be sharing at the summit
- 00:12:55next next month so hopefully you can all
- 00:12:58see my screen at the
- 00:13:00moment so the rise of large yeah the
- 00:13:03rise of large language models has led to
- 00:13:05a growing interest in their uses for
- 00:13:08data analytics within market research
- 00:13:11our paper titled the machines are here
- 00:13:13but will they take over what we did was
- 00:13:15we conducted a large research exercise
- 00:13:18involving um quarter of a million
- 00:13:21generated AI responses across a diverse
- 00:13:24set of scenarios and what we're doing is
- 00:13:26looking at their ability to replicate
- 00:13:28the results of previous conin and Maxi
- 00:13:32studies one of the main motivations
- 00:13:34behind this
- 00:13:40paper one of the main motivations behind
- 00:13:42this paper was a previous paper that
- 00:13:45explored the use of GPT 3.5 to mimic
- 00:13:48consumer behavior um they tested the
- 00:13:52consistencies of these models responses
- 00:13:54to these four economic theories and they
- 00:13:57concluded that large language models
- 00:13:59could broadly serve as a toour for
- 00:14:02understanding consumer preferences and
- 00:14:04and they found that the behaviors were
- 00:14:06consistent with economic theory however
- 00:14:09there were some other conclusions um two
- 00:14:12of the main conclu other conclusions
- 00:14:14were that there was a positional bias in
- 00:14:17that the first concept was selected more
- 00:14:19often than the other Concepts that were
- 00:14:21presented in these large language models
- 00:14:24and it was very sensitive to the prompts
- 00:14:26that it was
- 00:14:27given so it led just to you know
- 00:14:30although that paper laid a lot of the
- 00:14:32groundwork for us we actually had a lot
- 00:14:34of questions that we needed answers for
- 00:14:36so we put together these hypothesis to
- 00:14:38test can large language models handle
- 00:14:41more complex designs do different models
- 00:14:44impact Choice responses now that we've
- 00:14:47seen such a rise in the different types
- 00:14:49of large language
- 00:14:51models Within These models there's a
- 00:14:54temperature setting which controls for
- 00:14:57the randomness and the variability of
- 00:15:00the models responses so we wanted to
- 00:15:02better understand how this impacts
- 00:15:04performance and we also wanted to look
- 00:15:07at you know what's the best way to
- 00:15:08prompt the models and also understand
- 00:15:12how of a big issue the positional bias
- 00:15:16is we also wanted to look at you know if
- 00:15:19we ran analysis at a respondent
- 00:15:23individual respondent level could we
- 00:15:25achieve the same level of
- 00:15:26differentiation as we would get from
- 00:15:28real studies and what would happen if we
- 00:15:31were to train large language models with
- 00:15:33external kind of
- 00:15:36results ultimately as a commercial
- 00:15:38organization the critical hypothesis was
- 00:15:41that do the results derived from large
- 00:15:43language models provide the same
- 00:15:45commercial insights as a study with real
- 00:15:48respondents so with these hypothesis in
- 00:15:51mind we started off with exploratory
- 00:15:54first phase the this phase was focused
- 00:15:57around testing these experimental
- 00:15:59factors we took three commercial data
- 00:16:01sets you and took kind of a random
- 00:16:04sample of 500 respondents from each of
- 00:16:06those data sets and we looked at three
- 00:16:09different types of large language models
- 00:16:10we looked at GPT 4 claw 2 and Gemini Pro
- 00:16:13we also looked at temperature settings
- 00:16:15and what implications varing those
- 00:16:17temperature settings would do so here
- 00:16:19the lower value leads to more
- 00:16:21deterministic responses where higher
- 00:16:23values would then lead to more diverse
- 00:16:25responses and the prompts we fed were
- 00:16:28fixed prompts where we'd developed
- 00:16:30personas around some of their
- 00:16:32demographics and behaviors which were
- 00:16:34incorporated into the prompts and all
- 00:16:36the tasks were submitted as a single
- 00:16:38prompt in
- 00:16:40wo this first stage allowed us to narrow
- 00:16:43down the focus onto what we called the
- 00:16:46phase two of the research so what we did
- 00:16:49was we took the best performing large
- 00:16:51language model and parameter settings
- 00:16:53and we designed experiments around
- 00:16:56trying to refine the models to see if we
- 00:16:59can get better models better accuracy of
- 00:17:01the models the three main areas that we
- 00:17:03refined the models were refining The
- 00:17:05Prompt so looking
- 00:17:08at how could we change the prompts could
- 00:17:11we simplify the prompts in any way what
- 00:17:13would happen if we Chang the Persona
- 00:17:15text would that impact the responses
- 00:17:17that we were getting the second thing we
- 00:17:19looked at was training the models what
- 00:17:21we how we did that was we actually fed
- 00:17:23the um the the large language models
- 00:17:26answers from previous uh respond and we
- 00:17:29said this is how respondents have
- 00:17:31answered previously please take this
- 00:17:32into consideration when you are
- 00:17:34responding to the tasks that you're
- 00:17:36given and thirdly we then looked at the
- 00:17:39positional bias where we random
- 00:17:41randomized and rotated the order of the
- 00:17:44concepts to see if that would make a
- 00:17:46difference in the responses that we
- 00:17:49got so where are we today with answering
- 00:17:52you know large language models to
- 00:17:54answering Choice
- 00:17:55tasks large language models have shown
- 00:17:58impressive cap capabilities and there is
- 00:18:00still a lot of things that we need to
- 00:18:02consider and learn to unlock their use
- 00:18:04cases while mitigating their risks and
- 00:18:06limitation this is actually an AI
- 00:18:08generated image and you know it's done a
- 00:18:11really good job of generating the lead
- 00:18:14Terminator but if you look around the
- 00:18:16image you may find some imperfections my
- 00:18:19colleagues Chris and cam will be
- 00:18:20discussing these imperfections and also
- 00:18:23the answers to the hypothesis um that we
- 00:18:25looked at testing as well at the
- 00:18:28Analytics Insight Conference next month
- 00:18:30um so please do sign up if you want to
- 00:18:33hear some of the answers that we have um
- 00:18:36and some of the the insights that we
- 00:18:38want to share with you thank you
- 00:18:41fantastic Mula I got a question for you
- 00:18:44how how much are clients asking for AI
- 00:18:49solutions for marketing research or is
- 00:18:51it something you're seeing mainly being
- 00:18:54driven or recommended to clients by
- 00:18:56consulting firms and research providers
- 00:19:00um we definitely do have um clients that
- 00:19:03are more interested in this and want to
- 00:19:05partner with us um in in kind of almost
- 00:19:07co-developing Solutions so you know we
- 00:19:11are also developing our own Solutions as
- 00:19:14well so we have um we're looking at
- 00:19:17developing a an AI based chat bot um and
- 00:19:20we've got a number of initiatives that
- 00:19:21we're running um currently with clients
- 00:19:24so you know we're also looking at new
- 00:19:26types of work that can be looked out in
