November 18, 2025 • 33 mins
Is AI supposed to make things easier… so why does it feel like chaos?
You’re not alone. In this special episode of Leveraging AI, we invite you into the room where the real conversations happen — our live, unfiltered Friday AI Hangouts.
Hear real business people share what’s working, what’s breaking, and what’s making them want to throw their laptops across the room.
💡 Whether you’re a solopreneur, in a scaling team, or inside a major enterprise — these discussions will sound very familiar.
One takeaway? The learning curve is real. But the gold is in the room.
🎯 In this session, you'll discover:
- Why AI Projects are replacing Custom GPTs — and when you shouldn’t switch
- How business pros are using Claude skills, Comet browser, and Make.com to automate faster
- The difference between frustration and progress when implementing agents
- What goes wrong when AI makes up your part numbers (yes, really)
- How to choose between N8N, Make, and Zapier — without wasting weeks
- The one piece of advice for managing large AI-driven data systems
- Real talk: Why AI tools still fail — and how to debug smarter, not harder
This episode features you — the voices of our global AI business community. Want to be in the next one?
Join our next Friday AI Hangout, open to all.
👉 Register here (https://services.multiplai.ai/events0) — it’s free, it’s fun, and it’s the realest AI talk you'll find anywhere.
About Leveraging AI
- The Ultimate AI Course for Business People: https://multiplai.ai/ai-course/
- YouTube Full Episodes: https://www.youtube.com/@Multiplai_AI/
- Connect with Isar Meitis: https://www.linkedin.com/in/isarmeitis/
- Join our Live Sessions, AI Hangouts and newsletter: https://services.multiplai.ai/events
If you’ve enjoyed or benefited from some of the insights of this episode, leave us a five-star review on your favorite podcast platform, and let us know what you learned, found helpful, or liked most about this show!
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Speaker 43 (00:00):
Hello and welcome to the Leveraging AI Podcast, the
(00:02):
podcast that shares practical,ethical ways to leverage AI to
improve efficiency, grow yourbusiness, and advance your
career.
This Isar Metis, your host, andwe have a special episode for
you today.
For over a year now, a communityof amazing business people from
all around the world who are allAI enthusiasts, who wants to
learn together, have beenmeeting every Friday for what we
(00:24):
call the Friday AI Hangouts.
There's usually 15 to 25 peopleevery week.
As I mentioned, different placesaround the world, different
sizes of companies, from largecorporations to solopreneurs and
everything in between, andeverybody brings ideas,
questions, solutions, thingsthey've learned, systems that
they failed, and so on, and it'sjust a fascinating conversation.
(00:48):
Every single week, it'scompletely open to anybody to
join.
And there are endless number ofgolden nuggets from these
meetings, and I thought it wouldbe interesting for you to listen
to segments out of these AIFriday Hangouts to see if that's
something you might beinterested in joining as well.
So this episode is going to besnippets out of different
(01:11):
community meetings.
Many of them are extremelytactical and valuable and you're
gonna learn a lot from them justlike you learn from our regular
Tuesday episodes.
But it will also give you anidea of what it feels like, so
you can decide if you want tojoin us or not.
If you do wanna join us, thereis a link in the show notes
click on the link register.
It will appear on your calendar,and then obviously you don't
(01:32):
have to join, but wheneveryou're available on Friday at
1:00 PM Eastern, you can dothat.
I will obviously like to seemany of you and to get to know
you better and to welcome you toour amazing community, and now
to the episode.
Speaker 10 (01:45):
I wonder if anybody has any experience, um, sort of
taking what they did in a, likein a custom GPT into the new GPT
cus uh, chat GPT projects.
And the context is this, I builtsort of like a prototype for a
client.
Um, they had new apprenticestarting, they've never had an
apprentice before and they hadno clue how they were gonna do
with it, deal with it.
(02:07):
And so I ended up finding likewhat the apprentice's like three
year plan was and put that intothe GPT and then I found out
what systems they used in houseand created this thing that they
can talk to it, figure out howto teach the person, give it to
the kid to learn his things andstuff, and.
Now, now they love it and, um,are gonna wanna use it and build
on it.
And I realized that it would beso much better to have it in
(02:27):
this new, like, chat GBTprojects because then everything
stays in the same place and Ican control it, but have more
different users.
So I'm just wondering, do theinstructions still work as well?
Has anybody tried thatmigrating?
Speaker 3 (02:40):
I will let other people answer before I answer
Anyone?
No
Speaker 6 (02:47):
one.
Okay.
So a few, how many of you needan overview of projects in
general before we dive into whatare the differences?
Okay, so the, the biggest, uh,uh, two seconds on the biggest
difference between projects andcustom gpt.
Uh, custom gpt followinstructions and do them.
And that's it.
(03:07):
And that's the end of the thing.
And a project is a think about aproject as a mini private
version of Chachi pt, and inthat mini private version, you
can give it its own universalcontext.
You can give it special set ofinstructions, but very different
(03:30):
than a Chachi pt.
You can have conversation basic,basically a regular chat, just
like a regular chat with GPTonly inside your little
universe.
And it has also,
Speaker 4 (03:40):
it has its own memory.
So it has its own memory.
Speaker 6 (03:44):
Here's where I'm going.
Uh, it has many benefits rightnow over custom gpt.
Uh, the first one is you canhave free form conversations
with it versus in chat GPT, itjust follows the instructions,
right?
You can't just have a regularconversation in there.
Uh, so that's a big benefit.
The other big benefit is thatyou can upload more files than
(04:06):
you can upload into custom GPTsinside projects, which is
obviously a big deal.
Uh, the third thing, which ishuge is what Jonathan just said.
It has its own memory in its ownlittle universe, which means in
addition to the information thatyou gave it, when you trained it
to do the thing, and you gave itaccess to files, it learns over
(04:27):
time because you, you give itmore information as you're
having conversations with it.
So as of right now.
There's really no reason tocreate new custom gpt.
It makes a lot more sense tobuild them inside projects
because of all the things that Ijust said.
Uh, those who already exist incustom gpt, the answer is it
(04:49):
depends.
And the reason it depends.
If it's working, it's working.
Like there's no, there's noreason to move it over.
Like if it's doing the thingthat it needs to do, let it do
the thing that it needs to do.
But if it, if it doesn't, or ifyou wanted to have a, the
ability to have free formconversations with that in mind,
or you want it to learn overtime or all these kind of
things, then literally all youhave to do is upload the same
(05:12):
files on the project side, uh,copy the instructions to the
instruction sections, and, andthat's it.
And you're, you're off to theraces.
So you, it, it, it can work.
And if you think about it,projects in Claude.
That's what they've done for awhile.
Like in Claude, you don't havethese two different animals.
(05:33):
There's one thing which isprojects which you can use, kind
like projects in chat PT or likecustom GPS in chat pt, depending
how you want to use it.
And so I think when Chachi PTcame up with project was to
partially mimic, but they shouldhave just converted cha GP like
custom gpt into projects.
(05:54):
And I think they just wannaconfuse people go like, whoa,
where about Cha g pt?
Like custom gpt go.
So, uh, right now it's kindalike duplicate, but the only
Speaker 10 (06:04):
thing is that you can't, you can't do connectors
or something right at the momentwith these new projects.
So there's
Speaker 6 (06:09):
a few things.
There's a few things you cannotdo.
Uh, like the, the only realbenefit of custom gpt right now
is the whole section of thebottom where you can write code
and do all these kind of things.
Uh, that does not exist insideof projects yet.
Uh, but again, I, I, I thinkwhat makes sense is that they
would merge the two things intoone thing that has all this
(06:30):
functionality.
