May 7, 2025 • 31 mins
AI isn't magical but mathematical, and understanding how it works removes mystery while helping make informed business decisions. We break down the fundamentals of AI in a way that demystifies its inner workings and explains why it's transformative for businesses.
• Understanding how AI processes a simple request like "write a haiku about puppies" reveals its potential for business applications
• AI converts language into mathematical vectors to find relationships between words and concepts
• Context windows limit how much an AI can "remember" in a conversation, similar to human short-term memory
• Retrieval Augmented Generation (RAG) allows businesses to prioritize their own data when generating AI responses
• AI agents can perceive environments, make decisions, and take actions toward specific goals like searching and organizing data
• New business models like "service as software" are emerging, where AI handles operations behind traditional services
• Four principles for effective AI: treat it like a human, include it in brainstorming, expect continuous improvement, and keep humans in oversight roles
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Speaker 1 (00:01):
You're listening to Risk and Resolve, and now for
your hosts, ben Conner and ToddHufford.
Well, that made me sound reallygeeky, but I'm wearing a suit,
so I can't be that geeky, right?
So I really did want to wearjeans and a t-shirt, but you
know, it is what it is.
So here's what we're going todo today.
(00:22):
We are going to talk about howAI works.
I want to break it down becausehere's what we're going to do
today.
We are going to talk about howAI works.
I want to break it down becausehere's what I believe.
Let me first make sure thismoves to me.
There's me.
Why AI?
All right.
So what are we talking abouthere?
Why do we care about AI?
What is it that's actuallyhappening?
What is it that's going onunder the scenes, under the
covers?
It's not magic, so let'sunderstand that.
So, anyone like math?
(00:43):
There's not a lot of handsgoing up, all right.
There's some hands going up,all right.
We're going to get into some ofthe math.
Now, I'm not going to get toomathy, but here's what I think.
I think that if we understandhow things work, it takes the
mystery out of it and we canstart to make informed decisions
about it.
So that's where I want to behere today.
That's my goal is.
I want to take some of themystery away from it.
So we're not going to get intoall the calculus and all the
(01:05):
details, but I'm going to tryand explain this in such a way
that you can understand thatthis isn't magical, that there
are some decisions we can makeand how this all works.
So we're going to talk a littlebit about puppy poetry.
We're then going to jump intosomething about.
We're going to talk about thetools that exist.
We want to make sure weunderstand all those different
tools that exist inside of theAI stack, specifically about how
(01:28):
you would leverage that insideof a business.
Sound good?
All right, let's jump in.
Why all the hype?
All right, who has done this?
This is in Claude.
Who has gone in and said writeme a haiku about puppies.
It's kind of the starting pointof AI, right, like?
You come in and you're like hey, write me a haiku about puppies
.
It's kind of the starting pointof AI, right, like you come in
(01:49):
and you're like hey, give mesomething to like, show me that
it can do it.
And it did.
And we're like woohoo, right,we're all excited, we're
thrilled, we're like this isbusiness revolutionary, right?
Well, it is, and from acomputer science standpoint, I
want to break down why and Iwant to talk through that
standpoint.
So, first of all, what's goingon here?
(02:11):
This word haiku, what does itmean?
Well, we know what it means,right, we know that it's a
poetry, we know that it's threelines, we know that each line
has a certain number ofsyllables with it.
So we understand that there aresome rules around what that
word means.
But what's really cool here isthe computer understood that,
(02:36):
and that is something worthtaking note of.
The second thing is I said, hey,here are these set of rules
called a haiku, but I want youto do something.
I want you to apply those rulesto a goal, to this target
called puppies.
So I'm saying, hey, I want youto take this set of rules three
lines, syllables are hard.
(02:56):
I want you to figure that outand I want you to goal seek
based upon the outcome, therules and constraints in the
plan.
I want you to put together aplan and I want you then to
execute on that plan towardsthis outcome.
That's pretty cool.
Now, it's cool to see it do itfor something as simple as a
haiku right, that's neat.
But imagine if this said I wantyou to onboard this new
(03:21):
employee, or I want you tounderwrite this mortgage
information.
That's why we care, that's whywe get excited about it.
That's what's going on here isbecause we actually are at a
point where we're able to usenatural language to capture the
information that we have, tocapture the rules about how a
business works, and then sayhere's the goal we want you to
accomplish and it can work itsway towards it.
(03:42):
And there's special ways andthere's a way that it does that.
And it's pretty logical whenyou start to understand it.
But that's what's going on here.
That's why people get excitedabout it.
So when you see people go hey,look, it wrote, you know, look
at what it did for my email.
Or look at what it did for youknow it could write me a poem,
that's the neat part.
We're tracking.
Okay, all right.
(04:04):
So let's talk about some of thetools.
We'll talk a little bit abouttask automation, because right
now, ai and automation they'rekind of being talked about as
the same thing.
So we'll talk a little bitabout RPA.
I didn't put on their API stuffbecause that gets all super
geeky, but that's really most ofwhat's happening here.
If you don't know what APImeans, it stands for application
(04:28):
programming interface and it'skind of like the back door into
a lot of systems.
It's really kind of the way thesystems work.
But we'll talk a little bitabout RPA.
We'll then talk about whatknowledge workers are going on.
What is embedding,vectorization, llms, context
windows you might have heard alot of this stuff in the
marketing stuff that's out onthe news.
So let's talk about what theyare.
What's it really mean, how doesit actually apply to you and
why do you care?
And then we'll talk about whatlearning looks like.
And finally we'll jump intothis new thing called AI agents,
(04:52):
which kind of gets into some ofthe managerial side.
Let's start.
All right, task automation.
Anyone heard of CNRPA before?
All right, robotic processautomation.
I literally wrote this code.
You can't see it, which is good, because it's not beautiful
code, but I'm going to hit goand what you're going to watch
is this is going to log into awebsite and it's going to upload
(05:14):
like 50 orders.
Okay, notice, this is going fast.
That's because the computer'sdoing it.
That's because the computer'sdoing it completely without a
person touching it.
So that's called roboticprocess automation.
It's a technology that is kindof the foundation to a lot of
computer testing.
So software companies will comein and they're like hey, we
don't want to pay peoplemanually to go click through a
(05:34):
bunch of buttons, so we wrotecode to teach a computer how to
go do that for us.
And then all of a sudden werealize, hey, you know what
that's actually kind of usefuland there are some applications
that we can't get into throughdatabases or APIs.
Maybe we can go through thefront end the same way a human
does, and so that's what roboticprocess automation is.
You've maybe heard of UiPath orBlue Prism Automate, anywhere.
(06:00):
Those are some of the newtechnology, not new, they're old
now, it's like a whole threeyears ago.
So that's some of thetechnology that's been around
that's doing a lot of that.
But this is getting used a lotwith AI and you'll see here at
the end on the agent side wherethey're starting to apply some
of that.
But I want you to know that it'sout there.
And a lot of people talk aboutautomation and this is one of
the examples.
I did mention APIs, applicationprogramming, interfaces.
(06:20):
Frankly, that's how mostautomation is done, but it's not
as fun to show off.
So you know, showing code runisn't really that cool.
So, all right, let's move intothe fun stuff.
So AI is not magic, it's math.
So we're going to talk about itNow.
Bear with me, at first thisfeels heavy and you're like what
in the world are you doing?