- 00:19:29terms of um you know Vision AI where
- 00:19:33we're looking at AI to process images
- 00:19:36and extract relevant information such as
- 00:19:38branding um Ambiance and things like
- 00:19:41that so we definitely have a lot of
- 00:19:43interested parties um wanting to know
- 00:19:46where this is going um and I think you
- 00:19:49know with the the buzzword of everyone
- 00:19:51talking about um generative AI these
- 00:19:54large language models I think clients
- 00:19:56are definitely showing or coming to us
- 00:19:59um to help you know help to develop some
- 00:20:03Partnerships where they can better
- 00:20:05understand how they can move forward
- 00:20:06with within this
- 00:20:08landscape so I haven't seen your slides
- 00:20:11yet you you looked at three different
- 00:20:13conjoin analysis studies to try to see
- 00:20:16if uh llms can can do a decent job of
- 00:20:19replicating what real respondents did
- 00:20:21could could you give us kind of a thumbs
- 00:20:23up a thumbs to the side or a thumbs down
- 00:20:25about what about how well it's doing in
- 00:20:27its current technology
- 00:20:30State
- 00:20:33um how about that okay a little bit like
- 00:20:37that now at this point I think I think
- 00:20:40you know we we do need to understand
- 00:20:42that you know these models are trained
- 00:20:45on we don't know the source of the
- 00:20:47information in terms of where they're
- 00:20:48trained and and how they're trained and
- 00:20:50and that's all very much unknown so you
- 00:20:53know we understand that the models as
- 00:20:56they grow in size um sometimes that you
- 00:20:59know it may not be that the fact that we
- 00:21:01need larger models but more models that
- 00:21:04are more tailored towards
- 00:21:07specific um targeted to kind of more
- 00:21:10specific areas more maybe even
- 00:21:12categories especially within the the
- 00:21:14market research industry so right I I
- 00:21:17know your colleague I was having an
- 00:21:18email with Chris Moore and he commented
- 00:21:21to me he says well you know if you're
- 00:21:23talking about a a a widely known and
- 00:21:26wide widely talked about topic on
- 00:21:28internet such as electric vehicles in
- 00:21:30the USA uh then llms might be able to do
- 00:21:34pretty well but he said but what about
- 00:21:36light bulbs in Estonia and that's what
- 00:21:39he he brought up as an example is okay
- 00:21:42it can only answer based on what it's
- 00:21:44trained upon and if we're talking about
- 00:21:46light bulbs in Estonia how well is it
- 00:21:48going to do compared to a well-known and
- 00:21:50well addressed topic on the internet
- 00:21:52such as electric vehicles in the United
- 00:21:55States yeah fantastic thank you mangela
- 00:21:57let's move on to Dan
- 00:21:59Penny Dan penny is a research and
- 00:22:01insights director at Microsoft his team
- 00:22:04supports monetization and business
- 00:22:06planning which drives decisions on
- 00:22:08business models packaging and pricing
- 00:22:11he's worked in various research roles
- 00:22:12since joining Microsoft in 2004 from
- 00:22:15supporting Azure and Microsoft 365
- 00:22:18product marketing to corporate issues
- 00:22:20like public policy engagement before
- 00:22:22Microsoft he worked at research
- 00:22:24international ultimately part of canar
- 00:22:26TNS and he also has a doctorate in 17th
- 00:22:30century French religious history with
- 00:22:34that Dan please take the stage oh thank
- 00:22:37you Brian and it's great to be here and
- 00:22:40yes if only that doctorate was in
- 00:22:42generative Ai and conjoint I'd be much
- 00:22:45better off um but really looking forward
- 00:22:47to uh the Sal tooth conference uh we
- 00:22:50actually have two presenters at the
- 00:22:52conference um because in Microsoft I
- 00:22:54would say we're really experimenting in
- 00:22:55and seeing adoption of AI for a broad
- 00:22:58range of research use cases so those
- 00:23:00flow across surveying uh guide design
- 00:23:03surveys themselves with chatbots uh
- 00:23:05analysis of open-ended data like Kevin
- 00:23:07talked about audio and video as well uh
- 00:23:10and then cross project synthesis and
- 00:23:12insights and actually Barry Jennings is
- 00:23:13going to be talking about that bigger
- 00:23:14picture at the conference and really our
- 00:23:16overall construct thinking about AI uh
- 00:23:19which is really a key Initiative for our
- 00:23:20research team this year so AI generated
- 00:23:23synthetic data which I'll be talking
- 00:23:25about at the conference along with two
- 00:23:27of my colleagues is really one aspect of
- 00:23:29that Bader AI initiative and actually
- 00:23:31one of our speakers is uh is Jimbo brand
- 00:23:33who uh manua mentioned his paper earlier
- 00:23:36on on col joy in one of those early
- 00:23:38papers um so at the conference we'll be
- 00:23:40talking about our general approach uh
- 00:23:43here around AI synthetic data and in
- 00:23:45learnings from five different
- 00:23:46experiments that we've done um I would
- 00:23:48say um just like usually talked about
- 00:23:50like we're really on a journey here uh
- 00:23:52and for us sort of being on the client
- 00:23:54side that's really involved uh vendor
- 00:23:56Outreach uh partnering with several uh
- 00:23:59research firms and AI Specialists and
- 00:24:01then developing our own in-house dools
- 00:24:03um I'd say there's a lot of academic
- 00:24:04literature too and so we're trying to
- 00:24:06sort of take that into account so for
- 00:24:08instance the way that um in some papers
- 00:24:11uh have talked about uh the way that
- 00:24:12Chad gbt might mainly focus on
- 00:24:14maximizing expected payoffs rather than
- 00:24:17doing what humans do which is often
- 00:24:19acting uh in a risk averse way for gains
- 00:24:22uh and risk-seeking for
- 00:24:24losses so I would say it's very early
- 00:24:26days we've really barely left the sh
- 00:24:29uh We've we've barely left Tatooine uh
- 00:24:31pick your metaphor um I do think though
- 00:24:33it is helpful to have a framework for
- 00:24:35thinking about the different types of AI
- 00:24:37generated data experiments and we'll
- 00:24:39talk about that at the conference
- 00:24:41because we really think that a single
- 00:24:43experiment proves almost nothing um it's
- 00:24:46rather along the lines of what Chris
- 00:24:47Chapman actually has talked about Chris
- 00:24:49Chapman and Google has talked about um
- 00:24:51so I think having a framework where we
- 00:24:53have Dimensions like you know who so
- 00:24:56what is the audience is it already a
- 00:24:57mainstream audience is it that Niche
- 00:25:00audience of light bulb buyers in Estonia
- 00:25:03as Brian you were mentioning uh so the
- 00:25:05who the what is it a familiar issue or a
- 00:25:08simple question or is it testing new
- 00:25:10value really new product value or a
- 00:25:12complex packaging question they're
- 00:25:14really LED themselves to a conjoint and
- 00:25:16then there the how Dimension how are we
- 00:25:18going to use this uh prec is it is it
- 00:25:20just for general direction is it a
- 00:25:23business critical topic is it something
- 00:25:25where we need a ton of precision uh is
- 00:25:27it to help pre tal cross stimul like we
- 00:25:29have a lot of use cases so I really
