Latest / AI Top 5 / From Harry Potter Hacks to Agentic Shopping: A Busy Week in AI

Transcript

From Harry Potter Hacks to Agentic Shopping: A Busy Week in AI · AI Top 5

0:03
SPEAKER_01: Hi everyone, welcome to our fourth episode of the AI Top Five.
0:09
SPEAKER_01: I'm Chelsea.
0:11
SPEAKER_00: And I'm Eve.
0:12
SPEAKER_00: Welcome back.
0:13
SPEAKER_01: And if you haven't joined us for any of our other podcasts, we are big-time lovers of all things AI and tech, and we spend hours when we probably should be sleeping, scouring the internet, and bringing you guys the most important things that happen this week in the world of AI and why it matters to you.
0:32
SPEAKER_01: But just to get warmed up and loose, we like to do a series called Unprompted, where I just fire rapid fire questions at Eve that he has no idea are coming and he has to answer truthfully.
0:43
SPEAKER_01: So Eve, okay, let's go.
0:46
SPEAKER_00: You know what?
0:47
SPEAKER_00: This is actually a bit of a stressful moment in my week.
0:50
SPEAKER_00: So I'm I'm getting situated.
0:52
SPEAKER_00: All right, bring it.
0:53
SPEAKER_01: I love this for me.
0:55
SPEAKER_01: Um so we have to not obviously go any route of pop culture because Eve spends most of his time reading trade journals and not TMZ.
1:05
SPEAKER_01: It's fine.
1:07
SPEAKER_01: But the first question I have, and I was actually thinking about this randomly at three o'clock in the morning this morning.
1:13
SPEAKER_01: If our grandparents came back and saw AI, what do you think that they would say or do?
1:22
SPEAKER_00: Oh wow.
1:24
SPEAKER_00: Oh, that is a really, really good question.
1:27
SPEAKER_00: Um well look, I I don't know about grandparents, but I can tell you what my elderly parents are, how they're reacting to it.
1:35
SPEAKER_00: Um and it I think it's it's one of two ways.
1:39
SPEAKER_00: Uh way number one is complete acceptance and to the extent of over-reliance, I would say.
1:46
SPEAKER_00: And number two is full-on reject.
1:48
SPEAKER_00: There's not a lot of subtlety.
1:50
SPEAKER_00: I think, you know, I think people from that generation, if I go back on my point number one, probably come from a time where you were taught something, and what you saw written on the book was to be held as a gospel.
2:03
SPEAKER_00: It was true, and you know, you just accepted it.
2:06
SPEAKER_00: Um now I think now that you have the ability to ask questions um and get answers that sound extremely authoritative in a fraction of a second.
2:18
SPEAKER_00: I mean, I've seen it that like, oh look, AI says this, so it's gotta be true.
2:23
SPEAKER_00: Or look at this video of a I don't know, the dog dancing.
2:27
SPEAKER_00: It's like it's gotta be true.
2:29
SPEAKER_00: Like, no, that's not.
2:30
SPEAKER_00: Yeah.
2:31
SPEAKER_00: Um I I think you know, if if our grandparents would come back, number one, they would be absolutely floored by what's out there.
2:36
SPEAKER_00: That's number one.
2:37
SPEAKER_00: And I think you know, they would probably over-rely and not be able to distinguish and not they would probably not exercise the same sort of critical sense that we have in terms of what is true, what's not true.
2:48
SPEAKER_00: They may also, as I said, completely reject it.
2:50
SPEAKER_00: And you know, what's interesting, I was talking with a with a friend of mine who's a prof in the UK, and she's been doing some research on the acceptance on of AI amongst young people, and she was telling me how younger generation, the the the kids aged 12 to 16 now, actually have a full-on reject of AI.
3:15
SPEAKER_00: They will say, Oh my god, that's so AI, that's so bland, that's so boring.
3:19
SPEAKER_00: Yeah.
3:19
SPEAKER_00: And so it's interesting if you put that in conjunction with the question about the the grandparents.
3:25
SPEAKER_00: You you see that uh, you know, there's an interesting juxtaposition here to be to be had.
3:31
SPEAKER_01: 100%.
3:32
SPEAKER_01: Yeah, and it's wild to me.
3:34
SPEAKER_01: Anyone on social media has seen AI slot videos be out there of, you know, some dog going and saving its dog partner.
3:42
SPEAKER_01: And now there is one real video of this, but you know, a dog gets hit and this dog goes, you know, and dodges traffic and pulls it to the side.
3:49
SPEAKER_01: And it's wild to me how many people can't discern that's AI.
3:54
SPEAKER_01: And even with AI flags on content now on certain social platforms, people still are taking it as gospel.
4:00
SPEAKER_01: And it is a little terrifying.
4:01
SPEAKER_01: Speaking of gospel, my grandparents would probably say it's the market beast.
4:05
SPEAKER_01: Um they would love it and they would use it, but they would say, you know, obviously it's in times repent, just because, you know, we grew up in a holler in West Virginia.
4:15
SPEAKER_01: So anything like that would be that.
4:17
SPEAKER_01: But yeah, it is a little concerning how much people aren't using their discernment to really figure out what's AI and what's fact.
4:26
SPEAKER_01: And to say that, you know, this kind of goes into my next impromptu with you.
4:31
SPEAKER_01: Are we measuring the wrong things?
4:34
SPEAKER_01: So companies love talking about how many employees are using AI.
4:38
SPEAKER_01: They like putting benchmarks out there of, hey, you know, I want to see so many hours log of AI, blah, blah, blah, how many hours they've saved, how many prompts they've generated.
