May 9, 2024 • 25 mins
AI in supply chains isn't just about crunching data—it's about creating bridges between complex info and everyday decision-making.
In this episode, Lasse Holmstedt, VP of Engineering at Alloy.ai, joins us to discuss the dynamic role of generative AI and large language models within the consumer goods sector. Throughout the conversation, we delve not only into the innovative application of AI technologies that encompass forecasting and data harmonization but also highlight the critical importance of maintaining clean and accurate data for AI effectiveness.
Listen in for a deeper understanding of the potential and pitfalls of using pre-trained AI models. Along the way we’ll also explore strategies to safeguard sensitive business data and discover how AI can transform sales and demand planning processes.
Three Key Takeaways:
- Explore leveraging textual and qualitative data in addition to numerical approaches to enhance decision-making processes while mitigating risk
- Consider applications of AI for enhanced visibility and predictive capabilities in inventory management, forecasting, and cross-platform product matching for more efficient operations
- Establish a clear data governance framework to ensure that data fed into AI systems is accurate, up-to-date, and free from errors, laying a solid foundation for reliable AI outputs
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Speaker 1 (00:02):
From Alloy AI.
This is Shelf Life.
Is generative AI really a gamechanger for the industry, or is
(00:23):
it mostly hype?
What are the strengths and oris it mostly hype?
What are the strengths andshortcomings of AI platforms?
How risky is it to share yourcompany's data with an AI
platform?
On every episode of Shelf Life,we answer questions like these
and more, with the help ofleaders across the consumer
goods industry.
Today, our guest is LasseHalmsted, vp of Engineering at
(00:43):
Alloy AI.
Lasse has been with Alloy forseven years, during which time
he's been the primary architectof our data platform and
application.
Recently, he also led the workof integrating generative AI
into the Alloy product for thefirst time.
Previous to joining Alloy,lasse worked for Nokia, where he
focused on developing consumermobile apps.
(01:04):
I'm your host, abbey Carruthers,product manager at Alloy AI.
We'll be back with Lhasa rightafter this.
As a consumer brand, youconnect with dozens of external
partners and internal systems toget a complete picture of your
business.
Each one is different.
Alloy AI makes it easy toconnect data from retailers,
(01:30):
e-commerce, supply chainpartners and even your own ERP,
then easily surface insights todrive sales growth.
Every day, brands use Alloy AIto see POS trends, measure
promotion performance and makebetter replenishment decisions
with their retail partners.
That's why we're trusted by BIC, crayola, valvoline, melissa
Doug Bosch and many more.
Get a demo at alloyai today,lasse, welcome to Shelf Life.
(01:51):
Lasse, can you get us startedby telling us a little bit about
your background and the workyou've been doing with AI?
Speaker 2 (01:57):
Of course, and thank you so much for having me.
I think you covered a lot ofthe intro right there.
I've been a VP of Engineeringnow at Alloy for the last three
years.
Been of the intro right there.
I've been a VP of engineeringnow at Alloy for the last three
years, been in the softwareindustry overall for about 15.
And then, yeah, at Nokia.
What I worked on was being alead for one for the first
Android Maps project, where weshipped that to I don't know
like 30 million people orthereabouts.
(02:18):
That was really exciting.
But also shipping severalmobile phones and smartphones
back when Nokia still did them.
It was really exciting at thattime.
But I met the founders of Alloyaround 2016, very soon since
the company was founded.
And then, yeah, the rest ishistory.
My guys, I've been at Alloyever since.
(02:42):
A little bit on the ways that weuse AI today at Alloy, I think
requires a little bit ofexpansion of what we do at the
high level at Alloy, to beginwith.
So I think, fundamentally Iwould say we're a platform,
we're a data platform for thesupply chain with the vision of
really connecting it end-to-end,all the way from manufacturing
to the POS, e-commerce and tothe consumer, and so that
(03:05):
requires you to look at thingslike forecasting, external
signals and the like on the POSend to determine what's going to
happen next, but alsounderstanding how do you get the
data to be clean to begin with,whether it's from manufacturing
or from the POS side, from yourERP, and kind of making sense
of all of that.
So there's a lot of use casesfor advanced algorithms, machine
(03:28):
learning, deep learning and thelike, really across the whole
spectrum, and so the use casesthat we then cover and the
audience really focuses a lot onthe sales themes and kind of
revenue operations, salesoperations and also demand
planning teams, to name a few ofthose examples of things that
we focus on and who our currentusers are.
(03:50):
But so to expand a little biton the details there on how
Alloy works, it's roughly likethis we're essentially ingesting
and harmonizing data fromhundreds of different unique
sources, really globally.
There's a heavy focus on NorthAmerica from our business
perspective.
A lot of our clients are there,but we roll out Alloy globally
(04:18):
for some of our customers.
Yeah, so what we do isessentially looking at pulling
the data from hundreds ofdifferent integrations and as we
do that, we have to make senseof both the POS side, erp side
of things and then really buildall of that into a singular data
model together.
This is no small feat ofengineering.
We've worked on this for 70years now and there's a lot of
need for again, yeah, like deeplearning models and the like A
(04:41):
lot of, let's say,non-generative AI use cases that
we have leveraged now forseveral years already.
Then, on the user of, let's say, non-generative AI use cases
that we have leveraged now forseveral years already.
Then, on the user end, asyou're, let's say, a sales
leader and looking at actionableinsights into what should you
do about all of this data, we'reoffering metrics like phantom
inventory, blast sales and, ofcourse, sell-through forecasts
that we generate ourselves tomake the data actionable.
(05:04):
So then, going back to thequestion of how we use AI,
there's already a lot of thosedifferent use cases in there and
it's really, I think, fair tosay that we use it pervasively
across the platform.