- 00:25:32think of it sort of as a cube with these
- 00:25:33dimensions and we need to test in really
- 00:25:36all the different spaces in that Cube to
- 00:25:39figure out what are The Right Use cases
- 00:25:41and uh um and how can we actually use it
- 00:25:43for the
- 00:25:44business uh and I would say that when
- 00:25:46assessing results we need to look pretty
- 00:25:48closely at the kind of measures that a
- 00:25:50number of folks have talked about like
- 00:25:52is there stability is there validity in
- 00:25:55that it agrees with things it should
- 00:25:56agree with and disagree with things
- 00:25:58where it should differ and what kind of
- 00:26:00distribution do we see in the data um
- 00:26:02and actually I think there's sort of the
- 00:26:04question about whether we'd really
- 00:26:05expect similar results to a human given
- 00:26:08the way that we know that humans work
- 00:26:10with conjoins so for instance with
- 00:26:12attribute nonattendance does gbt
- 00:26:15essentially provide the attribute
- 00:26:16nonattendance in a similar way so having
- 00:26:18said that I'll just turn briefly to our
- 00:26:20findings uh your sort of Romanesque
- 00:26:23thumbs up thumbs down uh some of the
- 00:26:25work that we have done is sort of
- 00:26:26conjoint related to all touch on that uh
- 00:26:28first I guess on the good side we have
- 00:26:30seen gbt out of the box can can work
- 00:26:32well on those public known topics your
- 00:26:34car example Brian uh so topics where
- 00:26:37we're asking things like PC form factor
- 00:26:39preferences or expectations of device
- 00:26:41costs on the consumer side um secondly
- 00:26:44from some conjoint work that it can do a
- 00:26:46decent job in simulating willingness to
- 00:26:48pay and feature importances at least in
- 00:26:51a couple of scenarios with well-known
- 00:26:52consumer products um and actually third
- 00:26:55that gbt can give answers that
- 00:26:56correspond more closely to reality than
- 00:26:58some surveys where really humans can
- 00:27:00might get confused or find it hard to
- 00:27:02render judgment so that's sort of the
- 00:27:04good news I guess the the the thumbs up
- 00:27:06bit on the thumbs down bit I would say
- 00:27:08that we see that um gbt out of the box
- 00:27:11as particularly out of the box can be
- 00:27:12just too optimistic or Tech forward uh
- 00:27:14so for instance asking it on the
- 00:27:17question about whether AI will actually
- 00:27:19be a positive influence on people's
- 00:27:20lives uh gbt and a rather
- 00:27:22self-interested way says yes more so
- 00:27:24than people uh second that it struggles
- 00:27:27with topics that are really distant from
- 00:27:28The Prompt information or from what's in
- 00:27:31public domain uh so for instance brand
- 00:27:34attributes uh we tend to see a lot
- 00:27:36higher agreement with those than humans
- 00:27:39uh and canar talks about something
- 00:27:40similar and it was really unsuccessful
- 00:27:43with much more sophisticated scenarios
- 00:27:45where we're testing say new products
- 00:27:46value in the commercial space to a
- 00:27:49specific audience like security decision
- 00:27:51makers so I would say in general we
- 00:27:53really see sort of the most promise with
- 00:27:55particular scenarios and uh having the
- 00:27:58right horse for the course uh so what
- 00:28:00might work in a particular consumer
- 00:28:03scenario might not work in a commercial
- 00:28:05scenario where we might need a much more
- 00:28:08trained model leveraging internal data
- 00:28:10and Survey data to pre-build those uh
- 00:28:13that model and with fine-tuning maybe
- 00:28:15for specific tasks and so for instance
- 00:28:18in another experiment we did see that by
- 00:28:20leveraging rag uh we could see actually
- 00:28:22very similar utilities to our human data
- 00:28:24even in that somewhat more complex case
- 00:28:27um but I would say even there we're very
- 00:28:28much sort of on the journey and uh we
- 00:28:31expect to be doing a lot more
- 00:28:32experimenting I think we're we're really
- 00:28:34one of those firms which would be in the
- 00:28:36that bucket of uh being doing a lot of
- 00:28:39Outreach uh so to folks like manua to
- 00:28:42understand what they're doing and and
- 00:28:43looking for partnership um so yeah we're
- 00:28:46looking forward to continuing that
- 00:28:47Journey thank you so much so can can you
- 00:28:51give a an example of a specific instance
- 00:28:53in which AI contributed significantly to
- 00:28:55a strategy or research effort at
- 00:28:57Microsoft so far yeah it's it's an
- 00:29:00interesting question Brian because in
- 00:29:01some ways it's suffused almost
- 00:29:02everywhere I would say so almost
- 00:29:05everything that we do at some level I
- 00:29:06would say incorporate and not everything
- 00:29:08but a large proportion so for instance
- 00:29:11uh I would say an awful lot now of the
- 00:29:13cases where we have open-ended data
- 00:29:15we're either experimenting with a bot in
- 00:29:18the survey or we're using it for coding
- 00:29:20so for instance um our sort of main
- 00:29:24customer and partner satisfaction survey
- 00:29:2650,000 open then did twice a year we are
- 00:29:30using uh you know using it for coding in
- 00:29:32that case um I think we have something
- 00:29:34like 80 odd projects in the first half
- 00:29:37where we're using AI so it is suffused
- 00:29:40everywhere but on the other hand I think
- 00:29:41there are a lot of specific use cases
- 00:29:43where we haven't yet said uh like with
- 00:29:46synthetic data from con joints that
- 00:29:48we're confident enough to step away and
- 00:29:51not do the human survey we're much more
- 00:29:53in the experiment do parallels and I
- 00:29:56think really the question is
- 00:29:59um what's the what's the right series of
- 00:30:02tests for a particular scenario that we
- 00:30:04given us confidence to actually take
- 00:30:07that step away um and you know you can
- 00:30:09imagine that's not going to be the most
- 00:30:12the most business critical topics to
- 00:30:14begin with it's likely to be sort of
- 00:30:15lower Ling fruit um but yeah I would say
- 00:30:17otherwise AI especially in coding in
- 00:30:21qual sort of summarization is more
- 00:30:24everywhere than nowhere um actually yeah
- 00:30:27and when you see coding you're not
- 00:30:28meaning
- 00:30:30programming programming languages you're
- 00:30:32talking about coding of open-end content
- 00:30:35right yeah exactly like uh with video um
- 00:30:38and so on so you can imagine I think the
- 00:30:39typical call Project now is somewhat
- 00:30:41different than it used to be in the
- 00:30:42sense of the ability to do very quick
- 00:30:45turn summaries from each uh interview
- 00:30:47right uh the ability to then do thematic
- 00:30:50views like it it I think it's
- 00:30:53democratizing at least access to the
- 00:30:55material much more quickly uh giving us
- 00:30:58uh an ability to report earlier on what
- 00:31:00we're seeing so yeah I would say that's
- 00:31:02um yeah but that's what I meant by the
- 00:31:03coding so I mean we're doing we're using
- 00:31:05it for coding too but that's sort of a
- 00:31:06separate
- 00:31:07bucket thanks so much our last speaker
- 00:31:11is uh Dr Jeffrey doson he's a professor
- 00:31:15of marketing at the Marriott School of