4:47
SPEAKER_01: But they're not necessarily translating those to like better business outcomes.
4:52
SPEAKER_01: So should we stop measuring AI adoption and actually start measuring AI's actual contribute contribution to revenue, profitability, and quality?
5:02
SPEAKER_00: Oof.
5:03
SPEAKER_00: That's a good question.
5:05
SPEAKER_01: Um flex than that journalism degree over here.
5:10
SPEAKER_00: So look, I think there's a you know, there is a general acceptance in the industry that um well point one, AI is massive.
5:23
SPEAKER_00: Super helpful.
5:24
SPEAKER_00: Point two, everybody's using it.
5:26
SPEAKER_00: Point three, in a lot of businesses, the the expected impact has not fully realized yet.
5:33
SPEAKER_00: And some people are asking why that's the case.
5:36
SPEAKER_00: Um and to your first point about are we measuring the wrong thing when it comes to using AI, I think that that's right.
5:44
SPEAKER_00: But I think it's also because it's it is it remains hard to link a business, it can be hard to link a business outcome with AI unless you frame it properly.
5:56
SPEAKER_00: As in, um, unless you start a new process with AI, you can design a bit of an experiment, you've got control groups, right?
6:05
SPEAKER_00: Um, it's hard to come down with a I can prove in my control experiment with a p-value of 0.05, which is you know less than chance, less than 5% that the that the phenomena I'm seeing is due to a random event, um that it's gonna be the case.
6:24
SPEAKER_00: So I think what we're seeing though is probably a lot of people using it for personal productivity, right?
6:33
SPEAKER_00: That's what you know.
6:34
SPEAKER_00: Um I want to summarize my emails, I'm gonna write better, I want to I don't want to read an article, blah blah blah blah, right?
6:41
SPEAKER_00: Those are all personal things.
6:42
SPEAKER_00: I think um using AI for broader team-based activities is still in its infancy, and I think that's one of the main causes why we're not seeing full value being realized just yet.
6:56
SPEAKER_01: Yeah.
6:57
SPEAKER_01: No, 100% that makes a lot of sense.
7:00
SPEAKER_01: And then finally, the last the last question as far as unprompted goes.
7:07
SPEAKER_01: What is the wildest thing you have done with AI to date?
7:12
SPEAKER_00: Oh, the wildest thing I have done today.
7:16
SPEAKER_00: Oh, um.
7:21
SPEAKER_00: Okay, I mean I've got a good one.
7:23
SPEAKER_00: Um, so my kids are just getting into Harry Potter.
7:27
SPEAKER_00: Um they love it.
7:29
SPEAKER_00: Um and um we have by the way, what I'm gonna share is maybe like I would call it maybe a gray area in terms of legality.
7:40
SPEAKER_00: So just bear with me for a second.
7:42
SPEAKER_00: But we have those.
7:46
SPEAKER_00: So it's it's okay.
7:47
SPEAKER_00: I think it's gonna be okay.
7:48
SPEAKER_00: Yeah, yeah.
7:49
SPEAKER_00: I'm not gonna be going into too many specifics, but we have those boxes that store audio.
7:54
SPEAKER_00: Like there's like they're not like not audiobooks, but like they're like basically glorified MP3 players for kids.
8:02
SPEAKER_00: Okay, cool.
8:04
SPEAKER_00: Obviously, it's a closed sourced system, so you've got a bit of a you've got a USB plug, and you can connect to the internet, download an app, and then you can download audiobooks onto that.
8:13
SPEAKER_00: But of course, every audiobook is very, very expensive, right?
8:16
SPEAKER_00: So I'm like, you know what?
8:18
SPEAKER_00: You know, are there people out there that would allow me to actually put my own content onto that platform?
8:26
SPEAKER_00: And there was, you know, so I found one, right?
8:28
SPEAKER_00: I'm like, all right, let's try this.
8:29
SPEAKER_00: Try to run the program, and it gave me the first time a ton of errors.
8:33
SPEAKER_00: Like, oh, this is the wrong library, you've got the because it was like a year old, right?
8:36
SPEAKER_00: People had not maintained it, so it's not the wrong library, you're outdated, blah, blah.
8:39
SPEAKER_00: And I'm like, you know what?
8:41
SPEAKER_00: I'm not gonna go into the source code and try to modify.
8:44
SPEAKER_00: I'm just not gonna, I don't have the patience for that.
8:46
SPEAKER_00: But then I was like, you know what, hold on a second.
8:48
SPEAKER_00: Let me just pull cursor and just say, all right, cursor, here's here's a problem.
8:53
SPEAKER_00: This is the error I'm getting.
8:54
SPEAKER_00: This is where the library is.
8:56
SPEAKER_00: What do you think?
8:57
SPEAKER_00: And it actually went in there and just changed the source code for me, recompiled it.
9:01
SPEAKER_00: And before you knew it, I could I was actually able to upload my own content into those audiobooks.
9:07
SPEAKER_00: So my own Harry Potter books into those audiobooks that I could have paid for them once already.
9:11
SPEAKER_00: I don't want to pay for them again.
9:12
SPEAKER_00: It's stupid.
9:13
SPEAKER_00: Um, I mean, 40 bucks to just put uh an audiobook that I already have onto that box.
9:20
SPEAKER_00: No.
9:20
SPEAKER_00: Um, so I was able to take that content and just transfer it onto that box.
9:25
SPEAKER_00: And, you know, honestly, I would never ever have been able to do that before.
9:31
SPEAKER_00: Uh I mean, I could have done it, probably done it by hand if I if you had given me a month of like tinkering on the side because it's just painful to do that.