Uh, but I think that's the onlything right now that doesn't
exist.
Speaker 10 (06:36):
Okay.
And have you tried sharing itwith people and using the
different permissions and stuff?
Oh, yeah.
Speaker 6 (06:39):
Hold on.
So, so that's the biggest thingyou cannot share.
Oh, no.
Now you can share everything.
They changed it like a coupleweeks ago.
Yeah, exactly.
Yeah, it's brand new.
It just changed.
I share like two or three weeksago you couldn't share projects.
You could only share custom gptand now you can share projects
as well.
You can even share them.
(07:01):
I'm confusing, I'm using so muchof Claude recently in Claude.
You can share projects withanyone, like you can share
projects with the world and theydon't see the instructions and
the data behind it, but they canuse the project in chat.
GPTI don't remember how they didit, but I know you can now share
projects in chat GPT that youcouldn't a few weeks ago.
So yeah, you, you can shareprojects.
Speaker 10 (07:23):
Okay, well I'll, I'll, I'll keep working on this
and I, I will report back how itgoes.
Awesome.
'cause I also wanna like havedifferent permission levels.
I think it would be good if theApprentice can use it, but only
in study mode or something.
Right.
Whereas the people that have towork.
Yeah.
So I
Speaker 6 (07:37):
think in that case, you probably want to create
another version of it.
Maybe that would be for theApprentice.
And I think that would make it alot easier to control where you
give it slightly differentinstructions.
Speaker 7 (07:47):
Yeah.
We'll see.
Okay.
Thank you.
Speaker 10 (07:49):
And just before, before while I still have the
floor, I just wanted to saythank you to Ken for sending me
a podcast after the last, um,session, which was really
helpful to my last, uh, awesomela last, last week's issue.
Thank you, Ken.
Speaker 6 (08:03):
This is not about me.
This is about us all helpingeach other and there's nothing
that makes me happier than, youknow, uh, that if I die
tomorrow, this team can continuebeing helpful to another,
Speaker 38 (08:14):
Actually I've been listening to a lot of your, um,
podcasts, uh, every week and.
Trying to use the tools and, uh,the frustration level is
unbelievable.
Um, it took Gemini to write me ascript to merge a bunch of doc
Google docs into one doc 38script attempts, and everyone
(08:38):
said, I got it right now I gotit right.
It here to just cut and pastethis into the script.
It's the JSON, no problem.
Oh, it didn't work.
Oh, send me the error stuff.
Okay, here's the error.
Oh, I know what went wrong.
This'll fix it.
38 attempts.
So the frustration of thesethings just don't seem to be
ready for prime time every, Idid another one where I asked it
to create a prompt where I couldprepare for a meeting for
(09:01):
someone.
You, one of the use cases thatyou, right.
So I said, gave me this promptand, uh, so I gave me the
prompt, I gave it the person'sURL, uh, on LinkedIn, their
name, all this background, andit went and gave me a wonderful
brief.
Snapshot brief of the, you know,questions.
I could ask this person to startthe conversation except it was
(09:21):
the wrong person.
And then when I questioned itand said, oh, I can't look at
LinkedIn profiles, that's allprivate.
Well what if I gave my login?
Oh no, I can't do that either.
So the frustration level I'mhaving actually accomplishing
any task that I'm trying toautomate is unbelievable and
trying to learn any to end.
(09:43):
And your last podcast you hadlast week with the guest for the
French lady, like this is, uh,when she said the learning curve
is steep, it's unbelievable.
And I'm a relative.
I've been in, you know it a longtime, relatively savvy guy.
This is, but I'm not a coder.
It is unbelievably steeplearning curve.
And I don't know what thegroup's experience has been, but
(10:06):
the consensus seems to be thatdon't go with the easy ones like
Zapier make go with the harderone.
But it is a way lot morecomplex.
Speaker 39 (10:16):
Oh, um, okay.
So I mean, ma make.com issignificant, is still e even the
other ones have a bit of alearning curve.
So if, if, uh, NAN is consideredhard at that point, then, uh,
yeah, I'd definitely look at,uh, you know, make.com is
actually really nicely, uh, um,laid out and, uh, it's fairly
(10:41):
and it's very powerful.
Speaker 30 (10:43):
Yeah, so I agree.
I, I always tell people, getstarted, start with make, uh,
from a value to investment.
It is by far the best one rightnow.
Uh, it is not as flexible andcapable as NA 10.
It is way easier to use anddefinitely way easier to get
started.
Uh, even just connectingaccounts, like connecting your
(11:07):
Gmail account or your MicrosoftOutlook account make.
Is, you go in, you give it yourpassword, and you're done,
you're connected doing it in, inNA 10, you go to Google Cloud or
to Microsoft Azure, and you haveto open a project and you have
to assign permissions, and youhave like, it's freaking, it's
like 20 steps that you need tobe an manager to do.
(11:29):
So, a hundred percent.
Like if, if you're not exactly,if you're not now in the process
of I'm gonna automate my entirebusiness and it's gonna save me
a lot of money, uh, doing itthrough NA 10, just go with
make, uh, there, there, I'm, I'm
Speaker 39 (11:41):
actually in the process of automating a lot of
our business.
Uh, but I'm designing it all inMake first.
Yeah.
Yeah.
And once I have all of that inmake, I am going after the ones
that use the most operations andI'm gonna move those over, uh,
(12:02):
first and use, you know, use itas an optimization problem.
But designing it and make is soeasy and so fast that it's just
to get stuff out there as a, asa initial tool.
It's great.
Speaker 38 (12:17):
It seems now I understand why there's a whole
business out there for people tobuild your N eight N agent for
you.
'cause it's so complicated.
So there's the buy, buy, thebuild it, there's the build it
yourself.
Um, and then there's an easierway to build it yourself
apparently.
I'm wondering, is there anythingwhere it's pre-built, you just
saw, you know, we all want togo, I wanna check my emails, I
wanna monitor my emails, and ifthis email comes in, I wanna do
(12:39):
X with that email.
I wanna reply and I wannacustomize reply.
So I need a little bit of abrain.
So I might have agent kit, butlike what's been pre-written,
these seem like pretty standardbus business processes.
Business is simple.
Get the business, do thebusiness, get paid a lot of
similar processes.
Is there a library out there ofpre-written and in agents that
(13:01):
do specific things?
Speaker 30 (13:04):
A hundred percent.
So both on Make and on Zapierand on NA 10 and all of those,
there are thousands of existingtemplates.
Most of them are free.
Uh, and some of them will costyou, okay, pay me 30 bucks.
I'm like, okay, I'll pay you 30bucks to save myself.
The three hours of agony.
(13:24):
And what I do in many, manycases is I mix and match
segments of these templates.
So there are long templates.
It does all these things.
I'm like, okay, I just needthis.
But now I have this buildingblock that I can reuse.
And then there's this otherthing that does the other part
that I need, and all I have tofigure out is how to connect
(13:44):
them and make them talk to oneanother.
And the same is true in Make andin NA 10.
Uh, yeah.
Literally thousands of peopleare, are sharing them.
Oh, a lot more than thousands.
Probably tens of thousands ofpeople are sharing them on the
internet.
And you can, if you just search,uh, for template.
This.
Now, the, the thing that I'mstill puzzled by, and I'm sure
(14:06):
it's coming and there's, there'sthird party attempts, but not
inside of NA 10 and make,there's zero reason right now.
Why you shouldn't be able to sayin English what you want the
system to do, and then it willjust create the automation for
you.
A to Z soup to nuts, the wholething.
Test it out.
Verify, check that it's working.
(14:27):
Ask me for what you need fromme.
Like I need your API key forthis.
Here's the instructions on howto go and get it.