- 00:31:16Business at BYU Jeff received his PhD in
- 00:31:20quantitative marketing from the Fisher
- 00:31:21College of Business at Ohio State
- 00:31:24University his research focuses on the
- 00:31:26development and applic ation of basian
- 00:31:28statistical methods to a variety of
- 00:31:30theoretical and applied marketing and
- 00:31:32management problems Jeff has taught
- 00:31:35courses in marketing research marketing
- 00:31:37analytics pricing strategy customer
- 00:31:39relationship management survey research
- 00:31:41Advanced analytics and generative
- 00:31:43artificial intelligence so I suppose
- 00:31:45that's going to make you an expert on
- 00:31:47the topic go ahead
- 00:31:49Jeff well thanks Brian let me share my
- 00:31:52screen expert expert is a strong word I
- 00:31:55don't think I fall into the expertise
- 00:31:56category um so I'm I'm excited to
- 00:31:59present this I'm I'm chairing uh a
- 00:32:00session as part of the academic track at
- 00:32:03the conference um that includes three
- 00:32:05speakers all speaking about the use of
- 00:32:06generative artificial intelligence in
- 00:32:07conjoint uh we have IA Israeli whose
- 00:32:10paper on GPT for marketing research has
- 00:32:12me been mentioned a couple of times
- 00:32:14she's from the Harvard Business School
- 00:32:15she'll be presenting to us uh Nino Hart
- 00:32:17who used to be at Ohio State is now with
- 00:32:18skim is presenting um another paper on
- 00:32:21on using large language models to
- 00:32:23explore U the effects of product
- 00:32:25introduction exploring sort of
- 00:32:26interesting heterogen distrib tions and
- 00:32:28then I'm presenting something a little
- 00:32:29bit different and so my my paper is
- 00:32:31called creating experimental stimuli
- 00:32:33with generative AI um this is the
- 00:32:35outcome of a project that I've been
- 00:32:36working on for the past year or so uh
- 00:32:38which is to say I've been doing a thing
- 00:32:40I like that thing a lot but I'm not sure
- 00:32:42if it's the right thing to do uh it's
- 00:32:43co-authored with with Roger Bailey from
- 00:32:45Ohio State and so the um the project
- 00:32:47we've been working on um is uh that
- 00:32:49motivated this is related to this
- 00:32:51question of what is the the value of
- 00:32:52artistic style so these generative AI
- 00:32:55systems large language models and text
- 00:32:56to image generators they're trained on
- 00:32:58enormous data sets that were largely
- 00:33:00scraped from the internet so it has
- 00:33:02access to information that's part of the
- 00:33:03public domain some information that is
- 00:33:05uh that's probably privately held and it
- 00:33:07may be a violation of of intellectual
- 00:33:09property laws um artists in particular
- 00:33:11uh they're super concerned about the way
- 00:33:13that text to image generation uh may
- 00:33:15maybe misappropriate their style and
- 00:33:17potentially replace them in terms of the
- 00:33:19the work they do as artists and so our
- 00:33:21paper explores this issue lots of
- 00:33:23lawsuits going on right now uh we
- 00:33:25explore a variety of different potential
- 00:33:27remediations uh to this problem um and
- 00:33:29one of those is the potential to pay
- 00:33:30artists a royalty uh for for the use of
- 00:33:33their style and maybe a fee for for
- 00:33:35being included within the training data
- 00:33:37set used to create these these models um
- 00:33:39and so to to assess um a royalty
- 00:33:42structure we need to understand what is
- 00:33:43the incremental value of a particular
- 00:33:45artist style um as it relates to to like
- 00:33:48preference willingness to pay in a
- 00:33:49commercial context and to get there this
- 00:33:51is this is a great application of conoid
- 00:33:53analysis and so um so I could do this a
- 00:33:55couple of ways so one is I I could
- 00:33:57create a conid study where I use uh
- 00:33:59verbal descriptions of products and so
- 00:34:00in this case we're doing preference for
- 00:34:02like vinyl stickers that you might stick
- 00:34:03on like your bumper or like your Stanley
- 00:34:05or your C toop carrier and I could
- 00:34:07actually describe like what is the
- 00:34:08subject of the sticker so what's the
- 00:34:10topic who is the artist and then I can
- 00:34:12give a price point and so in this case I
- 00:34:13might say like the subject is a cat in a
- 00:34:15cup and the artist is alons Muka and the
- 00:34:18price point is 149 and I could collect
- 00:34:20data about this but it creates a lot of
- 00:34:22problems right because I I might know
- 00:34:23the subject I might not know the artist
- 00:34:25but I have no idea of like what the
- 00:34:26sticker actually looks like that matters
- 00:34:28a lot uh and maybe a better way to do
- 00:34:30this is to actually um create those
- 00:34:32Concepts themselves and so in this case
- 00:34:34I've got the same information embodied
- 00:34:36Within These images that i' I've created
- 00:34:38um in this case using mid Journey so
- 00:34:40I'll show you the sticker or uh a
- 00:34:41concept of the sticker uh a price point
- 00:34:43and I'll have you pick the one you like
- 00:34:44the best um so what we're effectively
- 00:34:47doing here is we're creating our
- 00:34:49experimental design in the space of the
- 00:34:50generative prompt and so I'm laying out
- 00:34:52my attributes and levels within the
- 00:34:54prompt that's being used to create these
- 00:34:55images so create an image of of create a
- 00:34:58sticker of a subject in the style of an
- 00:34:59artist so for example create a sticker
- 00:35:02of a cat in the cup in the style of
- 00:35:04alfons Muka and these systems are really
- 00:35:06phenomenal at generating this this uh
- 00:35:08this this this type of object this this
- 00:35:10this the stimulite the challenge is that
- 00:35:11if I just create one of these things um
- 00:35:14um I'm I'm Bound by basically the fixed
- 00:35:16effect of that object so the interaction
- 00:35:17of the topic with the artist and then
- 00:35:19that particular realization from from
- 00:35:21the stochastic generator and so my
- 00:35:23intuition around this is I have to
- 00:35:24create lots of versions of these things
- 00:35:26and so I can create many versions of of
- 00:35:29of stickers generated from the same
- 00:35:30basic prompt um and I can use that to
- 00:35:33sort of back into like what like what is
- 00:35:34the the value of the artistic style of
- 00:35:35of Muka for example um and so we're
- 00:35:38using gen AI to create stimuli um as
- 00:35:41I've been doing this uh it's starting to
- 00:35:43apply a lot of different contexts um
- 00:35:45it's applies to scenarios where verbal
- 00:35:47descriptions of products are are
- 00:35:48difficult or effortful for consumers to
- 00:35:50evaluate um humans as we know are better
- 00:35:52at responding to very specific things as
- 00:35:54opposed to these abstract Concepts and
- 00:35:56so we're finding applications in concept
- 00:35:58testing and package design and product
- 00:35:59listing photos uh social media posts
- 00:36:02advertising copy which would be a textto
- 00:36:04text example um lots of applications but
- 00:36:07this raises a bunch of of issues uh
- 00:36:10related to kind of the measurement of of
- 00:36:12these effects and so we're effectively
- 00:36:14creating a new measurement scale and we