So, for example, as we ingestdata in from these different POS
sources, we have to do productmatching, for, let's say,
walmart data is different fromthe target data, from the Amazon
(05:26):
data and so forth.
They keep products in differentways.
There's the Amazon ASINs,there's the Walmart UPCs, target
DPCIs, right.
And then multiply that by like800, right?
So you need to have veryflexible algorithms to do the
matching across that wholespectrum with full automation
(05:46):
and with confidence that itactually works.
And so that's a great example,I think, of a deep learning
algorithm that we use at Alloy,where we do this sort of cross
and transitive matching ofproducts across all of these
retailers, looking at thebrand's taxonomy, looking at all
of the retailer taxonomies, andthen iterating through the
layers and layers of additionalinsights, transforming the data
(06:09):
as we go and then taking thatdata in and matching those
products from the retailers tothe brands and vice versa, and
then, ultimately, the use casethat then we are able to serve
there for the end user is givingthem an omni-channel visibility
into their entire sales orinventory across all of the
(06:30):
channels that they sell through.
That's a lot to take in rightthere, but so that's just again
one of those use cases.
Speaker 1 (06:39):
Fantastic, so you touched on a lot of different
examples of AI there I'mparticularly interested in.
I think you briefly mentionedgenerative AI, so I know there's
been a lot of hype in theindustry around generative AI in
particular.
Can we just take a step backand do you mind explaining to
our listeners what exactly wemean by generative AI and how it
works?
Speaker 2 (06:57):
Yeah, there's definitely been a lot of buzz
around the AI in the last oneyear or so, and so it's
interesting because thesetechnologies have existed
already for some time, so thelarge language models
particularly but they havereally kicked off and made it
big this year.
There's been incredible amountof progress.
There's this one quote I reallylike from Arthur C Clarke that
goes just about like this so anysufficiently advanced
(07:19):
technology is indistinguishablefrom magic, and I feel a little
bit is that how?
That is how media uses the wordAI, because the results of some
of these technologies are so,so incredible, particularly with
these large language models andgenerative AI.
So Jack GPT in particulargetting a lot of press I think
deservedly so and like I thinkit's in a lot of people's minds
(07:42):
like how exactly this works.
But so I can try to take a stabat explaining it a little bit
and maybe take some of the magicout, because, at the heart,
generative AI is really a set oftechnologies that well, plainly
generates stuff.
It sounds kind of cheap, butthat is ultimately what it is,
(08:02):
and that stuff could be text,which is the case, of course,
for Google, BART, gpp and thelike, or it could be images,
like for DALI, for example, ormusic 3D models, it could be
PowerPoint presentations, itcould be really anything, but
that's roughly what that termmeans at a very high level,
right, and the underpinningtechnology then, behind the
(08:27):
generative AI as a term, isneural networks.
And so, to explain neuralnetworks really briefly, they
are essentially probabilisticalgorithms that take inputs like
so the text that you write as auser perhaps it's a question
like what is my last sales forproduct A and then they produce
(08:48):
outputs that they are trained toproduce.
Again, it could be text, images, music, whatever, and the
training then is used.
It's this one-time process thatis used to, in effect, ease out
the algorithm and ease outthose, let's say, the
probabilities that the algorithmgenerates based on the training
data.
So a very simple example of aneural network would be, for
(09:11):
example, to detect if there arecats in an image.
And this is not like superfar-fetched, because if you
think about captures, the kindof like a little bit irritating
human checks up all over theinternet.
Often you are asked as a humanto check if there's cats, if
there's traffic lights, there'smotorcycles and the like in
these pictures.
That's all part of neuralnetwork training, this data that
(09:34):
you use in CAPTCHAs is actuallyused to train image recognition
algorithms, but so a simpleneural network could be like
this Give me an answer yes or noif there are cats in this
picture and then the algorithmwill take a number of inputs,
like every single pixel in animage, and then the output is
just a yes or no.
(09:54):
It's a Boolean, like a truth ina statement, in a word, and so a
large language model or an LLMis not really any different, so
they are really just trainedwith a huge amount of content
online.
The model is massively morecomplex than this cat image
recognizer, and the training isthen biased towards assisting
(10:15):
the user by really doing onething by predicting the next
possible word, sorry, bypredicting the most likely next
word or a token, perhaps insequence, and they use the
context that the user thenprovided, meaning the text that
you wrote as a user, and perhapsadditional context that the
(10:37):
model has, let's say, from theavailable platform, anyways, and
so, to give an example aboutthe supply chain, you might have
the model trained to answerquestions on inventory levels
out of stock sales opportunitiesand the like.
But so that's kind of taking theexample from those I don't know
cat image recognitionalgorithms and using the neural
(11:00):
networks there into largelanguage models that are
massively more complicated andvery powerful and can generate
really accurate human-readabletext, right.
But so to kind of go back andstress on a couple of the points
earlier that are really key forunderstanding how these models
work, all that these languagemodels really do is not magic,
(11:21):
but rather predicting the nextword in the sequence as they
generate those, let's say,paragraphs, sentences or entire
documents that the user mightask them to create, and doing so
in a way that it assists theuser, and so that might be a
little bit eye-opening andremove some of the magic around
this technology.
(11:41):
And so there's no question thatI think the tech is incredibly
powerful, but it helpsunderstand, I think, some of the
issues also around the techthat we might want to discover a
little bit more as well.
Speaker 1 (11:52):
Yeah, absolutely.
It still does sound a littlebit like magic to me, but I can
recognize that of course theremust be limitations with any new
technology.
So, yeah, do you mind talkingto us a little bit around?
What are those kind ofshortcomings of these LLM
platforms?