I need your login to this.
Just gimme your login.
I need, and you will give it allthis stuff and it will run, uh,
the close, the two closestthings that I have right now
that I'm using.
And I shared that with you inthe past, but since you're new,
I'm gonna share it again.
Uh, one is I have a Claudeproject.
(14:50):
That has an NA 10 MCP serverwith a whole set of instructions
on how to use the MCP serverthat I never got to work a
hundred percent, but it gets me85% there, which is way better
than I can do on my own in fiveminutes.
Uh, it takes a little bit oftime, but it's not a big deal.
Like to get the MCP set up andconnecting it to Claude and, and
(15:13):
then running it.
Uh, so this is one goodshortcut.
Again, it's not perfect, butit's gonna save you X percent.
Um, the other one is I now do100% of my entertainment work in
the Comet browser that is nowfree and available to everybody,
so you don't have to pay or be amember of anything.
Uh.
(15:34):
And comment has been really goodat helping me solve NA 10 issues
that I'm running into.
Uh, sometimes it'll be similarto what you experienced, like,
oh yeah, I solved it for you.
Like, okay, it still doesn'twork.
Ah, okay.
I know I done wrong and kindalike that.
Uh, but usually within three orfour attempts with the right
prompting and the rightreference materials, which is a
(15:55):
very big deal.
I'm gonna get to your firstpoint in a minute.
Uh, it actually helps me a lot.
So I do all of my NA 10 work inComet and it helps me solve a
lot of problems I don't know howto solve on my own.
Uh, and it's kind of funnybecause despite the fact I'm
asking it to do the thing everytime, it's like, oh, here's how
you do this.
I'm like, no, just do it.
I'm like, ah, okay.
And then it go, it goes and doesit.
(16:16):
And you have to do it everysingle time.
But it, but it works.
Uh, I am, I, now going back towhat I said in the beginning,
I'm shocked, literally shockedthat anything doesn't have a
built-in tool like this already.
Uh, and it could be a paidseparate tool, like, pay me 20
bucks a month to build.
Yeah, sure, no problem.
I'll pay it.
(16:36):
Uh, just to have that tool runand build automations and test
them and verify and everything.
So basically run your own MCPserver, know how to test it,
know how to install it, and soon.
Uh, I'm, I'm shocked it doesn'texist yet, but I will be really,
really surprised if it doesn'thappen in the near future.
But I've been saying that for atleast six months
Speaker 6 (16:55):
So how many of you used, uh, Claude skills?
Oh.
Okay.
So about about a month ago,when, when, when the new Claude
came out.
I don't know when that was, likex number of weeks ago.
And then we'll go to you,Richard.
But, uh, they came out withskills.
(17:17):
So what are skills?
Skills are like little pieces ofinstructions.
That could be anything becauseit's basically like a fancy
prompt with knowledge.
So think about, okay, the bestway to describe it is like a
library of prepackaged, customgpt or broad projects in this
(17:38):
case that it can pull at willwhenever it needs to get access
to it.
So think about, I think I showedyou this many times before.
Do you know, I, I showed youthat in a regular chat.
In chat GPT, you can use the atsymbol to call a custom GPT.
Did you know that?
(18:00):
Yes.
No.
No.
Okay.
So I'll show very, very quicklyand then I'll show you, uh, I'll
show you the Claude thing.
Cool.
Speaker 11 (18:08):
Do you see Chachi PT or am I sharing something else?
Speaker 6 (18:13):
Chachi PT Chachi.
Speaker 3 (18:14):
Okay.
Speaker 6 (18:14):
Yeah, so if you, if you hit the, a symbol in a
regular chat in Chachi pt,usually maybe it needs a little
Speaker 3 (18:21):
refresh to wake up.
Obviously, when you want to demosomething, it doesn't work.
Speaker 12 (18:36):
Yeah.
Uh, in theory.
I had the same thing.
Um, if I was trying to do it infive or five, one, I could only,
I'd have to change to 4.0 if Icreated the gpt.
No, no.
I've
Speaker 6 (18:50):
did this a gazillion times in the past two weeks.
So not in five one, but Idefinitely did this on five.
So let's go to, uh, GPT five.
Uh,
Speaker 3 (18:59):
thinking at
Speaker 6 (19:08):
Interesting.
Well try it out.
It, it conceptually works.
Like it gives you a, like adropdown menu of your, that
never happened to me that itdidn't work, so I'm wondering
why it's not working.
Maybe the update took the friendaway again.
Yeah, maybe.
Uh, so usually what happens whenyou hit the add sign, it opens
(19:32):
like a dropdown menu that showsyou all your didn't work for me
either.
Speaker 7 (19:36):
I just tried it.
Doesn't work.
It never de know it existed, butit doesn't do anything for me.
Speaker 6 (19:43):
Weird.
So nevermind.
But what it used to do is toallow you to maybe
Speaker 7 (19:48):
still in an old model,
Speaker 6 (19:50):
it allows you to use several different, what you're
saying in four.
Oh, it works.
Let's try just, just to show youthat it's working.
Uh, four.
Oh, it definitely
Speaker 12 (19:57):
worked in, in four.
Oh, but they didn't
Speaker 6 (20:02):
in five.
So it is working.
I used it in five for sure.
Uh, do you, but, but you cankinda see here, right?
So you can, you can add whateverthis, and then you can give it
whatever you want to give it.
And then the, the idea behind itwas that you can use several of
them in the single conversation.
So if you had like a three stepprocess, you can use the step
(20:24):
one and then the step two andthen the step three, uh, just by
adding it.
And then it knows what happenedbefore and it uses that as the
input for the next step.
Um, the, so Claude.
Uh, scales is kind of the sameconcept, only it knows on its
own to use the scales.
(20:44):
So if you look at cloud desktop,which is the way to get them or
create them, uh, if you go toyour settings and you have to
install cloud desktop, but ifyou go to your settings and you
go to, uh, capabilities and youscroll down, then you have
(21:08):
scales and then you need to turnthem on.
And then you can have all thesethings that it already comes up
with out of the box and you needto turn each and every one of
them on.
Um, canvas, is this why it'sdoing
Speaker 10 (21:20):
so good on your Excel and Word documents?
Is that like a new skill that ithas?
Speaker 6 (21:24):
So, so, yes.
But I think some of them are,are, are built in and you don't
even have to turn them on.
Speaker 10 (21:31):
Okay.
Speaker 6 (21:33):
So, but you can see two very interesting ones.
One is called MCP Builder andthe other is called Skill
Creator.
So if you have Skill Creator onand you tell it that you want to
create new skills, you literallyhave a regular conversation with
it and it will give you apackage that you can download.
Once you download that, you canupload it here and create a new
(21:54):
skill.
So I have a new skill of mybrand guidelines, the tone, I
like to use examples of that,the actual documents with my
color coding and so on.
So if I, if I now go here and Istart a new chat and I say, uh,
(22:16):
please create a document thatwill promote the Friday AI
Hangouts, which is a greatcommunity of people talking
about ai.
It needs to be no more than fiveparagraphs long.
Describe the benefits and whypeople should join.
Create it as a well-designedMicrosoft document.
(22:44):
Oh, sorry.
Stop.
Wait, you know what?
I, I actually wonder if whatI've done it anyway.
Uh, use my brand guidelinesthere.
You'll see what it's going todo.
Thinking, thinking aboutreferencing brand guidelines,
da, da, da, da.
(23:05):
Identify task requirements,prepared to consult brand, and
now it's reading the file'causeit knows reading, multiply i
brand guidelines directory tounderstand the brand identity.
So it's actually, so these,these, uh, scales are, are
packages.