You're like talking crazy mathstuff with me.
(06:42):
It makes sense.
It's not that complicated.
What we want to do is and wetalked about it earlier we want
to convert these sentences intoways that we can understand them
.
All right, this is what'scalled embeddings.
So what AI did and kind of thebig breakthrough, is it said hey
, let's figure out a way tounderstand how a sentence is put
together so that we can compareit to other sentences.
(07:05):
So what we're going to do iswe're going to take every word
in the sentence in this caselet's take the word Italian and
we're going to look at all thewords that are on the left and
then we're going to weight them,basically put mathematical
numbers around the relationshipbetween all those words, so that
we have some indication of howthose words relate to each other
.
Then we're gonna take all thewords on the right, and we're
(07:25):
gonna weight those and we'regonna put numbers around them.
Okay, finally, we end up withwhat ends up being a really long
list of numbers.
So, yes, I went all geeky.
This is what's called a vector.
Remember physics class?
Right?
Vectors you had to calculateall that stuff.
Well, that's basically what wedid here, and so, while this is,
(07:45):
I'm showing you kind of asimple version, this is done on
parameters and billions andbillions of different weightings
in the way that it all works.
So you get a really, really,really complicated list of
numbers.
But for the purpose of this,let's just worry about two.
Let's just put them in atwo-dimensional array here,
right?
So a normal look.
So let's take our sentencelasagna is Italian food.
(08:06):
And now what we're going to dois we're going to go and we're
going to load that into ourlarge language model.
So we're going to first embedit, okay, so we run through our
embedding algorithm that givesus all of our numbers.
That creates a vector.
We now have this vector, that'sout there.
We drop that vector on on ourplatform.
Then I come along and I say, hey, ravioli is Italian food.
Let's do the same thing.
Now, it's not quite the samevector, right, it's just a
(08:31):
little bit different.
So we're still talking aboutItalian food.
We're still talking aboutravioli.
We're still talking.
We're talking about ravioli nowinstead of lasagna.
So we have very similar type ofthings happening here, but it's
not quite the exact same.
And then we're going to come inand say, super mario's italian.
So it's kind of the same, butit's not right.
So it's vector is going to be alittle bit different, because
(08:52):
we're not talking about food allof a sudden, we're not talking
about a video game and and so,like, everything's just gonna be
a little bit different, allright.
So why does that matter?
Now let's go load the entireinternet, like everything we
know about all of mankind,everything that's on the
internet.
Let's load it all up and let'sturn them all into these vectors
that are.
Now we're gonna get thismassive database of all these
vectors that are out there.
We come along and we say, hey,I need Italian recipes.
(09:17):
So what's it gonna do?
We do the exact same thing.
We literally turn it into avector and then we say give me
all the vectors that are kind ofclose.
So for the math, people, that'scosine, similarity is what
they're doing there.
But they say go give me all thevectors that are kind of close
to what it is that I'm doing orwhat it is that I've asked you
for, and I want you to bringthose back.
(09:37):
And now I have a set ofknowledge I can work with.
So see how we got there.
So the reason I want you tounderstand how the vectors work
is because when you understandthat and you understand how it
starts to pull its informationtogether, you start to learn how
to talk to it.
If I say don't think ofelephants, did you all think of
elephants?
(09:58):
So if I'm in a prompt and I saydon't do something, what did I
just tell it to do?
I just told it to pull it in.
So that's part of the way thatwe learn how to talk to it.
And that's why they say, whenyou're writing a prompt is to
tell it what you want it to do,not what you don't want it to do
.
So by understanding how this isbeing put together, we can
(10:18):
start to be intelligent in theway that we look at the way
we're going gonna apply all thiswe tracking, we good, okay, all
right.
So now let's talk a little bitmore about large language models
.
So we're gonna come in here andwe're gonna say, hey, I want
Italian recipes.
And it's gonna go out to ourlarge language model.
It's gonna pull in that lasagnawas Italian food.
(10:38):
It then pulls in that ravioliis Italian food.
But it didn't pull in SuperMario, right, because that
vector didn't fit.
It was outside of it.
So Super Mario being Italiandoesn't matter.
And then it says, hey, here area few Italian recipes that you
can try at home.
Sweet.
As was mentioned earlier, thisis a mathematical probability
list.
It's basically trying to figureout what you like and what you
(10:58):
want, and then it's basicallyusing a random number plus the
next most probable word for youand trying to pick that out for
you.
I really dumbed that down, butthat's really kind of what's
going on there.
So, anyways, again, it's notmagic, it's just using
probability theory, all right.
So we now have here are a fewItalian recipes you can do at
(11:21):
home.
Anybody ever been talking toyour AI and it suddenly like
forgets.
You're like what just happened.
Like we've been talking aboutI'm a Notre Dame football fan.
So like we've been talkingabout Notre Dame football and
all of a sudden you're bringingup stuff about Purdue.
How dare you Like what is goingon here.
(11:41):
So here's what happens.
As soon as I come in and I askand I give it a different prompt
I was thinking more oftraditional Italian, like maybe
Tuscan, food.
All of a sudden, this thingcalled the context comes into
play.
Everything that's outside ofthis box here gets forgotten,
and it's basically the stuffthat's oldest, it's the stuff
(12:04):
that came in at the top.
So why is that?
Essentially, what's going on iswe can only pull in so many
vectors at a time in order tomake a decision and we run out
of space.
Think of, like having a deskthat's got a certain amount of
usage on it and you starteventually just running out of
places to put things.
So what do you do?
You start taking the stuff youdon't use and you put it on the
(12:24):
floor.
That's what it's doing.
So when you're having aconversation with AI and all of
a sudden it goes, I don't knowwhat you're talking about.
This is what's happened.
Essentially, we've blown outthe context.
Now I want to be careful here,because this makes it look like
the context is really small.
It's really hard to comprehendthe amount of information that
(12:47):
we, as humans, hold in our headat any given moment, and the AI
context is basically areplication of our short-term
memory and the things that wejust note.
I mean, for instance, I'mtalking really fast because
that's just who I am and there'slots of words coming out of my
mouth.
I know all those words andthey're all just coming to me
really quickly.
That's all in my memorysomewhere and we're pulling it
(13:08):
up really really fast.
For AI to replicate that, itbasically has to have all that
information in its memory.
So it doesn't matter the factthat this context right now can
literally hold the entire HarryPotter series all seven books in
its context and have no trouble.
That's still not enough.
It's still too little when youthink about how we as humans
make decisions.
(13:28):
It's still going to run out ofspace.
So this is an important thing tounderstand because, as you are,
a practical tip that I like todo is I like to start new chats
at almost every time, so Isummarize what I've been saying
and I copy that into a new chatbecause it gives me the
information that I want in acontext that I know I have limit
(13:50):
, I know I have control around.
So if you ever find yourselfrunning out of room or run, or
it keeps forgetting things, andyou're like.
I told you that already.
So, if you ever find yourselfrunning out of room or it keeps
forgetting things and you'relike I told you that already why
are you forgetting?
It's because it's moving out ofthe context.
So summarize it, create a newchat, and it'll pick up right
where it left off.
Does that make sense?
Okay, so that's what's going oninside the large language model
(14:11):
.
So this is essentially how we'regetting our knowledge.
We have the automation tools.
We now have the knowledge.
How do we get our knowledge?