Speaker 2 (12:06):
Yeah, in the context of our domain so the supply
chain in particular we have anincredible amount of data that
we need to work with, and so, inorder for our users at Ally to
make the best possible decisions, we need that data to be clean
and accurate, and so this is abig focus for us internally, and
we spent already a lot of timebuilding algorithms and models
that really allow us to providethat data in raw format and also
(12:30):
in these predictive more, let'ssay, insightful, predictive
versions that help the userunderstand what's going to
happen next, and so, as a partof understanding the
technologies or the tech aroundLLMs, we put the large language
models to the test as well, andso what we discovered and I
think we're not the only oneshere in terms of tech
organizations is there's a lotof complexities with the LLMs
(12:56):
when you start to deal withnumbers, and so they are not
good today in doing math andthey are not really built for
that.
The kind of the name largelanguage models are just.
They are really good withlanguage and this is a hundred
percent true.
They are incredible atproducing English or Japanese or
Finnish or, like the languageyou named, like a human language
(13:18):
is just as easy for them as aprogramming language, but asking
them to do math, the technologyis not there yet, and the kind
of problems that you might runinto are simply, let's say,
asking for the largest possiblenumber, so what were the largest
lost sales instances at myregion?
(13:39):
It's not going to work out toowell Sometimes.
It's just going to pick thelongest number, and the longest
number might be the one with themost decimal points.
This is the kind of problemsthat we have seen, and so, while
there are ways to work aroundthis with technology, it's not
something that you necessarilywant to rely on when you're
making business criticaldecisions.
There are just bettertechnologies for that type of
(14:00):
jobs.
So what I want to say is thisis not a silver bullet.
The other thing and this is, Ithink, more pervasive for LLMs
with any use case is theirtendency to hallucinate, making
things up, and so what thatmeans is you're asking a
question, let's use the samewhat were my last sales in
(14:21):
California for product A, againas an example, and we will not
give the LLM the language model,the data about sales or lost
sales or anything.
It should not be able to answerthis question.
It should say I don't know.
But given the rightcircumstances, occasionally what
you will see is that ChatGPT orBARD or LLM2 or any number of
(14:44):
these off-the-shelf LLMs theywill just make answers up and
that's not too great.
And so again the same thing.
Looking at that, 100% accuracywhen making business critical
decisions is just superimportant, and the problem there
doesn't go away with pushingmore data to the language model
(15:04):
either.
In fact, in some cases it mightmake things worse.
So I think the rule of the thumbhere that at least we have
discovered so far is when you doneed to do whether it's math or
other, when you need to dodecisions within your platform
or within your business thatrequire 100% accuracy or close
to 100% accuracy.
(15:25):
The LLM is not necessarily theright tool for that.
Now can it help Absolutely, andthere's a lot of ways in which
they are really really powerful,but it's not the foundation
that you can build, let's say,mathematical algorithms on, and
it's also really reallyexpensive to run, just
computationally and thereforealso just in terms of pure
(15:46):
dollars.
Speaker 1 (15:48):
So the fact that these LLMs aren't good at math
seems like it would limit howthey're applied to supply chain
problems in particular.
I'm understanding that you'resaying you have to make sure
you're using these LLMs in theright way, right, Because
sometimes you might get ananswer but it might not be the
right answer.
So is that fair that thislimits how you can apply these
technologies in the supply chainspace specifically, or are
(16:10):
there creative ways that we canuse them?
Speaker 2 (16:12):
There's absolutely creative ways that we can make
the most use of those, and so,if you think about those couple
of key facets of these largelanguage models that I was
highlighting earlier, so lookingat the aspect of generating
text and assisting the user, themodels have truly been trained
(16:32):
with that assistance aspect inmind.
So, of course, the number onething then kind of goes that
makes sense is, yeah, reallylooking at, well, building an
assistant of some sort where youget to speak in English or
again, like in Finnish orJapanese or Polish or whatever
language you speak, and then thelanguage model has the
(16:52):
conversation with the platformthat you work with.
Let's take, yeah, like Alloy foran example, so you can just go
ahead and ask the question andthis is true today and this is
how we integrated OpenAI intoAlloy you can go ahead and ask
what were my last sales forproduct CH005, being that your
SKU ID, and you'll get theanswer in a few seconds.
(17:15):
You could build a dashboardmanually for this and that's
probably what the power userwould do.
They probably have a whole lotof other questions that they
know exactly how to get theanswers to within a platform
like Alloy, but for the user whojust dropped in to the platform
.
It's going to be a lot harderfor them to go about their
business that way readingdocumentation and all of that
(17:37):
than just going ahead and askingthe question in plain English,
and so I think there's realpower to that in democratizing
technology and so making it sothat, ultimately, the data that
you make decisions on is trulythe asset that is democratized,
and the whole team, not only thepeople who are really tech
savvy get to have those answersthat they're looking for as
(18:00):
they're doing their work,whether it's business analysts
or sales leaders.
Speaker 1 (18:05):
Yeah, absolutely.
I think that's particularlyexciting from a software
perspective, as you say, helpingnew users or less technical
teams really have access to thatdata without having to
understand how to configure allthese complicated insights in
some data analysis tool.
Being able to just ask thosequestions in plain English, I
think, as you say, really reallyopens us up to more users being
(18:27):
able to access these datainsights.
And so what about data security?
I know that's a big concernthat we hear about a lot of the
time, particularly in CPGs.
A lot of companies you knowclosely guard their data.
It's kind of I guess it'sconcerning or worrying right to
think about just handing overlarge sets of these data to
these platforms and not knowinghow it's going to be used.
(18:48):
So how can our listeners thinkabout that when it comes to data
security and also leveragingthese new technologies?
Speaker 2 (18:56):
This is really interesting because there's been
a lot of development, even inthe last few weeks, really, so
in the beginning of the year.
