August 12, 2025 • 33 mins
GPT-5 has arrived with meaningful improvements in logical reasoning, processing speed, and the ability to dynamically switch between models based on task complexity! The technology excels at pattern recognition in data analysis while software developers are finding new ways to integrate it into their workflow without sacrificing quality.
Mike sits down with Punchmark Back-End Developer, Andy Szoke, to discuss what the arrival of this new version means, how to balance your humanity with productivity, and the basics of prompt generation.
Learn more about our sponsor, Jewel-Craft: jewel-craft.com
Send feedback or learn more about the podcast: punchmark.com/loupe
Learn about Punchmark's website platform: punchmark.com
Inquire about sponsoring In the Loupe and showcase your business on our next episode: podcast@punchmark.com
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Speaker 1 (00:18):
Welcome to In the Loop.
Two years since the last modelrelease and this one is, of
course, a huge step up and, Imust admit, I don't even fully
understand it.
So I brought on my friend andalso my coworker, andy Zoki, who
is a backend developer forPunchmark.
He's also one of the mostplugged in guys with regards to,
(00:39):
you know, chatgpt and thedevelopment of AI, and he stays
on top of this stuff and uses itin ways that I couldn't even
imagine and I asked him.
We had a real nice conversationdiscussing what the new changes
mean, what it's capable of, theways that we use it in our own
life, and I think it's just amore casual conversation,
(01:01):
because Andy's one of my bestfriends and he's also my
co-worker and hearing how heuses it maybe informs me on how
I can use it better.
If you're not comfortable withusing AI, no problem.
This might not be the episodefor you, but if you are kind of
dabbling, this, one kind of getsinto some deeper weeds and
maybe you enjoy that.
Speaker 2 (01:21):
So everybody enjoy and maybe you enjoy that.
So everybody enjoy.
This episode is brought to youby Punchmark, the jewelry
industry's favorite websiteplatform and digital growth
agency.
Our mission reaches way beyondtechnology.
With decades of experience andlong-lasting industry
relationships, punchmark enablesjewelry businesses to flourish
(01:41):
in any marketplace.
We consider our clients ourfriends, as many of them have
been friends way before becomingclients.
Punchmark's own success comesfrom the fact that we have a
much deeper need and obligationto help our friends succeed.
Whether you're looking forbetter e-commerce performance,
business growth or campaignsthat drive traffic and sales,
punchmark's website andmarketing services were made
(02:03):
just for you.
It's never too late totransform your business and
stitch together your digital andphysical worlds in a way that
achieves tremendous growth andresults.
Schedule a guided demo today atpunchmarkcom.
Slash go.
Speaker 1 (02:19):
And now back to the show.
What is up everybody?
I'm joined by Andy Zoki, mygood buddy and also back-end
developer at Punchmark.
How are you doing today, andy?
I'm doing pretty good.
What's up, mike?
Doing so well.
Really cool getting a chance tohave you on.
You were on when we weretalking about VR and you've
(02:43):
always been very tech forwardwhen it comes to, you know, I
guess, adopting new technologiesand kind of.
You follow the state of AI verywell.
But, big news, it actuallyhappened yesterday when we were
recording this.
But when this releases it'llhave been a few days, but GPT-5
just released.
Can you give like a pseudooverview on what GPT-5 is for
(03:06):
people listening?
Speaker 3 (03:08):
Sure, yeah.
So GPT-5 is OpenAI's latestiteration of their chat, gpt,
llm, large language model andit's you know.
You've heard of AI, you'veheard of all these things that
it can do.
Gpt-5 is the latest iteration,looking at the initial rollout,
(03:28):
and this is bleeding edge.
I literally just got access toit this morning, so I haven't
had too much of a chance to playaround with it, and that's
going to be coming over the nextcouple of weeks, but it looks
like it's moderately better atcoding tasks.
It's slightly better logicalreasoning Maybe not as much of
an exponential rise as somepeople might have been hoping
(03:49):
for.
One thing that we've seen withthe most recent releases is it's
just taking a lot more time andmoney than people were hoping
for.
So it's one barrier thatthey're actively working on,
obviously, but, yeah, it's justa little bit of a better thinker
.
One thing that people aresaying off the bat that they
aren't necessarily fans of isthe personality is a bit drier
(04:11):
and, uh, it could be more curt.
Um, I've heard, uh I saw oneperson on reddit describe it as
a overworked secretary.
That's just kind of exasperated.
Talking to you, it's maybe notthat bad, but it definitely has
a little bit more of attentionto that than uh its predecessors
, uh gpt 4.5 and 03 yeah,because so there was.
Speaker 1 (04:33):
There's different gpts, uh, that are tailored to
different types of tasks, likethere's 03, and then there's 03,
mini, and then there's, uh,there's four.
Um, I have to be honest,because we were talking about
this probably about a month agoand you're like, oh yeah, you're
, you're using four, you're noteven using the advanced version
that you're paying for.
(04:54):
I was like, oh, I didn'trealize you had to do that.
Apparently, with gpt5, it willdynamically switch between the
models, uh, depending on whatyour task requires.
So if you need a deep thinkingone, it can do that, and if you
need a you know, just a quickanswer, then it'll do that as
well.
Do you switch between models,or were you switching between
(05:14):
the models when you're using it?
Speaker 3 (05:16):
Yeah, I've always been a big fan of O3.
I think that's just been themost capable model for most
complex tasks.
I use it as a developer.
I know lots of developers useChatGPT.
You've probably heard vibecoding, you know it can be an
insult.
On the far end it's justtelling ChatGPT to write
everything and then you justcopy and paste it into your code
(05:37):
base and you ship a millionbugs and, you know, make a lot
of people unhappy.
We would never do that, ofcourse not.
No, and we don't.
Obviously, the correct way touse it is almost as if it's a
colleague where you have ideasabout how you want to do
something and you say I havethese classes, I'm trying to put
this architecture together, butI want to optimize it for low
(05:59):
cost, low latency, that kind ofthing.
It's really good at taking allthose different pieces and
churning through them andputting it together, and 5 is a
little bit better at that.
There are still a few moremodels that you can split out.
So if you're on a paid version,you still get the model picker
at the top left where you canopen GPT-5, it says flagship and
(06:20):
then for other models you canpick between uh, or if you're a
pro user you can do pro.
So it does split it out by costa bit, um, but all of them are
a bit faster, a bit smarter and,uh, the base one, the gpt5, uh,
non-thinking, it's super fastand it's still very good at
thinking.
I'm impressed.
Like I have my, I send a promptand before I pull my uh finger
(06:42):
off the enter key, it's alreadystarted writing the first
response in some cases.
So it's crazy.
Speaker 1 (06:47):
I think it's been really interesting the way that
I've been using it.
Is it used to be that I wouldgo to ChatGPT and I would like,
oh this is something you couldhelp me with, and then I would
kind of interface with it andnow, like an oracle, you know,
I'd have to go to it and I wouldask it and then it would answer
(07:07):
me.
Now what I do is I just have itdownloaded to my desktop.
I just keep one of the windowsopen.
I have a dual window set up inmy workstation, so I have a
laptop, a second screen.
On my second screen, I haveSlack for work, I have ChatGPT
and I have Spotify.
Like.
Those are the essentials atthis point, and things that it's
(07:28):
good for are drawingconclusions and drawing
inferences from data sheets.
Especially is what I've beenhaving really good success with.
I just did the episode before.
This one was very much helpedby chat GBT, where you know, oh,
you're the one that built methe leaderboard exports.
(07:49):
Remember that back in back inthe day.
Oh, yeah, the finance.
That's right For the e-commerceperformance reports, and what's
super funny is I can look at itand, you know, draw inference
on how e-commerce is doing.