Well, if we can put things intovectors and we can then have
context, like we're talking to ahuman that has memory, we now
can pull information back andforth and we can have a
conversation with our data.
Here's a cool thing, though wecan do this with what's called
(14:35):
unstructured data, and that'salso kind of a revolutionary
thought.
Like all your emails aresuddenly now accessible to some
computer system in order to makesome decisions around, because
we can understand the context ofthe language.
We can understand the way theinformation is being
communicated, way theinformation is being
communicated, so this isn't justall the stuff in your database,
(14:55):
though.
That's important and, trust me,I want all that stuff because
that will make better decisions.
But it's also all the stuff inyour PDFs, all the stuff in your
Word documents, all the stuffon your emails.
If it's unstructured data, wedon't have access to it and we
can start to build models aroundhow to communicate and interact
on that front.
That's pretty cool from abusiness standpoint, right?
(15:16):
All of a sudden, stuff that waslocked away in Excel files is
now available to us.
Make sense, all right.
So what if we take all of ourorganizational data and we load
it into a local database?
We take all of our Word file,all of our documents, all of our
(15:37):
emails and we load it up into adatabase that we control.
That's on our premises, that is, in our world.
We load it all into there.
We call it a vector database,and then we adjust the way that
the chatbots work so that when Igive it a prompt, the first
thing it does is go out to ourdatabase and pull information in
, and then it goes out to theshared LLM to pull in general
(16:00):
data before it generates theresponse.
This is what's referred to asretrieval, augmented generation,
and essentially what it's doingis it's saying hey, let's let
your data be your data, let'sput it in its own space and
let's reference it.
When we're having aconversation, let's reference
that first.
When we're having aconversation around how you want
(16:21):
or the prompt that you want,let's make sure that we're
getting what you want about yourdata first, but let's not try
and load the entire internetinto our local database.
Let's leverage that too.
So we're trying to get the bestof both worlds here, make sense
.
So what this lets us do now is Ican now load up special
information into a centralenvironment, into my environment
(16:42):
that I control, and I can thenaccess it inside of a prompt
engine.
Now, you're not going to,probably.
Well, actually, you know whatAnyone use the projects in chat,
gpt or Claude.
You've seen those.
That's basically what they'redoing here.
So when you load up a projecton the right side I'm thinking
(17:02):
in Claude right now but there'sa place where you can load up
documents, right, you can loadup information.
Well, when you're loading upthat information, what's it
doing?
It's vectorizing it.
It's embedding it, vectorizingit and loading it into its own
little environment, its ownlittle database.
Then when you make a prompt, itgoes well, let's look there
first.
So if you haven't done this, wedo it with our RoboSource brand
(17:23):
.
So, like, how do you talk?
Like, what's the voice of ourcompany sound like?
So my wife is the voice of ourcompany and so she has all the
documents and all the way thatwe communicate.
She's a much better writer thanI am.
It sounds really good when shesays it.
It sounds kind of corny when Isay it.
So we put all of her examplesin place and then I can go in
and say hey, I'm trying to writea blog about this, but can you
(17:45):
make it sound like the way thatyour RoboSource would talk?
What does it do?
It goes to our database, itgoes to our local project, pulls
in examples about all of theway that she communicates.
It then takes my prompt and itgoes all right, let's go
leverage all the information wehave on the Internet as a whole
and let's now model that patternin a way that makes sense.
(18:07):
And so we now have now I have away to make it sound like the
way that our company shouldsound, which is awesome.
And what do I do with that?
I can now take that and I canload it back into my document.
We just created a learning loop.
Does that make sense?
So we now went from I justcreated something, it's good.
I drop it back into myorganizational specific data.
(18:27):
It gets loaded into my database.
It's now organizationalknowledge.
We're learning, we're gettingsmarter.
The business is starting topick up information.
That's essentially what's goingon within the learning on the
AI at a simple level for abusiness, kind of cool huh.
That's a new opportunity for uswhere we can actually apply
(18:50):
learning and, as we're learningthings, we can actually apply it
and continue to feed that in away that we can keep operating.
We can keep apply it andcontinue to feed that in a way
that we can keep operating.
We can keep getting better andbetter as a business in the way
that we operate on a day-to-daybasis.
Neat I'm used to people talkingto me while I'm doing this, so
this is a little bit hard for me.
All right, let's see how muchtime I got here.
(19:13):
All right, we are.
See how much time I got here.
We are doing well.
Agents anyone heard of them?
You seen them online.
Most people not heard of them.
It looks like All right.
So what's going on with agents?
Agents are essentially a way fora computer to do your work for
you in some ways.
So I'm going to show an examplehere.
(19:35):
This is using Claude's computeruse model, so it's a form of an
agent that's actually makingsome decisions.
What you're going to watch itdo is it's going to perceive the
environment.
By that I mean it's going toactually understand the computer
desktop.
It's going to, essentially,you're going to watch it.
It goes really fast, which iswhy I'm describing it now.
You're going to watch it takepictures of the desktop
(19:56):
environment, identify where,like, firefox is.
Then you'll see it move themouse down, click Firefox and
open it.
It's going to actually gothrough, make decisions, take
actions and accomplish a goal.
So in this case, what I'm sayingis I ask it down here.
You'll see me type here now say, go get the top 10 mops for
(20:16):
sale on Amazon and put it into aCSV file, comma separated,
value for file for me, and hereare the fields that I want you
to get, and I tell it to go.
So what does it do?
Well, watch, it just took ascreenshot and it identified
where Firefox was.
Oh yeah, it just opened Firefox.
It's then going to figure outwhere this.
Oh, we just went to Amazon.
(20:37):
It's going to type in thesearch box mops and we're going
to search for mops.
Now, this was just a prompt,right?
You just said, please do thisfor me.
And it basically created a plan.
It goal-sought, right?
It said oh, this is the goalyou want, let's work my way
backwards, figure out all thesteps that need to happen.
I will create each of thosesteps along the way and I will
(21:04):
then go execute on those stepson your behalf here.
Notice, my screen just went todark mode so you can tell what
time of day I was doing that.
And here we're done.
We actually have here's a file.
It gives me the output of thefile, the top 10.
You can't read it.
I can read it up here.
The second, the last one, saysSwiffer PowerMop Multi-Surface
Mop Kit, which I think is thisone over here.
(21:24):
So it literally looked at thescreen, pulled out the data,
found out what it needed to do,all on your behalf, off of one
sentence, right?
And this is where I usually getquestions about Skynet,
terminator, things along thoselines.
So in this case, it's cool,it's not Skynet.
(21:52):
But again, let's break down howit did it.
Right, we already know itvectorizes it.
Right, we know how to createthe vectors.
We talked about that and, bythe way, doing vectors on images
is basically the same thingLittle nuance to it, but it's
basically the same thing.
So we create a vector.
We then are able to create aplan.
Well, we know that that canhappen because when we ask it to
(22:14):
create plans, it goes out andit finds all the things.
Because we say create a plan,and it goes out to the world of
the internet and says how do youcreate a plan?
And it gets back all thatinformation and it then creates
those models for you.
So it's doing the same thinghere.
It's really doing nothingdifferent than what we talked
about the first three slides.
It's just that we gave itpermission to go ahead and do
things on our behalf using tools.
So why does this matter?
(22:41):
Let's get into kind of thetheory of work then.
