YT YouTube2Text

YouTube2Text → Transcripts → How to Actually Build & Sell Software with AI as a Non-Techie

How to Actually Build & Sell Software with AI as a Non-Techie — Transcript

by Nate Herk | AI Automation · 13,311 words · 1,890 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00So you've been coding in Python for over
0:0310 years now. How different do you view
0:05the software engineering space now that
0:07we have all these AI tools?
0:08>> I don't write a single line of code
0:11anymore. Not a single. So everything
0:13goes through codecs or through cloud
0:16code. Think about any skill you would
0:18learn. You're not going to start at the
0:19elite level. But right now we all start
0:22with these same really great tools that
0:24can help us produce code.
0:25>> If you had to put a number to it, how
0:27much faster do you think you're able to
0:28move now? Well, compared to when I was
0:30just starting out, probably 100x, you
0:32can build useful stuff with one prompt.
0:34Heck, you can even build entire software
0:36products that you could sell even if you
0:39have a limited technical background.
0:41>> Do you have any examples of maybe
0:42landmines that you've stepped on when
0:44trying to scale something,
0:45>> the stakes get bigger and bigger and
0:46bigger as you start to build software
0:48that's used by more and more people,
0:50there are certain obligations when it
0:51comes to data privacy and protection,
0:53security breach. That's one thing that
0:55you cannot take back. Once those emails
0:57and that personal data is leaked, you
0:59can't go back.
0:59>> If I was starting today with no
1:02technical experience, I would be most
1:03excited about this.
1:08>> All right, Dave. So, you've been coding
1:11in Python for over 10 years now. Your
1:14team has delivered 50 plus AI solutions
1:17with Glido. You are shipping new
1:18features and updates like every week. I
1:21just want to know how different do you
1:23view the software engineering space now
1:25that we have all these AI tools.
1:28>> Oh man, what a question to kick this off
1:29with. Well, obviously it's completely
1:32different. It's completely different.
1:34What happened over the past three years
1:36and especially what happened in the last
1:38year is just wild. It's just crazy in
1:41terms of what we can do right now. And
1:43we're all experiencing this, right?
1:44We're all using these tools. We're all
1:46producing lots of code. We're building.
1:48And so first of all, it's like really
1:50exciting. So if you ask me like what is
1:54mainly different from when I started out
1:57back in 2013, so like over 10 years ago
2:00already, well, first of all, coding was
2:02really freaking hard [laughter]
2:03and learn and learning and learning to
2:05code as well. And you would just spend
2:08so much time working on this tidy little
2:11program, looking up Stack Overflow, and
2:13then trying to make stuff work, right?
2:17So you could build small little things
2:20and then if you worked on something long
2:22enough you could build something useful
2:25maybe right now fast forward in the
2:29world that we live in today you can
2:31build useful stuff with one prompt and
2:33we all know that we see it on your
2:35channel like every week just one prompt
2:38and you you can build stuff. So what
2:40that then also means is that we can now
2:43just think bigger as developers, as
2:47builders, and we can now do stuff and
2:50build stuff that just wasn't realistic
2:54or possible a couple of years ago. So
2:57that means we can build automations for
3:00yourself, for personal productivity, for
3:02your company, inside the company, for
3:03clients. Heck, you can even build entire
3:06software products that you could sell
3:08even if you have limited a limited
3:11technical background. That's now all
3:13possible and it's getting faster every
3:15week.
3:16>> If you had to put a number to it, how
3:19much faster do you think you're able to
3:20move now?
3:23>> Well, compared to when I was just
3:24starting out, probably 100x. Well,
3:27that's not fair because I was a complete
3:29noob. But let's say let's [laughter] do
3:30a little bit more of a of a realistic
3:32take. Let's say we go back to 2022.
3:38At that time, I was already working as a
3:42freelance data scientist, data analyst
3:45for like two or three years at that
3:47point. So, I was professionally writing
3:49code mostly Python code almost every day
3:52for already three years.
3:55>> Okay. If I look at that point right now
3:57and see like how much I could ship and
3:59produce and do and all of that compared
4:01to what I can do right now, I would I
4:04would still put it like a solid 10x like
4:07a solid 10x of actual useful like
4:10production ready code that you can ship
4:13and on some projects is maybe even more
4:15because I can now also work on projects
4:19and work with languages that before were
4:22just not possible. So I come from a
4:24Python background, data scientist, later
4:26AI engineering. So I used to make
4:28backend systems. But if you were to ask
4:30me to build like a a shiny like front
4:32end or dashboard, I have like I have
4:34never written a single line of
4:37JavaScript like actually in my life
4:40pretty much. I only started to pick that
4:41up once the coding agents were there. I
4:43was like, "Yeah, like let's just try
4:45it." Right? So that is I think a
4:47distinction like like a solid 10x on on
4:50the work that I already did
4:52professionally but like probably even
4:54more than that on other like
4:55possibilities that I just couldn't do
4:57before.
4:59>> Gotcha. So before kind of preai
5:02engineers would kind of typically
5:03specialize in either like more of the
5:05back end at a very high level like a
5:06back end or a front end and it was a
5:08very different way you thought about
5:09writing that code.