So JetGPT, the one gettingreally the most press here,
didn't have a suck to audit done.
They essentially said,implicitly or explicitly, that
they are using all of the datafor training, and that put your
(19:16):
data ultimately at real risk.
Because imagine the situationwhere, if a third party like
Chachipiti uses their data, usesyour questions to train their
model, the next time that themodel is trained and that's an
intense process, so it doesn'thappen all the time but the next
time it is trained it's goingto be learning things about your
(19:37):
business that only very fewpeople should actually have
access to, and that's just amassive business risk.
Your competitors in the worstcase, they could be just going
ahead and asking those questionsabout like, for example, now
has this enterprise scheme wherethey say they will not use the
(20:08):
data anywhere for training andonly for compliance reasons.
They keep the data on theirservers.
I think this is a fantasticstep forward, but we take it a
little bit further than that,and our implementation is not to
actually send any of thenumerical data under any
circumstances to any of theseoff-the-shelf models that we
don't control.
And so we kind of kill twobirds with one stone, if you
(20:33):
will, where we don't have toworry about the accuracy
problems we do the math and wekeep the math on our side but we
also don't have to then worryabout and our customers don't
have to worry about thecompliance, legality problems
and potentially exposing theinformation to competitors,
maybe some months or a yearlater when this retraining does
happen.
But so there's definitely a lotof risks there.
(20:57):
When you, let's say, as a CIO,are looking at AI strategy and
evaluating the options ofwhether you build or buy.
If you are building, youprobably don't have that sort of
staffing to go ahead andretrain the models yourself.
This costs tens of millions ofdollars.
So JetGPT costs about $100million to train the latest
(21:19):
version, gpt-4.
Few companies will have thebudget to do something of that
sort.
So the chances are higher thatyou would be leveraging some of
these off-the-shelf models andso you have to be really careful
about all of the legalese,different jurisdictions doing
different things and potentiallypersisting your data and maybe
using it.
It's easier to kind of skip onthat part altogether.
(21:42):
And yeah, look at software,look at platforms that say we
are never submitting sensitivedata to these platforms.
Just the easier conversation tohave.
Speaker 1 (21:52):
So, more broadly, how should companies be thinking
about leveraging not justgenerative AI, but AI in general
?
We think about strategy and howcompanies are making sure that
they're capitalizing on allthese opportunities.
What would you recommend,particularly for CPG companies
and companies in the supplychain space?
How can they be thinking aboutbuilding this into their
business strategy?
Speaker 2 (22:13):
Yeah, I think it's really important to have a
holistic perspective on yeah,essentially AI as a whole,
machine learning and the keyenabler technologies that then
ultimately enable the ITdepartments to deliver as much
value as possible for thosestakeholder teams, because, if
you look at LLMs, for example,they have only made it big this
year and, while the pace ofprogress is really intense,
(22:37):
there's going to be a lot morerobust, a lot more performant
models in 2024, 2025 than whatwe have today, and so there's a
lot more mature technologies outthere that I think would be
beneficial to look at seriouslyand not just at the LLM side of
the spectrum.
So, looking at what I saidearlier about the deep learning
(22:58):
and machine learning and kind oflike, let's say, forecasting
algorithms, I think one of thekey enablers really for sales
teams is accurate forecasts, andaccurate forecasts is a great
way of then looking at whetherit's more accurate shipment
plans or just more actionablesales outcomes, but it all
starts with clean data, and soin order for you to produce
(23:21):
accurate forecasts whether it'swith Ally or whether it's with,
let's say, sap, ibp or whetherit's some other technology all
of those different forecastingalgorithms and platforms they
rely ultimately on the data tobe there and for the data to be
clean, and so what our focus hasreally been is very heavily on
making the data as clean aspossible, because that is the
(23:44):
core enabler for all of thefuture AI steps that are going
to be taken in the next yearsand as the LLM's progress, I'm
sure that the math problems anda lot of the performance
problems that we see today costissues.
They will eventually disappear,but they need to integrate with
your data warehouse and yoursingle source of truth for your,
(24:06):
whether it's demand data, foryour selling data.
What have you?
Speaker 1 (24:11):
And I think that the AIs will be limited by the data
cleanliness at the end of theday, so I think what you're
talking about is really thecompanies need to be I've heard
the term AI ready right, even ifthey're not necessarily
building solutions around thesetechnologies.
Today, there's a lot to be donein terms of making sure that
the data and your dataarchitecture and the way you
(24:31):
store and model that data isready to be able to make use of
these systems in the future.
Speaker 2 (24:37):
Absolutely.
I think.
Still today, many of thecompanies are still looking at
moving into a single source oftruth data warehousing solution,
looking at really commoditizingand democratizing the data as a
part of the decision makingprocess.
We see that we help ourcustomers with that, but there's
a lot of companies out therethat are not doing this in the
(24:57):
same sort of way and notfollowing suit.
Not following suit andespecially with these large
language models, generative AIas that technology develops
further, the companies whoalready have that clean data
will be able to take a betteradvantage of the market.
I think that's just the facts.
We are not there today, notwith all pieces of the LLM
technologies, but we'll getthere soon enough, and so next
(25:20):
year the year after will be very, very interesting indeed.
Speaker 1 (25:23):
Absolutely Really excited to see how things
develop in the coming weeks andmonths.
That's all we have time fortoday.
Lasse, Thank you so much forjoining us here and thank you
for all the insight.
Speaker 2 (25:32):
Thank you for having me.
Speaker 1 (25:34):
You've been listening to Lasse Holmsted, VP of
Engineering at Alloy AI.
That's all for this week.
See you next time on Shelf Life.
From Alloy AI.
This is Shelf Life.
Is generative AI really a gamechanger for the industry, or is
(00:23):
it mostly hype?