So management often basicallycomes in and says, hey, here's
our strategic goals, here's whatwe're trying to set up, here's
what we're trying to make happen.
Then we're going to go hireknowledge workers and we're
going to say we're going to giveyou a role, we're going to call
you chief operating officer andwe want to make sure that you
hit all of these KPIseffectively.
And what does that person do?
Well, they set up a set ofdifferent goals of what it means
(23:05):
in order to accomplish that jobeffectively.
And then, in order toaccomplish the job, they then
break that down into tasks inorder to get it done.
Traditionally, what happens is,when we start talking about
automating a company, we come inand we say, well, hey, let's
look at these tasks.
Right, let's look at the tasksthat are at the lowest level,
let's figure out how we canautomate those for you, because,
at the end of the day, you'restill responsible for figuring
(23:27):
out what the goal should be inthe first place.
So you tell me what the goal isand then I'll go and I'll
figure out how to execute eachof those tasks.
Well, where the world isstarting to head is a world
where the management can now sayto an AI agent hey, here's the
KPI, and the KPI can figure out,or the agent can figure out
(23:48):
what's that goal and what tasksneed to happen, and start to
execute on those tasks.
Now, are we there yet?
No, not even close.
But you can see we're gettingthere and you can see the
technology is, and you can seethat we're going to get there
pretty quick.
So this is what has peoplegoing whoa.
(24:08):
This could be revolutionary,because we now can actually
because you can see how we canget from basic vectorization of
data to I'm doing your entirefreaking job.
Now is it going to replace jobs?
It'll replace some.
Is it going to replace all ofthem?
No, humans are way too.
Ingenuity Is that the word?
I don't know.
Smart Humans are way too smartto not come up with new things
(24:33):
on their own right.
We're going to invent new ideas, we're going to invent new
thoughts and really, at the endof the day, most of business is
relationship anyways, andcomputers are not going to be
great at that.
It makes it okay, but it's notgreat at it.
And so you're still going to.
Business is going to be donehuman to human, and that's still
going to be the strategicadvantage.
But a lot of the operationalday-to-day is going to start.
(24:57):
You're going to start seeing AIagents be able to execute on
this.
Has anyone heard the conceptthat's come out recently, called
service as software?
Yeah, has anyone heard theconcept that's come out recently
, called service as software?
Yeah, so this is basically whatthey're getting at.
They're saying, hey, if an AIagent can take what used to be
service only offering, say maybepayroll, and we can automate it
, can automate it, then can wenow go to businesses and offer
(25:25):
them this service, but behindthe scenes, have the AI run the
operations in its entirety andthen we'll charge you.
Basically, our business modelwill essentially be for the
outcome that we produce.
We'll bill you for a successfuloutcome.
That's a different way ofthinking about things, but it's
getting a lot of traction,specifically in the investing
world and on the coasts.
Pay for the outcome that youget, pay for the service
(25:47):
delivery that is delivered foryou.
So pay for every payroll runand then back it with AI agents
that do a lot of that work foryou.
Now I'm going to put a wholebunch of caveats around things.
I think I don't remember what'snext.
Ah, good, hurdles to adoption.
So it's slow.
It's getting faster, but it'sslow and by that.
(26:14):
So I kind of lied a little bit.
It doesn't actually breakthings down by words.
It breaks things down by whatare called tokens, and tokens
are parts of words.
I just no one really knows howto explain what a token is or
when a token is, so I justturned it into words to make it
easier to understand.
But when you're reading thedocumentation, they talk about
tokens per second or tokens, andso total side note, anyone see
(26:36):
the thing about how AI couldn'ttell you how many R's were in
strawberry?
Yeah, so that's because itisn't actually reading words and
it isn't actually countingcharacters.
It breaks it into tokens and so, because of the way that it
breaks it down into tokens, itsaw two tokens that had an R in
it, so it said there were two,even though Barry has two R's in
it by itself.
So that's kind of what's goingon there.
(26:58):
But, as a result, this is stilltoo slow and therefore it feels
kind of awkward, especially whenyou're doing voice stuff.
Anyone talked with a voice AI,yet You're like, wow, it can
understand me, wow, it gives meinformation I want.
Eventually it's a little slow.
(27:19):
There's like these awkwardpauses in between every sentence
, right, whereas processingwhat's going on, that's just
because it's just not fastenough yet.
So that's got to get fixed.
Tooling needs to be improved.
Context window size still needsto be improved.
We talked about that a littlebit.
We just can't remember enoughin order to make really informed
(27:42):
, long-term, good decisions.
We have to keep breaking themdown into smaller pieces, which
is impromptu engineering, ifyou're looking it up.
That's chain of thoughtreasoning it's what they call
that, but that's essentiallywhat's going on there.
The tooling isn't quite ready tojust have conversations with
you yet.
And then usability this iswhere I don't know about you,
(28:02):
but I'm kind of already burnedout on chatbots.
Like the next person that comesout with a brand new,
revolutionary AI thing and putsa chatbot in front of me, I'm
going to scream.
It's like, yes, I understandthe value behind it, but at the
same time, we've got to come upwith better ways to interact
with the system.
The models haven't beeninvented yet.
They're still being created.
(28:22):
We don't know how to interactwith.
Like the only intelligent thingwe interact with is people and
we interact with them throughwords, so that's the only way we
know how to interact with theAI.
But we're also on a computerand so there's new models that
need to be created around it.
We just haven't figured themout yet.
So I think that's coming.
(28:49):
Anyone read Malik's bookCo-Intelligence Fascinating book
.
I think he's out of Wharton.
He has four principles that Ilike, and I like to talk about.
The first is that you shouldtreat AI like a human.
So when you talk to AI andthat's not because they're going
to become our AI overlords, butwe treat AI like a human.
(29:09):
Why?
Because AI was trained on datathat was created by humans.
So if the patterns that AI isreplicating are speech patterns
of humans, then by talking to itthe same way we would talk to a
human, we're going to get thebetter pattern out of it.
It's going to recognize thosepatterns more effectively.
So talk to it like it's a human.
(29:31):
Second rule and I'm doing theseoff the top of my head, so
correct.
I'm sorry if I mess them up alittle bit but a second rule is
to bring AI to the table,because if you've played with it
, you know it really solves theblank sheet syndrome problem,
where you're like I don't knowhow to create a marketing
(29:52):
strategy for a new product onXYZ.
Well, ask it, it'll give you 50ideas.
Half of them might be terrible,but you no longer have a blank
sheet.
You now have something to gowith.
So it actually does acceleratethe brainstorming and problem
solving side.
So, yeah, bring AI to the table.
The third one is that it's thedumbest it's ever going to be.
(30:14):
Ai is just getting smarter andsmarter and smarter and is going
to be continually advancing, sobe prepared for that.
And then the fourth, andprobably my favorite one, is
human in the loop.
By that they mean don't trustit.
My favorite analogy that I sayis anyone have a toddler?
(30:36):
Okay, sometimes you're blownaway by the toddler's ability to
manipulate you to get thatcookie and you just look at them
and you're like that's it.
You're going to Harvard, it'shappening, you're a genius, it's
over.
And now, five seconds later,you're screaming don't put your
hand on the stove, right.
That's kind of AI right nowLike it's brilliant and it will
(31:00):
blow your mind and then it'll dothe dumbest stuff and you're
like what just happened?