9:41
SPEAKER_00: And it was literally done in 10 minutes.
9:44
SPEAKER_00: And to me, that just blew my mind that it was so efficient and so good.
9:48
SPEAKER_01: Yeah.
9:48
SPEAKER_01: That's really exciting.
9:50
SPEAKER_01: I love that.
9:51
SPEAKER_01: For those of you who are new to our podcasts, Eve scours the internet every week to figure out what is the biggest information you should know as far as AI goes.
9:59
SPEAKER_01: And I think that what we do best here is then take those nuggets and break it down and let you know what it means from a CPG perspective.
10:07
SPEAKER_01: The first one that popped out this week, Eve, is the Mistral Time Stroll, if I'm saying that right.
10:13
SPEAKER_01: The AI power casting.
10:15
SPEAKER_00: Mistral.
10:15
SPEAKER_01: Yes.
10:16
SPEAKER_00: It's French.
10:17
SPEAKER_01: Oh, I'm sorry.
10:17
SPEAKER_01: What is it?
10:18
SPEAKER_00: Mistral.
10:19
SPEAKER_00: It's the name of a wind in southern France.
10:23
SPEAKER_01: Mistral.
10:24
SPEAKER_00: Mistral.
10:25
SPEAKER_01: Okay.
10:25
SPEAKER_01: I'll let you say that one.
10:27
SPEAKER_01: So basically, that company introduced Times Times that a family of smaller specialized forecasting modules, right?
10:36
SPEAKER_01: Capable of predicting future outcomes, incorporating variables such as holidays and schedules, which is very exciting, and generating uncertainty ranges.
10:45
SPEAKER_01: Talk a little bit about why you think this is revolutionary, because in CPG, obviously, especially category management, we do a lot of forecasting.
10:53
SPEAKER_01: And AI to date has not really been that great as far as giving us, you know, a positive R square.
10:59
SPEAKER_01: So what do you think this could do from a revolutionary perspective for AI?
11:05
SPEAKER_00: Yeah, so I think you know, this model is I would call it the the newest entry, entrant in the broader domain of pre-trained forecasting model.
11:20
SPEAKER_00: So very much like a ChatGPT or whatnot is a pre-trained model for text.
11:28
SPEAKER_00: Those folks were like, well, if you can do it for text, you should be, if you can predict the next word in a sentence, you should be able to predict longer time series.
11:35
SPEAKER_00: Just makes sense.
11:36
SPEAKER_00: Um now, what it means though is you are basically taking a model and you're training that forecasting model on a huge wide range of data, right?
11:49
SPEAKER_00: Um now to be clear, it it it this means it is not trained on your data, it's trained on generic data.
11:56
SPEAKER_00: Um but where things get interesting is when you and you may think, well, then it's completely useless because it's just noise.
12:03
SPEAKER_00: Well, actually, what it's able to do is find out some patterns and see some regularities into that uh background data so that when you put in your own data and say, hey, look, I've just launched a new product, I've got I don't have a lot of history.
12:20
SPEAKER_00: Can you give me some sort of forecast for that?
12:23
SPEAKER_00: It will be able to leverage what it thinks could be relevant to then extend the time series.
12:28
SPEAKER_00: And just to be clear, that is there's a number of those um models that exist out there.
12:35
SPEAKER_00: Some of them are quite good, and just to be clear, they tend to be, you know, you tend to get some pretty good results for basically no effort, which is why they're interesting.
12:48
SPEAKER_00: But you will never get as good an output or as good a prediction as if you build your own forecasting model from scratch.
12:57
SPEAKER_00: There's a limit to physics, I would say.
12:59
SPEAKER_00: Um, but still, that model from Istal is now top of the class in terms of being able to run forecasting based on your data, but also on a large corpus of data that it's seen in the past.
13:14
SPEAKER_01: That's amazing.
13:15
SPEAKER_01: So, Eve, do you think that we will get to a point where a category manager or anyone in CPG, rather New Grove, whatever category they're in from a data perspective, can generate a reliable demand forecast without having a data science team behind them?
13:31
SPEAKER_00: Yeah, absolutely.
13:32
SPEAKER_00: I think, look, the the I think we're getting very much to the point now where you can actually interact with um, and I'm saying that because we're we're building one right now, but you can interact with a model and say, hey, look, please use here's a here's a bit here's some data, please forecast this quantity for me.
13:53
SPEAKER_00: Please forecast my sales, my revenue, my opex, my whatever, right?
13:56
SPEAKER_00: And then if you say, and I'm in this industry and I would like you to take those things into consideration, it's gonna come back and say, okay, well, because you're in this industry, these are the most typical covariates that I should look into.
14:07
SPEAKER_00: So if you're in the um the CPG industry, some of the interesting covariates could be uh inflation, could be uh some of the disposable income, could be um uh average income in this in a specific zip code, right?
14:24
SPEAKER_00: Those are all things that you could have.
14:26
SPEAKER_00: And it can decide by itself to incorporate those covariates into the model, which is kind of cool.
14:32
SPEAKER_00: Run the model, give you an output, and you can actually see, like, right, this is how the model did on the back testing, this is what we're forecasting, and then you can go back and say, Oh, okay, I like the model, but what if I change this?
14:44
SPEAKER_00: Rerun the model by adding a covariate or this or that, or you can go back and say, oh, you know, I see that there was a big error in my model back in you know March 2024.
15:00
SPEAKER_00: And then you can say, Oh, I know to you can say to the model, oh, I know why there was a big delta, why you got it wrong.