5:11>> Oh yeah for sure. And you of course you
5:13have full stack engineers as well. But
5:15then even with full stack it would it
5:18was probably focused on like specific
5:21languages and you know you back then you
5:23also had these like 10x engineers right
5:25like the the old legends of these and
5:28you have you had these like just crazy
5:30developers that could do it all but
5:32generally speaking mostly people would
5:35either do back end front end security
5:38infrastructure they would have like one
5:40or two languages that they're
5:41comfortable with and that is now um that
5:44that is now completely like flipped on
5:46its head.
5:47>> So now instead of really having to
5:50specialize, everyone that's building can
5:52sort of be a jack of all trades there,
5:54more of a generalist and and feel
5:56confident that they are kind of able to
5:59write high quality code in maybe some of
6:01the areas that they weren't super
6:03proficient in beforehand.
6:05>> Yes, for sure. And I also would highly
6:08recommend
6:09everyone to try and do that because why
6:12is that the case? If you don't if you
6:15stick to specialization where of course
6:18like there there's always value in
6:20specialization right so maybe people
6:22watching if you're working in enterprise
6:24let's say you're working for meta or
6:25Google or whatever and you're part of
6:27like a highly specialized team
6:29>> you need to be top 1% of what you're
6:32doing right you need to be a hyper
6:33specialist [clears throat]
6:34>> but let's be real most people don't work
6:37in the top 1% at mad at Google and all
6:39of that most of us are like we're
6:42building we like to build stuff and
6:44there you can better nowadays be a
6:47generalist because if you specialize
6:50means you focus on one part of the
6:52project let's say means you also need
6:55other people to deliver something end to
6:57end to ship something and now your AI
7:00agents can move way faster than you can
7:04communicate and keep can keep up with
7:05teammates so if you cannot work across
7:08the entire stack you are quite quickly
7:10going to hit a bottleneck where your
7:13agent is all already ready, right? It's
7:15ready to ship. It's ready to work on the
7:16next part. But if for example, I don't
7:18know, there is not an integration with
7:20the front end because you only do back
7:21end or um you don't know how to work
7:23with the databases and adjust your back
7:25end to the database, right? there is a
7:27dependency and that creates problems and
7:31I've seen really in the first of all the
7:34projects that I've worked on the teams
7:35that I've worked with but also now what
7:37we do in glido and how we build glido
7:39that we have very of course still we
7:41have separate roles and responsibilities
7:42and we have more people but we give
7:44people complete ownership on a
7:48particular aspect of the product where
7:50they can really ship from something from
7:52like start to finish on their own
7:55without needing someone else other than
7:57like maybe a review from someone in the
7:59process, right? But not like actual work
8:00or code that needs to be developed. So I
8:03think being a generalist is the best way
8:05to go. It's now possible. So you can
8:07just get better at um engineering, learn
8:12engineering principles and then which
8:13language or framework you apply that to
8:17uh starts to matter less and less.
8:19>> Real quick guys, I just have to take a
8:20second to tell you about the sponsor of
8:22today's video, Code Rabbit. So, I know
8:23that a lot of you guys are building with
8:25coding agents now. And the code
8:26definitely piles up faster than you can
8:28actually read it because the agent
8:29doesn't just touch one file at a time.
8:30It goes through your routes, your
8:32schemas, your tests, and your config in
8:34a single pass. And then you open the
8:36pull request and GitHub gives you an
8:38alphabetical list of files. So, you're
8:40rebuilding the logic in your own head.
8:41So, Chainstack is Code Rabbit's review
8:43interface for that exact problem. What
8:45it does is it splits the pull request
8:46into cohorts of related work and orders
8:49them by what depends on what. So, it
8:51reads it in the order that it was
8:52actually built. It opens on an overview
8:54page with the summary, the walkthrough,
8:55and what's blocking merge, merge
8:57conflicts, and failing CI checks you can
8:59fix from the page in one click. There's
9:01also a timeline of every change,
9:03approval, and comment. So, you can see
9:05why each decision got made. And the
9:07semantic diff view is the part that I
9:08would probably use the most because when
9:10code gets moved, a normal diff shows it
9:12deleted on one side and added back on
9:14the other, bearing the real change. The
9:16semantic view shows you what actually
9:18changed. And you can ask the agent
9:20questions about the PR right on that
9:21page. So, if you're shipping more agent
9:23code than you can read, there's a free
9:2514-day trial. No credit cards required.
9:27The link is in the description. And huge
9:29thanks to Code Rabbit for sponsoring
9:30this part of the video. Now, let's get
9:32back to it. I hear you. And I think that
9:34I'm really excited for this conversation
9:36because my audience is typically people
9:38that are coming from a nontechnical
9:40background. And it seems like your
9:41audience is like the exact opposite. And
9:45>> one one thing that I hear from my
9:47audience like when you know I'm building
9:48stuff and I always sometimes wonder is
9:51this right? You know cuz the only really
9:52source of truth that I can trust is my
9:54agent and that's not always going to be
9:57the source of truth that you want to
9:58trust.
9:58>> I'm curious you know I want to zoom out
10:00and I want to start from the very
10:01beginning but real quick before we
10:03transition into that. I'm just curious
10:05how did you feel about the whole vibe
10:08coding era like when that term really
10:09started to pop up and you as a software
10:12engineer there was probably so many
10:14alarms going off in your head when
10:16everyone was just picking up these tools
10:17and started you know building apps and
10:19stuff like that.