But there's you know howevermany columns there's like 70
(08:10):
columns in that thing and theneach column has like 120 rows
and you know, blah, blah, blah.
There's all these things.
It's every single month.
I can only draw so muchanalysis from it, just because
my brain has like a speed cap onit.
You know it can only do so muchanalysis from it, just because
my brain has like a speed cap onit.
You know it can only do so much.
But if you export one of thoseas a CSV and you feed it into
this thing, it is so good.
(08:32):
If you just give it a promptlike hey, I have this report,
this is the industry.
Here's the relationship betweenthe columns.
Can you give me some feedbackon it?
Then it does.
And then you can just go onestep further and be like can you
draw some correlations that Imight not have thought of?
And it does?
And the one that was reallycool it was in the last episode,
(08:53):
but it was about the number oflogins correlates very closely.
It correlates at like 0.77correlation, which is out of one
.
It correlates between number ofsales and also larger sales.
So you could see that as peoplelog in more, they're probably
(09:15):
using their website more andthey are making more sales, and
I thought that was really cool,because I never once thought to
correlate that column to onethat's completely unrelated to
it, but it was able to see thatand I think that's what a robot
is better at doing than thehuman mind.
Speaker 3 (09:31):
For sure, and that's where the biggest and, to your
point, that's the biggestselling point I think, of these
higher GPT models is theirability to pattern recognize
Because, like you say, you canfeed it, these csvs with tens of
thousands of cells, and you canfeed it across multiple months
and you can spend your timegoing through each individual
client and looking at, you know,doing the month over month.
(09:53):
Even if you, you know, havethose analytics broken out, it's
still going to take you, youknow, a good chunk of your
morning and you give it a chatgbt and it does the work for it.
And obviously, obviously, youknow you have to have privacy
concerns.
There is a setting where youcan say don't train future
models on my data because youobviously don't want it.
You know GPT-6 to startspitting out customers,
(10:14):
e-commerce data.
But that's one of the thingsthat is architecture lends
itself to super well is it'sliterally, it's a neural network
.
Well, is it's literally, it's aneural network.
It's built out of all theseneurons and they take the input
that you give it and theyprocess it and they feed into
other layers and when theselayers light up in a specific
way, the gpt model sees that, oh, this node is lit up over here.
(10:37):
What is, oh, this node is themonth over month node, and it
just does all that internally.
Um, so it's.
It's crazy.
I mean, we're at the pointwhere, uh it even the lead
developers aren't 100 on how itpicks up some of the patterns
that it picks up.
Right now it's just doing itsown thing what?
Speaker 1 (10:51):
uh, what do you when you're talking about how you're
using it?
Uh, because, as I've started totalk to people, it's very, it
almost feels very personal toask like, hey, how are you using
this new tool?
But to to me, it's so sandboxright now that I think we should
learn from each other.
But no one seems to be talkingabout it, and the people who are
(11:12):
talking about it are the kindof people I don't want to listen
to right now.
You know, like people that areway too into it, and so, as a
result, I'm like asking myfriends, like what are you doing
?
So here's what I've been usingchat gpt for and I'm going to
read, you know, like the, thesummarization plots.
Uh, on the sidebar you know I'mtalking about for the different
chats.
So here is like the last coupleuh, rewrite product data copy.
(11:36):
So this is, I have it.
Write a lot of copy for me.
Basically, I like take bulletpoints, I put them in it,
reformulates it.
Email regarding vendor removalwe removed a vendor.
I didn't really know how tolike formulate that thought
without sounding like too casualor too serious.
It's like okay, can you justwrite this for me.
Grilled jalapeno poppers recipeAll hands.
(11:57):
Summary I fed it all of thebullet points from all hands and
had it write a paragraph aboutit.
It was kind of interesting.
I probably won't do it though.
Sharpening a single bevel knife.
I had a question about likewhat degree should you sharpen a
single bevel knife?
Percentage calculations that'sa big one for me.
I'm not very good with math and, as a result, I have a hard
(12:19):
time with percentages A lot ofthe times.
I you remember one time I firedoff the alarms.
I was like something is wrong.
This number is totally bad.
And then Andy was like oh yeah,you still have to multiply by a
hundred on that one orsomething like that.
You need to move the decimalplace, and it's like.
Those are the kinds of things.
It just fact checks me now.
(12:40):
And what kind of things are youusing for both at Punchmark and
maybe in?
Speaker 3 (12:45):
your day-to-day Sure.
Yeah, so for Punchmark, that'slike I mentioned.
Gpt-5 and its processors aresuper good at taking, especially
for coding, taking largeamounts of data and figuring out
how to process it.
I mentioned you can talk to itas a software guy.
You can talk to it about ideasfor architectures and it'll kind
(13:07):
of roast you if it's a bad oneor help you build it up, and
it'll kind of roast you if it'sa bad one or help you build it
up.
And between that and there's abunch of just day-to-day like
I'm trying to put a littlecouple of commands together to
say, copy this file from oneserver to the other and then
change these perms on it, dowhatever.
And I could definitely sit downand think about how to do each
(13:28):
and every one of those and lookat Google and get the exact
syntax right for that.
Or I can just say, hey, gbt,give me the commands to run.
And again, you don't blindlyjust run it and take down the
prod server or something.
You look through it and makesure that it's actually doing
what you want.
But just those little thingsadd up to a decent size serving
every day of savings, just nothaving to do that kind of low
(13:51):
level stuff Is it?
Speaker 1 (13:53):
does it know the punch mark code base, like?
Does it know like a substantialamount of like, because it has
to have that relationship to thesource of truth?
Is there, like at one point didsomeone have to like, you know,
in like a private instance,share with it like, hey, here's
the architecture for this space.
(14:16):
That way it knows how to answer?
Or is it just kind of guessing?
Speaker 3 (14:20):
So the reason we haven't done that so far,
besides the privacy concernswell, I guess it's mitigated if
you all agree to do that setting, which I think we all have at
Punchmark.
But it's just a ton of code,right?
I mean, you look at our codebase.
It's, I think, in the gigsrange, so it's way beyond what
you can just give in a casualprompt to ChatGPT and ask it to
(14:42):
tie it together.
One way I am using it, though,is similar to that.
There is something called IDE.
If you're a software guy, itstands for Integrated
Development Environment, andit's basically just a fancy term
for the fancy notepad that youuse to do all your coding.
So it's basically a text editorthat will help you autocomplete
(15:04):
lines.
It'll show you if your codingsyntax is messed up.
It's basically what mostenterprise-level developers use
to do all their coding.
There's a new one that's comeout in the last few months
called Cursor, which I've beenplaying around with, and that
basically does what you'retalking about, where you install
it on your computer, you loadit, you load it up, you load in
(15:25):
the code base like you would forany other IDE, and then on the
side it has a little bar whereyou can say you can drag in like
specific files.
You literally just drag anddrop it from parts of your IDE
into like a little side paneland you say hey, this file is
giving this response, this fileis not liking it, blah, blah,
blah.
And then it'll go and it'llthink and it'll actually change
(15:46):
it in your IDE, right in frontof you, and it'll basically just
at the end say do you approveor not?
And then you can say no, fixthis.
And it'll fix it.
Or you can say yes, looks great, Now let's move on to this.
So that's kind of the new waythat AI is getting into.
Coding is through those IDEplugins, Really interesting.
Speaker 1 (16:05):
It's so it hasn't really made its way into design
yet.
Just for example, I have an XDfile here, adobe XD.
That's what we designinterfaces with and I'm trying
to work on this, this new designfor the vendor portal.
And what's really interestingis like I could feel myself
being, like I wish he could justput me on second base and
(16:27):
basically let me start furtherdown the road than me.