What are the strengths and oris it mostly hype?
What are the strengths andshortcomings of AI platforms?
How risky is it to share yourcompany's data with an AI
platform?
On every episode of Shelf Life,we answer questions like these
and more, with the help ofleaders across the consumer
goods industry.
Today, our guest is LasseHalmsted, vp of Engineering at
(00:43):
Alloy AI.
Lasse has been with Alloy forseven years, during which time
he's been the primary architectof our data platform and
application.
Recently, he also led the workof integrating generative AI
into the Alloy product for thefirst time.
Previous to joining Alloy,lasse worked for Nokia, where he
focused on developing consumermobile apps.
(01:04):
I'm your host, abbey Carruthers,product manager at Alloy AI.
We'll be back with Lhasa rightafter this.
As a consumer brand, youconnect with dozens of external
partners and internal systems toget a complete picture of your
business.
Each one is different.
Alloy AI makes it easy toconnect data from retailers,
(01:30):
e-commerce, supply chainpartners and even your own ERP,
then easily surface insights todrive sales growth.
Every day, brands use Alloy AIto see POS trends, measure
promotion performance and makebetter replenishment decisions
with their retail partners.
That's why we're trusted by BIC, crayola, valvoline, melissa
Doug Bosch and many more.
Get a demo at alloyai today,lasse, welcome to Shelf Life.
(01:51):
Lasse, can you get us startedby telling us a little bit about
your background and the workyou've been doing with AI?
Speaker 2 (01:57):
Of course, and thank you so much for having me.
I think you covered a lot ofthe intro right there.
I've been a VP of Engineeringnow at Alloy for the last three
years.
Been of the intro right there.
I've been a VP of engineeringnow at Alloy for the last three
years, been in the softwareindustry overall for about 15.
And then, yeah, at Nokia.
What I worked on was being alead for one for the first
Android Maps project, where weshipped that to I don't know
like 30 million people orthereabouts.
(02:18):
That was really exciting.
But also shipping severalmobile phones and smartphones
back when Nokia still did them.
It was really exciting at thattime.
But I met the founders of Alloyaround 2016, very soon since
the company was founded.
And then, yeah, the rest ishistory.
My guys, I've been at Alloyever since.
(02:42):
A little bit on the ways that weuse AI today at Alloy, I think
requires a little bit ofexpansion of what we do at the
high level at Alloy, to beginwith.
So I think, fundamentally Iwould say we're a platform,
we're a data platform for thesupply chain with the vision of
really connecting it end-to-end,all the way from manufacturing
to the POS, e-commerce and tothe consumer, and so that
(03:05):
requires you to look at thingslike forecasting, external
signals and the like on the POSend to determine what's going to
happen next, but alsounderstanding how do you get the
data to be clean to begin with,whether it's from manufacturing
or from the POS side, from yourERP, and kind of making sense
of all of that.
So there's a lot of use casesfor advanced algorithms, machine
(03:28):
learning, deep learning and thelike, really across the whole
spectrum, and so the use casesthat we then cover and the
audience really focuses a lot onthe sales themes and kind of
revenue operations, salesoperations and also demand
planning teams, to name a few ofthose examples of things that
we focus on and who our currentusers are.
(03:50):
But so to expand a little biton the details there on how
Alloy works, it's roughly likethis we're essentially ingesting
and harmonizing data fromhundreds of different unique
sources, really globally.
There's a heavy focus on NorthAmerica from our business
perspective.
A lot of our clients are there,but we roll out Alloy globally
(04:18):
for some of our customers.
Yeah, so what we do isessentially looking at pulling
the data from hundreds ofdifferent integrations and as we
do that, we have to make senseof both the POS side, erp side
of things and then really buildall of that into a singular data
model together.
This is no small feat ofengineering.
We've worked on this for 70years now and there's a lot of
need for again, yeah, like deeplearning models and the like A
(04:41):
lot of, let's say,non-generative AI use cases that
we have leveraged now forseveral years already.
Then, on the user of, let's say, non-generative AI use cases
that we have leveraged now forseveral years already.
Then, on the user end, asyou're, let's say, a sales
leader and looking at actionableinsights into what should you
do about all of this data, we'reoffering metrics like phantom
inventory, blast sales and, ofcourse, sell-through forecasts
that we generate ourselves tomake the data actionable.
(05:04):
So then, going back to thequestion of how we use AI,
there's already a lot of thosedifferent use cases in there and
it's really, I think, fair tosay that we use it pervasively
across the platform.
So, for example, as we ingestdata in from these different POS
sources, we have to do productmatching, for, let's say,
walmart data is different fromthe target data, from the Amazon
(05:26):
data and so forth.
They keep products in differentways.
There's the Amazon ASINs,there's the Walmart UPCs, target
DPCIs, right.
And then multiply that by like800, right?
So you need to have veryflexible algorithms to do the
matching across that wholespectrum with full automation
(05:46):
and with confidence that itactually works.
And so that's a great example,I think, of a deep learning
algorithm that we use at Alloy,where we do this sort of cross
and transitive matching ofproducts across all of these
retailers, looking at thebrand's taxonomy, looking at all
of the retailer taxonomies, andthen iterating through the
layers and layers of additionalinsights, transforming the data
(06:09):
as we go and then taking thatdata in and matching those
products from the retailers tothe brands and vice versa, and
then, ultimately, the use casethat then we are able to serve
there for the end user is givingthem an omni-channel visibility
into their entire sales orinventory across all of the
(06:30):
channels that they sell through.
That's a lot to take in rightthere, but so that's just again
one of those use cases.
Speaker 1 (06:39):
Fantastic, so you touched on a lot of different
examples of AI there I'mparticularly interested in.
I think you briefly mentionedgenerative AI, so I know there's
been a lot of hype in theindustry around generative AI in
particular.