15:05
SPEAKER_00: It's because back in those days, we did X, and I'm gonna give you some data to be able to use that information.
15:12
SPEAKER_00: Uh, for example, we we launched a new product.
15:15
SPEAKER_00: Okay, cool.
15:15
SPEAKER_00: That explains where you see a bump, right?
15:18
SPEAKER_00: And then you can say, well, um, I also happen to have a schedule for my future launches over the next two years.
15:26
SPEAKER_00: And let's assume that my launches are about as successful as the previous ones.
15:31
SPEAKER_00: Use that to your for to use to do the forecast, and you should be able to get a much better outcome.
15:36
SPEAKER_00: And you don't, the beauty of that is you don't need to go and monkey up with like a ton of ton of config files and parameter tuning, a lot of that stuff can be abstracted away, which is very, very helpful.
15:47
SPEAKER_01: Very helpful.
15:48
SPEAKER_01: Speaking of helpful, my favorite recent topic, MetaMuse.
15:53
SPEAKER_01: And I'm at the point where personally I'm ready to call this a category management disruption.
15:58
SPEAKER_01: I don't know if you are kind of aligned with that.
16:00
SPEAKER_00: Oh, yeah.
16:01
SPEAKER_00: Absolutely.
16:01
SPEAKER_01: You know, it's it's definitely a product discovery disruption.
16:05
SPEAKER_01: And what happened this week is that Meta introduced new Muse models, introducing or including downloadable open weight technology and more advanced agent capabilities, which is pretty big, right?
16:17
SPEAKER_01: And really, we're talking about the disruption of product discovery.
16:21
SPEAKER_01: So this can really fundamentally change the path to purchase, which you and I have talked about a couple of times.
16:26
SPEAKER_01: Instead of consumers actually searching retailer websites or learning from word of mouth like traditional path to purchase, browsing digital shelves or comparing products themselves, AI agents are may increasingly recommend and select products for them.
16:42
SPEAKER_01: So talk a little bit about this, Eve, because it really raises an enormous question question, which is how do you win at a digital shelf when a human is not shopping it?
16:53
SPEAKER_00: Yeah, I mean that's that is probably the million-dollar question for the industry right now, right?
16:57
SPEAKER_00: Definitely.
16:58
SPEAKER_00: Um and so maybe if I take a bit of a step back.
17:01
SPEAKER_00: So, you know, Meta has launched Muse, that agent, that personal agent that does the shopping.
17:06
SPEAKER_00: And I think if you've watched the previous videos, you'll have heard that I've got some significant reservations about the use of that, especially if you know, because it's free and the amount of data and some of the behaviors that I've been exhibiting uh from that model.
17:24
SPEAKER_00: Um what's interesting in this week at least is that Meta has released uh another set of models that may power this Muse agent, but that can also power other things.
17:34
SPEAKER_00: So there are models that are equivalent to um Opus, to Sonnet, to GPT and whatnot, right?
17:42
SPEAKER_00: Or a Deep Seek, right?
17:43
SPEAKER_00: So they're quite generic, um, but they can be used to run agents.
17:47
SPEAKER_00: Now to the to the other question, you know, what does happen when agents do the shopping and the research for you?
17:57
SPEAKER_00: Well, I mean, I actually just yesterday I was talking with a with someone who is into in marketing, digital marketing, and what he was telling me is that um he's now built a whole business specifically for helping brands optimize their their reach in the world of agentic commerce, right?
18:18
SPEAKER_00: So that means very often, you know, you don't need necessarily to have all those fancy pictures, but very often it's a lot of FAQs, a lot of text for the model to be able to ingest, a lot of user reviews, fake or not, or real.
18:32
SPEAKER_00: You know, that there's always a bit of a debate there.
18:34
SPEAKER_00: Um, some you know, some recommendation from third parties, right?
18:38
SPEAKER_00: Those are the things that a model will go out and search and ingest to actually give the recommendation.
18:44
SPEAKER_00: Um but it is true that you know all of that stuff happens without you ever needing to do the research.
18:52
SPEAKER_00: Now, now, what is also true though is before you can tell Muse to go buy a specific item, very often you need to be aware of what the what is it the item that you want to buy.
19:03
SPEAKER_00: So I still think that that very early in the funnel, there still is an opportunity to say, oh, and by the way, these are this is a new type of boots that you may want to buy, and then you go may go to Muse and say, Yes, go buy me this boot, which which could be a different, slightly different use case from hey, my feet are wet, just buy me a pair of boots, the best pair of boots you can possibly find for that price, in which case it's gonna buy whatever, and it's gonna come back to you with probably options that you can choose from.
19:29
SPEAKER_00: Um, but then again, you know, if it gives you options for you to choose from, you will make you may need to do a bit of your own decisions and your own research to narrow down what's the final trigger that should that you should pull.
19:43
SPEAKER_00: So um massive changes.
19:46
SPEAKER_00: I still think though that you know awareness of what can be purchased is still there's still a lot of juice to be to be had there.
19:56
SPEAKER_01: Definitely.
19:57
SPEAKER_01: And I think this is going to be an ongoing conversation on this podcast.
20:01
SPEAKER_01: You know, we'll have put a dip a pen in it, it'll be one to watch and something that's definitely on our radar to see how it really shapes the marketplace because this is a true disruptor, like we've talked about.
20:12
SPEAKER_01: And I don't know that us in the category management field are comfortable yet with agentic shopping.
20:18
SPEAKER_01: So how do you get there?
20:20
SPEAKER_01: You know, I think that you and I are working on some information behind the scenes, some direction there.