- 00:36:15need to be able to evaluate the
- 00:36:16properties of that scale specifically
- 00:36:19with respect to like the validity and
- 00:36:20reliability of what we're measuring um
- 00:36:23how do we deal with the fact that gen is
- 00:36:24intrinsically stochastic which is to say
- 00:36:26there's a one to many mapping between
- 00:36:28the semantic information in a prompt and
- 00:36:30then the images or text that's generated
- 00:36:32accordingly I think I think
- 00:36:32theoretically it's a one to infinite
- 00:36:34mapping between uh The Prompt and what
- 00:36:36can be created um these these images and
- 00:36:38text they generated from distributions
- 00:36:40of unknown form can we learn something
- 00:36:41about those
- 00:36:42distributions um is it even possible to
- 00:36:45orthogonally manipulate multiple
- 00:36:46features within the same prompt like
- 00:36:47images are are multi-dimensional they're
- 00:36:49evocative they're easy to respond to but
- 00:36:51can I really manipulate artistic style
- 00:36:53and and the base image uh simultaneously
- 00:36:56um if not can I find ways to estimate
- 00:36:58what is the correlation between those
- 00:37:00things and find some way to sort of
- 00:37:01debias by by understanding how those
- 00:37:03things are associated with with uh with
- 00:37:05each other in the embeding space so the
- 00:37:07paper we're presenting is uh is
- 00:37:09addressing these and and potentially
- 00:37:11many other issues um our goal is
- 00:37:13basically to provide a protocol that can
- 00:37:15be used by researchers to implement and
- 00:37:16to also justify this this type of
- 00:37:18measurement system where we're building
- 00:37:20the experimental design in the space of
- 00:37:21the generative prompt and now we're
- 00:37:22creating these really interesting
- 00:37:24evocative things that consumers can
- 00:37:25respond to um really excited to talk
- 00:37:27about
- 00:37:29it fantastic Jeff I got a followup
- 00:37:32question for you I I imagine that
- 00:37:34universities are rapidly adjusting to
- 00:37:36the influence of AI and education what
- 00:37:38are some of the major initiatives for
- 00:37:42professors surrounding Ai and teaching
- 00:37:44students about marketing strategy and
- 00:37:46analytics so I I don't know if we have
- 00:37:48any university-wide uh strategies I
- 00:37:50think there's a lot of experimentation
- 00:37:51that's being done by by individual
- 00:37:53instructors and I'd say in the business
- 00:37:54school we're maybe a little more excited
- 00:37:56about this than my my colleagues like in
- 00:37:57English they're they're kind of worried
- 00:37:58about what this does to their ability to
- 00:38:00teach writing for example um for an
- 00:38:03illustration like I I teach a class
- 00:38:04right now on um it's the Core Business
- 00:38:07analytics class required for all of our
- 00:38:08our first year MBA students and we've
- 00:38:10been doing um all of the the coding so
- 00:38:12actual computer coding and and execution
- 00:38:15within the premium chat GPT platform and
- 00:38:17so I'm having students uh train complex
- 00:38:20machine learning and statistical models
- 00:38:22and then to be able to query those
- 00:38:23models to do kind of like uh decision
- 00:38:25theoretic stuff after the fact and so
- 00:38:27that's been an experiment that I've done
- 00:38:29for the first time this year it's been
- 00:38:31it's been really effective to be honest
- 00:38:32it's worked really well although there
- 00:38:33are some challenges challenges with it
- 00:38:36um it's a nice way to get sort of
- 00:38:37non-technical students up to speed with
- 00:38:39respect to coding um I'm also teaching a
- 00:38:41class on generative artificial
- 00:38:42intelligence for marketing productivity
- 00:38:44and in that class our students are just
- 00:38:45basically building a business using gen
- 00:38:47tools including text to text generators
- 00:38:49and text to image generators and a
- 00:38:51little bit of coding on top of it and
- 00:38:53that's also been a really interesting
- 00:38:54experiment so to answer your question um
- 00:38:56lots of people are trying lots of
- 00:38:57different things I don't think we've
- 00:38:58settled yet on what the best approach is
- 00:39:01um but there is a lot of opportunity and
- 00:39:02prob probably a lot of drawbacks too but
- 00:39:04um but I think most of us are pretty
- 00:39:05pretty excited about
- 00:39:07it super thanks so much why don't all of
- 00:39:10you go ahead and and all of you
- 00:39:12panelists go ahead and turn on your
- 00:39:14videos and it we've been going pretty
- 00:39:16well here I think we have about five
- 00:39:18minutes for you guys maybe to have a
- 00:39:21follow-up comment or to address a
- 00:39:23question to another panelist and then
- 00:39:25with that we'll turn it over to just who
- 00:39:27will pull some questions from the
- 00:39:29audience
- 00:39:33Q&A well I'm happy to start Brian with
- 00:39:35maybe with a question uh maybe to manua
- 00:39:38and Kevin in particular I guess um one
- 00:39:40one of the things that we're interested
- 00:39:41in is obviously we have sort of in the
- 00:39:43context of conjoint because we do a lot
- 00:39:45of conjoint um we're semi- obsessed with
- 00:39:48conjoint I would say in uh in our work
- 00:39:50at Microsoft um uh whether you're doing
- 00:39:53much work where it's not about synthetic
- 00:39:56design but about hey how AI might be
- 00:39:58leveraged in the conjoint itself so for
- 00:40:00instance with asking through a chatbot
- 00:40:04uh more open-ended data uh slowing
- 00:40:07somebody down as part of the conjoint
- 00:40:09itself so that AI can do a better job
- 00:40:12maybe in actually improving the data
- 00:40:13quality from the conjoint and
- 00:40:15potentially the open-ended it's being
- 00:40:16folded into the actual model itself as
- 00:40:19as part of those parameters or where for
- 00:40:21instance AI might be part of the actual
- 00:40:23model building or the design generation
- 00:40:25in the first place I was just kind of
- 00:40:26curious if you've done much any sort of
- 00:40:28work around that area as part of the
- 00:40:30conjoin
- 00:40:33work uh you're you're on
- 00:40:36mute oh thank you in terms of directly
- 00:40:39the open ends into con the conjoint
- 00:40:41model directly I'd say no not quite yet
- 00:40:44um there has been a fair amount of work
- 00:40:47in terms of sort of AI interactive why
- 00:40:49and dynamic probing using Ai and the
- 00:40:51impact that has on the quality of data
- 00:40:53not just of the response which is
- 00:40:55significant but also on actually the
- 00:40:56rest of the survey where people realize
- 00:40:59that wow you're actually paying
- 00:41:00attention to me ask me why and they then
- 00:41:03start paying more attention as you go
- 00:41:05through um and there have been
- 00:41:07significant improvements in like the
- 00:41:08overall quality of the rest of the
- 00:41:10survey and I don't think I've seen that
- 00:41:12specifically in the context of condrin
- 00:41:14yet but I have to imagine that would be
- 00:41:16one of the areas where it would have a
- 00:41:18huge Improvement precisely because
- 00:41:20condrin was an ative sometimes
- 00:41:22repetitive task as you go through Choice
- 00:41:24sets yeah we had done one experiment we