10:20>> Yeah. So first of all I found it very
10:23exciting because the thing is I was also
10:25very much part of that whole vibe coders
10:29journey even though I have uh experience
10:32as an engineer like I've said I was now
10:34also creating front-end applications
10:36fully custom code websites
10:38>> and I wasn't reading the code and I
10:40couldn't care so like that would also
10:42put me in in a category of like five
10:44coder
10:45>> and even though that I technically know
10:48how to like set up an architecture and
10:51how to secure your applications and all
10:53of that, right? That of course helps and
10:54that adds a little bit of um just extra
10:57experience to it. But I'm very much pro
11:00pro vibe coding. You just need to know
11:03what the stakes are depending on what
11:05you're working on, right? Because
11:07>> if you're building if you're building
11:08your personal website,
11:10>> by all means like vibe away. There's
11:12like zero risk, right? But it it's
11:14always about like risk versus reward
11:16trade-off. So I think it's awesome that
11:19everyone can build right now. I can
11:20build more. Uh that's super fun. Anyone
11:23can build. And then the question is you
11:25just need to start being careful like
11:26once you let other users, other people
11:29use your software and especially when
11:31personal data is involved, right? So
11:33that's that's I think where you need to
11:35draw kind of like a fine line in terms
11:37of like how far can you really just push
11:40things without really understanding
11:42what's going on and just going off of
11:44the looks or like checking whether
11:46something works and diving deeper into
11:49hm let's see if this actually is secure
11:51because now it's not your only your own
11:53data that you're working with but it's
11:55also other people's right so I think
11:57that is um yeah so that is kind of like
12:00my view on the whole vibe coding uh turn
12:04more or trend really that was going on.
12:06>> I hear you. Yeah. I mean it's very cool
12:08because we see all of these stories of
12:10people, you know, from whatever sort of
12:13background that just have this idea and
12:16then they start talking to an AI agent
12:18and they're able to actually run with
12:19it. And we've seen some pretty cool
12:20exits. We've seen people just change
12:23their lives because they've been able to
12:24take their idea and turn it into
12:25something. So, let's actually like kind
12:27of walk through that process a little
12:28bit. If someone has an idea and they
12:30want to sit down this weekend and they
12:32want to start building that out, what
12:34how would you do that? What what sort of
12:36steps and planning do you go through in
12:37order to start building some sort of app
12:39or software?
12:42>> Great question. So, so let let's walk
12:45through this. So,
12:47first of all,
12:49you need to get a rough idea
12:52of how big the thing is that you want to
12:57build.
12:58Because when it comes to building
12:59products and let's focus on software
13:02products uh right now to make it a
13:05little bit simpler right um there's
13:07different level different levels to
13:09things
13:10>> okay
13:10>> so first of all I would say ask yourself
13:12the question is this something is this a
13:16software tool that just you want to use
13:19like because you want to have some type
13:22of tool either something that you're
13:24currently paying for right now that you
13:25don't want to have a subscription for
13:26anymore or just something that doesn't
13:28exist or you want to combine m multiple
13:30tools into one like is it something that
13:32you just want to use. So, so that that's
13:34like one question because another level
13:36could be uh for example, yes, it's for
13:39me, but it's also for example for people
13:41within my company. So, maybe it's like a
13:43team thing, a company thing or maybe
13:46it's something where you say I want to
13:48build something and I want to sell that
13:49to other people. So, I want to have this
13:52as like an AI agency offer maybe, right?
13:55So, I sell it to other companies or I
13:57sell it to other people. Mhm.
13:58>> And that could be partly productized
14:02meaning you build something you package
14:03it and you sell it or it can be partly a
14:05surface right so there's you build
14:07something but then there's a surface
14:08element to it as well right so you help
14:10implement it into businesses this is all
14:13related to for example like AI
14:14automation right as you know no single
14:16automation is ever 100% the same so
14:19there's also an a customization
14:21component to it and then the another
14:23level is do you want to build like a
14:25software product like fully handoff that
14:27people can just jump on a subscription.
14:29They can just go to a website, create an
14:31account, download it, get on a
14:32subscription, and use it. For example,
14:34what we're doing with Glido, right? So,
14:37there's these four different categories
14:40uh that that your ID could fall into.
14:43And why is that important? Because as
14:45you progress up that ladder, I say I
14:47would I would position it as a ladder.
14:49Um it gets more complex and the stakes
14:52are bigger. Again, summarize, if you
14:55just build something for you, you use
14:57it. Stakes are very low. Just just go
15:00and build. Just open cloud code and go
15:02for it.
15:03>> If you use it inside your team, other
15:06people use it. It's still considered
15:07internal, right? But you may already
15:09want to be a little bit more careful
15:10like who's going to use this? What are
15:12what kind of like data is going to be in
15:13this application? Now then if you sell
15:16it to to uh a company there's even like
15:20a legal aspect involved right because
15:23you probably have a contractor's terms
15:25of service there is certain there are
15:27certain obligations when it comes to
15:28data privacy and protection so there's a
15:31bigger risk
15:32>> and then the and the other level is the
15:34like when you actually build the
15:36software product and you let um
15:37consumers use it we have the whole idea
15:39of data privacy and and personal rights
15:42and all of that when it comes to data so
15:43and and probably more people using it.