20:25
SPEAKER_01: But, you know, it it didn't necessarily evolve overnight, but it came on quickly.
20:30
SPEAKER_01: And, you know, it's one of those things where some organizations are better at pivoting than others.
20:36
SPEAKER_01: And it's really time, I think, that our recommendation would be pivot and really try to figure out agentic shopping, you know, throw some research behind it and have start having those conversations about what does our digital shelf look like?
20:50
SPEAKER_01: Is it as optimized as it could be and really start throwing focus there?
20:54
SPEAKER_01: So I think that's really important for our constituency, Eve, to really focus on.
21:00
SPEAKER_00: I mean, look, it is not going away.
21:02
SPEAKER_01: No.
21:03
SPEAKER_00: Period end of story.
21:04
SPEAKER_00: This is not a battle that can be won.
21:07
SPEAKER_00: I mean, the the ship has sailed.
21:08
SPEAKER_00: So you just have to get on with it.
21:11
SPEAKER_01: Poor sharp.
21:12
SPEAKER_01: Next one, ramp glass.
21:14
SPEAKER_01: There's no French ramp à glass.
21:18
SPEAKER_00: Just wanted to check.
21:19
SPEAKER_00: It's not French.
21:20
SPEAKER_00: No.
21:21
SPEAKER_01: So this is really about organizational intelligence from what you spoke about.
21:26
SPEAKER_01: So ramp launch glass, it's an AI workspace designed to capture reusable prompts, workflows, and institutional knowledge while managing AI's spending and model selection, which is really exciting because we get into conversations about the expense behind AI, tokens, you know, burning, all that information that we've talked about a couple of times.
21:46
SPEAKER_01: So I think, you know, on top of getting your input about what this could mean, the question that I have for you is how do we make sure, you know, if we're going to implement AI to really learn some organizational Knowledge because right now, like a lot of organizations, category management knowledge lives within the person, not the organization.
22:07
SPEAKER_01: However, how do we make sure that we're scaling good thinking rather than tribal knowledge or bad habits with something like this?
22:17
SPEAKER_00: Wow, that's a very good question.
22:19
SPEAKER_00: Um what I would tell you is that it can be hard to distinguish one from the other.
22:24
SPEAKER_00: Definitely.
22:24
SPEAKER_00: Then very often the proof will be in the pudding.
22:28
SPEAKER_00: Good thinking tends to lead to good outcomes.
22:31
SPEAKER_00: Not always, but it tends to at a it's gonna have it's gonna have a higher likelihood to getting you to a good outcome.
22:36
SPEAKER_00: So I would probably start from like, you know, what have been the things that you've done or the decisions you've taken that have led to something that was successful, a successful launch, a successful promo, a successful promotion, and whatnot, and then work backwards from there.
22:49
SPEAKER_00: Um what is what is true though is that a lot of the tools today, like the GPTs, the clouds, you know, uh have been tailored and built for personal productivity, right?
23:04
SPEAKER_00: So they've been built so that I've got my own chat, I've got my space, and that's it, right?
23:10
SPEAKER_00: They've not been built for team collaborations.
23:15
SPEAKER_00: It is somewhat changing to some extent with some of the work that's being done on Slack and whatnot.
23:20
SPEAKER_00: But to a large extent, that is true.
23:23
SPEAKER_00: So you will build your skills, you will build your own things.
23:25
SPEAKER_00: And no, if you even go back um to some of what we've talked about, which was you know, like you can build your own software, your own self-well.
23:34
SPEAKER_00: Like I will build my own replica of OneNote, I will build my own replica of this or that, right?
23:41
SPEAKER_00: Um, what it really means is a if you look at it from a systemic perspective, it's a massively wasteful spend because you're all recreating the same thing at 50, 100 places in the organization, right?
23:54
SPEAKER_00: And so you could imagine that you know you will build some skills that are good for you to do your job, but a colleague of yours may have to redo the same thing slightly differently, and you will not benefit from the cross-learning here.
24:07
SPEAKER_00: And yes, maybe you'll send the skill to the other person, cool, but then you'll keep evolving yours, and that other person will not but get the benefit of the new versions or the other way around.
24:18
SPEAKER_00: And so what ramp is really aiming for is that you know, is to get all of the knowledge and the skills out of your own personal chat and put it into the broader platform, the broader, sorry, the broader enterprise platform so that those can be shared and you know the the you know you and the company can benefit from this better collaboration and this shared knowledge and shared skills, which is you know, I think a still a gap today in the market.
24:48
SPEAKER_01: Definitely.
24:50
SPEAKER_01: Definitely.
24:50
SPEAKER_01: Now, speaking of something that I don't necessarily know is a gap in the market, Google deep review, right?
24:56
SPEAKER_01: So AI that challenges your analysis.
24:58
SPEAKER_01: Now, in talking to you, it kind of feels like Google might be a little bit behind the eight on this because we do have a couple of platforms who have already launched deep research modules.
25:11
SPEAKER_01: Talk to me a little bit about what your thoughts are about Google Deep Review and then compare it a little bit to like a Claude Deep Research.
25:19
SPEAKER_01: You know, what is it doing differently?
25:22
SPEAKER_01: Who wins out if I need to have a really kind of critique on what I'm putting out before I go to a via presentation?
25:32
SPEAKER_00: Right.
25:33
SPEAKER_00: So look, a lot of the deep research agents out there were built on the premise that you what you want at the end is some sort of a white paper that should be well informed with lots of citations, you know, anywhere between five to ten pages long.
25:49
SPEAKER_00: And you know, you're gonna the deep and you expect the deep research to take whatever 10 minutes, whatever it is, right?