So understand that when you'reworking with it and that's why
we put human in the loop so askit to do something for you, then
have a human review it and makesure that it's all good to go.
So this is me.
I'm a RoboSource.
(31:21):
We build automation tools forintelligent companies.
I had fun.
Thanks for letting me share mygeekiness with you, and, leo,
I'll be around all day.
Let me know if I can help withanything.
You're listening to Risk and Resolve, and now for
your hosts, ben Conner and ToddHufford.
Well, that made me sound reallygeeky, but I'm wearing a suit,
so I can't be that geeky, right?
So I really did want to wearjeans and a t-shirt, but you
know, it is what it is.
So here's what we're going todo today.
(00:22):
We are going to talk about howAI works.
I want to break it down becausehere's what we're going to do
today.
We are going to talk about howAI works.
I want to break it down becausehere's what I believe.
Let me first make sure thismoves to me.
There's me.
Why AI?
All right.
So what are we talking abouthere?
Why do we care about AI?
What is it that's actuallyhappening?
What is it that's going onunder the scenes, under the
covers?
It's not magic, so let'sunderstand that.
So, anyone like math?
(00:43):
There's not a lot of handsgoing up, all right.
There's some hands going up,all right.
We're going to get into some ofthe math.
Now, I'm not going to get toomathy, but here's what I think.
I think that if we understandhow things work, it takes the
mystery out of it and we canstart to make informed decisions
about it.
So that's where I want to behere today.
That's my goal is.
I want to take some of themystery away from it.
So we're not going to get intoall the calculus and all the
(01:05):
details, but I'm going to tryand explain this in such a way
that you can understand thatthis isn't magical, that there
are some decisions we can makeand how this all works.
So we're going to talk a littlebit about puppy poetry.
We're then going to jump intosomething about.
We're going to talk about thetools that exist.
We want to make sure weunderstand all those different
tools that exist inside of theAI stack, specifically about how
(01:28):
you would leverage that insideof a business.
Sound good?
All right, let's jump in.
Why all the hype?
All right, who has done this?
This is in Claude.
Who has gone in and said writeme a haiku about puppies.
It's kind of the starting pointof AI, right, like?
You come in and you're like hey, write me a haiku about puppies
.
It's kind of the starting pointof AI, right, like you come in
(01:49):
and you're like hey, give mesomething to like, show me that
it can do it.
And it did.
And we're like woohoo, right,we're all excited, we're
thrilled, we're like this isbusiness revolutionary, right?
Well, it is, and from acomputer science standpoint, I
want to break down why and Iwant to talk through that
standpoint.
So, first of all, what's goingon here?
(02:11):
This word haiku, what does itmean?
Well, we know what it means,right, we know that it's a
poetry, we know that it's threelines, we know that each line
has a certain number ofsyllables with it.
So we understand that there aresome rules around what that
word means.
But what's really cool here isthe computer understood that,
(02:36):
and that is something worthtaking note of.
The second thing is I said, hey,here are these set of rules
called a haiku, but I want youto do something.
I want you to apply those rulesto a goal, to this target
called puppies.
So I'm saying, hey, I want youto take this set of rules three
lines, syllables are hard.
(02:56):
I want you to figure that outand I want you to goal seek
based upon the outcome, therules and constraints in the
plan.
I want you to put together aplan and I want you then to
execute on that plan towardsthis outcome.
That's pretty cool.
Now, it's cool to see it do itfor something as simple as a
haiku right, that's neat.
But imagine if this said I wantyou to onboard this new
(03:21):
employee, or I want you tounderwrite this mortgage
information.
That's why we care, that's whywe get excited about it.
That's what's going on here isbecause we actually are at a
point where we're able to usenatural language to capture the
information that we have, tocapture the rules about how a
business works, and then sayhere's the goal we want you to
accomplish and it can work itsway towards it.
(03:42):
And there's special ways andthere's a way that it does that.
And it's pretty logical whenyou start to understand it.
But that's what's going on here.
That's why people get excitedabout it.
So when you see people go hey,look, it wrote, you know, look
at what it did for my email.
Or look at what it did for youknow it could write me a poem,
that's the neat part.
We're tracking.
Okay, all right.
(04:04):
So let's talk about some of thetools.
We'll talk a little bit abouttask automation, because right
now, ai and automation they'rekind of being talked about as
the same thing.
So we'll talk a little bitabout RPA.
I didn't put on their API stuffbecause that gets all super
geeky, but that's really most ofwhat's happening here.
If you don't know what APImeans, it stands for application
(04:28):
programming interface and it'skind of like the back door into
a lot of systems.
It's really kind of the way thesystems work.
But we'll talk a little bitabout RPA.
We'll then talk about whatknowledge workers are going on.
What is embedding,vectorization, llms, context
windows you might have heard alot of this stuff in the
marketing stuff that's out onthe news.
So let's talk about what theyare.
What's it really mean, how doesit actually apply to you and
why do you care?
And then we'll talk about whatlearning looks like.
And finally we'll jump intothis new thing called AI agents,
(04:52):
which kind of gets into some ofthe managerial side.
Let's start.
All right, task automation.
Anyone heard of CNRPA before?
All right, robotic processautomation.
I literally wrote this code.
You can't see it, which is good, because it's not beautiful
code, but I'm going to hit goand what you're going to watch
is this is going to log into awebsite and it's going to upload
(05:14):
like 50 orders.
Okay, notice, this is going fast.
That's because the computer'sdoing it.
That's because the computer'sdoing it completely without a
person touching it.
So that's called roboticprocess automation.
It's a technology that is kindof the foundation to a lot of
computer testing.
So software companies will comein and they're like hey, we
don't want to pay peoplemanually to go click through a
(05:34):
bunch of buttons, so we wrotecode to teach a computer how to
go do that for us.
And then all of a sudden werealize, hey, you know what
that's actually kind of usefuland there are some applications
that we can't get into throughdatabases or APIs.
Maybe we can go through thefront end the same way a human
does, and so that's what roboticprocess automation is.
You've maybe heard of UiPath orBlue Prism Automate, anywhere.
(06:00):
Those are some of the newtechnology, not new, they're old
now, it's like a whole threeyears ago.
So that's some of thetechnology that's been around
that's doing a lot of that.
But this is getting used a lotwith AI and you'll see here at
the end on the agent side wherethey're starting to apply some
of that.
But I want you to know that it'sout there.
And a lot of people talk aboutautomation and this is one of
the examples.
I did mention APIs, applicationprogramming, interfaces.
(06:20):
Frankly, that's how mostautomation is done, but it's not
as fun to show off.
So you know, showing code runisn't really that cool.
So, all right, let's move intothe fun stuff.
So AI is not magic, it's math.
So we're going to talk about itNow.
Bear with me, at first thisfeels heavy and you're like what
in the world are you doing?
You're like talking crazy mathstuff with me.
(06:42):
It makes sense.
It's not that complicated.
What we want to do is and wetalked about it earlier we want
to convert these sentences intoways that we can understand them
.
All right, this is what'scalled embeddings.
So what AI did and kind of thebig breakthrough, is it said hey
, let's figure out a way tounderstand how a sentence is put
together so that we can compareit to other sentences.
(07:05):
So what we're going to do iswe're going to take every word
in the sentence in this caselet's take the word Italian and
we're going to look at all thewords that are on the left and
then we're going to weight them,basically put mathematical
numbers around the relationshipbetween all those words, so that
we have some indication of howthose words relate to each other
.