They have instructions andknowledge base, just like all
the other stuff.
So in the knowledge base, it hasmy brand guidelines, the color
(23:28):
coding, the fonts that I use,all these kind of things.
It has my logo, so it can use mylogo because it's gonna write
HTML code and it can take thelogo and place it inside the
html.
It's like stuff like that.
Uh, but you can create them forliterally anything like you can
have, and it knows how to pullthem in real time.
Now, the biggest benefit ofputting them in real time is
that until it needs them, theydon't use tokens.
(23:49):
They don't take part of yourcontext window.
So it's freaking absolutelybrilliant.
So that's plot skills and it's,it's, again, relatively new, but
you can create them with thecreation scale inside of scale,
(24:10):
which is kind of meta.
This is, this is how it works.
And to call that scale, youliterally want to tell it.
You want to create a new skilland then you'll say, oh, I have
a skill to create skills.
Then you will, you'll load thatand then it knows how to create
the package.
But then you can download andre-upload, which it by itself is
a real stupid, like if I toldyou I wanna create a skill, I
don't need to download it andupload it, just save it wherever
(24:32):
it needs to be saved.
But that's a whole, a wholeother thing.
Uh, so here we go.
Let's see what it did.
And I will see that's gonna havethe colors of my shirt and stuff
like that in my logo and prettycool.
Right.
Speaker 8 (24:48):
And how did it get the context of that data?
Did it, you uploaded somethingbefore or it's going to your
website?
Yeah.
So
Speaker 6 (24:55):
you, by the way, you can see it here.
So Perfect.
I've created the f say, here'syour document.
The document includes brandalign, so it knows how to pull
the colors.
So these things come from thatscale that I created.
And the way I created the scaleis having a conversation, asking
it to create the scale.
(25:15):
And then I gave it the brandguidelines by tone of voice,
stuff like that.
It creates a package.
You re-upload the package,you're done.
And you can do this forliterally anything specific
formatting of invoices, like itshould literally, anything it
needs to know, you can create ascale and we'll call the scale
(25:35):
when it needs it, which isfreaking amazing.
GMT20251107-174847_Recordin (25:39):
Um, all right.
I need help from all of yougeniuses that work, uh,
developing and doing all thecool coding stuff and that
happen to know the evils ofMicrosoft.
So, I know I probably lost halfof you guys, but the other rest
that are interested in, um, wedevelop, we had to go back and
(26:00):
develop our own chat bot, andit's a technical chat bot for
all of our part numbers and allof our stuff.
And we're doing it in co-pilotstudio, which is it, it's not as
bad as it sounds.
It's, it's pretty good.
What we're finding is that I seeyou laughing Isa.
I'm, I'm not, I'm not, I, I hadgood and bad experiences with,
(26:23):
with Microsoft copilot, so, sowhat I'm finding is everything
went fantastic.
We uploaded the whole catalogin.
This is, uh, thrilling stuff.
I know all of you guys arelosing your minds over how
exciting this looks like.
I know.
(26:43):
I'll send you guys one.
Um, it has all of our partnumbers and each part number has
different variations of, weuploaded everything and then
halfway of reading the chart, itstarts to hallucinate and create
part numbers by adding a one.
(27:04):
So for example, part number 1 01 6 4, instead of oh four, it
says oh five.
The weirdest thing.
So we got in with the Microsoftfolks, uh, with the co-pilot
studio team.
They looked at it, they gave usa couple suggestions, it didn't
work.
And then they sent us to dealwith OCR.
(27:25):
We've been playing with OCR, butthat thing is monstrosity.
Uh, we're way over our headswith this.
Any suggestions?
Has anybody had experience withOCR?
So I, I do, but my question is,don't you have a digital version
of this thing?
(27:45):
Why do you need OCR?
Well, according to the, theco-pilot studio folks from
co-pilot studio, they say thatwe need to add or add this
toggle button in co-pilot studioto allow OCR as a connector so
that it can read that tablebetter.
(28:11):
But Oh, okay.
I'm, I'm, I'm going back.
Like I, you're talking about themaster document, right?
Yeah.
Like if the document, if youhave the raw data of the
document versus the finalversion, kinda like print ready
copy of the document.
Yeah.
Just use the raw data then post,yeah.
(28:32):
If we uploaded that, yeah.
We have the digital catalog.
For some reason it cannot read.
It stops reading like halfwaythe table, and that's when it
starts hallucinating.
Uh, how much data is on there?
Again, I, I don't, I don't, Iknow these are, uh, 450
(28:57):
something pages.
Yeah, but it shouldn't, youknow, I, I would, I would assume
it has nothing to do with OCRand again, I, I don't know
enough to answer, but OCRrelates to, I have a page that I
wanna scan, or I have a PDF filein, in its final flattened
(29:21):
format, that the data may not beas easy.
Yeah.
And I wanna read what's on thefile versus.
I have the source data in adigital format, like a CSV or
something like that, or directlyfor the team.
Yeah.
I assume you guys have the rawdata and not just the final
(29:41):
formatting for the book.
Yeah, and if that is the case,then it has nothing to do with
OCR because it doesn't need toread the page.
It actually consumes the data.
Mm-hmm.
And so my guess would be.
That so, so the way all thesetools work mm-hmm.
The very first step is calledchunking.
Mm-hmm.
(30:02):
And I don't want get tootechnical, but it's basically
taking your data and cutting itinto smaller pieces, pieces that
then the AI knows how toretrieve based on the vectors
that it's generating.
Yeah.
Knowing how to do chunking forlarge data is art.
Yeah.
Okay.
And if you don't know how to dothat, it is not going to work.
(30:25):
I literally just read a veryinteresting paper that does not
necessarily relate to whatyou're doing, but it relates to
what you're talking about on abigger picture of somebody that
was trying to get the AI to alsoknow what's in actually charts
and graphs outside of a fulldocument versus just the text
and what he did in step one.
(30:48):
Is that he created a metadatafor each and every one of the
charts, and it still didn'twork.
And what he did then he changedthe chunking to include the two
paragraphs before and the twoparagraphs after each chart.
So now it has the context ofwhat the chart is, of what it's,
(31:09):
you think about a book.
It's like, okay, you're talkingabout something, they're gonna
show us something.
Then sometimes you'll analyzethe thing.
It gives it more context onwhat's in the chart, so it could
actually now understand thechart better.
And so knowing how to dochunking, especially if you're
talking about 400 pages of rawdata, is the trick.
(31:29):
I don't know how to do that, butthere's people that, that's what
they do.
And if I had to guess, this isyour problem and not, mm-hmm.
The OCR aspect, because if youhave the raw data, then OCR is
not even a part of the equation.
Mm-hmm.
Okay.
So this reminds me of theproblem that you saw that you
and I, uh, connected on when Iwas trying to process that 2,500
(31:50):
page word, uh, PPDF document.
And even with the raw data, itwas almost impossible'cause we
couldn't find a platform thatwas able to handle the, the
amount of data.
So can, you can definitely chunkdata that is way, way bigger
than that.
Like people are creatingdatabases of like the entire
fricking universe, you know,think about open ai, right?
(32:12):
They're like, what the hell ischa pt?
It's like they, they took thedata off the internet and they
chunked it.
So it's, it's doable.
You just need to know how to doit.
And, and again, there.
People who, that's what theyknow how to do.
Uh, and the, the basic out ofthe box tools are great if you
wanna upload 20 pages, 30 pages,50 pages, a hundred pages.
(32:34):
Once you start getting to morethan that, or you start getting
into very, very specificinformation.
Like, again, think about thedifference between just creating
a chat bot that talks about, uh,HR knowledge.
How do I, you know, sign up fora PTO.