Like right now, what did I haveto do?
I had to go and like frigging,draw things on a piece of paper
to organize my thoughts.
I had to meet with Brian.
We had to bullet out all ofthese you know, key features and
interactions and things likethat and then start designing
(16:47):
them.
And then we're going to gothrough like know, uh, review
and I kind of like wish I couldjust be like, hey, what is this
interface supposed to look like?
Can you give me ideas?
But it hasn't really seemed toget there yet.
But the functionality it doesseem like there's a right and
wrong way and that's why it canaccess that.
All right, everybody.
We're going to take a quickbreak and hear a word from our
(17:07):
sponsor.
This episode is brought to youin part by Jewelcraft.
Running a jewelry store is nosmall feat.
That's where Jewelcraft comesin.
They're a fourth generationfamily owned business based in
Kentucky that's built theirreputation on being the reliable
behind the scenes partner thatjewelers across the country
trust, whether it's jewelryrepair, manufacturing, custom
(17:28):
design or appraisals.
They can handle anything thatyour jewelers across the country
trust, whether it's jewelryrepair, manufacturing, custom
design or appraisals.
They can handle anything thatyour jewelers don't have the
time for and help scale yourprojects.
And because their work is doneright here in the USA, you avoid
the hassle of overseasproduction, delays or tariffs.
That means faster turnarounds,consistent quality and peace of
mind.
With a team of seasoned expertsand decades of experience,
(17:51):
jewelcraft is your one-stop shopfor everything you need to keep
your jewelry business runningsmoothly so you can stay focused
on what matters most yourcustomers.
Learn more at Jewelcraftcomthat's Jewel-Craftcom.
Learn more about them in theshow notes below.
And now back to the show.
And we're back, so let's gointo what your GPT-5 like
(18:21):
history is and nothing, onlyonly good stuff.
Speaker 3 (18:22):
Oh man, that's going to take some digging to find
that.
All right, no, just kidding.
So yeah, just some of the moremore recent.
So I'm a little bit of ahobbyist investor.
I like to buy stock in somecompanies I believe in.
So there's a lot of grunt workthat goes in that.
You have to look at their books, their cash flow, pl statements
(18:43):
, earnings calls, see what themarket thinks of them.
All that yada, yada.
Obviously that's super easy toautomate through GPT-5.
Speaker 1 (18:49):
They can book for that.
They can do all that for you.
Speaker 3 (18:51):
Oh, that's really interesting and that's what some
of the.
So one of the big problems thatAI is trying to get past is
called hallucination, which iswhere it just makes up something
completely out of the blue.
That's completely not true.
You've seen some, you know youcan have some benign example of
that where it just you know getswhatever hat on something.
Like if of that, where it justyou know gets whatever hat on
(19:13):
something.
Like if something happened onsome day and actually happened
on another one, you're like no,it happened on saturday.
You're like, okay, but on a bigscale.
Like there's some lawyers thathave tried to use chat gpt in
like filing their legal briefs,like paralegals, and it just
makes up court cases out of oldcloth sometimes.
Um, I like some of the earlierones.
So, yeah, there's.
You know, every industry hasits version of vibe coding, I
guess.
So I guess that's the legalprofessions.
(19:34):
But the most recent versionshave something called a chain of
thought, which is where itactually it gives itself a
little internal answer to whatyour prompt is, and it'll, if
you expand it, while GPT chatGPT is thinking, it'll say
something like okay, the user isasking blah, blah, blah, and
it'll talk through itself,thinking it'll say something
like okay, the user is asking,uh, blah, blah, blah, um, and
it'll talk through itself andit'll search the web and it'll
(19:57):
find uh citations for exactlywhat it's saying and then it'll
add it in line.
So, uh, it doesn't completelysolve the problem and it
definitely makes it a lot slower, so you can't use it for
everything, um, but yeah, yeah,exactly almost like a little
wikipedia thing'll have a likethe name of where it pulled it
from and you can click on it andit'll take you right there.
So as far as things that arevery sensitive to like, I don't
(20:20):
want you to hallucinate whattheir PNL statement was like
it'll, if you want to doublecheck it, and it'll bring you
right to it.
So it's definitely gotten a lotbetter in that respect.
Speaker 1 (20:28):
So let's just let's just talk about one of these
examples.
Can you pull one up?
And I'm very interested in howyou're formulating a query,
because people have given meexamples before but sometimes
they don't.
Like I'm curious, like whatdoes one actually look like to
someone?
For example, for mine aboutthis episode, I said I'd love to
(20:51):
talk about GPT-5 on my podcastand highlight the main
differences from four to fiveand new features and use cases
for GPT-5 and how the widerpublic might be able to leverage
it, but also make itinteresting and tell me what I
should be focused on.
So that's like I'm just I don'tknow anything about GPT-5 yet.
So it's like OK, give me like a, put me on second base, and I
(21:14):
think that's kind of how Iformulate my queries.
What are you kind of startingwith?
Sure.
Speaker 3 (21:20):
I mean I won't recite my entire opening prompt
because it's pretty long, but Igive it a bunch of.
I kind of give it a backgroundabout what's going on in general
.
I'll say kind of what's going onin the world the geopolitical
side, the macroeconomic side.
I'll tell it specifically to doits own research and put its
own model together about how itsees the world.
(21:42):
So there's a bunch ofconsulting companies out there
that'll publish reports onsomething that's their opinion
and I basically tell it todiscard all that and build its
own internal model from scratchso it has its own idea about
where the world's at and whereit might move.
And then I tell it to thinkabout how all these different
(22:02):
factors in the world mightinteract and how these
interactions might have impactsthat are knock-on effects that
impact other sectors in thefuture, have impacts that are
knock-on effects that impactother sectors in the future.
And when I say that the GPTneural network is very good at
handling all that informationand processing, it is really
good.
So it's obviously like I said,you don't want to vibe code your
way through finance, obviously,but as far as just to use your
(22:26):
term get you on second base,it's very strong.
Speaker 1 (22:29):
Do you think it is effective?
Is it working for you?
Because sometimes we talk aboutit, it's very strong.
Do you think it is effective?
Is it working for you?
Because sometimes we talk aboutit it's like oh, and this is
what I use it for, but no onetells you that it's like, it's
not.
It's definitely not perfect,but what about for this use case
?
Speaker 3 (22:42):
It's very good, like I said, but what I'm using it
for, which is doing the deepdive in the background, I'm not
asking it to go out and make me$100,000.
I'm coming at it with an ideaof hey, I think this opportunity
might be here in the market.
Just give me a quick read, seeif it's worth pouring more
investment in and if it'ssomething I want to look into.
Speaker 1 (23:05):
Yeah, I guess one of the examples that we had talked
about when I was like, hey, doyou mind coming on and just
chatting about this with me?
Was a couple of years ago, Ithink it was in 2020.
In 2018, or 2019, I wrote astory.
I tried my hardest to write abook and the big reason being I
(23:28):
read so many books but I'venever tried writing one.
So I was like, okay, I shouldtry this out, and this is
obviously a while back, and mygoal was I just wanted to write
50,000 words and it eventuallygot to, I think it's.
I just pulled it up like 75,000words or so at this point and a
an average size book is like inthe, you know, 90 to 150,000
(23:52):
word range.
And what's funny is I didn'tfinish the book but I wrote
enough of it that I felt prettygood about that and I thought it
was a really interestingexercise.
And I wrote it over the courseof a year and I thought that
that was like a fun experiment.
But since I'd never finished it,I one time was like, should I
feed this story into GPT, likechapter by chapter?
(24:16):
So feed a chapter one, be like,hey, what are your thoughts on
this and then go chapter two allthe way to the end and then be
like, well, that's the end ofthe story.