15:45So just the risk gets bigger. So um
15:48let's pause there for a second because
15:49that that's what I think uh it's good to
15:51start there.
15:52>> Um so was that all clear?
15:55>> Yeah. So it sounds like before you start
15:57building you want to kind of have an
15:58idea of essentially you said how big and
16:01that makes me think like how many how
16:03many people will use this? How many
16:04hands will this touch? And would it be
16:07fair to say that a good way to get
16:09started is you kind of start on that
16:11first rung of the ladder where it's like
16:12I'm going to build this first just so I
16:14can use it. And that's sort of the way
16:16I'm thinking about it. And that doesn't
16:18mean that later if I wanted to have my
16:19team use it, I couldn't scale up the
16:21back end because, you know, it sounds
16:23like as you move up the ladder of
16:26requirements and people, you have to
16:28think about other things like the
16:29architecture a little bit more and the
16:31database and just making sure that it
16:33can handle this stuff, the privacy. But
16:35it's totally fine to start off on that
16:37first rung and then when you need to
16:38move up, you can start to do that,
16:41right? It's not like you choose, okay,
16:42this is going to be a personal tool and
16:43then later, maybe six months later, you
16:45realize you want to scale it up. You're
16:46not like stuck, right? You're always
16:48able to continue to move up the ladder.
16:50>> Yeah, I think that's a that's a very
16:51good clarification because of course you
16:53can always like make the ambition
16:55bigger, right, of the project. And I
16:58think one of the tr uh one thing to
17:00watch out for is for example like
17:02skipping levels and being someone for
17:04example no technical background and
17:06jumping straight into building custom
17:07software and then selling that to people
17:09and letting companies rely on
17:10[clears throat] it. Right? So there you
17:12skipped a couple of steps in terms of
17:13like your homework in order to figure
17:15out okay what does this look like? What
17:17does this mean? What data is flowing
17:18through it? So I think that is a a very
17:20like natural way to progress in there
17:22because you're totally right. So when we
17:24make that a little bit more kind of like
17:26tactical as to what's different right
17:28there is not much different nowadays in
17:31terms of the process it is still even
17:34for my projects like I don't write a
17:37single line of code anymore not a single
17:40so everything goes through nowadays
17:42either through through codeex or through
17:44cloud code or even through grot
17:46depending on kind of like who has the
17:48best models so everything goes through
17:50the coding agents so that's also
17:52interesting like for the audience uh
17:55watching that I think this is the very
17:58first time in history where someone with
18:02zero coding skills and someone who has
18:04been doing this for over a decade use
18:06the exact same tools and the exact same
18:08process which is quite quite wild to
18:12think about it right now just like now
18:13that I'm explaining this I'm just like
18:15making this up but now I think about it
18:17it's actually pretty wild because think
18:18about any skill that you would learn
18:22right if you just start with, I don't
18:25know, playing basketball or anything
18:27like that. You're not going to start at
18:28like the elite level like your exercises
18:30and what you do is going to be way
18:32different, right? But right now, we all
18:34start with these same really really
18:37great tools that can help us produce
18:39code. So the only thing that is like
18:40different is what instructions do you
18:42give it and how do you review the
18:43output, right? And then that's the
18:45skill. That is where like the reps come
18:47in and that's where also the stakes get
18:49bigger and bigger and bigger as you
18:50start to build software that's used by
18:52by more and more people. But I think
18:54that is like very motivating for like
18:56people to watch. Like you have these
18:58tools you can jump on like a $20
19:00subscription. You probably want to get
19:02to like $100 and $200 otherwise you know
19:05you run out.
19:05>> You'd be sitting there all week like man
19:07when is this reset coming? You probably
19:09want a little bit more, but like that
19:10for most for most people that is within
19:12reach and from that you can start you
19:15can start building and you can start to
19:16build really really cool things and
19:20>> yeah start building with something that
19:23you would use that you think is cool and
19:25then start to make it make it better and
19:27then over time um and we can we can get
19:31into this deeper if if you want to but
19:33there are like for for for the most part
19:36VIP coding is totally fine. But there is
19:39one aspect of it that is security where
19:41you just cannot afford to make the risks
19:43because let's say even if you build an
19:45application and it's like vibe coded and
19:47the architecture is bad like what's the
19:50worst thing that can happen like people
19:51will use it and at some point like it
19:53will like get a little bit slower or it
19:56will break but you'll notice right
19:58you'll start fixing it. So there's
20:00nothing really like so like and that is
20:02something you'll be aware of because
20:03people start complaining it's like huh
20:05maybe maybe we need to revisit that but
20:07like a security a security breach that's
20:10one thing that you like cannot take back
20:12out sorry we'll fix it in the next
20:14version you know like that's that's out
20:16once once those emails and that personal
20:18data is leaked you can't go back so that
20:20is one aspect that uh we could get into
20:22that if you if you want to but that is I
20:24think the most important thing as you
20:25level up security
20:27>> totally totally and I definitely want to
20:28get into that and I think we'll just
20:30kind of keep working our way up here.