Versus I need the exact partnumber.
(32:55):
Mm-hmm.
It's a very different kind ofretrieval, which means it
requires a very different kindof chunking and data set up.
And, and again, sadly, I don't,I I, I'm open to anybody else if
you know somebody or a tool thatdoes it, but I don't
Hello and welcome to the Leveraging AI Podcast, the
(00:02):
podcast that shares practical,ethical ways to leverage AI to
improve efficiency, grow yourbusiness, and advance your
career.
This Isar Metis, your host, andwe have a special episode for
you today.
For over a year now, a communityof amazing business people from
all around the world who are allAI enthusiasts, who wants to
learn together, have beenmeeting every Friday for what we
(00:24):
call the Friday AI Hangouts.
There's usually 15 to 25 peopleevery week.
As I mentioned, different placesaround the world, different
sizes of companies, from largecorporations to solopreneurs and
everything in between, andeverybody brings ideas,
questions, solutions, thingsthey've learned, systems that
they failed, and so on, and it'sjust a fascinating conversation.
(00:48):
Every single week, it'scompletely open to anybody to
join.
And there are endless number ofgolden nuggets from these
meetings, and I thought it wouldbe interesting for you to listen
to segments out of these AIFriday Hangouts to see if that's
something you might beinterested in joining as well.
So this episode is going to besnippets out of different
(01:11):
community meetings.
Many of them are extremelytactical and valuable and you're
gonna learn a lot from them justlike you learn from our regular
Tuesday episodes.
But it will also give you anidea of what it feels like, so
you can decide if you want tojoin us or not.
If you do wanna join us, thereis a link in the show notes
click on the link register.
It will appear on your calendar,and then obviously you don't
(01:32):
have to join, but wheneveryou're available on Friday at
1:00 PM Eastern, you can dothat.
I will obviously like to seemany of you and to get to know
you better and to welcome you toour amazing community, and now
to the episode.
Speaker 10 (01:45):
I wonder if anybody has any experience, um, sort of
taking what they did in a, likein a custom GPT into the new GPT
cus uh, chat GPT projects.
And the context is this, I builtsort of like a prototype for a
client.
Um, they had new apprenticestarting, they've never had an
apprentice before and they hadno clue how they were gonna do
with it, deal with it.
(02:07):
And so I ended up finding likewhat the apprentice's like three
year plan was and put that intothe GPT and then I found out
what systems they used in houseand created this thing that they
can talk to it, figure out howto teach the person, give it to
the kid to learn his things andstuff, and.
Now, now they love it and, um,are gonna wanna use it and build
on it.
And I realized that it would beso much better to have it in
(02:27):
this new, like, chat GBTprojects because then everything
stays in the same place and Ican control it, but have more
different users.
So I'm just wondering, do theinstructions still work as well?
Has anybody tried thatmigrating?
Speaker 3 (02:40):
I will let other people answer before I answer
Anyone?
No
Speaker 6 (02:47):
one.
Okay.
So a few, how many of you needan overview of projects in
general before we dive into whatare the differences?
Okay, so the, the biggest, uh,uh, two seconds on the biggest
difference between projects andcustom gpt.
Uh, custom gpt followinstructions and do them.
And that's it.
(03:07):
And that's the end of the thing.
And a project is a think about aproject as a mini private
version of Chachi pt, and inthat mini private version, you
can give it its own universalcontext.
You can give it special set ofinstructions, but very different
(03:30):
than a Chachi pt.
You can have conversation basic,basically a regular chat, just
like a regular chat with GPTonly inside your little
universe.
And it has also,
Speaker 4 (03:40):
it has its own memory.
So it has its own memory.
Speaker 6 (03:44):
Here's where I'm going.
Uh, it has many benefits rightnow over custom gpt.
Uh, the first one is you canhave free form conversations
with it versus in chat GPT, itjust follows the instructions,
right?
You can't just have a regularconversation in there.
Uh, so that's a big benefit.
The other big benefit is thatyou can upload more files than
(04:06):
you can upload into custom GPTsinside projects, which is
obviously a big deal.
Uh, the third thing, which ishuge is what Jonathan just said.
It has its own memory in its ownlittle universe, which means in
addition to the information thatyou gave it, when you trained it
to do the thing, and you gave itaccess to files, it learns over
(04:27):
time because you, you give itmore information as you're
having conversations with it.
So as of right now.
There's really no reason tocreate new custom gpt.
It makes a lot more sense tobuild them inside projects
because of all the things that Ijust said.
Uh, those who already exist incustom gpt, the answer is it
(04:49):
depends.
And the reason it depends.
If it's working, it's working.
Like there's no, there's noreason to move it over.
Like if it's doing the thingthat it needs to do, let it do
the thing that it needs to do.
But if it, if it doesn't, or ifyou wanted to have a, the
ability to have free formconversations with that in mind,
or you want it to learn overtime or all these kind of
things, then literally all youhave to do is upload the same
(05:12):
files on the project side, uh,copy the instructions to the
instruction sections, and, andthat's it.
And you're, you're off to theraces.
So you, it, it, it can work.
And if you think about it,projects in Claude.
That's what they've done for awhile.
Like in Claude, you don't havethese two different animals.
(05:33):
There's one thing which isprojects which you can use, kind
like projects in chat PT or likecustom GPS in chat pt, depending
how you want to use it.
And so I think when Chachi PTcame up with project was to
partially mimic, but they shouldhave just converted cha GP like
custom gpt into projects.
(05:54):
And I think they just wannaconfuse people go like, whoa,
where about Cha g pt?
Like custom gpt go.
So, uh, right now it's kindalike duplicate, but the only
Speaker 10 (06:04):
thing is that you can't, you can't do connectors
or something right at the momentwith these new projects.
So there's
Speaker 6 (06:09):
a few things.
There's a few things you cannotdo.
Uh, like the, the only realbenefit of custom gpt right now
is the whole section of thebottom where you can write code
and do all these kind of things.
Uh, that does not exist insideof projects yet.
Uh, but again, I, I, I thinkwhat makes sense is that they
would merge the two things intoone thing that has all this
(06:30):
functionality.
Uh, but I think that's the onlything right now that doesn't
exist.
Speaker 10 (06:36):
Okay.
And have you tried sharing itwith people and using the
different permissions and stuff?
Oh, yeah.
Speaker 6 (06:39):
Hold on.
So, so that's the biggest thingyou cannot share.
Oh, no.
Now you can share everything.
They changed it like a coupleweeks ago.
Yeah, exactly.
Yeah, it's brand new.
It just changed.
I share like two or three weeksago you couldn't share projects.
You could only share custom gptand now you can share projects
as well.
You can even share them.
(07:01):
I'm confusing, I'm using so muchof Claude recently in Claude.
You can share projects withanyone, like you can share
projects with the world and theydon't see the instructions and
the data behind it, but they canuse the project in chat.
GPTI don't remember how they didit, but I know you can now share
projects in chat GPT that youcouldn't a few weeks ago.
So yeah, you, you can shareprojects.
Speaker 10 (07:23):
Okay, well I'll, I'll, I'll keep working on this
and I, I will report back how itgoes.
Awesome.
'cause I also wanna like havedifferent permission levels.
I think it would be good if theApprentice can use it, but only
in study mode or something.
Right.
Whereas the people that have towork.
Yeah.
So I
Speaker 6 (07:37):
think in that case, you probably want to create
another version of it.
Maybe that would be for theApprentice.
And I think that would make it alot easier to control where you
give it slightly differentinstructions.
Speaker 7 (07:47):
Yeah.
We'll see.
Okay.
Thank you.