25:56
SPEAKER_00: But at the end of the day, it is a give me a topic and I will get you a report.
26:00
SPEAKER_00: That's what deep research is.
26:02
SPEAKER_00: Deep review is different.
26:04
SPEAKER_00: Deep review is the next step, which is okay, cool, you've done a you I have a report now, but I want to verify.
26:11
SPEAKER_00: That's cool.
26:12
SPEAKER_00: That's a deep review part of the process, right?
26:15
SPEAKER_00: And this review is way more difficult because even today, sometimes if you do a deep research, you will see some hallucinations, you will see that some of the citations are wrong, right?
26:26
SPEAKER_00: Don't exist, or misquoted, right?
26:29
SPEAKER_00: Or misinterpreted.
26:30
SPEAKER_00: So, what deep review is doing is feeling the next step in the evolution from I've got a question, I've got a report, now I'm gonna take that report and I'm gonna go back to the review stage and say, like, I let me peel the onion and check every single sentence, every single assertion.
26:52
SPEAKER_00: I will check the proof, I will make sure that everything is well structured.
26:56
SPEAKER_00: Um, will deep research eventually be merged into deep deep review be merged into deep research?
27:03
SPEAKER_00: Maybe uh it would make sense, but I think for um for a lot of um companies, especially companies that have a lot of technical documentation, right?
27:15
SPEAKER_00: Um that can be massively helpful.
27:18
SPEAKER_00: And I think that can be especially helpful if you is for companies that have um a changing product.
27:25
SPEAKER_00: So if you have, hey, I've got a new version of the product, oh, I've got a review and redo all of the documentation I've ever written, all of the user manuals, all of those things have to be rewritten from scratch.
27:36
SPEAKER_00: And then they also they also need to be checked, right?
27:39
SPEAKER_00: And so I think those are the sort of use cases where this new and it's it's brand, I mean it's quite recent.
27:46
SPEAKER_00: This deep review agent just came out last week.
27:48
SPEAKER_00: Yeah.
27:49
SPEAKER_00: Um so I I think there's a I think there's a real proper gap there that this could fill.
27:56
SPEAKER_01: Absolutely.
27:57
SPEAKER_01: In in your conversation just now, there's two things that kind of came to my mind.
28:01
SPEAKER_01: First off, the reminder, which I think is so important, is that even if AI sources something, do not have false confidence that that's true, right?
28:10
SPEAKER_01: Don't take that as Bible.
28:11
SPEAKER_01: Definitely go into the source, read it yourself, because there have been a couple of times to your point where there have been some pretty significant, significant hallucinizations on quotes.
28:24
SPEAKER_01: And it's like, how did you how exactly did you decipher this to mean that?
28:28
SPEAKER_01: But okay, so just constantly remember that you should be validating quotes, sources, anything that AI brings back to you.
28:37
SPEAKER_01: I think that's a good thing.
28:37
SPEAKER_00: And by the way, and there have been examples, public examples of I think it was either EY or KPMG or one of those companies, like a big, big company, right?
28:46
SPEAKER_01: Definitely not McKinsey.
28:47
SPEAKER_00: Putting up course not.
28:49
SPEAKER_00: Um but that that essentially put out a report on AI, and that report had been obviously, you know, made with AI, but some of the some of the citations were wrong.
29:02
SPEAKER_00: And there were there were some hallucinations in there, and there was some level of irony to put out an AI report that show that explicitly sh you know exhibited some of the flaws that you can have in using AI.
29:16
SPEAKER_00: So, you know.
29:17
SPEAKER_01: Definitely.
29:18
SPEAKER_01: And this actually brings another conversation that kind of just came to me.
29:23
SPEAKER_01: But as far, because in category management and CPG, we work so much, so much of our our workload is analytics, right?
29:32
SPEAKER_01: Of the main platforms, so call it anthropic, open AI, Gemini with Google, um, Deep Seek, Perplexity, you name it.
29:41
SPEAKER_01: I feel like I'm missing one of the big ones.
29:43
SPEAKER_01: Copilot.
29:44
SPEAKER_01: Who's really leading the race in analytics right now?
29:48
SPEAKER_00: In analytics specifically.
29:50
SPEAKER_01: Analytics specifically.
29:55
SPEAKER_00: I mean, look, I think the there are two main horses in the in the race right now.
29:59
SPEAKER_00: It's probably OpenAI and uh and anthropic.
30:03
SPEAKER_00: I would say probably right now the leaders.
30:05
SPEAKER_00: Now, to be fair, Jimmy and I just released a new model called Argon.
30:09
SPEAKER_00: That's um probably they released it a few weeks ago.
30:11
SPEAKER_00: It's now been it's in public access as of a few days ago now.
30:15
SPEAKER_00: So we don't fully exactly know.
30:18
SPEAKER_00: Now you still need to do a lot of testing to get some real-world evidence.
30:22
SPEAKER_00: Um, but in terms of pure raw analytics, um some of the examples that we saw with you know the the likes of OpenAI or GPT or GPT or Anthropic essentially solving mathematical puzzles that or puzzle, mathematical proofs that have that have been completely unsolved for like hundreds of years, and not just one or two, but if we're talking hundreds at a time, that is just absolutely mental.
30:52
SPEAKER_00: And you know, I also there's also like a new paradigm there, which is that a lot of the discoveries happen when you take a method from one place of one domain and then you apply it to another domain.
31:08
SPEAKER_00: So in my own work, uh back in my PhD thesis a long time ago, um, we used some um some techniques that were made for um genome sequencing, and you use them on time series, uh we use them for time series analysis because at the end of the day, the genome is just a series of stuff, and you can just use that.