Then we're gonna take all thewords on the right, and we're
(07:25):
gonna weight those and we'regonna put numbers around them.
Okay, finally, we end up withwhat ends up being a really long
list of numbers.
So, yes, I went all geeky.
This is what's called a vector.
Remember physics class?
Right?
Vectors you had to calculateall that stuff.
Well, that's basically what wedid here, and so, while this is,
(07:45):
I'm showing you kind of asimple version, this is done on
parameters and billions andbillions of different weightings
in the way that it all works.
So you get a really, really,really complicated list of
numbers.
But for the purpose of this,let's just worry about two.
Let's just put them in atwo-dimensional array here,
right?
So a normal look.
So let's take our sentencelasagna is Italian food.
(08:06):
And now what we're going to dois we're going to go and we're
going to load that into ourlarge language model.
So we're going to first embedit, okay, so we run through our
embedding algorithm that givesus all of our numbers.
That creates a vector.
We now have this vector, that'sout there.
We drop that vector on on ourplatform.
Then I come along and I say, hey, ravioli is Italian food.
Let's do the same thing.
Now, it's not quite the samevector, right, it's just a
(08:31):
little bit different.
So we're still talking aboutItalian food.
We're still talking aboutravioli.
We're still talking.
We're talking about ravioli nowinstead of lasagna.
So we have very similar type ofthings happening here, but it's
not quite the exact same.
And then we're going to come inand say, super mario's italian.
So it's kind of the same, butit's not right.
So it's vector is going to be alittle bit different, because
(08:52):
we're not talking about food allof a sudden, we're not talking
about a video game and and so,like, everything's just gonna be
a little bit different, allright.
So why does that matter?
Now let's go load the entireinternet, like everything we
know about all of mankind,everything that's on the
internet.
Let's load it all up and let'sturn them all into these vectors
that are.
Now we're gonna get thismassive database of all these
vectors that are out there.
We come along and we say, hey,I need Italian recipes.
(09:17):
So what's it gonna do?
We do the exact same thing.
We literally turn it into avector and then we say give me
all the vectors that are kind ofclose.
So for the math, people, that'scosine, similarity is what
they're doing there.
But they say go give me all thevectors that are kind of close
to what it is that I'm doing orwhat it is that I've asked you
for, and I want you to bringthose back.
(09:37):
And now I have a set ofknowledge I can work with.
So see how we got there.
So the reason I want you tounderstand how the vectors work
is because when you understandthat and you understand how it
starts to pull its informationtogether, you start to learn how
to talk to it.
If I say don't think ofelephants, did you all think of
elephants?
(09:58):
So if I'm in a prompt and I saydon't do something, what did I
just tell it to do?
I just told it to pull it in.
So that's part of the way thatwe learn how to talk to it.
And that's why they say, whenyou're writing a prompt is to
tell it what you want it to do,not what you don't want it to do
.
So by understanding how this isbeing put together, we can
(10:18):
start to be intelligent in theway that we look at the way
we're going gonna apply all thiswe tracking, we good, okay, all
right.
So now let's talk a little bitmore about large language models
.
So we're gonna come in here andwe're gonna say, hey, I want
Italian recipes.
And it's gonna go out to ourlarge language model.
It's gonna pull in that lasagnawas Italian food.
(10:38):
It then pulls in that ravioliis Italian food.
But it didn't pull in SuperMario, right, because that
vector didn't fit.
It was outside of it.
So Super Mario being Italiandoesn't matter.
And then it says, hey, here area few Italian recipes that you
can try at home.
Sweet.
As was mentioned earlier, thisis a mathematical probability
list.
It's basically trying to figureout what you like and what you
(10:58):
want, and then it's basicallyusing a random number plus the
next most probable word for youand trying to pick that out for
you.
I really dumbed that down, butthat's really kind of what's
going on there.
So, anyways, again, it's notmagic, it's just using
probability theory, all right.
So we now have here are a fewItalian recipes you can do at
(11:21):
home.
Anybody ever been talking toyour AI and it suddenly like
forgets.
You're like what just happened.
Like we've been talking aboutI'm a Notre Dame football fan.
So like we've been talkingabout Notre Dame football and
all of a sudden you're bringingup stuff about Purdue.
How dare you Like what is goingon here.
(11:41):
So here's what happens.
As soon as I come in and I askand I give it a different prompt
I was thinking more oftraditional Italian, like maybe
Tuscan, food.
All of a sudden, this thingcalled the context comes into
play.
Everything that's outside ofthis box here gets forgotten,
and it's basically the stuffthat's oldest, it's the stuff
(12:04):
that came in at the top.
So why is that?
Essentially, what's going on iswe can only pull in so many
vectors at a time in order tomake a decision and we run out
of space.
Think of, like having a deskthat's got a certain amount of
usage on it and you starteventually just running out of
places to put things.
So what do you do?
You start taking the stuff youdon't use and you put it on the
(12:24):
floor.
That's what it's doing.
So when you're having aconversation with AI and all of
a sudden it goes, I don't knowwhat you're talking about.
This is what's happened.
Essentially, we've blown outthe context.
Now I want to be careful here,because this makes it look like
the context is really small.
It's really hard to comprehendthe amount of information that
(12:47):
we, as humans, hold in our headat any given moment, and the AI
context is basically areplication of our short-term
memory and the things that wejust note.
I mean, for instance, I'mtalking really fast because
that's just who I am and there'slots of words coming out of my
mouth.
I know all those words andthey're all just coming to me
really quickly.
That's all in my memorysomewhere and we're pulling it
(13:08):
up really really fast.
For AI to replicate that, itbasically has to have all that
information in its memory.
So it doesn't matter the factthat this context right now can
literally hold the entire HarryPotter series all seven books in
its context and have no trouble.
That's still not enough.
It's still too little when youthink about how we as humans
make decisions.
(13:28):
It's still going to run out ofspace.
So this is an important thing tounderstand because, as you are,
a practical tip that I like todo is I like to start new chats
at almost every time, so Isummarize what I've been saying
and I copy that into a new chatbecause it gives me the
information that I want in acontext that I know I have limit
(13:50):
, I know I have control around.
So if you ever find yourselfrunning out of room or run, or
it keeps forgetting things, andyou're like.
I told you that already.
So, if you ever find yourselfrunning out of room or it keeps
forgetting things and you'relike I told you that already why
are you forgetting?
It's because it's moving out ofthe context.
So summarize it, create a newchat, and it'll pick up right
where it left off.
Does that make sense?
Okay, so that's what's going oninside the large language model
(14:11):
.
So this is essentially how we'regetting our knowledge.
We have the automation tools.
We now have the knowledge.
How do we get our knowledge?
Well, if we can put things intovectors and we can then have
context, like we're talking to ahuman that has memory, we now
can pull information back andforth and we can have a
conversation with our data.
Here's a cool thing, though wecan do this with what's called
(14:35):
unstructured data, and that'salso kind of a revolutionary
thought.
Like all your emails aresuddenly now accessible to some
computer system in order to makesome decisions around, because
we can understand the context ofthe language.
We can understand the way theinformation is being
communicated, way theinformation is being
communicated, so this isn't justall the stuff in your database,
(14:55):
though.
That's important and, trust me,I want all that stuff because
that will make better decisions.