Speaker 10 (07:49):
And just before, before while I still have the
floor, I just wanted to saythank you to Ken for sending me
a podcast after the last, um,session, which was really
helpful to my last, uh, awesomela last, last week's issue.
Thank you, Ken.
Speaker 6 (08:03):
This is not about me.
This is about us all helpingeach other and there's nothing
that makes me happier than, youknow, uh, that if I die
tomorrow, this team can continuebeing helpful to another,
Speaker 38 (08:14):
Actually I've been listening to a lot of your, um,
podcasts, uh, every week and.
Trying to use the tools and, uh,the frustration level is
unbelievable.
Um, it took Gemini to write me ascript to merge a bunch of doc
Google docs into one doc 38script attempts, and everyone
(08:38):
said, I got it right now I gotit right.
It here to just cut and pastethis into the script.
It's the JSON, no problem.
Oh, it didn't work.
Oh, send me the error stuff.
Okay, here's the error.
Oh, I know what went wrong.
This'll fix it.
38 attempts.
So the frustration of thesethings just don't seem to be
ready for prime time every, Idid another one where I asked it
to create a prompt where I couldprepare for a meeting for
(09:01):
someone.
You, one of the use cases thatyou, right.
So I said, gave me this promptand, uh, so I gave me the
prompt, I gave it the person'sURL, uh, on LinkedIn, their
name, all this background, andit went and gave me a wonderful
brief.
Snapshot brief of the, you know,questions.
I could ask this person to startthe conversation except it was
(09:21):
the wrong person.
And then when I questioned itand said, oh, I can't look at
LinkedIn profiles, that's allprivate.
Well what if I gave my login?
Oh no, I can't do that either.
So the frustration level I'mhaving actually accomplishing
any task that I'm trying toautomate is unbelievable and
trying to learn any to end.
(09:43):
And your last podcast you hadlast week with the guest for the
French lady, like this is, uh,when she said the learning curve
is steep, it's unbelievable.
And I'm a relative.
I've been in, you know it a longtime, relatively savvy guy.
This is, but I'm not a coder.
It is unbelievably steeplearning curve.
And I don't know what thegroup's experience has been, but
(10:06):
the consensus seems to be thatdon't go with the easy ones like
Zapier make go with the harderone.
But it is a way lot morecomplex.
Speaker 39 (10:16):
Oh, um, okay.
So I mean, ma make.com issignificant, is still e even the
other ones have a bit of alearning curve.
So if, if, uh, NAN is consideredhard at that point, then, uh,
yeah, I'd definitely look at,uh, you know, make.com is
actually really nicely, uh, um,laid out and, uh, it's fairly
(10:41):
and it's very powerful.
Speaker 30 (10:43):
Yeah, so I agree.
I, I always tell people, getstarted, start with make, uh,
from a value to investment.
It is by far the best one rightnow.
Uh, it is not as flexible andcapable as NA 10.
It is way easier to use anddefinitely way easier to get
started.
Uh, even just connectingaccounts, like connecting your
(11:07):
Gmail account or your MicrosoftOutlook account make.
Is, you go in, you give it yourpassword, and you're done,
you're connected doing it in, inNA 10, you go to Google Cloud or
to Microsoft Azure, and you haveto open a project and you have
to assign permissions, and youhave like, it's freaking, it's
like 20 steps that you need tobe an manager to do.
(11:29):
So, a hundred percent.
Like if, if you're not exactly,if you're not now in the process
of I'm gonna automate my entirebusiness and it's gonna save me
a lot of money, uh, doing itthrough NA 10, just go with
make, uh, there, there, I'm, I'm
Speaker 39 (11:41):
actually in the process of automating a lot of
our business.
Uh, but I'm designing it all inMake first.
Yeah.
Yeah.
And once I have all of that inmake, I am going after the ones
that use the most operations andI'm gonna move those over, uh,
(12:02):
first and use, you know, use itas an optimization problem.
But designing it and make is soeasy and so fast that it's just
to get stuff out there as a, asa initial tool.
It's great.
Speaker 38 (12:17):
It seems now I understand why there's a whole
business out there for people tobuild your N eight N agent for
you.
'cause it's so complicated.
So there's the buy, buy, thebuild it, there's the build it
yourself.
Um, and then there's an easierway to build it yourself
apparently.
I'm wondering, is there anythingwhere it's pre-built, you just
saw, you know, we all want togo, I wanna check my emails, I
wanna monitor my emails, and ifthis email comes in, I wanna do
(12:39):
X with that email.
I wanna reply and I wannacustomize reply.
So I need a little bit of abrain.
So I might have agent kit, butlike what's been pre-written,
these seem like pretty standardbus business processes.
Business is simple.
Get the business, do thebusiness, get paid a lot of
similar processes.
Is there a library out there ofpre-written and in agents that
(13:01):
do specific things?
Speaker 30 (13:04):
A hundred percent.
So both on Make and on Zapierand on NA 10 and all of those,
there are thousands of existingtemplates.
Most of them are free.
Uh, and some of them will costyou, okay, pay me 30 bucks.
I'm like, okay, I'll pay you 30bucks to save myself.
The three hours of agony.
(13:24):
And what I do in many, manycases is I mix and match
segments of these templates.
So there are long templates.
It does all these things.
I'm like, okay, I just needthis.
But now I have this buildingblock that I can reuse.
And then there's this otherthing that does the other part
that I need, and all I have tofigure out is how to connect
(13:44):
them and make them talk to oneanother.
And the same is true in Make andin NA 10.
Uh, yeah.
Literally thousands of peopleare, are sharing them.
Oh, a lot more than thousands.
Probably tens of thousands ofpeople are sharing them on the
internet.
And you can, if you just search,uh, for template.
This.
Now, the, the thing that I'mstill puzzled by, and I'm sure
(14:06):
it's coming and there's, there'sthird party attempts, but not
inside of NA 10 and make,there's zero reason right now.
Why you shouldn't be able to sayin English what you want the
system to do, and then it willjust create the automation for
you.
A to Z soup to nuts, the wholething.
Test it out.
Verify, check that it's working.
(14:27):
Ask me for what you need fromme.
Like I need your API key forthis.
Here's the instructions on howto go and get it.
I need your login to this.
Just gimme your login.
I need, and you will give it allthis stuff and it will run, uh,
the close, the two closestthings that I have right now
that I'm using.
And I shared that with you inthe past, but since you're new,
I'm gonna share it again.
Uh, one is I have a Claudeproject.
(14:50):
That has an NA 10 MCP serverwith a whole set of instructions
on how to use the MCP serverthat I never got to work a
hundred percent, but it gets me85% there, which is way better
than I can do on my own in fiveminutes.
Uh, it takes a little bit oftime, but it's not a big deal.
Like to get the MCP set up andconnecting it to Claude and, and
(15:13):
then running it.
Uh, so this is one goodshortcut.
Again, it's not perfect, butit's gonna save you X percent.
Um, the other one is I now do100% of my entertainment work in
the Comet browser that is nowfree and available to everybody,
so you don't have to pay or be amember of anything.
Uh.
(15:34):
And comment has been really goodat helping me solve NA 10 issues
that I'm running into.
Uh, sometimes it'll be similarto what you experienced, like,
oh yeah, I solved it for you.
Like, okay, it still doesn'twork.
Ah, okay.
I know I done wrong and kindalike that.
Uh, but usually within three orfour attempts with the right
prompting and the rightreference materials, which is a
(15:55):
very big deal.
I'm gonna get to your firstpoint in a minute.
Uh, it actually helps me a lot.
So I do all of my NA 10 work inComet and it helps me solve a
lot of problems I don't know howto solve on my own.
Uh, and it's kind of funnybecause despite the fact I'm
asking it to do the thing everytime, it's like, oh, here's how
you do this.