Can we just take a step backand do you mind explaining to
our listeners what exactly wemean by generative AI and how it
works?
Speaker 2 (06:57):
Yeah, there's definitely been a lot of buzz
around the AI in the last oneyear or so, and so it's
interesting because thesetechnologies have existed
already for some time, so thelarge language models
particularly but they havereally kicked off and made it
big this year.
There's been incredible amountof progress.
There's this one quote I reallylike from Arthur C Clarke that
goes just about like this so anysufficiently advanced
(07:19):
technology is indistinguishablefrom magic, and I feel a little
bit is that how?
That is how media uses the wordAI, because the results of some
of these technologies are so,so incredible, particularly with
these large language models andgenerative AI.
So Jack GPT in particulargetting a lot of press I think
deservedly so and like I thinkit's in a lot of people's minds
(07:42):
like how exactly this works.
But so I can try to take a stabat explaining it a little bit
and maybe take some of the magicout, because, at the heart,
generative AI is really a set oftechnologies that well, plainly
generates stuff.
It sounds kind of cheap, butthat is ultimately what it is,
(08:02):
and that stuff could be text,which is the case, of course,
for Google, BART, gpp and thelike, or it could be images,
like for DALI, for example, ormusic 3D models, it could be
PowerPoint presentations, itcould be really anything, but
that's roughly what that termmeans at a very high level,
right, and the underpinningtechnology then, behind the
(08:27):
generative AI as a term, isneural networks.
And so, to explain neuralnetworks really briefly, they
are essentially probabilisticalgorithms that take inputs like
so the text that you write as auser perhaps it's a question
like what is my last sales forproduct A and then they produce
(08:48):
outputs that they are trained toproduce.
Again, it could be text, images, music, whatever, and the
training then is used.
It's this one-time process thatis used to, in effect, ease out
the algorithm and ease outthose, let's say, the
probabilities that the algorithmgenerates based on the training
data.
So a very simple example of aneural network would be, for
(09:11):
example, to detect if there arecats in an image.
And this is not like superfar-fetched, because if you
think about captures, the kindof like a little bit irritating
human checks up all over theinternet.
Often you are asked as a humanto check if there's cats, if
there's traffic lights, there'smotorcycles and the like in
these pictures.
That's all part of neuralnetwork training, this data that
(09:34):
you use in CAPTCHAs is actuallyused to train image recognition
algorithms, but so a simpleneural network could be like
this Give me an answer yes or noif there are cats in this
picture and then the algorithmwill take a number of inputs,
like every single pixel in animage, and then the output is
just a yes or no.
(09:54):
It's a Boolean, like a truth ina statement, in a word, and so a
large language model or an LLMis not really any different, so
they are really just trainedwith a huge amount of content
online.
The model is massively morecomplex than this cat image
recognizer, and the training isthen biased towards assisting
(10:15):
the user by really doing onething by predicting the next
possible word, sorry, bypredicting the most likely next
word or a token, perhaps insequence, and they use the
context that the user thenprovided, meaning the text that
you wrote as a user, and perhapsadditional context that the
(10:37):
model has, let's say, from theavailable platform, anyways, and
so, to give an example aboutthe supply chain, you might have
the model trained to answerquestions on inventory levels
out of stock sales opportunitiesand the like.
But so that's kind of taking theexample from those I don't know
cat image recognitionalgorithms and using the neural
(11:00):
networks there into largelanguage models that are
massively more complicated andvery powerful and can generate
really accurate human-readabletext, right.
But so to kind of go back andstress on a couple of the points
earlier that are really key forunderstanding how these models
work, all that these languagemodels really do is not magic,
(11:21):
but rather predicting the nextword in the sequence as they
generate those, let's say,paragraphs, sentences or entire
documents that the user mightask them to create, and doing so
in a way that it assists theuser, and so that might be a
little bit eye-opening andremove some of the magic around
this technology.
(11:41):
And so there's no question thatI think the tech is incredibly
powerful, but it helpsunderstand, I think, some of the
issues also around the techthat we might want to discover a
little bit more as well.
Speaker 1 (11:52):
Yeah, absolutely.
It still does sound a littlebit like magic to me, but I can
recognize that of course theremust be limitations with any new
technology.
So, yeah, do you mind talkingto us a little bit around?
What are those kind ofshortcomings of these LLM
platforms?
Speaker 2 (12:06):
Yeah, in the context of our domain so the supply
chain in particular we have anincredible amount of data that
we need to work with, and so, inorder for our users at Ally to
make the best possible decisions, we need that data to be clean
and accurate, and so this is abig focus for us internally, and
we spent already a lot of timebuilding algorithms and models
that really allow us to providethat data in raw format and also
(12:30):
in these predictive more, let'ssay, insightful, predictive
versions that help the userunderstand what's going to
happen next, and so, as a partof understanding the
technologies or the tech aroundLLMs, we put the large language
models to the test as well, andso what we discovered and I
think we're not the only oneshere in terms of tech
organizations is there's a lotof complexities with the LLMs
(12:56):
when you start to deal withnumbers, and so they are not
good today in doing math andthey are not really built for
that.
The kind of the name largelanguage models are just.
They are really good withlanguage and this is a hundred
percent true.
They are incredible atproducing English or Japanese or
Finnish or, like the languageyou named, like a human language
(13:18):
is just as easy for them as aprogramming language, but asking
them to do math, the technologyis not there yet, and the kind
of problems that you might runinto are simply, let's say,
asking for the largest possiblenumber, so what were the largest
lost sales instances at myregion?
(13:39):
It's not going to work out toowell Sometimes.
It's just going to pick thelongest number, and the longest
number might be the one with themost decimal points.