- 00:41:27certainly have been seeing that in the
- 00:41:28experiment that it at least in that one
- 00:41:30experiment that it did improve the sort
- 00:41:32of the the overall data quality because
- 00:41:34it I think because of that exact issue
- 00:41:36that it slowed people down and they feel
- 00:41:38like somebody's actually sort of rather
- 00:41:39like in a if it was sort of more
- 00:41:41qualitative interview that somebody's
- 00:41:43actually sort of paying attention to
- 00:41:44them um in a dynamic way
- 00:41:51so thanks again my pleasure we have
- 00:41:55question we have time for yet another
- 00:41:57question among the panelists or comment
- 00:42:01I'll throw up there has anyone um been
- 00:42:03experimenting on or with with simulated
- 00:42:06respondents on emotional responses um
- 00:42:10and comparison between say functional
- 00:42:12and emotional traits you know whether or
- 00:42:14not um simulated respondents behave
- 00:42:16differently than real people when
- 00:42:17they're evaluating things with you know
- 00:42:19in a more emotional context
- 00:42:26specifically
- 00:42:30I can I can I can maybe touch on that I
- 00:42:33I think I mean I know that there's been
- 00:42:34quite a lot of literature around that I
- 00:42:36think har mine for instance I've seen
- 00:42:38some work they've done I think actually
- 00:42:39abos has done some work on that too I
- 00:42:41think um I think the thing that we have
- 00:42:44seen at least is that I'm not sure if
- 00:42:46this quite in that emotional bucket but
- 00:42:48the the the emotional reaction to
- 00:42:51Brands is sort of is a little different
- 00:42:54with gbt versus human so we do see in
- 00:42:56general this sort of much higher level
- 00:42:58of sort of emotional brand
- 00:43:01Association um with gbt as opposed to
- 00:43:04human so in general like our brand
- 00:43:06scores if all our brand surveys were via
- 00:43:09gbt our brand scores would be through
- 00:43:11the roof um but so with everybody else's
- 00:43:13and so it would be a great way of uh
- 00:43:15really meeting your brand targets for
- 00:43:16the year um so yeah I think that's like
- 00:43:20how you sort of can fine-tune the
- 00:43:23emotional response there I think I think
- 00:43:25yeah that's something we need to sort of
- 00:43:26do more
- 00:43:30more well super maybe at this point we
- 00:43:33we let Justin dig into the bag of
- 00:43:36questions and start pulling out a few to
- 00:43:39submit to the panelists for their
- 00:43:41thoughts yeah before I do just a quick
- 00:43:45plug here to check out our our software
- 00:43:48um we are working hard to make an
- 00:43:51amazing system we'd like you to take a
- 00:43:53look at it at discover. software.com and
- 00:43:56also this analytics and insight Summit
- 00:43:58right uh April 29th through May 3rd San
- 00:44:01Antonio if you can't make it down there
- 00:44:04uh check out these uh this virtual
- 00:44:05access option and you can see that at
- 00:44:08satus
- 00:44:09software.com onference
- 00:44:12okay well uh we have a bunch of
- 00:44:15questions let's see here um uh mangela
- 00:44:19there's a question for you uh in their
- 00:44:20paper did they only focus on temperature
- 00:44:23as a prompt parameter or also explore
- 00:44:26other prompt components parameters such
- 00:44:29as top P best of frequency or uh
- 00:44:33presence penalty if not what was the
- 00:44:35reason to focus on temperature as a
- 00:44:37parameter
- 00:44:40only so um I
- 00:44:44think
- 00:44:46it's the temperature setting is quite
- 00:44:49well known from the original paper that
- 00:44:51we we looked at which um they only
- 00:44:54looked at one temperature setting um I
- 00:44:57think they use a temperature setting of
- 00:44:58one um in that paper so what we wanted
- 00:45:01to do was we wanted to really understand
- 00:45:04if you varied that temperature setting
- 00:45:07you know what impact would that have
- 00:45:08then on the results so we ran you know
- 00:45:11over kind of 50 experiments just looking
- 00:45:13at those different temperature
- 00:45:17settings okay thank you uh Jeff one for
- 00:45:20you um how competent will students using
- 00:45:23generative AI to code be at debugging or
- 00:45:26Rec izing errors are we engineering a
- 00:45:29black box culture yeah that's a great
- 00:45:31question so these are MBA students and
- 00:45:33so they tend not to be super competent
- 00:45:34with respected coding to begin with um
- 00:45:36the way I view it so so in in chat GPD
- 00:45:39with her code interpreter what it what
- 00:45:40it will do is it it writes code in
- 00:45:41Python it does execution in browser and
- 00:45:43then allows them to query those results
- 00:45:45using natural language and so I would
- 00:45:46view this uh as maybe a way to get
- 00:45:48people into coding uh maybe not a
- 00:45:50replacement for coding but a way to get
- 00:45:51them into coding it feels a lot like
- 00:45:53learning to code by recording like
- 00:45:54macros in VBA and then doing Ting after
- 00:45:57the fact within within Excel um and so
- 00:45:59anything I can do to get my students
- 00:46:01interested in analytics and and coding
- 00:46:04and and Analysis I think is I think is
- 00:46:06useful um but again I I there there are
- 00:46:09lots of like intended consequences lots
- 00:46:10of unintended consequences I think at
- 00:46:12this point we we don't know
- 00:46:15um great thank you another question for
- 00:46:18the group uh do we know if how llms
- 00:46:23prioritize newer
- 00:46:25information
- 00:46:32maybe what do you mean by newer
- 00:46:37information I would think that it might
- 00:46:40be talking about okay was this was this
- 00:46:44information that I scraped information
- 00:46:47that was written this year or previous
- 00:46:48year and is our LM our llm is going to
- 00:46:52going to wait the more recent
- 00:46:54information that it scrapes higher than
- 00:46:56older
- 00:46:59information I I don't believe that's the
- 00:47:01way they function right so the
- 00:47:02construction of the llm itself I don't
- 00:47:03think prioritizes the fre the recency of
- 00:47:05the of the data a lot of these systems
- 00:47:07now they they interact with the internet
- 00:47:09and so when they produce responses
- 00:47:10they'll do an internet search and
- 00:47:11provide things that maybe are a little
- 00:47:12more contemporaneous but that's not part
- 00:47:14of the construction of the of the large
- 00:47:16language model
- 00:47:18itself one of the second that one of the
- 00:47:21caveat though is that um llms when
- 00:47:23they're referencing the training data um
- 00:47:26they do recognize that some of the
- 00:47:28training data references previous data
- 00:47:29sufferance that there's a medical
- 00:47:31article that says this is old data and
- 00:47:33refers to a prior article right it's an
- 00:47:36establishing that the newer article is
- 00:47:38you know more recent and more relevant
- 00:47:40so in that sense even though it may not
- 00:47:42be specifically trained to to prioritize
- 00:47:45newer information the content itself may
- 00:47:48have the effect of prioritizing newer
- 00:47:51information great thank you uh question
- 00:47:54for Dan what are some of the tools you
- 00:47:56have used to help you distill
- 00:47:58qualitative findings are there