I'm like, no, just do it.
I'm like, ah, okay.
And then it go, it goes and doesit.
(16:16):
And you have to do it everysingle time.
But it, but it works.
Uh, I am, I, now going back towhat I said in the beginning,
I'm shocked, literally shockedthat anything doesn't have a
built-in tool like this already.
Uh, and it could be a paidseparate tool, like, pay me 20
bucks a month to build.
Yeah, sure, no problem.
I'll pay it.
(16:36):
Uh, just to have that tool runand build automations and test
them and verify and everything.
So basically run your own MCPserver, know how to test it,
know how to install it, and soon.
Uh, I'm, I'm shocked it doesn'texist yet, but I will be really,
really surprised if it doesn'thappen in the near future.
But I've been saying that for atleast six months
Speaker 6 (16:55):
So how many of you used, uh, Claude skills?
Oh.
Okay.
So about about a month ago,when, when, when the new Claude
came out.
I don't know when that was, likex number of weeks ago.
And then we'll go to you,Richard.
But, uh, they came out withskills.
(17:17):
So what are skills?
Skills are like little pieces ofinstructions.
That could be anything becauseit's basically like a fancy
prompt with knowledge.
So think about, okay, the bestway to describe it is like a
library of prepackaged, customgpt or broad projects in this
(17:38):
case that it can pull at willwhenever it needs to get access
to it.
So think about, I think I showedyou this many times before.
Do you know, I, I showed youthat in a regular chat.
In chat GPT, you can use the atsymbol to call a custom GPT.
Did you know that?
(18:00):
Yes.
No.
No.
Okay.
So I'll show very, very quicklyand then I'll show you, uh, I'll
show you the Claude thing.
Cool.
Speaker 11 (18:08):
Do you see Chachi PT or am I sharing something else?
Speaker 6 (18:13):
Chachi PT Chachi.
Speaker 3 (18:14):
Okay.
Speaker 6 (18:14):
Yeah, so if you, if you hit the, a symbol in a
regular chat in Chachi pt,usually maybe it needs a little
Speaker 3 (18:21):
refresh to wake up.
Obviously, when you want to demosomething, it doesn't work.
Speaker 12 (18:36):
Yeah.
Uh, in theory.
I had the same thing.
Um, if I was trying to do it infive or five, one, I could only,
I'd have to change to 4.0 if Icreated the gpt.
No, no.
I've
Speaker 6 (18:50):
did this a gazillion times in the past two weeks.
So not in five one, but Idefinitely did this on five.
So let's go to, uh, GPT five.
Uh,
Speaker 3 (18:59):
thinking at
Speaker 6 (19:08):
Interesting.
Well try it out.
It, it conceptually works.
Like it gives you a, like adropdown menu of your, that
never happened to me that itdidn't work, so I'm wondering
why it's not working.
Maybe the update took the friendaway again.
Yeah, maybe.
Uh, so usually what happens whenyou hit the add sign, it opens
(19:32):
like a dropdown menu that showsyou all your didn't work for me
either.
Speaker 7 (19:36):
I just tried it.
Doesn't work.
It never de know it existed, butit doesn't do anything for me.
Speaker 6 (19:43):
Weird.
So nevermind.
But what it used to do is toallow you to maybe
Speaker 7 (19:48):
still in an old model,
Speaker 6 (19:50):
it allows you to use several different, what you're
saying in four.
Oh, it works.
Let's try just, just to show youthat it's working.
Uh, four.
Oh, it definitely
Speaker 12 (19:57):
worked in, in four.
Oh, but they didn't
Speaker 6 (20:02):
in five.
So it is working.
I used it in five for sure.
Uh, do you, but, but you cankinda see here, right?
So you can, you can add whateverthis, and then you can give it
whatever you want to give it.
And then the, the idea behind itwas that you can use several of
them in the single conversation.
So if you had like a three stepprocess, you can use the step
(20:24):
one and then the step two andthen the step three, uh, just by
adding it.
And then it knows what happenedbefore and it uses that as the
input for the next step.
Um, the, so Claude.
Uh, scales is kind of the sameconcept, only it knows on its
own to use the scales.
(20:44):
So if you look at cloud desktop,which is the way to get them or
create them, uh, if you go toyour settings and you have to
install cloud desktop, but ifyou go to your settings and you
go to, uh, capabilities and youscroll down, then you have
(21:08):
scales and then you need to turnthem on.
And then you can have all thesethings that it already comes up
with out of the box and you needto turn each and every one of
them on.
Um, canvas, is this why it'sdoing
Speaker 10 (21:20):
so good on your Excel and Word documents?
Is that like a new skill that ithas?
Speaker 6 (21:24):
So, so, yes.
But I think some of them are,are, are built in and you don't
even have to turn them on.
Speaker 10 (21:31):
Okay.
Speaker 6 (21:33):
So, but you can see two very interesting ones.
One is called MCP Builder andthe other is called Skill
Creator.
So if you have Skill Creator onand you tell it that you want to
create new skills, you literallyhave a regular conversation with
it and it will give you apackage that you can download.
Once you download that, you canupload it here and create a new
(21:54):
skill.
So I have a new skill of mybrand guidelines, the tone, I
like to use examples of that,the actual documents with my
color coding and so on.
So if I, if I now go here and Istart a new chat and I say, uh,
(22:16):
please create a document thatwill promote the Friday AI
Hangouts, which is a greatcommunity of people talking
about ai.
It needs to be no more than fiveparagraphs long.
Describe the benefits and whypeople should join.
Create it as a well-designedMicrosoft document.
(22:44):
Oh, sorry.
Stop.
Wait, you know what?
I, I actually wonder if whatI've done it anyway.
Uh, use my brand guidelinesthere.
You'll see what it's going todo.
Thinking, thinking aboutreferencing brand guidelines,
da, da, da, da.
(23:05):
Identify task requirements,prepared to consult brand, and
now it's reading the file'causeit knows reading, multiply i
brand guidelines directory tounderstand the brand identity.
So it's actually, so these,these, uh, scales are, are
packages.
They have instructions andknowledge base, just like all
the other stuff.
So in the knowledge base, it hasmy brand guidelines, the color
(23:28):
coding, the fonts that I use,all these kind of things.
It has my logo, so it can use mylogo because it's gonna write
HTML code and it can take thelogo and place it inside the
html.
It's like stuff like that.
Uh, but you can create them forliterally anything like you can
have, and it knows how to pullthem in real time.
Now, the biggest benefit ofputting them in real time is
that until it needs them, theydon't use tokens.
(23:49):
They don't take part of yourcontext window.
So it's freaking absolutelybrilliant.
So that's plot skills and it's,it's, again, relatively new, but
you can create them with thecreation scale inside of scale,
(24:10):
which is kind of meta.
This is, this is how it works.
And to call that scale, youliterally want to tell it.
You want to create a new skilland then you'll say, oh, I have
a skill to create skills.
Then you will, you'll load thatand then it knows how to create
the package.
But then you can download andre-upload, which it by itself is
a real stupid, like if I toldyou I wanna create a skill, I
don't need to download it andupload it, just save it wherever
(24:32):
it needs to be saved.
But that's a whole, a wholeother thing.
Uh, so here we go.
Let's see what it did.
And I will see that's gonna havethe colors of my shirt and stuff
like that in my logo and prettycool.
Right.
Speaker 8 (24:48):
And how did it get the context of that data?
Did it, you uploaded somethingbefore or it's going to your
website?
Yeah.
So
Speaker 6 (24:55):
you, by the way, you can see it here.
So Perfect.
I've created the f say, here'syour document.