I never wrote the rest of it.
Here's the outline in the storysculpture.
Can you write me the rest ofthe story?
Do you think it could do thatand match, like, my writing
style?
Or do you think it wouldhallucinate and like completely
(24:40):
get off off target, or what doyou think the result of that
would be?
Cause I'm tempted to try andthe only thing that's stopping
me is, like you know, ethics.
Speaker 3 (24:50):
Yeah, those pesky ethics.
Yeah, I mean I, I would say ifyou, if you wanted to do that,
it would probably do a very goodjob.
Uhaders may or may not be ableto spot the difference from when
you stopped and when GPTstarted.
That being said, we're movingto a place in the world where so
much of everything is AI.
You know the term AI slop isout there for just stuff that
(25:13):
people create that just sparksengagement on the internet and
you know people profit from it.
But there's so much AI contentout there that it's actually
kind of interesting.
Reddit as a company is valuedvery highly right now because
it's a place where obviously,there's a bunch of bots but
there's a bunch of people andthe people are interacting.
(25:33):
Their end goal is to use thatand train AI models off the
people interacting, which is tosay, there's just going to be
more AI stuff on the internetand out there, right?
So I think there's a specialdignity in not just waving away
the ethics side and saying Iwant GPT to finish this story
and it'll probably be a goodstory, but doing it yourself,
putting it out there, goingthrough the creative process and
(25:53):
you're a creative guy, I mean,you're an artist, you understand
all that, but I think using itas an editor more than a
co-author would be the best wayto do it.
Speaker 1 (26:01):
Yeah, I think that that's definitely one of those
personal journeys that you haveto go on, and I definitely know
that some people do not havethose scruples.
I'm sure that there are so manybooks out there that have
already been published andpeople are profiting off of it
that were written, you know,maybe entirely or primarily by a
(26:24):
bot, but yeah, I think it'skind of.
There's something special, Ithink, about the tactile nature
of human error.
You know, the fact that it'slike it's, it's pretty real, and
that's what I always thinkabout with my paintings is I
think that people don't evenreally care about the image.
It's more that it's on a pieceof paper and if you hold it up
(26:46):
close enough, you can see where,my like, paint bled into
something else and in the errorsin the pencil markings, and I
think that's almost what peoplevalue more than they value the
actual image of itself.
It's that it was.
You know it was paper and nowit's a painting With development
and coding.
(27:06):
There is not necessarily that.
You know there is no like thisolder form, this innate like
source of truth.
What do you see?
Is it just like a like?
Where do your morals kind ofcome in and prevent you from
using Chachi Petit to run, likeyou know, the wider part of your
life.
Is that something you try tokeep it at arm's length for,
(27:28):
like, yeah, like what you shouldeat for the day.
Speaker 3 (27:31):
Well, I do use it to spin up a decent amount of
recipes, so it does tell me whatto eat for the day, but in
general, yeah, I think it's it'snot healthy to get super
plugged into it.
You hear stories of people withAI boyfriends and girlfriends
and with ChatGPT 5 just gettingrolled out.
They pulled out all those oldmodels that people had their
(27:53):
boyfriends and girlfriends on,so they're probably not having a
good time right now, but Ithink it's important, especially
as, I say, more AI is going to.
Ai is just going to keepexponentially rising more and
more and it's definitely atemptation and a risk to let it
run more of your life becauseit's it's so efficient and it's
(28:14):
so optimized and it it's notnecessarily wrong when it tells
you how to handle your problems.
But there's a risk that you andthe GPT personality just kind
of merge right and then you knowyou start talking like GPT-5 on
the internet and we're alreadyseeing that.
How many MDASHs have you comeacross since GPT-5 rolled out?
I mean that's a.
I still.
I'm at the point now where Ismile a little every time I see
(28:35):
someone with a typo in somethingthat they write because I know
it's written by a human being.
Speaker 1 (28:48):
It's isn't that so interesting because the m dash
in particular so good peoplelistening, uh, insight, the m
dash is, if you, if you go dashdash, it usually forms into a
longer one, but it's a hallmarkof something written by um, by a
bot.
And what's really interestingis, uh, for a really long time I
actually didn't know how tomake an m dash and and I was
like, so if I don't know and Iwrite on the internet all the
time, I'm pretty sure that thisrandom person didn't know and
this is something that they justcopied out of somewhere.
(29:10):
And I just think that you know,there's the whole you know dead
internet theory.
You know, with like, we're justthere's only so many people on
the internet and we're justreacting to information.
That's not that was, you know,created by bots.
And it's just like no one isactually on the internet anymore
.
We're not even talking to eachother.
But I'm not about you're a bot,exactly, people.
(29:31):
It's something I've beenthinking about.
It's kind of funny.
Why do we have this episode?
Chat tpt5 came out, but to meit's just.
I think it's a reasonablemoment to check in and have us
have like a real like, let'slook at this thing.
You know that whole.
You were the one that showed methe south park episode about uh
chachi vt and that was back.
I was back.
Uh, remember we were in raleighwhen we watched that.
(29:53):
We went and, um, uh, went tothat event.
Yeah, that was a bunch of yearsago and at this point it's so
crazy to think that that ishowever many models ago you know
things that this thing is ableto be way more complex.
I think it's reasonable forpeople to just have a moment to
analyze their own usage inrelationship with this, with
(30:14):
this tool.
You know, if you only got ahammer, you're going to look for
ways to use a hammer.
So it's about being a littlebit of having a little.
It's about being a little bitof having a little bit of touch
and a little bit of I don't knowcreativity with how you're
using it, and not to lose thehumanity of it all.
You know what I mean For sure.
Speaker 3 (30:32):
It's a tool like any other.
You know, it's a tool that'sdeveloped a lot quicker than I
think most people were expecting, even people who thought years
ago that AI might be coming soon.
I think it has caught a lot ofpeople off guard with how
quickly it's gotten here, whichis kind of how the last few
innovations happened.
No one had a smartphone intheir pocket until everyone did
(30:53):
right.
No one had an email addressuntil everyone did, so it's just
one of those things where it'sjust it's here now.
No one rode a Lime Scooteruntil they dropped those by the
hundreds in cities all over theplace.
But yeah, it's a tool.
It's not one that you shouldthrow your entire life at, even
if it's tempting.
Use it for its purpose ofhelping you figure out searches.
(31:13):
Help it analyze whatever yourbusiness need is, but don't give
too much of your personalityaway to it.
And also be careful whatinformation that you give them,
because even if you tell themdon't train my information on
future models, there's noguarantee that they're not going
to do other selling of yourinformation to advertisers or
their own internal analytics.
(31:35):
I wouldn't be surprised if theyhad a GPT that combed through
everyone's account just to findthe fun little nuggets of
information to pull out.
So be careful with that side.
But yeah, if you can do allthat and you can use it as a
tool and be effective with it.
Speaker 1 (31:49):
I think it'll be great for you.
Yeah, more power to you.
Yeah, just, buyer beware, justbe smart, andy.
I think maybe that's wherewe'll end it Just kind of maybe
a little more of a kind of athoughtful, introspective
episode.
Normally I'm doing interviews,but sometimes I just think it's
kind of fun to kind of, you know, muse about something and kind
of think, but, andy, I reallyappreciate it.
I'll let you get back to theend of your day.
(32:11):
Thanks, everybody.
I'll see you next time.
New episodes every Tuesday.
Bye.
All right, everybody.
That's another show.
Thanks so much for listening.
My guest this week was AndyZoki, the backend developer at
Punchmark.
He's also one of my bestfriends.
This episode was brought to youby Punchmark and produced and
(32:33):
hosted by me, michael Burpo.
This episode was edited by PaulSuarez with music by Ross
Cockrum.