- 00:48:00limitations to what kind of qual data
- 00:48:02you are willing to feed into the AI for
- 00:48:05example pre-release marketing
- 00:48:09content uh yeah so we um yeah we do
- 00:48:12there are some particular tools so we
- 00:48:14actually uh for instance have a
- 00:48:15partnership with Vox pop me uh and so we
- 00:48:19um we're basically leveraging Fox pop me
- 00:48:22as a platform and uh Ro Romani uh from
- 00:48:26uh RT is actually talking a number of
- 00:48:28conferences I think with Vox pop me so
- 00:48:30yeah so I would say there is sort of a
- 00:48:32particular platform that we we we're
- 00:48:33using a lot for um for that um for video
- 00:48:37audio and so on um and then in terms of
- 00:48:40material yeah we we're obviously we're
- 00:48:43also very very careful about so what
- 00:48:45material we use and where we use it so
- 00:48:47for instance uh we do have some guidance
- 00:48:50uh uh this is for anything that we do
- 00:48:52actually about what which llm it's it's
- 00:48:56used on uh so for instance you know we
- 00:48:58don't want to put anything uh on an llm
- 00:49:01that's sort of public like we'll want to
- 00:49:03run it on Azure um in a confidential
- 00:49:06setting so that nothing goes to a to
- 00:49:10another external public LM so we're very
- 00:49:12actually careful uh about we have a lot
- 00:49:15of guidance that we've given to people
- 00:49:17about what you can and can't test and
- 00:49:19how you can and can't test it um so yeah
- 00:49:22I'd say we're actually sort of very
- 00:49:23careful about that what we don't want is
- 00:49:25a whole bunch of material going
- 00:49:27including from respondents right going
- 00:49:29out into a public domain so everything
- 00:49:31is in contained
- 00:49:33environments great thanks another
- 00:49:35question here about AI hallucinations it
- 00:49:38was mentioned that qual qualitative
- 00:49:40analysis has been more efficient now
- 00:49:43that we have gen AI have there been any
- 00:49:46issues with AI hallucinating when
- 00:49:49creating summaries or identifying themes
- 00:49:52from your
- 00:49:55knowledge
- 00:49:58yes there are a lot of stories about
- 00:50:00that um qualitative researchers who've
- 00:50:02seen that basically summarize it that um
- 00:50:05and the common refrain is that the llm
- 00:50:07models are trained to be very helpful
- 00:50:09and they're so helpful they're so eager
- 00:50:12to make you happy that when there's
- 00:50:13nothing in there to justify a particular
- 00:50:15finding like find a quote that justifies
- 00:50:17this finding it will happily make up
- 00:50:19that quote um now there are ways to
- 00:50:22address that and a lot of it is prompt
- 00:50:23engineering so if you engineer your
- 00:50:25prompt to specifically tell it not to do
- 00:50:27that you can protect yourself to some
- 00:50:29degree from certain types of
- 00:50:30hallucination but it's certainly
- 00:50:32something you have to be cautious
- 00:50:34about yeah and I think just to add to
- 00:50:36that on the sort of helpful side I mean
- 00:50:38I think an example is where if you if
- 00:50:40for instance you might say in part of
- 00:50:41the prompt you know choose one to three
- 00:50:43things it will tend to pick three things
- 00:50:46uh it will it will try and be as helpful
- 00:50:47as possible and the human won't do that
- 00:50:49and so you get sort of a different
- 00:50:50distribution of counts uh in in cases
- 00:50:54like that so yeah so I think there are
- 00:50:56the the those kind of differences around
- 00:50:57helpfulness I'm not sure about I mean
- 00:50:59I'm sure we have examples of
- 00:51:00hallucination but certainly that sort of
- 00:51:02helpfulness aspect uh is is definitely
- 00:51:06that okay uh with the development of gen
- 00:51:08AI tools do you think there will be
- 00:51:11potentially a data collection tool that
- 00:51:14can replace
- 00:51:20surveys I
- 00:51:24mean
- 00:51:25maybe
- 00:51:28if you want I I'll just answer so so
- 00:51:30partly yes I mean I think the question
- 00:51:32is going to be a little different right
- 00:51:34um because we can gather brand equity
- 00:51:37and we can run a lot of the surveys we
- 00:51:38run through opened analysis or just
- 00:51:40conversational tools and generate
- 00:51:42incredible data and quantify it at
- 00:51:44Scales of n equals a th000 n equals
- 00:51:462,000 the question is what will surveys
- 00:51:48look like and one of the big questions I
- 00:51:50think we're going to be asking ourselves
- 00:51:52is not can I gather this data through an
- 00:51:55open and end method because previously
- 00:51:57you know for 50 years has been what I
- 00:51:58you know I can't do this too costly to
- 00:52:00code open ends Etc but should I um and I
- 00:52:04think that that the comparison will be
- 00:52:06much more closer to something like um
- 00:52:08when you're trying to gather brand
- 00:52:10awareness and you do aided awareness
- 00:52:11versus uned awareness right uned
- 00:52:13awareness is what's top of mind what do
- 00:52:15people think about aided awareness is
- 00:52:17here's here's 20 different claims I
- 00:52:19could make which of these claims is most
- 00:52:22compelling and and you you know that's P
- 00:52:25that's a close in an answer so the
- 00:52:26question is what are we trying to do
- 00:52:28with it so it's no longer what are we
- 00:52:30capable of doing or what is cost
- 00:52:32effective because the cost is going to
- 00:52:34be reduced across the board but what do
- 00:52:36we want to do what type of information
- 00:52:39how do we want to engage people what do
- 00:52:40we want to elicit from their minds again
- 00:52:43just a that's a very high LEL thought
- 00:52:46thank you I do think as well that we're
- 00:52:48so far away from that that it's I I
- 00:52:52understand the question but it's also
- 00:52:53this it's such a sort of theoretical
- 00:52:55question that I think it's sort of not
- 00:52:57really top of mind for us at least right
- 00:52:59now I I think for us it's more about um
- 00:53:02what what are the tools that can make us
- 00:53:04more effective um in terms of you know
- 00:53:07uh Speed and Agility and so on and so on
- 00:53:09uh and then sort of this exploration
- 00:53:11around synthetic because um because
- 00:53:14coming back to Brian's question earlier
- 00:53:16it's not like we're remotely in a
- 00:53:18position where we would have a single
- 00:53:19use case I think the synthetic where
- 00:53:21right now we would we would say we'll do
- 00:53:23this as opposed to the human um you know
- 00:53:25it be so limited and so yeah we just
- 00:53:29quite a long way I think from
- 00:53:31that great uh another question there's a
- 00:53:34lot of them here by the way and um we're
- 00:53:36probably not going to be able to answer
- 00:53:38all of them uh except for maybe our
- 00:53:40panelists will be willing to type in
- 00:53:42some answers uh if you interact with the
- 00:53:44Q&A section there uh towards the end
- 00:53:46anyway here's another one do any of
- 00:53:49y'all have recommendations for a solid
- 00:53:52primer one two hours worth of learning
- 00:53:55on prompt engine engering for beginners
- 00:53:58there are tons of resources out there
- 00:54:00but it's a bit overwhelming to put it
- 00:54:04mildly I can make a recommendation so
- 00:54:07Ethan mullik at at Wharton has done a