The document includes brandalign, so it knows how to pull
the colors.
So these things come from thatscale that I created.
And the way I created the scaleis having a conversation, asking
it to create the scale.
(25:15):
And then I gave it the brandguidelines by tone of voice,
stuff like that.
It creates a package.
You re-upload the package,you're done.
And you can do this forliterally anything specific
formatting of invoices, like itshould literally, anything it
needs to know, you can create ascale and we'll call the scale
(25:35):
when it needs it, which isfreaking amazing.
GMT20251107-174847_Recordin (25:39):
Um, all right.
I need help from all of yougeniuses that work, uh,
developing and doing all thecool coding stuff and that
happen to know the evils ofMicrosoft.
So, I know I probably lost halfof you guys, but the other rest
that are interested in, um, wedevelop, we had to go back and
(26:00):
develop our own chat bot, andit's a technical chat bot for
all of our part numbers and allof our stuff.
And we're doing it in co-pilotstudio, which is it, it's not as
bad as it sounds.
It's, it's pretty good.
What we're finding is that I seeyou laughing Isa.
I'm, I'm not, I'm not, I, I hadgood and bad experiences with,
(26:23):
with Microsoft copilot, so, sowhat I'm finding is everything
went fantastic.
We uploaded the whole catalogin.
This is, uh, thrilling stuff.
I know all of you guys arelosing your minds over how
exciting this looks like.
I know.
(26:43):
I'll send you guys one.
Um, it has all of our partnumbers and each part number has
different variations of, weuploaded everything and then
halfway of reading the chart, itstarts to hallucinate and create
part numbers by adding a one.
(27:04):
So for example, part number 1 01 6 4, instead of oh four, it
says oh five.
The weirdest thing.
So we got in with the Microsoftfolks, uh, with the co-pilot
studio team.
They looked at it, they gave usa couple suggestions, it didn't
work.
And then they sent us to dealwith OCR.
(27:25):
We've been playing with OCR, butthat thing is monstrosity.
Uh, we're way over our headswith this.
Any suggestions?
Has anybody had experience withOCR?
So I, I do, but my question is,don't you have a digital version
of this thing?
(27:45):
Why do you need OCR?
Well, according to the, theco-pilot studio folks from
co-pilot studio, they say thatwe need to add or add this
toggle button in co-pilot studioto allow OCR as a connector so
that it can read that tablebetter.
(28:11):
But Oh, okay.
I'm, I'm, I'm going back.
Like I, you're talking about themaster document, right?
Yeah.
Like if the document, if youhave the raw data of the
document versus the finalversion, kinda like print ready
copy of the document.
Yeah.
Just use the raw data then post,yeah.
(28:32):
If we uploaded that, yeah.
We have the digital catalog.
For some reason it cannot read.
It stops reading like halfwaythe table, and that's when it
starts hallucinating.
Uh, how much data is on there?
Again, I, I don't, I don't, Iknow these are, uh, 450
(28:57):
something pages.
Yeah, but it shouldn't, youknow, I, I would, I would assume
it has nothing to do with OCRand again, I, I don't know
enough to answer, but OCRrelates to, I have a page that I
wanna scan, or I have a PDF filein, in its final flattened
(29:21):
format, that the data may not beas easy.
Yeah.
And I wanna read what's on thefile versus.
I have the source data in adigital format, like a CSV or
something like that, or directlyfor the team.
Yeah.
I assume you guys have the rawdata and not just the final
(29:41):
formatting for the book.
Yeah, and if that is the case,then it has nothing to do with
OCR because it doesn't need toread the page.
It actually consumes the data.
Mm-hmm.
And so my guess would be.
That so, so the way all thesetools work mm-hmm.
The very first step is calledchunking.
Mm-hmm.
(30:02):
And I don't want get tootechnical, but it's basically
taking your data and cutting itinto smaller pieces, pieces that
then the AI knows how toretrieve based on the vectors
that it's generating.
Yeah.
Knowing how to do chunking forlarge data is art.
Yeah.
Okay.
And if you don't know how to dothat, it is not going to work.
(30:25):
I literally just read a veryinteresting paper that does not
necessarily relate to whatyou're doing, but it relates to
what you're talking about on abigger picture of somebody that
was trying to get the AI to alsoknow what's in actually charts
and graphs outside of a fulldocument versus just the text
and what he did in step one.
(30:48):
Is that he created a metadatafor each and every one of the
charts, and it still didn'twork.
And what he did then he changedthe chunking to include the two
paragraphs before and the twoparagraphs after each chart.
So now it has the context ofwhat the chart is, of what it's,
(31:09):
you think about a book.
It's like, okay, you're talkingabout something, they're gonna
show us something.
Then sometimes you'll analyzethe thing.
It gives it more context onwhat's in the chart, so it could
actually now understand thechart better.
And so knowing how to dochunking, especially if you're
talking about 400 pages of rawdata, is the trick.
(31:29):
I don't know how to do that, butthere's people that, that's what
they do.
And if I had to guess, this isyour problem and not, mm-hmm.
The OCR aspect, because if youhave the raw data, then OCR is
not even a part of the equation.
Mm-hmm.
Okay.
So this reminds me of theproblem that you saw that you
and I, uh, connected on when Iwas trying to process that 2,500
(31:50):
page word, uh, PPDF document.
And even with the raw data, itwas almost impossible'cause we
couldn't find a platform thatwas able to handle the, the
amount of data.
So can, you can definitely chunkdata that is way, way bigger
than that.
Like people are creatingdatabases of like the entire
fricking universe, you know,think about open ai, right?
(32:12):
They're like, what the hell ischa pt?
It's like they, they took thedata off the internet and they
chunked it.
So it's, it's doable.
You just need to know how to doit.
And, and again, there.
People who, that's what theyknow how to do.
Uh, and the, the basic out ofthe box tools are great if you
wanna upload 20 pages, 30 pages,50 pages, a hundred pages.
(32:34):
Once you start getting to morethan that, or you start getting
into very, very specificinformation.
Like, again, think about thedifference between just creating
a chat bot that talks about, uh,HR knowledge.
How do I, you know, sign up fora PTO.
Versus I need the exact partnumber.
(32:55):
Mm-hmm.
It's a very different kind ofretrieval, which means it
requires a very different kindof chunking and data set up.
And, and again, sadly, I don't,I I, I'm open to anybody else if
you know somebody or a tool thatdoes it, but I don't
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.
Crime Junkie
Does hearing about a true crime case always leave you scouring the internet for the truth behind the story? Dive into your next mystery with Crime Junkie. Every Monday, join your host Ashley Flowers as she unravels all the details of infamous and underreported true crime cases with her best friend Brit Prawat. From cold cases to missing persons and heroes in our community who seek justice, Crime Junkie is your destination for theories and stories you won’t hear anywhere else. Whether you're a seasoned true crime enthusiast or new to the genre, you'll find yourself on the edge of your seat awaiting a new episode every Monday. If you can never get enough true crime... Congratulations, you’ve found your people. Follow to join a community of Crime Junkies! Crime Junkie is presented by Audiochuck Media Company.
The Brothers Ortiz
The Brothers Ortiz is the story of two brothers–both successful, but in very different ways. Gabe Ortiz becomes a third-highest ranking officer in all of Texas while his younger brother Larry climbs the ranks in Puro Tango Blast, a notorious Texas Prison gang. Gabe doesn’t know all the details of his brother’s nefarious dealings, and he’s made a point not to ask, to protect their relationship. But when Larry is murdered during a home invasion in a rented beach house, Gabe has no choice but to look into what happened that night. To solve Larry’s murder, Gabe, and the whole Ortiz family, must ask each other tough questions.