31:32
SPEAKER_00: Um so that's just an example of like, oh, look, cool technique being used here, maybe we can apply it somewhere else and have the same results.
31:42
SPEAKER_00: And now what AI is uniquely suited for doing is actually to say, hey, look, there's been a breakthrough here.
31:50
SPEAKER_00: Let's look at all of the other potential areas in science where you're seeing a similar problem and to see if that breakthrough could potentially be helpful.
32:00
SPEAKER_00: And so that's one of the ways now where you know those new front frontier models, like the the the absolute latest, the Astoras or the Fables are you know essentially kicking ass.
32:14
SPEAKER_01: Yeah, for sure.
32:16
SPEAKER_01: For sure.
32:17
SPEAKER_01: Very expensive, but very expensive, yes.
32:20
SPEAKER_01: You're gonna burn some coin, but you're gonna get some good answers.
32:24
SPEAKER_01: So our last one really of our podcast is JP Morgan and the agenc payments.
32:31
SPEAKER_01: So AI.
32:32
SPEAKER_01: Obviously, we keep on talking about agentic AI shopping, right?
32:36
SPEAKER_01: So JP Morgan established restrictions on third-party AI agents that were conducting financial track transactions, highlighting the growing tensions between what AI agents can do and what they're authorized to do.
32:50
SPEAKER_01: This really is very interesting, but I want you to opine on it first before I ask you a question.
32:56
unknown: Okay.
32:57
SPEAKER_00: I mean, look, I think what you what you're seeing is, and and I think is this goes back to the question we talked about with regards to the agentic shopping, right?
33:04
SPEAKER_00: It's all of those companies do not want to be disambiguated, right?
33:09
SPEAKER_00: Right.
33:10
SPEAKER_00: Um, and so if you're a payment company, uh well, sure you will take a fee off the transaction.
33:17
SPEAKER_00: That's part of that's your core business model.
33:19
SPEAKER_00: But the other, the other core business model is getting and acquiring all of the data on the transaction, on the user, what was bought, right?
33:28
SPEAKER_00: That has probably as much, if not more, value than just the cut that they're taking off the top.
33:35
SPEAKER_00: Um, and in a world of agentic purchases, you may end up in a place where that link is broken, is is cut.
33:47
SPEAKER_00: Yeah.
33:47
SPEAKER_00: And if I'm a payment company, I do not like that one bit because I'm losing the ability to collect that information, right?
33:55
SPEAKER_00: And so I think that's why JP Morgan is currently saying, oh, we'll we don't allow it because of all the risks, which is true.
34:01
SPEAKER_00: There are absolutely risks, but I think beyond that, it's the hey, we are losing, we're being disintermediated, and we don't like that.
34:10
SPEAKER_01: Definitely.
34:11
SPEAKER_01: Yeah.
34:11
SPEAKER_01: And that was my first thought, too, is that when a consumer starts delegating their purchase decision, who actually wins that purchase decision, right?
34:22
SPEAKER_01: Is it the AI agent?
34:24
SPEAKER_01: Is it the retailer?
34:25
SPEAKER_01: Is it the manufacturer?
34:26
SPEAKER_01: Like it's going to get very messy as to like who can actually claim the purchase decision in that world.
34:34
SPEAKER_01: And I don't think we have a good answer for it yet, to be 100% honest.
34:38
SPEAKER_01: It still needs to flush itself out a little bit.
34:42
SPEAKER_01: But you know, how I think there's going to be a question of how do CPG brands protect their relevance when the consumer is delegating their purchase decision.
34:52
SPEAKER_00: Yeah.
34:52
SPEAKER_00: And look, I think, you know, when you when you have a a purchase decision, there's also some sort of liability in case things go wrong.
35:00
SPEAKER_00: So let me go back to my example, right?
35:03
SPEAKER_00: I won't, because it's been raining here, so uh, this is front of mind.
35:07
SPEAKER_00: Um my feet are still wet to some extent.
35:09
SPEAKER_00: Um I want to buy some boots.
35:12
SPEAKER_00: Okay, well, the agent is gonna go out and get me some boots.
35:15
SPEAKER_00: But what if I under specify what I want to buy?
35:18
SPEAKER_01: Right.
35:19
SPEAKER_00: And the wrong thing gets bought.
35:22
unknown: Right.
35:23
SPEAKER_00: Who's to blame?
35:23
SPEAKER_00: Is it me because I gave the wrong instruction?
35:25
SPEAKER_00: Is it the agent because he did not ask me?
35:28
SPEAKER_00: It's like, hey, look, your question is kind of stupid.
35:31
SPEAKER_00: You should give me more information.
35:33
SPEAKER_01: Right.
35:33
SPEAKER_00: Right?
35:34
SPEAKER_00: Or is it because the manufacturer did not give enough information to the agent to make the right decision?
35:42
SPEAKER_00: Yeah, maybe all of the above.
35:44
SPEAKER_00: Um, but at the end of the day, no, money will still be taken out of your pocket and will not probably not be refunded.
35:51
SPEAKER_01: So definitely.
35:53
SPEAKER_01: Definitely.
35:54
SPEAKER_01: Well, I hope you all enjoyed our five questions.
35:57
SPEAKER_01: We do have a viewer write in.
36:00
SPEAKER_01: I'm super excited that these are coming in.
36:02
SPEAKER_01: So, Eve, this is for us.
36:04
SPEAKER_01: And this is a category manager from a Midwest CPG manufacturer.