20:31That's why I love the ladder analogy
20:33because you really do have to earn your
20:34way up to the next rung. You can't just
20:36jump from the ground to rung four or
20:38you're probably going to fall and you
20:40you just have to start from the bottom
20:41again. So
20:42>> yeah,
20:43>> it sounds like the first piece is kind
20:44of you figure out how big and you you
20:46think about high level like how big am I
20:48building this and how much risk is there
20:50involved?
20:50>> Yeah,
20:51>> once you kind of have a good idea of
20:52that,
20:53>> what is the next step to start that
20:55building process? What does it look like
20:56for you to sort of map this out in your
20:57head to sort of plan out the
20:59requirements? How does that process
21:01look? Or if that's not the next step,
21:03what is the next step?
21:04>> Yeah. No, that is that is definitely the
21:06next step. So, uh, when you have this
21:08idea and when you want to start just
21:10just building, um, what I've observed
21:13over the, uh, past year, pretty much
21:17one year ago,
21:20everyone was really like pushing
21:25um, spec driven development, the
21:27importance of plans, and you can
21:29probably remember this, right? Um, and
21:32I've just found that as the models get
21:34better and better and better and the
21:35models that we have to right now like
21:38the the planning phase and the
21:40generating the specs becomes like it
21:42becomes less important I found.
21:44>> So what does that mean spec planning?
21:46>> Uh yeah. So spec spec different
21:48development was is uh refers to kind of
21:52like a way of building software where
21:53you first use AI not to produce code but
21:56to produce specs and what are specs?
21:59specifications. So you describe how the
22:03project or the product should work in
22:05natural language. So you and and you
22:07could and you would use AI for this. So
22:09you would pretty much say like I want to
22:11build XYZ, I want to build this and then
22:14let's create some specifications for
22:16this. So it would say the AI would start
22:18to create a document. Okay, it product
22:20needs to do this, it needs to do that,
22:22it needs to do that, it needs to do
22:23that, etc. And you would create this
22:25this plan
22:27um and the specifications. And then once
22:29you had all of these documents
22:31only then you would read those and you
22:33would approve those and then only then
22:35you would like go into uh development
22:37mode and you would actually ask your
22:40agents to start uh writing the code and
22:43to start building the thing. That was
22:46pretty much the meta um one year ago
22:49when we just got these like big
22:51breakthroughs with cloud code open 4.5
22:544.6 six, right? Like the good old glory
22:57days of AI [laughter] AI agents. Um, and
23:00now it's just like getting crazier and
23:01crazier. So, um, the thing is there's
23:04still value in that and I don't think
23:06it's wrong to go through that process
23:08even just to help you think about what
23:09it is that you want. But the thing is
23:12what I find myself doing most of the
23:14time right now, I just like I go
23:16straight into building and like I let
23:18the model figure out what it like how it
23:21should structure things. And there are
23:23of course certain kind of like standards
23:27that I kind of like steer the model
23:29towards when I'm for example working
23:32with the Python programming language
23:34just because I've so much experience
23:36with that. But like the summary of this
23:39it gets easier and easier and easier to
23:41just tell it what you want and it will
23:44start to it will start to build it will
23:46start to build it. So coming back to you
23:49started this question. So let's say
23:50people have this idea and what is the
23:53next step? I would say right now get on
23:55codeex, get on cloud code, use GPT6
23:58Astra or Opus 5.5 which are currently
24:01like two top frontier models. Opus 5.5
24:05really good so far.
24:07>> Super good.
24:08>> Super good. And just tell it tell it
24:11what you want to build.
24:12>> And and the Yeah. And as as simple as it
24:16sounds for people just starting out,
24:18there's not much more to it. And then
24:20you just look at the output and you tell
24:22it what you like, what you don't like,
24:25and that will already get you really,
24:28really far.
24:30>> Yeah. And and the cool thing about that
24:32is that I think humans are really good
24:33at explaining what they want. Like, you
24:35know, it's easy to say, I want this, I
24:36want this, but you really just have to
24:38sort of go through that grill me process
24:40where you have AI interview you about
24:42>> what specifically do you want? What do
24:45you not want? What does this look like?
24:46What is your vision? And I was just
24:49talking to a few Google engineers and
24:50and they were explaining this as saying
24:52like
24:53>> the plumbing, the boring stuff on the
24:55back end that that we used to have to
24:57think about
24:58>> where humans used to have to set that
24:59all up before they could start to be
25:00creative and think about the vision.
25:02That plumbing is pretty much just taken
25:03care of now because you explain what you
25:05want and the agents know how to go set
25:06up all that stuff so that you can really
25:08just they kept calling it intent driven
25:10automation or intent driven engineering
25:13and I thought that was pretty cool. I
25:14wanted to see do you agree with that? Do
25:16you agree with that narrative or that
25:18mindset?
25:18>> Yeah, for sure. I I really like that. I
25:20like that term intent driven. And um
25:23what does it mean? You just tell the
25:25model what you want. Like you you have a
25:27certain intent as a builder, you have an
25:29idea and you want to make that a
25:31reality. And the thing is often
25:34when you start with something even
25:37though you might have like
25:40a rough idea of where this thing should
25:43go
25:45until you actually see it in front of
25:47you and use it. Um
25:51you cannot really give the the best
25:54feedback I found. So you can like then
25:56that's why I've moved away from like big
25:58plans and specs and all of that. just
26:00like just just build me this and then I
26:02see it and I look at it and I test it
26:04and it's like hm this feels a little bit
26:06off or can we make that faster or this
26:09doesn't really look right like whether
26:11whether it's visual or actual logic
26:13right or the output here is wrong like
26:15can we fix that right and um that's also
26:18now the pretty much the like the
26:21development process with glido the tool
26:24that we're building right now it's
26:25mostly like ideas and then just using
26:29natural language like it's actually fun.