But it's also all the stuff inyour PDFs, all the stuff in your
Word documents, all the stuffon your emails.
If it's unstructured data, wedon't have access to it and we
can start to build models aroundhow to communicate and interact
on that front.
That's pretty cool from abusiness standpoint, right?
(15:16):
All of a sudden, stuff that waslocked away in Excel files is
now available to us.
Make sense, all right.
So what if we take all of ourorganizational data and we load
it into a local database?
We take all of our Word file,all of our documents, all of our
(15:37):
emails and we load it up into adatabase that we control.
That's on our premises, that is, in our world.
We load it all into there.
We call it a vector database,and then we adjust the way that
the chatbots work so that when Igive it a prompt, the first
thing it does is go out to ourdatabase and pull information in
, and then it goes out to theshared LLM to pull in general
(16:00):
data before it generates theresponse.
This is what's referred to asretrieval, augmented generation,
and essentially what it's doingis it's saying hey, let's let
your data be your data, let'sput it in its own space and
let's reference it.
When we're having aconversation, let's reference
that first.
When we're having aconversation around how you want
(16:21):
or the prompt that you want,let's make sure that we're
getting what you want about yourdata first, but let's not try
and load the entire internetinto our local database.
Let's leverage that too.
So we're trying to get the bestof both worlds here, make sense
.
So what this lets us do now is Ican now load up special
information into a centralenvironment, into my environment
(16:42):
that I control, and I can thenaccess it inside of a prompt
engine.
Now, you're not going to,probably.
Well, actually, you know whatAnyone use the projects in chat,
gpt or Claude.
You've seen those.
That's basically what they'redoing here.
So when you load up a projecton the right side I'm thinking
(17:02):
in Claude right now but there'sa place where you can load up
documents, right, you can loadup information.
Well, when you're loading upthat information, what's it
doing?
It's vectorizing it.
It's embedding it, vectorizingit and loading it into its own
little environment, its ownlittle database.
Then when you make a prompt, itgoes well, let's look there
first.
So if you haven't done this, wedo it with our RoboSource brand
(17:23):
.
So, like, how do you talk?
Like, what's the voice of ourcompany sound like?
So my wife is the voice of ourcompany and so she has all the
documents and all the way thatwe communicate.
She's a much better writer thanI am.
It sounds really good when shesays it.
It sounds kind of corny when Isay it.
So we put all of her examplesin place and then I can go in
and say hey, I'm trying to writea blog about this, but can you
(17:45):
make it sound like the way thatyour RoboSource would talk?
What does it do?
It goes to our database, itgoes to our local project, pulls
in examples about all of theway that she communicates.
It then takes my prompt and itgoes all right, let's go
leverage all the information wehave on the Internet as a whole
and let's now model that patternin a way that makes sense.
(18:07):
And so we now have now I have away to make it sound like the
way that our company shouldsound, which is awesome.
And what do I do with that?
I can now take that and I canload it back into my document.
We just created a learning loop.
Does that make sense?
So we now went from I justcreated something, it's good.
I drop it back into myorganizational specific data.
(18:27):
It gets loaded into my database.
It's now organizationalknowledge.
We're learning, we're gettingsmarter.
The business is starting topick up information.
That's essentially what's goingon within the learning on the
AI at a simple level for abusiness, kind of cool huh.
That's a new opportunity for uswhere we can actually apply
(18:50):
learning and, as we're learningthings, we can actually apply it
and continue to feed that in away that we can keep operating.
We can keep apply it andcontinue to feed that in a way
that we can keep operating.
We can keep getting better andbetter as a business in the way
that we operate on a day-to-daybasis.
Neat I'm used to people talkingto me while I'm doing this, so
this is a little bit hard for me.
All right, let's see how muchtime I got here.
(19:13):
All right, we are.
See how much time I got here.
We are doing well.
Agents anyone heard of them?
You seen them online.
Most people not heard of them.
It looks like All right.
So what's going on with agents?
Agents are essentially a way fora computer to do your work for
you in some ways.
So I'm going to show an examplehere.
(19:35):
This is using Claude's computeruse model, so it's a form of an
agent that's actually makingsome decisions.
What you're going to watch itdo is it's going to perceive the
environment.
By that I mean it's going toactually understand the computer
desktop.
It's going to, essentially,you're going to watch it.
It goes really fast, which iswhy I'm describing it now.
You're going to watch it takepictures of the desktop
(19:56):
environment, identify where,like, firefox is.
Then you'll see it move themouse down, click Firefox and
open it.
It's going to actually gothrough, make decisions, take
actions and accomplish a goal.
So in this case, what I'm sayingis I ask it down here.
You'll see me type here now say, go get the top 10 mops for
(20:16):
sale on Amazon and put it into aCSV file, comma separated,
value for file for me, and hereare the fields that I want you
to get, and I tell it to go.
So what does it do?
Well, watch, it just took ascreenshot and it identified
where Firefox was.
Oh yeah, it just opened Firefox.
It's then going to figure outwhere this.
Oh, we just went to Amazon.
(20:37):
It's going to type in thesearch box mops and we're going
to search for mops.
Now, this was just a prompt,right?
You just said, please do thisfor me.
And it basically created a plan.
It goal-sought, right?
It said oh, this is the goalyou want, let's work my way
backwards, figure out all thesteps that need to happen.
I will create each of thosesteps along the way and I will
(21:04):
then go execute on those stepson your behalf here.
Notice, my screen just went todark mode so you can tell what
time of day I was doing that.
And here we're done.
We actually have here's a file.
It gives me the output of thefile, the top 10.
You can't read it.
I can read it up here.
The second, the last one, saysSwiffer PowerMop Multi-Surface
Mop Kit, which I think is thisone over here.
(21:24):
So it literally looked at thescreen, pulled out the data,
found out what it needed to do,all on your behalf, off of one
sentence, right?
And this is where I usually getquestions about Skynet,
terminator, things along thoselines.
So in this case, it's cool,it's not Skynet.
(21:52):
But again, let's break down howit did it.
Right, we already know itvectorizes it.
Right, we know how to createthe vectors.
We talked about that and, bythe way, doing vectors on images
is basically the same thingLittle nuance to it, but it's
basically the same thing.
So we create a vector.
We then are able to create aplan.
Well, we know that that canhappen because when we ask it to
(22:14):
create plans, it goes out andit finds all the things.
Because we say create a plan,and it goes out to the world of
the internet and says how do youcreate a plan?
And it gets back all thatinformation and it then creates
those models for you.
So it's doing the same thinghere.
It's really doing nothingdifferent than what we talked
about the first three slides.
It's just that we gave itpermission to go ahead and do
things on our behalf using tools.
So why does this matter?
(22:41):
Let's get into kind of thetheory of work then.
So management often basicallycomes in and says, hey, here's
our strategic goals, here's whatwe're trying to set up, here's
what we're trying to make happen.
Then we're going to go hireknowledge workers and we're
going to say we're going to giveyou a role, we're going to call
you chief operating officer andwe want to make sure that you
hit all of these KPIseffectively.
And what does that person do?
Well, they set up a set ofdifferent goals of what it means
(23:05):
in order to accomplish that jobeffectively.
And then, in order toaccomplish the job, they then
break that down into tasks inorder to get it done.
Traditionally, what happens is,when we start talking about
automating a company, we come inand we say, well, hey, let's
look at these tasks.