36:09
SPEAKER_01: Thank you out there.
36:10
SPEAKER_01: I have been in category management for almost 15 years, and it feels like every week there's a new AI tool promising to automate another part of my job.
36:19
SPEAKER_01: While I'm excited about technology, I'm also wondering what skills I should be actually developing to stay relevant.
36:26
SPEAKER_01: If you were a category manager today, where would you focus your learning on in the next year?
36:33
SPEAKER_01: Ooh, that's a really great question.
36:35
SPEAKER_00: That is a very, very good question.
36:37
SPEAKER_00: Yeah.
36:37
SPEAKER_00: Um I mean, look, the reality is things move so fast that you cannot, it's very hard to keep up with the industry.
36:46
SPEAKER_00: So it's that's just a reality.
36:48
SPEAKER_00: So the best thing you can do is um look at your current workflows and try to you know, try to try to try to build some automation yourself.
37:01
SPEAKER_00: Try to build, try to use an try to create an agent to do this or that.
37:05
SPEAKER_00: Um and it may work, it may not work, you know, you may or may not have the tools available to do that.
37:12
SPEAKER_00: And if you don't have the tools available within your company, then you should do it outside of the company's wall, not with the company's data, with synthetic data, obviously.
37:21
SPEAKER_00: Never, never, ever, ever put company data outside of the of those walls.
37:27
SPEAKER_00: Um, but I think uh that's the the only way to to at least keep on discovering is to is to try things.
37:37
SPEAKER_00: And you've got to be intentional about it, right?
37:38
SPEAKER_00: And that means it's not a oh wow, we've got a deadline tomorrow.
37:42
SPEAKER_00: Let me try something, because that is guaranteed to fail, right?
37:44
SPEAKER_01: Must as roll the dice.
37:46
SPEAKER_00: That's right.
37:46
SPEAKER_00: Uh what it what it should be is literally be intentional, as in, well, guess what?
37:50
SPEAKER_00: I am going to book two hours every week.
37:52
SPEAKER_00: I'm going to take every Friday from three to five, whatever it might be, to try to read on AI, to try to use AI in my day-to-day job, right?
38:03
SPEAKER_00: That's the only way for you to do it.
38:04
SPEAKER_00: But again, it's got to be intentional.
38:06
SPEAKER_00: It cannot just be, you know, ad hoc.
38:08
SPEAKER_00: Otherwise, it's not going to work.
38:11
SPEAKER_01: Absolutely.
38:11
SPEAKER_01: And I think for me, my first recommendation, which will also include a couple of shameless plugs that are shameless, make sure you have the foundations right.
38:20
SPEAKER_01: Make sure you know the anatomy of a good prompt.
38:22
SPEAKER_01: Make sure you know how to discern a hallucination from fact, right?
38:27
SPEAKER_01: Make sure that you know how to mask your data, you know.
38:32
SPEAKER_01: So if you're if you are tied to using Copilot, have Copilot mask your data and then you can use a different platform.
38:38
SPEAKER_01: I didn't say that out loud, but I just did.
38:40
SPEAKER_01: Uh, but I think, you know, the other thing is that if you are a member of the CMA, there are some amazing resources out there that'll help you learn from an AI perspective.
38:50
SPEAKER_01: Eve himself has done what, three now?
38:53
SPEAKER_00: Yeah, three or four trainings.
38:55
SPEAKER_01: And I think there's a fourth in the hopper, if I'm not mistaken.
38:57
SPEAKER_01: If not, I just told you.
38:59
SPEAKER_01: Uh, but three really great, amazing trainings on the functionality and understanding AI, what's an LLM, you know, all of these kind of buzzwords that people throw out, how do neural networks work?
39:10
SPEAKER_01: So, like all of that information are is in those trainings, which is fantastic.
39:14
SPEAKER_01: On top of that, I've cut 13 series recently of, you know, how to make how to use like application for AI to your work.
39:23
SPEAKER_01: So prompts that you can use today to better understand how to sweep the internet for reviews or how to do some advanced analytical work.
39:31
SPEAKER_01: So there are resources out there for you from a CMA perspective.
39:35
SPEAKER_01: Definitely a great reason to be a member and join us at conference where we'll be talking about a lot of these things.
39:41
SPEAKER_01: But I think, you know, you should be, I understand there's trepidation.
39:46
SPEAKER_01: We all do, right?
39:47
SPEAKER_01: Um, and you have the right attitude from a perspective of how do I learn AI?
39:51
SPEAKER_01: That is the 100% right attitude to have.
39:54
SPEAKER_01: Do not fight against the grain on this one.
39:56
SPEAKER_01: It you'll be left behind, right?
39:58
SPEAKER_01: So definitely sharpening your skills.
40:00
SPEAKER_01: Continue to listen to our podcasts where we bring you the most updated information.
40:05
SPEAKER_01: But there are resources out there that can assist you.
40:08
SPEAKER_01: You don't have to do it alone.
40:09
SPEAKER_01: So I hope that that does help.
40:11
SPEAKER_01: But Eve, is there anything you want to add as a last hurrah on this one?
40:16
SPEAKER_00: No, I think we've covered quite a bit of ground already.
40:19
SPEAKER_00: So it's been a busy week in the world of AI.
40:21
SPEAKER_00: So yes.
40:22
SPEAKER_01: So super excited to see what happens in another two weeks in the world of AI.
40:27
SPEAKER_01: For now, we're signing out, Eve and Chelsea out, and we will see you in two weeks on the flip side.
40:33
SPEAKER_00: Thank you.
40:34
SPEAKER_00: Bye-bye.