Don't forget to rate thepodcast on Spotify and Apple
Podcasts and leave us feedbackon punchmarkcom slash loop.
That's L-O-U-P-E.
Thanks.
We'll be back next week,tuesday, with another episode.
Cheers, bye.
Welcome to In the Loop.
Two years since the last modelrelease and this one is, of
course, a huge step up and, Imust admit, I don't even fully
understand it.
So I brought on my friend andalso my coworker, andy Zoki, who
is a backend developer forPunchmark.
He's also one of the mostplugged in guys with regards to,
(00:39):
you know, chatgpt and thedevelopment of AI, and he stays
on top of this stuff and uses itin ways that I couldn't even
imagine and I asked him.
We had a real nice conversationdiscussing what the new changes
mean, what it's capable of, theways that we use it in our own
life, and I think it's just amore casual conversation,
(01:01):
because Andy's one of my bestfriends and he's also my
co-worker and hearing how heuses it maybe informs me on how
I can use it better.
If you're not comfortable withusing AI, no problem.
This might not be the episodefor you, but if you are kind of
dabbling, this, one kind of getsinto some deeper weeds and
maybe you enjoy that.
Speaker 2 (01:21):
So everybody enjoy and maybe you enjoy that.
So everybody enjoy.
This episode is brought to youby Punchmark, the jewelry
industry's favorite websiteplatform and digital growth
agency.
Our mission reaches way beyondtechnology.
With decades of experience andlong-lasting industry
relationships, punchmark enablesjewelry businesses to flourish
(01:41):
in any marketplace.
We consider our clients ourfriends, as many of them have
been friends way before becomingclients.
Punchmark's own success comesfrom the fact that we have a
much deeper need and obligationto help our friends succeed.
Whether you're looking forbetter e-commerce performance,
business growth or campaignsthat drive traffic and sales,
punchmark's website andmarketing services were made
(02:03):
just for you.
It's never too late totransform your business and
stitch together your digital andphysical worlds in a way that
achieves tremendous growth andresults.
Schedule a guided demo today atpunchmarkcom.
Slash go.
Speaker 1 (02:19):
And now back to the show.
What is up everybody?
I'm joined by Andy Zoki, mygood buddy and also back-end
developer at Punchmark.
How are you doing today, andy?
I'm doing pretty good.
What's up, mike?
Doing so well.
Really cool getting a chance tohave you on.
You were on when we weretalking about VR and you've
(02:43):
always been very tech forwardwhen it comes to, you know, I
guess, adopting new technologiesand kind of.
You follow the state of AI verywell.
But, big news, it actuallyhappened yesterday when we were
recording this.
But when this releases it'llhave been a few days, but GPT-5
just released.
Can you give like a pseudooverview on what GPT-5 is for
(03:06):
people listening?
Speaker 3 (03:08):
Sure, yeah.
So GPT-5 is OpenAI's latestiteration of their chat, gpt,
llm, large language model andit's you know.
You've heard of AI, you'veheard of all these things that
it can do.
Gpt-5 is the latest iteration,looking at the initial rollout,
(03:28):
and this is bleeding edge.
I literally just got access toit this morning, so I haven't
had too much of a chance to playaround with it, and that's
going to be coming over the nextcouple of weeks, but it looks
like it's moderately better atcoding tasks.
It's slightly better logicalreasoning Maybe not as much of
an exponential rise as somepeople might have been hoping
(03:49):
for.
One thing that we've seen withthe most recent releases is it's
just taking a lot more time andmoney than people were hoping
for.
So it's one barrier thatthey're actively working on,
obviously, but, yeah, it's justa little bit of a better thinker
.
One thing that people aresaying off the bat that they
aren't necessarily fans of isthe personality is a bit drier
(04:11):
and, uh, it could be more curt.
Um, I've heard, uh I saw oneperson on reddit describe it as
a overworked secretary.
That's just kind of exasperated.
Talking to you, it's maybe notthat bad, but it definitely has
a little bit more of attentionto that than uh its predecessors
, uh gpt 4.5 and 03 yeah,because so there was.
Speaker 1 (04:33):
There's different gpts, uh, that are tailored to
different types of tasks, likethere's 03, and then there's 03,
mini, and then there's, uh,there's four.
Um, I have to be honest,because we were talking about
this probably about a month agoand you're like, oh yeah, you're
, you're using four, you're noteven using the advanced version
that you're paying for.
(04:54):
I was like, oh, I didn'trealize you had to do that.
Apparently, with gpt5, it willdynamically switch between the
models, uh, depending on whatyour task requires.
So if you need a deep thinkingone, it can do that, and if you
need a you know, just a quickanswer, then it'll do that as
well.
Do you switch between models,or were you switching between
(05:14):
the models when you're using it?
Speaker 3 (05:16):
Yeah, I've always been a big fan of O3.
I think that's just been themost capable model for most
complex tasks.
I use it as a developer.
I know lots of developers useChatGPT.
You've probably heard vibecoding, you know it can be an
insult.
On the far end it's justtelling ChatGPT to write
everything and then you justcopy and paste it into your code
(05:37):
base and you ship a millionbugs and, you know, make a lot
of people unhappy.
We would never do that, ofcourse not.
No, and we don't.
Obviously, the correct way touse it is almost as if it's a
colleague where you have ideasabout how you want to do
something and you say I havethese classes, I'm trying to put
this architecture together, butI want to optimize it for low
(05:59):
cost, low latency, that kind ofthing.
It's really good at taking allthose different pieces and
churning through them andputting it together, and 5 is a
little bit better at that.
There are still a few moremodels that you can split out.
So if you're on a paid version,you still get the model picker
at the top left where you canopen GPT-5, it says flagship and
(06:20):
then for other models you canpick between uh, or if you're a
pro user you can do pro.
So it does split it out by costa bit, um, but all of them are
a bit faster, a bit smarter and,uh, the base one, the gpt5, uh,
non-thinking, it's super fastand it's still very good at
thinking.
I'm impressed.
Like I have my, I send a promptand before I pull my uh finger
(06:42):
off the enter key, it's alreadystarted writing the first
response in some cases.
So it's crazy.
Speaker 1 (06:47):
I think it's been really interesting the way that
I've been using it.
Is it used to be that I wouldgo to ChatGPT and I would like,
oh this is something you couldhelp me with, and then I would
kind of interface with it andnow, like an oracle, you know,
I'd have to go to it and I wouldask it and then it would answer
(07:07):
me.
Now what I do is I just have itdownloaded to my desktop.
I just keep one of the windowsopen.
I have a dual window set up inmy workstation, so I have a
laptop, a second screen.
On my second screen, I haveSlack for work, I have ChatGPT
and I have Spotify.
Like.
Those are the essentials atthis point, and things that it's
(07:28):
good for are drawingconclusions and drawing
inferences from data sheets.
Especially is what I've beenhaving really good success with.
I just did the episode before.
This one was very much helpedby chat GBT, where you know, oh,
you're the one that built methe leaderboard exports.
(07:49):
Remember that back in back inthe day.
Oh, yeah, the finance.
That's right For the e-commerceperformance reports, and what's
super funny is I can look at itand, you know, draw inference
on how e-commerce is doing.
But there's you know howevermany columns there's like 70
(08:10):
columns in that thing and theneach column has like 120 rows
and you know, blah, blah, blah.
There's all these things.
It's every single month.
I can only draw so muchanalysis from it, just because
my brain has like a speed cap onit.
You know it can only do so muchanalysis from it, just because
my brain has like a speed cap onit.
You know it can only do so much.
But if you export one of thoseas a CSV and you feed it into
this thing, it is so good.
(08:32):
If you just give it a promptlike hey, I have this report,
this is the industry.