Right, let's look at the tasksthat are at the lowest level,
let's figure out how we canautomate those for you, because,
at the end of the day, you'restill responsible for figuring
(23:27):
out what the goal should be inthe first place.
So you tell me what the goal isand then I'll go and I'll
figure out how to execute eachof those tasks.
Well, where the world isstarting to head is a world
where the management can now sayto an AI agent hey, here's the
KPI, and the KPI can figure out,or the agent can figure out
(23:48):
what's that goal and what tasksneed to happen, and start to
execute on those tasks.
Now, are we there yet?
No, not even close.
But you can see we're gettingthere and you can see the
technology is, and you can seethat we're going to get there
pretty quick.
So this is what has peoplegoing whoa.
(24:08):
This could be revolutionary,because we now can actually
because you can see how we canget from basic vectorization of
data to I'm doing your entirefreaking job.
Now is it going to replace jobs?
It'll replace some.
Is it going to replace all ofthem?
No, humans are way too.
Ingenuity Is that the word?
I don't know.
Smart Humans are way too smartto not come up with new things
(24:33):
on their own right.
We're going to invent new ideas, we're going to invent new
thoughts and really, at the endof the day, most of business is
relationship anyways, andcomputers are not going to be
great at that.
It makes it okay, but it's notgreat at it.
And so you're still going to.
Business is going to be donehuman to human, and that's still
going to be the strategicadvantage.
But a lot of the operationalday-to-day is going to start.
(24:57):
You're going to start seeing AIagents be able to execute on
this.
Has anyone heard the conceptthat's come out recently, called
service as software?
Yeah, has anyone heard theconcept that's come out recently
, called service as software?
Yeah, so this is basically whatthey're getting at.
They're saying, hey, if an AIagent can take what used to be
service only offering, say maybepayroll, and we can automate it
, can automate it, then can wenow go to businesses and offer
(25:25):
them this service, but behindthe scenes, have the AI run the
operations in its entirety andthen we'll charge you.
Basically, our business modelwill essentially be for the
outcome that we produce.
We'll bill you for a successfuloutcome.
That's a different way ofthinking about things, but it's
getting a lot of traction,specifically in the investing
world and on the coasts.
Pay for the outcome that youget, pay for the service
(25:47):
delivery that is delivered foryou.
So pay for every payroll runand then back it with AI agents
that do a lot of that work foryou.
Now I'm going to put a wholebunch of caveats around things.
I think I don't remember what'snext.
Ah, good, hurdles to adoption.
So it's slow.
It's getting faster, but it'sslow and by that.
(26:14):
So I kind of lied a little bit.
It doesn't actually breakthings down by words.
It breaks things down by whatare called tokens, and tokens
are parts of words.
I just no one really knows howto explain what a token is or
when a token is, so I justturned it into words to make it
easier to understand.
But when you're reading thedocumentation, they talk about
tokens per second or tokens, andso total side note, anyone see
(26:36):
the thing about how AI couldn'ttell you how many R's were in
strawberry?
Yeah, so that's because itisn't actually reading words and
it isn't actually countingcharacters.
It breaks it into tokens and so, because of the way that it
breaks it down into tokens, itsaw two tokens that had an R in
it, so it said there were two,even though Barry has two R's in
it by itself.
So that's kind of what's goingon there.
(26:58):
But, as a result, this is stilltoo slow and therefore it feels
kind of awkward, especially whenyou're doing voice stuff.
Anyone talked with a voice AI,yet You're like, wow, it can
understand me, wow, it gives meinformation I want.
Eventually it's a little slow.
(27:19):
There's like these awkwardpauses in between every sentence
, right, whereas processingwhat's going on, that's just
because it's just not fastenough yet.
So that's got to get fixed.
Tooling needs to be improved.
Context window size still needsto be improved.
We talked about that a littlebit.
We just can't remember enoughin order to make really informed
(27:42):
, long-term, good decisions.
We have to keep breaking themdown into smaller pieces, which
is impromptu engineering, ifyou're looking it up.
That's chain of thoughtreasoning it's what they call
that, but that's essentiallywhat's going on there.
The tooling isn't quite ready tojust have conversations with
you yet.
And then usability this iswhere I don't know about you,
(28:02):
but I'm kind of already burnedout on chatbots.
Like the next person that comesout with a brand new,
revolutionary AI thing and putsa chatbot in front of me, I'm
going to scream.
It's like, yes, I understandthe value behind it, but at the
same time, we've got to come upwith better ways to interact
with the system.
The models haven't beeninvented yet.
They're still being created.
(28:22):
We don't know how to interactwith.
Like the only intelligent thingwe interact with is people and
we interact with them throughwords, so that's the only way we
know how to interact with theAI.
But we're also on a computerand so there's new models that
need to be created around it.
We just haven't figured themout yet.
So I think that's coming.
(28:49):
Anyone read Malik's bookCo-Intelligence Fascinating book
.
I think he's out of Wharton.
He has four principles that Ilike, and I like to talk about.
The first is that you shouldtreat AI like a human.
So when you talk to AI andthat's not because they're going
to become our AI overlords, butwe treat AI like a human.
(29:09):
Why?
Because AI was trained on datathat was created by humans.
So if the patterns that AI isreplicating are speech patterns
of humans, then by talking to itthe same way we would talk to a
human, we're going to get thebetter pattern out of it.
It's going to recognize thosepatterns more effectively.
So talk to it like it's a human.
(29:31):
Second rule and I'm doing theseoff the top of my head, so
correct.
I'm sorry if I mess them up alittle bit but a second rule is
to bring AI to the table,because if you've played with it
, you know it really solves theblank sheet syndrome problem,
where you're like I don't knowhow to create a marketing
(29:52):
strategy for a new product onXYZ.
Well, ask it, it'll give you 50ideas.
Half of them might be terrible,but you no longer have a blank
sheet.
You now have something to gowith.
So it actually does acceleratethe brainstorming and problem
solving side.
So, yeah, bring AI to the table.
The third one is that it's thedumbest it's ever going to be.
(30:14):
Ai is just getting smarter andsmarter and smarter and is going
to be continually advancing, sobe prepared for that.
And then the fourth, andprobably my favorite one, is
human in the loop.
By that they mean don't trustit.
My favorite analogy that I sayis anyone have a toddler?
(30:36):
Okay, sometimes you're blownaway by the toddler's ability to
manipulate you to get thatcookie and you just look at them
and you're like that's it.
You're going to Harvard, it'shappening, you're a genius, it's
over.
And now, five seconds later,you're screaming don't put your
hand on the stove, right.
That's kind of AI right nowLike it's brilliant and it will
(31:00):
blow your mind and then it'll dothe dumbest stuff and you're
like what just happened?
So understand that when you'reworking with it and that's why
we put human in the loop so askit to do something for you, then
have a human review it and makesure that it's all good to go.
So this is me.
I'm a RoboSource.
(31:21):
We build automation tools forintelligent companies.
I had fun.
Thanks for letting me share mygeekiness with you, and, leo,
I'll be around all day.
Let me know if I can help withanything.
Popular Podcasts
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!
The Breakfast Club
The World's Most Dangerous Morning Show, The Breakfast Club, With DJ Envy And Charlamagne Tha God!
The Joe Rogan Experience
The official podcast of comedian Joe Rogan.