26:30We're using glider to build glider,
26:32right? So, we use glider to talk to it.
26:34It's like, hey, can you try this, you
26:36know, like let's build it out. And often
26:38I do that straight on the um like
26:42production codebase. I'll do it on a
26:44separate branch, right? So, it's not
26:46like immediately shipped, but like there
26:47there's there's not even like a
26:48playground. I just create a new branch
26:50and it's like build it and then I can
26:52directly see what it looks like in the
26:54product itself and then I can iterate on
26:56it. And then that's I think uh a perfect
27:00description of like intent driven
27:02development. You just tell it yeah
27:03>> what what you want and it will build it.
27:06>> By the way guys, I've got this
27:07completely free SOP for you about
27:08getting your first a automation client.
27:10It's going to go over the exact steps
27:12that has been proven for hundreds of our
27:13AIS Plus members to get their first paid
27:16gigs. It goes over the one sentence
27:17service pitch that can get you started
27:18today. Why your first client should cost
27:20you money. The 5-minute video that
27:22answers can this person actually deliver
27:23before you've actually received any
27:25money. what to do when you have zero
27:26case studies. There's so many good
27:28things in here that are going to help
27:30you out. Even if you already do have
27:31clients, I would recommend grabbing this
27:33because like I said, it's yours
27:34completely free. So, if you want to grab
27:35this, there's a link for it down in the
27:37description. Let's get back to the
27:38video. That's so cool. It It's really
27:41interesting how you kind of pointed out
27:43how the meta a year ago was doing all
27:47this kind of planning, but now you say
27:48you're you're you're kind of building
27:50first more and then really just getting
27:51it in your hands and playing around with
27:52it. I remember when I used to use plan
27:54mode first for everything even not even
27:56if I was building automation but I would
27:58just use plan mode default just to sort
28:00of brainstorm and it sounds like maybe
28:02do you think the models are just getting
28:03better and the models and harnesses are
28:05getting better so that's why the med has
28:06shifted or
28:07>> yeah it's just and it's it's for sure
28:10the combination so that's also important
28:11to um to understand um and I think even
28:17like people that are actively working on
28:19clawed code on the development I think I
28:21even saw this on X like entropic
28:23engineers mentioning this as well that
28:25plan mode is not even really needed
28:26anymore because yes like models get
28:29better but um it is essentially also
28:32within the harness because what what
28:34what is really a plan right it is first
28:37>> essentially getting you give it you
28:39always start with a goal or a task I
28:42want to build this and then before it
28:46would help the model to first plan out
28:48so be before just getting into like uh
28:51code generation
28:52it would first like zoom out a little
28:54bit. Okay, let's just first reason over
28:57this like let's see if everything makes
28:58sense and then once we have the plan I
29:02can now execute on of it and it feels
29:04like um the engineers working at OpenAI
29:08and Tropic is essentially by making
29:11those models and the harnesses better
29:13and more capable they just merge this
29:15process as like a trait that the model
29:18and the harness combination already have
29:20without like turning the plan mode on H
29:24that makes sense. So now let's say we
29:27have started to throw in some of the I
29:30want statements and now we're getting
29:32some stuff back.
29:33>> Yeah.
29:34>> In automation,
29:36>> you know, we run a lot of evals and we
29:38see, okay, cool. We have this, you know,
29:39we have these inputs, we have these
29:40outputs, this is, you know, 90% correct.
29:43Let's make some iterations and and see
29:44if we can get this to like 98 99. How do
29:47you think about from a, you know, an app
29:49perspective or a software perspective
29:51essentially the eval process? What does
29:53it look like to verify? What does it
29:54look like to see if the things actually
29:56work rather than just assuming it does
29:58or having you, you know, go in there and
30:00click through a few times?
30:01>> Yeah. [snorts]
30:03So,
30:05this all this of course depends very
30:07much on what you're uh building, of
30:09course, because uh output quality. So,
30:13what what does good look like is always
30:17tied to the application that you're
30:19building, right? So, let's
30:20[clears throat] talk about glido. What
30:21does good look like? Well, um, Glido
30:25works well. If you press a button,
30:27people can talk. That will be processed
30:29in under 500 milliseconds and the text
30:32will be pasted back into your
30:33application like exactly the way you
30:37wanted to see there, ready to paste
30:38text. So, it will correct some errors
30:40and all of that while leaving generally
30:43most of your speech intact. Like that's
30:45what people want from a dictation app.
30:47So it starts with that getting clear on
30:50what does good look like.