Here's the relationship betweenthe columns.
Can you give me some feedbackon it?
Then it does.
And then you can just go onestep further and be like can you
draw some correlations that Imight not have thought of?
And it does?
And the one that was reallycool it was in the last episode,
(08:53):
but it was about the number oflogins correlates very closely.
It correlates at like 0.77correlation, which is out of one
.
It correlates between number ofsales and also larger sales.
So you could see that as peoplelog in more, they're probably
(09:15):
using their website more andthey are making more sales, and
I thought that was really cool,because I never once thought to
correlate that column to onethat's completely unrelated to
it, but it was able to see thatand I think that's what a robot
is better at doing than thehuman mind.
Speaker 3 (09:31):
For sure, and that's where the biggest and, to your
point, that's the biggestselling point I think, of these
higher GPT models is theirability to pattern recognize
Because, like you say, you canfeed it, these csvs with tens of
thousands of cells, and you canfeed it across multiple months
and you can spend your timegoing through each individual
client and looking at, you know,doing the month over month.
(09:53):
Even if you, you know, havethose analytics broken out, it's
still going to take you, youknow, a good chunk of your
morning and you give it a chatgbt and it does the work for it.
And obviously, obviously, youknow you have to have privacy
concerns.
There is a setting where youcan say don't train future
models on my data because youobviously don't want it.
You know GPT-6 to startspitting out customers,
(10:14):
e-commerce data.
But that's one of the thingsthat is architecture lends
itself to super well is it'sliterally, it's a neural network
.
Well, is it's literally, it's aneural network.
It's built out of all theseneurons and they take the input
that you give it and theyprocess it and they feed into
other layers and when theselayers light up in a specific
way, the gpt model sees that, oh, this node is lit up over here.
(10:37):
What is, oh, this node is themonth over month node, and it
just does all that internally.
Um, so it's.
It's crazy.
I mean, we're at the pointwhere, uh it even the lead
developers aren't 100 on how itpicks up some of the patterns
that it picks up.
Right now it's just doing itsown thing what?
Speaker 1 (10:51):
uh, what do you when you're talking about how you're
using it?
Uh, because, as I've started totalk to people, it's very, it
almost feels very personal toask like, hey, how are you using
this new tool?
But to to me, it's so sandboxright now that I think we should
learn from each other.
But no one seems to be talkingabout it, and the people who are
(11:12):
talking about it are the kindof people I don't want to listen
to right now.
You know, like people that areway too into it, and so, as a
result, I'm like asking myfriends, like what are you doing
?
So here's what I've been usingchat gpt for and I'm going to
read, you know, like the, thesummarization plots.
Uh, on the sidebar you know I'mtalking about for the different
chats.
So here is like the last coupleuh, rewrite product data copy.
(11:36):
So this is, I have it.
Write a lot of copy for me.
Basically, I like take bulletpoints, I put them in it,
reformulates it.
Email regarding vendor removalwe removed a vendor.
I didn't really know how tolike formulate that thought
without sounding like too casualor too serious.
It's like okay, can you justwrite this for me.
Grilled jalapeno poppers recipeAll hands.
(11:57):
Summary I fed it all of thebullet points from all hands and
had it write a paragraph aboutit.
It was kind of interesting.
I probably won't do it though.
Sharpening a single bevel knife.
I had a question about likewhat degree should you sharpen a
single bevel knife?
Percentage calculations that'sa big one for me.
I'm not very good with math and, as a result, I have a hard
(12:19):
time with percentages A lot ofthe times.
I you remember one time I firedoff the alarms.
I was like something is wrong.
This number is totally bad.
And then Andy was like oh yeah,you still have to multiply by a
hundred on that one orsomething like that.
You need to move the decimalplace, and it's like.
Those are the kinds of things.
It just fact checks me now.
(12:40):
And what kind of things are youusing for both at Punchmark and
maybe in?
Speaker 3 (12:45):
your day-to-day Sure.
Yeah, so for Punchmark, that'slike I mentioned.
Gpt-5 and its processors aresuper good at taking, especially
for coding, taking largeamounts of data and figuring out
how to process it.
I mentioned you can talk to itas a software guy.
You can talk to it about ideasfor architectures and it'll kind
(13:07):
of roast you if it's a bad oneor help you build it up, and
it'll kind of roast you if it'sa bad one or help you build it
up.
And between that and there's abunch of just day-to-day like
I'm trying to put a littlecouple of commands together to
say, copy this file from oneserver to the other and then
change these perms on it, dowhatever.
And I could definitely sit downand think about how to do each
(13:28):
and every one of those and lookat Google and get the exact
syntax right for that.
Or I can just say, hey, gbt,give me the commands to run.
And again, you don't blindlyjust run it and take down the
prod server or something.
You look through it and makesure that it's actually doing
what you want.
But just those little thingsadd up to a decent size serving
every day of savings, just nothaving to do that kind of low
(13:51):
level stuff Is it?
Speaker 1 (13:53):
does it know the punch mark code base, like?
Does it know like a substantialamount of like, because it has
to have that relationship to thesource of truth?
Is there, like at one point didsomeone have to like, you know,
in like a private instance,share with it like, hey, here's
the architecture for this space.
(14:16):
That way it knows how to answer?
Or is it just kind of guessing?
Speaker 3 (14:20):
So the reason we haven't done that so far,
besides the privacy concernswell, I guess it's mitigated if
you all agree to do that setting, which I think we all have at
Punchmark.
But it's just a ton of code,right?
I mean, you look at our codebase.
It's, I think, in the gigsrange, so it's way beyond what
you can just give in a casualprompt to ChatGPT and ask it to
(14:42):
tie it together.
One way I am using it, though,is similar to that.
There is something called IDE.
If you're a software guy, itstands for Integrated
Development Environment, andit's basically just a fancy term
for the fancy notepad that youuse to do all your coding.
So it's basically a text editorthat will help you autocomplete
(15:04):
lines.
It'll show you if your codingsyntax is messed up.
It's basically what mostenterprise-level developers use
to do all their coding.
There's a new one that's comeout in the last few months
called Cursor, which I've beenplaying around with, and that
basically does what you'retalking about, where you install
it on your computer, you loadit, you load it up, you load in
(15:25):
the code base like you would forany other IDE, and then on the
side it has a little bar whereyou can say you can drag in like
specific files.
You literally just drag anddrop it from parts of your IDE
into like a little side paneland you say hey, this file is
giving this response, this fileis not liking it, blah, blah,
blah.
And then it'll go and it'llthink and it'll actually change
(15:46):
it in your IDE, right in frontof you, and it'll basically just
at the end say do you approveor not?
And then you can say no, fixthis.
And it'll fix it.
Or you can say yes, looks great, Now let's move on to this.
So that's kind of the new waythat AI is getting into.
Coding is through those IDEplugins, Really interesting.
Speaker 1 (16:05):
It's so it hasn't really made its way into design
yet.
Just for example, I have an XDfile here, adobe XD.
That's what we designinterfaces with and I'm trying
to work on this, this new designfor the vendor portal.
And what's really interestingis like I could feel myself
being, like I wish he could justput me on second base and
(16:27):
basically let me start furtherdown the road than me.
Like right now, what did I haveto do?
I had to go and like frigging,draw things on a piece of paper
to organize my thoughts.
I had to meet with Brian.
We had to bullet out all ofthese you know, key features and
interactions and things likethat and then start designing
(16:47):
them.
And then we're going to gothrough like know, uh, review
and I kind of like wish I couldjust be like, hey, what is this
interface supposed to look like?
Can you give me ideas?
But it hasn't really seemed toget there yet.
But the functionality it doesseem like there's a right and
wrong way and that's why it canaccess that.
All right, everybody.