This is the kind of problemsthat we have seen, and so, while
there are ways to work aroundthis with technology, it's not
something that you necessarilywant to rely on when you're
making business criticaldecisions.
There are just bettertechnologies for that type of
(14:00):
jobs.
So what I want to say is thisis not a silver bullet.
The other thing and this is, Ithink, more pervasive for LLMs
with any use case is theirtendency to hallucinate, making
things up, and so what thatmeans is you're asking a
question, let's use the samewhat were my last sales in
(14:21):
California for product A, againas an example, and we will not
give the LLM the language model,the data about sales or lost
sales or anything.
It should not be able to answerthis question.
It should say I don't know.
But given the rightcircumstances, occasionally what
you will see is that ChatGPT orBARD or LLM2 or any number of
(14:44):
these off-the-shelf LLMs theywill just make answers up and
that's not too great.
And so again the same thing.
Looking at that, 100% accuracywhen making business critical
decisions is just superimportant, and the problem there
doesn't go away with pushingmore data to the language model
(15:04):
either.
In fact, in some cases it mightmake things worse.
So I think the rule of the thumbhere that at least we have
discovered so far is when you doneed to do whether it's math or
other, when you need to dodecisions within your platform
or within your business thatrequire 100% accuracy or close
to 100% accuracy.
(15:25):
The LLM is not necessarily theright tool for that.
Now can it help Absolutely, andthere's a lot of ways in which
they are really really powerful,but it's not the foundation
that you can build, let's say,mathematical algorithms on, and
it's also really reallyexpensive to run, just
computationally and thereforealso just in terms of pure
(15:46):
dollars.
Speaker 1 (15:48):
So the fact that these LLMs aren't good at math
seems like it would limit howthey're applied to supply chain
problems in particular.
I'm understanding that you'resaying you have to make sure
you're using these LLMs in theright way, right, Because
sometimes you might get ananswer but it might not be the
right answer.
So is that fair that thislimits how you can apply these
technologies in the supply chainspace specifically, or are
(16:10):
there creative ways that we canuse them?
Speaker 2 (16:12):
There's absolutely creative ways that we can make
the most use of those, and so,if you think about those couple
of key facets of these largelanguage models that I was
highlighting earlier, so lookingat the aspect of generating
text and assisting the user, themodels have truly been trained
(16:32):
with that assistance aspect inmind.
So, of course, the number onething then kind of goes that
makes sense is, yeah, reallylooking at, well, building an
assistant of some sort where youget to speak in English or
again, like in Finnish orJapanese or Polish or whatever
language you speak, and then thelanguage model has the
(16:52):
conversation with the platformthat you work with.
Let's take, yeah, like Alloy foran example, so you can just go
ahead and ask the question andthis is true today and this is
how we integrated OpenAI intoAlloy you can go ahead and ask
what were my last sales forproduct CH005, being that your
SKU ID, and you'll get theanswer in a few seconds.
(17:15):
You could build a dashboardmanually for this and that's
probably what the power userwould do.
They probably have a whole lotof other questions that they
know exactly how to get theanswers to within a platform
like Alloy, but for the user whojust dropped in to the platform
.
It's going to be a lot harderfor them to go about their
business that way readingdocumentation and all of that
(17:37):
than just going ahead and askingthe question in plain English,
and so I think there's realpower to that in democratizing
technology and so making it sothat, ultimately, the data that
you make decisions on is trulythe asset that is democratized,
and the whole team, not only thepeople who are really tech
savvy get to have those answersthat they're looking for as
(18:00):
they're doing their work,whether it's business analysts
or sales leaders.
Speaker 1 (18:05):
Yeah, absolutely.
I think that's particularlyexciting from a software
perspective, as you say, helpingnew users or less technical
teams really have access to thatdata without having to
understand how to configure allthese complicated insights in
some data analysis tool.
Being able to just ask thosequestions in plain English, I
think, as you say, really reallyopens us up to more users being
(18:27):
able to access these datainsights.
And so what about data security?
I know that's a big concernthat we hear about a lot of the
time, particularly in CPGs.
A lot of companies you knowclosely guard their data.
It's kind of I guess it'sconcerning or worrying right to
think about just handing overlarge sets of these data to
these platforms and not knowinghow it's going to be used.
(18:48):
So how can our listeners thinkabout that when it comes to data
security and also leveragingthese new technologies?
Speaker 2 (18:56):
This is really interesting because there's been
a lot of development, even inthe last few weeks, really, so
in the beginning of the year.
So JetGPT, the one gettingreally the most press here,
didn't have a suck to audit done.
They essentially said,implicitly or explicitly, that
they are using all of the datafor training, and that put your
(19:16):
data ultimately at real risk.
Because imagine the situationwhere, if a third party like
Chachipiti uses their data, usesyour questions to train their
model, the next time that themodel is trained and that's an
intense process, so it doesn'thappen all the time but the next
time it is trained it's goingto be learning things about your
(19:37):
business that only very fewpeople should actually have
access to, and that's just amassive business risk.
Your competitors in the worstcase, they could be just going
ahead and asking those questionsabout like, for example, now
has this enterprise scheme wherethey say they will not use the
(20:08):
data anywhere for training andonly for compliance reasons.
They keep the data on theirservers.
I think this is a fantasticstep forward, but we take it a
little bit further than that,and our implementation is not to
actually send any of thenumerical data under any
circumstances to any of theseoff-the-shelf models that we
don't control.
And so we kind of kill twobirds with one stone, if you
(20:33):
will, where we don't have toworry about the accuracy
problems we do the math and wekeep the math on our side but we
also don't have to then worryabout and our customers don't
have to worry about thecompliance, legality problems
and potentially exposing theinformation to competitors,
maybe some months or a yearlater when this retraining does
happen.