30:53Then the second question is can we find
30:56examples of that? Can we find examples
31:00of cases uh where it went well? And in
31:03glido for example uh we don't use user
31:07data because we like really designed
31:09this application to be private by design
31:12but we use the data of uh uh our uh our
31:16accounts internally meaning the
31:17developers that work on glido right so
31:19we use our own kind of like data sets
31:21that we collect we have a bunch of
31:23people so that data set then also grows
31:25and in there we can run these
31:27experiments so we have data sets in
31:29there like when we put this into our
31:31system we expect this to come out and
31:35and that essentially then becomes
31:36becomes an epha, right? Uh that you can
31:39run and that you can check against.
31:40>> Uh the cool thing is that as models and
31:44harnesses get better and capable,
31:47>> whenever you have some data and when you
31:50don't have some data, you can even ask
31:52the model to generate some realistic
31:54data for your use case. And again,
31:56whatever you're building, whether that's
31:57a chatbot, an automation of any sort.
32:01And once you give these models an
32:04objective in terms of like here's what I
32:06want this to look like. Here's what I
32:08want you to improve on. I want it to be
32:09faster. I want it to look more like
32:11this. I want it to be better. I want to
32:12have less errors. Once you give that and
32:16let that run in a loop and that pretty
32:18much means you tell it to go and improve
32:20it. Use the data, generate more if
32:22needed, do research, use sub agents, run
32:25this in the loop until it gets better.
32:27like it's a very messy prompt, but even
32:28just saying giving that prompt to a
32:32model like Opus 5.5 given the thing that
32:34you're building will kick off this like
32:37research experiment well where it will
32:39go on it and we'll try to find ways to
32:41to make it better. So what we find over
32:46and over again is like the better models
32:49and harnesses get, the easier all the
32:53things that we have to do as developers
32:54get. planning, spec different
32:57development, quality assurance,
32:59debugging, security, evaluations,
33:02you know, all of these things also get
33:05easier as the models get better. So, I
33:09think that's a really interesting uh
33:11property that that I found uh and also
33:13just something that you need to be aware
33:15of because you do need to ask the model,
33:17right? So, you do need to be aware of
33:19like what are Eiles? How can I set up a
33:21loop in order to make my application
33:23better? So just that realization
33:25>> and asking [clears throat] it
33:27>> is is the right starting point.
33:29>> Yeah. So it sounds like
33:34defining what does good look like and
33:36that's something that you would do with
33:38a human as well. You know, you're
33:39setting the expectations of what do you
33:41want? Because if you don't align with a
33:43human or an agent on what you're looking
33:45for, then when it comes back with
33:47something that isn't aligned, it's maybe
33:48not that it failed. Maybe it's just that
33:50you didn't you weren't clear enough on
33:52that. Yeah.
33:52>> And then from there,
33:54>> giving it a way to essentially verify
33:57that, you know, whether that's you
33:58verifying or or the agent verifying. I
34:00think kind of a combination of both. But
34:02>> what is good? Here's how you verify it.
34:04>> And
34:06um what you were saying about this whole
34:07like loop, it made me think of when
34:08Carpathy's like auto research came out.
34:10And I think that's when I remember
34:11having this big moment of like, wow,
34:13like agents can actually
34:15>> see their goal, prove it, and then keep
34:17researching and keep trying again until
34:18they hit it. And I thought that was
34:20incredible. Now, I've got an interesting
34:21one for you to kind of stem off that,
34:24which is a lot of the best auto research
34:28or goal prompts or loops are when you
34:30have an objective milestone to make them
34:32hit.
34:33>> What about how do you think about it if
34:35it's subjective? If it is something
34:37that's more like taste or, you know, if
34:39the definition of good isn't something
34:41that can be objectively proven.
34:43>> Yeah. Um, that's very true. Whenever you
34:46have something that's very easy like
34:48it's either good or bad but because you
34:51have a data set with examples those are
34:53the best scenarios or making things
34:55faster right so for glido we also have a
34:57lot of infrastructure that we need to
34:59manage uh models that we deploy um and
35:03there also code code and architecture
35:05around that if you just give it an
35:06objective like hey here's what we're at
35:09right now in terms of latency you can
35:11test that by running a test now go make
35:13it faster
35:14>> you know super clear because every time
35:17it runs an iteration, it can see am I
35:19above or below? Am I going to in the
35:21right direction? So, what if you have a
35:23use case that doesn't have this? And uh
35:27let's think about an example of this to
35:30uh make it a little bit more tangible.
35:32So,
35:33>> my most common is like video editing and
35:35I want it to like I'm saying like yeah,
35:37make it feel professional, make it feel
35:38engaging. It's like how does it know?
35:40Yeah.
35:40>> Yeah. Yeah. Yeah. So, yeah. Yeah. So
35:44like that that really is a tricky it's
35:46really a tricky example. I I'll give you
35:48one more and then I'll also consider
35:50this one.
35:50>> Okay.
35:51>> Think about this. But um we have for
35:53example also done a lot of work in
35:57customer support in customer care. And
36:01the thing is when a uh customer asks a
36:06question and the AI agent gives it a
36:10reply, right?
36:13There are hundreds, thousands, probably
36:16even like millions of different ways
36:19that you could reply to a customer and
36:21they could all be correct or they could
36:24all be wrong because the final goal in
36:27this case is that the customer is happy,
36:30right? So like they got a good reply.