We're going to take a quickbreak and hear a word from our
(17:07):
sponsor.
This episode is brought to youin part by Jewelcraft.
Running a jewelry store is nosmall feat.
That's where Jewelcraft comesin.
They're a fourth generationfamily owned business based in
Kentucky that's built theirreputation on being the reliable
behind the scenes partner thatjewelers across the country
trust, whether it's jewelryrepair, manufacturing, custom
(17:28):
design or appraisals.
They can handle anything thatyour jewelers across the country
trust, whether it's jewelryrepair, manufacturing, custom
design or appraisals.
They can handle anything thatyour jewelers don't have the
time for and help scale yourprojects.
And because their work is doneright here in the USA, you avoid
the hassle of overseasproduction, delays or tariffs.
That means faster turnarounds,consistent quality and peace of
mind.
With a team of seasoned expertsand decades of experience,
(17:51):
jewelcraft is your one-stop shopfor everything you need to keep
your jewelry business runningsmoothly so you can stay focused
on what matters most yourcustomers.
Learn more at Jewelcraftcomthat's Jewel-Craftcom.
Learn more about them in theshow notes below.
And now back to the show.
And we're back, so let's gointo what your GPT-5 like
(18:21):
history is and nothing, onlyonly good stuff.
Speaker 3 (18:22):
Oh man, that's going to take some digging to find
that.
All right, no, just kidding.
So yeah, just some of the moremore recent.
So I'm a little bit of ahobbyist investor.
I like to buy stock in somecompanies I believe in.
So there's a lot of grunt workthat goes in that.
You have to look at their books, their cash flow, pl statements
(18:43):
, earnings calls, see what themarket thinks of them.
All that yada, yada.
Obviously that's super easy toautomate through GPT-5.
Speaker 1 (18:49):
They can book for that.
They can do all that for you.
Speaker 3 (18:51):
Oh, that's really interesting and that's what some
of the.
So one of the big problems thatAI is trying to get past is
called hallucination, which iswhere it just makes up something
completely out of the blue.
That's completely not true.
You've seen some, you know youcan have some benign example of
that where it just you know getswhatever hat on something.
Like if of that, where it justyou know gets whatever hat on
(19:13):
something.
Like if something happened onsome day and actually happened
on another one, you're like no,it happened on saturday.
You're like, okay, but on a bigscale.
Like there's some lawyers thathave tried to use chat gpt in
like filing their legal briefs,like paralegals, and it just
makes up court cases out of oldcloth sometimes.
Um, I like some of the earlierones.
So, yeah, there's.
You know, every industry hasits version of vibe coding, I
guess.
So I guess that's the legalprofessions.
(19:34):
But the most recent versionshave something called a chain of
thought, which is where itactually it gives itself a
little internal answer to whatyour prompt is, and it'll, if
you expand it, while GPT chatGPT is thinking, it'll say
something like okay, the user isasking blah, blah, blah, and
it'll talk through itself,thinking it'll say something
like okay, the user is asking,uh, blah, blah, blah, um, and
it'll talk through itself andit'll search the web and it'll
(19:57):
find uh citations for exactlywhat it's saying and then it'll
add it in line.
So, uh, it doesn't completelysolve the problem and it
definitely makes it a lot slower, so you can't use it for
everything, um, but yeah, yeah,exactly almost like a little
wikipedia thing'll have a likethe name of where it pulled it
from and you can click on it andit'll take you right there.
So as far as things that arevery sensitive to like, I don't
(20:20):
want you to hallucinate whattheir PNL statement was like
it'll, if you want to doublecheck it, and it'll bring you
right to it.
So it's definitely gotten a lotbetter in that respect.
Speaker 1 (20:28):
So let's just let's just talk about one of these
examples.
Can you pull one up?
And I'm very interested in howyou're formulating a query,
because people have given meexamples before but sometimes
they don't.
Like I'm curious, like whatdoes one actually look like to
someone?
For example, for mine aboutthis episode, I said I'd love to
(20:51):
talk about GPT-5 on my podcastand highlight the main
differences from four to fiveand new features and use cases
for GPT-5 and how the widerpublic might be able to leverage
it, but also make itinteresting and tell me what I
should be focused on.
So that's like I'm just I don'tknow anything about GPT-5 yet.
So it's like OK, give me like a, put me on second base, and I
(21:14):
think that's kind of how Iformulate my queries.
What are you kind of startingwith?
Sure.
Speaker 3 (21:20):
I mean I won't recite my entire opening prompt
because it's pretty long, but Igive it a bunch of.
I kind of give it a backgroundabout what's going on in general
.
I'll say kind of what's going onin the world the geopolitical
side, the macroeconomic side.
I'll tell it specifically to doits own research and put its
own model together about how itsees the world.
(21:42):
So there's a bunch ofconsulting companies out there
that'll publish reports onsomething that's their opinion
and I basically tell it todiscard all that and build its
own internal model from scratchso it has its own idea about
where the world's at and whereit might move.
And then I tell it to thinkabout how all these different
(22:02):
factors in the world mightinteract and how these
interactions might have impactsthat are knock-on effects that
impact other sectors in thefuture, have impacts that are
knock-on effects that impactother sectors in the future.
And when I say that the GPTneural network is very good at
handling all that informationand processing, it is really
good.
So it's obviously like I said,you don't want to vibe code your
way through finance, obviously,but as far as just to use your
(22:26):
term get you on second base,it's very strong.
Speaker 1 (22:29):
Do you think it is effective?
Is it working for you?
Because sometimes we talk aboutit, it's very strong.
Do you think it is effective?
Is it working for you?
Because sometimes we talk aboutit it's like oh, and this is
what I use it for, but no onetells you that it's like, it's
not.
It's definitely not perfect,but what about for this use case
?
Speaker 3 (22:42):
It's very good, like I said, but what I'm using it
for, which is doing the deepdive in the background, I'm not
asking it to go out and make me$100,000.
I'm coming at it with an ideaof hey, I think this opportunity
might be here in the market.
Just give me a quick read, seeif it's worth pouring more
investment in and if it'ssomething I want to look into.
Speaker 1 (23:05):
Yeah, I guess one of the examples that we had talked
about when I was like, hey, doyou mind coming on and just
chatting about this with me?
Was a couple of years ago, Ithink it was in 2020.
In 2018, or 2019, I wrote astory.
I tried my hardest to write abook and the big reason being I
(23:28):
read so many books but I'venever tried writing one.
So I was like, okay, I shouldtry this out, and this is
obviously a while back, and mygoal was I just wanted to write
50,000 words and it eventuallygot to, I think it's.
I just pulled it up like 75,000words or so at this point and a
an average size book is like inthe, you know, 90 to 150,000
(23:52):
word range.
And what's funny is I didn'tfinish the book but I wrote
enough of it that I felt prettygood about that and I thought it
was a really interestingexercise.
And I wrote it over the courseof a year and I thought that
that was like a fun experiment.
But since I'd never finished it,I one time was like, should I
feed this story into GPT, likechapter by chapter?
(24:16):
So feed a chapter one, be like,hey, what are your thoughts on
this and then go chapter two allthe way to the end and then be
like, well, that's the end ofthe story.
I never wrote the rest of it.
Here's the outline in the storysculpture.
Can you write me the rest ofthe story?
Do you think it could do thatand match, like, my writing
style?
Or do you think it wouldhallucinate and like completely
(24:40):
get off off target, or what doyou think the result of that
would be?
Cause I'm tempted to try andthe only thing that's stopping
me is, like you know, ethics.
Speaker 3 (24:50):
Yeah, those pesky ethics.
Yeah, I mean I, I would say ifyou, if you wanted to do that,
it would probably do a very goodjob.