- 00:54:09lot of work on on providing sort of
- 00:54:12practical um tips on how how to use
- 00:54:14these systems so I would take a look at
- 00:54:15like gu Twitter feed is really useful um
- 00:54:17he he's published a book just recently
- 00:54:19that might be useful uh he was recently
- 00:54:21on a podcast with Ezra Klein where they
- 00:54:23talk about ways to use GPT um to use
- 00:54:26large language models in general um I
- 00:54:28think his stuff is is pretty fantastic
- 00:54:30can you say the name again slowly Ethan
- 00:54:33mullik don't ask I don't e t h an I
- 00:54:37think it's m o l l i CK
- 00:54:40maybe he's on the faculty at at the
- 00:54:42Wharton School at the University of
- 00:54:44Pennsylvania another kind of related
- 00:54:46question about you know kind of coming
- 00:54:48up to speed for researchers new to this
- 00:54:50topic are there any review papers or
- 00:54:53knowledge SharePoint that introduces or
- 00:54:56summarizes the different research
- 00:54:58efforts used in developing simulated
- 00:55:05respondents I mean I think there's a lot
- 00:55:07of papers out there I think one one
- 00:55:09that's quite good uh I just have it on
- 00:55:11another screen actually um is maybe I'll
- 00:55:14put it into the chat but there was one
- 00:55:16about using large language models to
- 00:55:17generate silicon samples challenges
- 00:55:19opportunities and guidelines I thought
- 00:55:21that was a pretty helpful sort of review
- 00:55:22of academic literature um but there are
- 00:55:25quite a few of those sort of academic
- 00:55:26papers I think challenge for somebody
- 00:55:28like me on the client side right without
- 00:55:30that muscle a historic muscle around
- 00:55:32this stuff is just sort of understanding
- 00:55:34it and wrapping your head around that
- 00:55:35but that but that I thought was actually
- 00:55:37quite a good sort of overview
- 00:55:39paper great um another question here
- 00:55:42about uh can anyone rub over the topic
- 00:55:45about ethics in
- 00:55:51AI there's a massive conversation about
- 00:55:53it at every single level the insights
- 00:55:55asso
- 00:55:56within companies I'm sure Microsoft has
- 00:55:58it so it's it's just huge it's too big
- 00:56:01of a topic to do in one
- 00:56:04minute agreed I mean my my paper that I
- 00:56:07I alluded to in my presentation is
- 00:56:09motivated around one of these ethical
- 00:56:10issues which is you know um where where
- 00:56:13does this training data come from who
- 00:56:14owns it what is owed as a result of
- 00:56:16owning it I think there are tons of
- 00:56:18lawsuits going on right now to try to
- 00:56:19settle these issues and that's that
- 00:56:21those are like the legal questions like
- 00:56:22the ethical questions are even more
- 00:56:23nuanced to to deal with so lots of
- 00:56:26questions lots of great conversation um
- 00:56:29but too much too much for a minute I
- 00:56:30agree with Kevin uh Jeff quick question
- 00:56:33for you how do you handle disclosure of
- 00:56:36AI generated images for the respondent
- 00:56:39do you tell them previous to the task
- 00:56:42and do you think there would be a
- 00:56:43different a difference in responses to
- 00:56:45the choice tasks between people that
- 00:56:47know it is AI generated versus those
- 00:56:50that don't uh yes I mean there are
- 00:56:52people that um that like AI generated
- 00:56:54stuff and some that dislike it um a lot
- 00:56:55of people that don't care um in that
- 00:56:57paper we do three different studies in
- 00:56:59the first study I don't believe we tell
- 00:57:00them that the photos are AI generated in
- 00:57:02the other two studies we we do tell them
- 00:57:04um and in terms of the dominant effect
- 00:57:06dominant effect that we care about which
- 00:57:07is like the willingness to pay for
- 00:57:08artistic style uh that's preserved
- 00:57:10across those different uh permutations
- 00:57:12of the study um but again it depends on
- 00:57:15what we're trying to capture the last
- 00:57:16study we're doing we want to find out
- 00:57:17like are people willing to pay for uh
- 00:57:20like artistic compensation so would they
- 00:57:21would they pay royalty to be able to to
- 00:57:23to compensate an artist if if the the
- 00:57:25artist generated using their style
- 00:57:27within a a text to image Ai and the
- 00:57:29answer to that question is is pretty
- 00:57:31strongly yes um and then we do some
- 00:57:33exploration around how much they would
- 00:57:34be willing to pay and that's also kind
- 00:57:35of
- 00:57:36interesting okay I'll do one last
- 00:57:39question and then if if you can stay
- 00:57:41along along around a little bit longer
- 00:57:43to kind of type in some answers to
- 00:57:45questions we have a lot more in here uh
- 00:57:47that would be helpful but we understand
- 00:57:49everybody has uh stuff coming up last
- 00:57:51question from our friend Leonard kale
- 00:57:53how do you get segments interpreted with
- 00:58:00AI so we did a we did a project in my
- 00:58:02class um a couple weeks ago on
- 00:58:04segmentation where I had the students uh
- 00:58:06you know I think we just applied K means
- 00:58:08within Chachi BT so it writes python
- 00:58:10code it does execution in browser
- 00:58:11generates segments and then we just uh
- 00:58:13we just asked chat GPT like characterize
- 00:58:16each of these segments and it goes into
- 00:58:18the verba and responses it goes into
- 00:58:19like the actual uh actual data um and
- 00:58:22does does a fairly fairly decent job um
- 00:58:25at at kind of a high level of describing
- 00:58:27who these individuals are and how they
- 00:58:28differ from one another um and so I
- 00:58:30guess maybe lender's question just just
- 00:58:31ask it um and do some experimentation
- 00:58:33around it and see how it
- 00:58:38performs go ahead see there's there's
- 00:58:41also been applications for people have
- 00:58:42asked it simply to build segments from
- 00:58:44qualitative data it's like here's a
- 00:58:46bunch of qualitative data give me three
- 00:58:48or four segments there's more Randomness
- 00:58:50in that in terms of if you were to like
- 00:58:52send that request a second time you
- 00:58:53could get different segments so just re
- 00:58:56nice you know there's it's not it's not
- 00:58:58you know it is there's some arbitrary
- 00:59:00arbitr and it's not
- 00:59:03magical okay with that I again if you
- 00:59:05can uh panelists kind of hang around to
- 00:59:07type in some answers that would be
- 00:59:09helpful uh Brian anything you want to
- 00:59:11say to wrap this up oh just just a huge
- 00:59:14Applause for these four individuals who
- 00:59:16spent time and and and this has been
- 00:59:18fascinating very much look forward to
- 00:59:21seeing you at uh the Ani Summit uh the
- 00:59:24the main core of the conference is going
- 00:59:26to be May 1st through 3rd and then the
- 00:59:29previous two days April 29th and 30th
- 00:59:32are going to be some optional workshops
- 00:59:33and tutorials that you can come to if if
- 00:59:36if you'd like to but thank you thank you
- 00:59:38so much this has been super
- 00:59:42interesting
- AI
- ricerca di marketing
- qualità dei dati
- analisi qualitativa
- modelli di linguaggio
- summit
- etica
- prompt engineering
- sistemi di raccolta dati
- innovazione