But so there's definitely a lotof risks there.
(20:57):
When you, let's say, as a CIO,are looking at AI strategy and
evaluating the options ofwhether you build or buy.
If you are building, youprobably don't have that sort of
staffing to go ahead andretrain the models yourself.
This costs tens of millions ofdollars.
So JetGPT costs about $100million to train the latest
(21:19):
version, gpt-4.
Few companies will have thebudget to do something of that
sort.
So the chances are higher thatyou would be leveraging some of
these off-the-shelf models andso you have to be really careful
about all of the legalese,different jurisdictions doing
different things and potentiallypersisting your data and maybe
using it.
It's easier to kind of skip onthat part altogether.
(21:42):
And yeah, look at software,look at platforms that say we
are never submitting sensitivedata to these platforms.
Just the easier conversation tohave.
Speaker 1 (21:52):
So, more broadly, how should companies be thinking
about leveraging not justgenerative AI, but AI in general
?
We think about strategy and howcompanies are making sure that
they're capitalizing on allthese opportunities.
What would you recommend,particularly for CPG companies
and companies in the supplychain space?
How can they be thinking aboutbuilding this into their
business strategy?
Speaker 2 (22:13):
Yeah, I think it's really important to have a
holistic perspective on yeah,essentially AI as a whole,
machine learning and the keyenabler technologies that then
ultimately enable the ITdepartments to deliver as much
value as possible for thosestakeholder teams, because, if
you look at LLMs, for example,they have only made it big this
year and, while the pace ofprogress is really intense,
(22:37):
there's going to be a lot morerobust, a lot more performant
models in 2024, 2025 than whatwe have today, and so there's a
lot more mature technologies outthere that I think would be
beneficial to look at seriouslyand not just at the LLM side of
the spectrum.
So, looking at what I saidearlier about the deep learning
(22:58):
and machine learning and kind oflike, let's say, forecasting
algorithms, I think one of thekey enablers really for sales
teams is accurate forecasts, andaccurate forecasts is a great
way of then looking at whetherit's more accurate shipment
plans or just more actionablesales outcomes, but it all
starts with clean data, and soin order for you to produce
(23:21):
accurate forecasts whether it'swith Ally or whether it's with,
let's say, sap, ibp or whetherit's some other technology all
of those different forecastingalgorithms and platforms they
rely ultimately on the data tobe there and for the data to be
clean, and so what our focus hasreally been is very heavily on
making the data as clean aspossible, because that is the
(23:44):
core enabler for all of thefuture AI steps that are going
to be taken in the next yearsand as the LLM's progress, I'm
sure that the math problems anda lot of the performance
problems that we see today costissues.
They will eventually disappear,but they need to integrate with
your data warehouse and yoursingle source of truth for your,
(24:06):
whether it's demand data, foryour selling data.
What have you?
Speaker 1 (24:11):
And I think that the AIs will be limited by the data
cleanliness at the end of theday, so I think what you're
talking about is really thecompanies need to be I've heard
the term AI ready right, even ifthey're not necessarily
building solutions around thesetechnologies.
Today, there's a lot to be donein terms of making sure that
the data and your dataarchitecture and the way you
(24:31):
store and model that data isready to be able to make use of
these systems in the future.
Speaker 2 (24:37):
Absolutely.
I think.
Still today, many of thecompanies are still looking at
moving into a single source oftruth data warehousing solution,
looking at really commoditizingand democratizing the data as a
part of the decision makingprocess.
We see that we help ourcustomers with that, but there's
a lot of companies out therethat are not doing this in the
(24:57):
same sort of way and notfollowing suit.
Not following suit andespecially with these large
language models, generative AIas that technology develops
further, the companies whoalready have that clean data
will be able to take a betteradvantage of the market.
I think that's just the facts.
We are not there today, notwith all pieces of the LLM
technologies, but we'll getthere soon enough, and so next
(25:20):
year the year after will be very, very interesting indeed.
Speaker 1 (25:23):
Absolutely Really excited to see how things
develop in the coming weeks andmonths.
That's all we have time fortoday.
Lasse, Thank you so much forjoining us here and thank you
for all the insight.
Speaker 2 (25:32):
Thank you for having me.
Speaker 1 (25:34):
You've been listening to Lasse Holmsted, VP of
Engineering at Alloy AI.
That's all for this week.
See you next time on Shelf Life.
Popular Podcasts
Las Culturistas with Matt Rogers and Bowen Yang
Ding dong! Join your culture consultants, Matt Rogers and Bowen Yang, on an unforgettable journey into the beating heart of CULTURE. Alongside sizzling special guests, they GET INTO the hottest pop-culture moments of the day and the formative cultural experiences that turned them into Culturistas. Produced by the Big Money Players Network and iHeartRadio.
On Purpose with Jay Shetty
I’m Jay Shetty host of On Purpose the worlds #1 Mental Health podcast and I’m so grateful you found us. I started this podcast 5 years ago to invite you into conversations and workshops that are designed to help make you happier, healthier and more healed. I believe that when you (yes you) feel seen, heard and understood you’re able to deal with relationship struggles, work challenges and life’s ups and downs with more ease and grace. I interview experts, celebrities, thought leaders and athletes so that we can grow our mindset, build better habits and uncover a side of them we’ve never seen before. New episodes every Monday and Friday. Your support means the world to me and I don’t take it for granted — click the follow button and leave a review to help us spread the love with On Purpose. I can’t wait for you to listen to your first or 500th episode!
Dateline NBC
Current and classic episodes, featuring compelling true-crime mysteries, powerful documentaries and in-depth investigations. Follow now to get the latest episodes of Dateline NBC completely free, or subscribe to Dateline Premium for ad-free listening and exclusive bonus content: DatelinePremium.com