Uhaders may or may not be ableto spot the difference from when
you stopped and when GPTstarted.
That being said, we're movingto a place in the world where so
much of everything is AI.
You know the term AI slop isout there for just stuff that
(25:13):
people create that just sparksengagement on the internet and
you know people profit from it.
But there's so much AI contentout there that it's actually
kind of interesting.
Reddit as a company is valuedvery highly right now because
it's a place where obviously,there's a bunch of bots but
there's a bunch of people andthe people are interacting.
(25:33):
Their end goal is to use thatand train AI models off the
people interacting, which is tosay, there's just going to be
more AI stuff on the internetand out there, right?
So I think there's a specialdignity in not just waving away
the ethics side and saying Iwant GPT to finish this story
and it'll probably be a goodstory, but doing it yourself,
putting it out there, goingthrough the creative process and
(25:53):
you're a creative guy, I mean,you're an artist, you understand
all that, but I think using itas an editor more than a
co-author would be the best wayto do it.
Speaker 1 (26:01):
Yeah, I think that that's definitely one of those
personal journeys that you haveto go on, and I definitely know
that some people do not havethose scruples.
I'm sure that there are so manybooks out there that have
already been published andpeople are profiting off of it
that were written, you know,maybe entirely or primarily by a
(26:24):
bot, but yeah, I think it'skind of.
There's something special, Ithink, about the tactile nature
of human error.
You know, the fact that it'slike it's, it's pretty real, and
that's what I always thinkabout with my paintings is I
think that people don't evenreally care about the image.
It's more that it's on a pieceof paper and if you hold it up
(26:46):
close enough, you can see where,my like, paint bled into
something else and in the errorsin the pencil markings, and I
think that's almost what peoplevalue more than they value the
actual image of itself.
It's that it was.
You know it was paper and nowit's a painting With development
and coding.
(27:06):
There is not necessarily that.
You know there is no like thisolder form, this innate like
source of truth.
What do you see?
Is it just like a like?
Where do your morals kind ofcome in and prevent you from
using Chachi Petit to run, likeyou know, the wider part of your
life.
Is that something you try tokeep it at arm's length for,
(27:28):
like, yeah, like what you shouldeat for the day.
Speaker 3 (27:31):
Well, I do use it to spin up a decent amount of
recipes, so it does tell me whatto eat for the day, but in
general, yeah, I think it's it'snot healthy to get super
plugged into it.
You hear stories of people withAI boyfriends and girlfriends
and with ChatGPT 5 just gettingrolled out.
They pulled out all those oldmodels that people had their
(27:53):
boyfriends and girlfriends on,so they're probably not having a
good time right now, but Ithink it's important, especially
as, I say, more AI is going to.
Ai is just going to keepexponentially rising more and
more and it's definitely atemptation and a risk to let it
run more of your life becauseit's it's so efficient and it's
(28:14):
so optimized and it it's notnecessarily wrong when it tells
you how to handle your problems.
But there's a risk that you andthe GPT personality just kind
of merge right and then you knowyou start talking like GPT-5 on
the internet and we're alreadyseeing that.
How many MDASHs have you comeacross since GPT-5 rolled out?
I mean that's a.
I still.
I'm at the point now where Ismile a little every time I see
(28:35):
someone with a typo in somethingthat they write because I know
it's written by a human being.
Speaker 1 (28:48):
It's isn't that so interesting because the m dash
in particular so good peoplelistening, uh, insight, the m
dash is, if you, if you go dashdash, it usually forms into a
longer one, but it's a hallmarkof something written by um, by a
bot.
And what's really interestingis, uh, for a really long time I
actually didn't know how tomake an m dash and and I was
like, so if I don't know and Iwrite on the internet all the
time, I'm pretty sure that thisrandom person didn't know and
this is something that they justcopied out of somewhere.
(29:10):
And I just think that you know,there's the whole you know dead
internet theory.
You know, with like, we're justthere's only so many people on
the internet and we're justreacting to information.
That's not that was, you know,created by bots.
And it's just like no one isactually on the internet anymore
.
We're not even talking to eachother.
But I'm not about you're a bot,exactly, people.
(29:31):
It's something I've beenthinking about.
It's kind of funny.
Why do we have this episode?
Chat tpt5 came out, but to meit's just.
I think it's a reasonablemoment to check in and have us
have like a real like, let'slook at this thing.
You know that whole.
You were the one that showed methe south park episode about uh
chachi vt and that was back.
I was back.
Uh, remember we were in raleighwhen we watched that.
(29:53):
We went and, um, uh, went tothat event.
Yeah, that was a bunch of yearsago and at this point it's so
crazy to think that that ishowever many models ago you know
things that this thing is ableto be way more complex.
I think it's reasonable forpeople to just have a moment to
analyze their own usage inrelationship with this, with
(30:14):
this tool.
You know, if you only got ahammer, you're going to look for
ways to use a hammer.
So it's about being a littlebit of having a little.
It's about being a little bitof having a little bit of touch
and a little bit of I don't knowcreativity with how you're
using it, and not to lose thehumanity of it all.
You know what I mean For sure.
Speaker 3 (30:32):
It's a tool like any other.
You know, it's a tool that'sdeveloped a lot quicker than I
think most people were expecting, even people who thought years
ago that AI might be coming soon.
I think it has caught a lot ofpeople off guard with how
quickly it's gotten here, whichis kind of how the last few
innovations happened.
No one had a smartphone intheir pocket until everyone did
(30:53):
right.
No one had an email addressuntil everyone did, so it's just
one of those things where it'sjust it's here now.
No one rode a Lime Scooteruntil they dropped those by the
hundreds in cities all over theplace.
But yeah, it's a tool.
It's not one that you shouldthrow your entire life at, even
if it's tempting.
Use it for its purpose ofhelping you figure out searches.
(31:13):
Help it analyze whatever yourbusiness need is, but don't give
too much of your personalityaway to it.
And also be careful whatinformation that you give them,
because even if you tell themdon't train my information on
future models, there's noguarantee that they're not going
to do other selling of yourinformation to advertisers or
their own internal analytics.
(31:35):
I wouldn't be surprised if theyhad a GPT that combed through
everyone's account just to findthe fun little nuggets of
information to pull out.
So be careful with that side.
But yeah, if you can do allthat and you can use it as a
tool and be effective with it.
Speaker 1 (31:49):
I think it'll be great for you.
Yeah, more power to you.
Yeah, just, buyer beware, justbe smart, andy.
I think maybe that's wherewe'll end it Just kind of maybe
a little more of a kind of athoughtful, introspective
episode.
Normally I'm doing interviews,but sometimes I just think it's
kind of fun to kind of, you know, muse about something and kind
of think, but, andy, I reallyappreciate it.
I'll let you get back to theend of your day.
(32:11):
Thanks, everybody.
I'll see you next time.
New episodes every Tuesday.
Bye.
All right, everybody.
That's another show.
Thanks so much for listening.
My guest this week was AndyZoki, the backend developer at
Punchmark.
He's also one of my bestfriends.
This episode was brought to youby Punchmark and produced and
(32:33):
hosted by me, michael Burpo.
This episode was edited by PaulSuarez with music by Ross
Cockrum.
Don't forget to rate thepodcast on Spotify and Apple
Podcasts and leave us feedbackon punchmarkcom slash loop.
That's L-O-U-P-E.
Thanks.
We'll be back next week,tuesday, with another episode.
Cheers, bye.
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!
Stuff You Should Know
If you've ever wanted to know about champagne, satanism, the Stonewall Uprising, chaos theory, LSD, El Nino, true crime and Rosa Parks, then look no further. Josh and Chuck have you covered.
The Joe Rogan Experience
The official podcast of comedian Joe Rogan.