36:32But whether you use a different word,
36:34yes or no, that is all um subjective
36:38meaning there can be there can be
36:39differences. So one of the techniques
36:42that you can use for that is what you
36:44would call an LLM as a judge to run that
36:48at scale because and how how does this
36:51work right so like the um the best way
36:55to to check the outputs of your AI
36:58agents when it's subjective uh the best
37:01way which is also the most expens
37:02expensive and non-scalable way is you
37:04you refle
37:08the AI do the work a human refuses it
37:10and says this is not good, this is good,
37:12etc. But then what's the point of
37:14automating it? Because in the case of
37:16like customer care, if a human needs to
37:18review every ticket, they could probably
37:19just reply themselves, right? Because
37:21it's probably easier to reply than to
37:23review and give good comments. So that's
37:26unscalable. But if you do that for a
37:30little while, you could create a little
37:32data set around it. when the uh customer
37:35asked this and the AI said that the
37:38human said this is good yes or no
37:41because you know you start to collect
37:43data and then this whole concept of an
37:46LLM as a judge is you use another
37:50language model so another like
37:52independent language model call or
37:54multiple calls or however you want to
37:56set it up in order to review the output.
37:59So you let an LLM look at the let's say
38:04the reply to the customer and you let
38:06the LLM say is this good yes or no and
38:09why now the challenge is in the
38:13beginning as you know like an LLM might
38:15say yes this is good or no this is not
38:18good but how but then how do you know
38:20that is good right so you like shift the
38:21problem from like an LLM that you need
38:23to check to another so there's a little
38:25trick for that and that is uh
38:27essentially called creating alignment
38:28between the human reviewer and the LLM.
38:32And you can do that by let's say you
38:34have 100 examples and you let the human
38:38review 100 examples and you you list all
38:41of this. You have an Excel spreadsheet
38:42for this whatever. This is like manual
38:43work. Nobody wants to do this. It's
38:45unscalable. But you need to do this
38:47because this is where you can
38:48essentially capture the taste. So human
38:50will say I like that. I don't like that.
38:54That should be better. And now what you
38:56want to do is you let an LLM run over
38:59all of that and you see where they where
39:01do they agree where do both the human
39:04and the LLM say this is good and this is
39:06bad and you'll get a score. So you'll
39:08have a percentage. So let's say the
39:10alignment is 80%. Or maybe maybe even
39:12like 50%. So 50% of the time the LLM
39:16agrees with the human yes or no. This is
39:17baseline. And now once you have that
39:20data this is where the interesting thing
39:22comes in. Now you can use AI again. So
39:24now you can say hey here's human data
39:25here's LLM data let's optimize the
39:28system prompt so we can create more
39:30alignment you know now you can do loops
39:32so then the LLM is a judge is it's a
39:35system prompt and a model that's pretty
39:37much it and then um if you let an LLM
39:42review that like oh yeah I can see the
39:43LM things a little bit different about
39:45this and this and that and I thought
39:46this was good but actually the human
39:48thought this was not correct so I'll
39:49change that in the system pro you run it
39:51again alignment is at 75 5%. Cool. Let's
39:54do another round. 80% 90% 95 99%.
39:59And now if you have a large enough
40:01sample size and you do this occasionally
40:04to avoid model drift, you'll have an LLM
40:08that with two degree of certainty agrees
40:12with how a human would review it. That's
40:15generally how it is done on real world
40:18projects. Coming back to your video
40:20editing example, you would follow a same
40:23process,
40:25but I would say it's probably even
40:27trickier because assessing what good
40:31looks like in an edit or an animation or
40:34a cut, there are some things that are
40:36obvious like some you can make an error,
40:38but if like a title has like a certain
40:41animation
40:43pattern or something like that and you
40:45don't really like that, that's tricky.
40:48But it's mostly because of the LLM's
40:50ability to inspect the visuals,
40:54>> right? I think that's what makes it more
40:55tricky. So, as multimodal capabilities
40:58get better, you can apply this exact
41:01same process even to optimizing and
41:04training your video editing agents.
41:07>> Mhm. Yeah. Yeah, I think the interesting
41:08thing there too is like let's say you
41:10have worked on refining this LLM as a
41:12judge
41:13>> and it's really really good
41:15>> and then if you switch the model it
41:16might not be as good anymore because the
41:18model might interpret it differently.
41:20But what's cool about that is you're
41:22putting in that time which is a tedious
41:24process because you're essentially
41:25creating these standards but then you
41:28can use sort of that loop that you know
41:30auto research loop on it to have it
41:32continuously optimized which I think is
41:34>> very very cool
41:35>> and
41:37it's it makes me think of it makes me
41:40think of like sports teams. They'll have
41:42these scouts that will go out and watch,
41:44you know, high school games or college
41:45games or whatever it is. And there are
41:47stats like height, weight, you know, how
41:49many times can they bench press 225.
41:51There's stats like that that are
41:52objective. But then that's also like the
41:54team trust the scouts opinion to say
41:56this person has like this guy has
41:59potential. He's explosive. He has great
42:01game sense. Like that's stuff that you
42:03can't pick up on paper as much. And it's
42:05like how do you get your LLM to have the
42:08taste and the judgment of that scout or
42:10of you? You know what I mean? And I
42:11think that that is
42:13>> super super cool. Um, now as we
42:17transition into
42:20now we've been running some
42:21verification, we have something that we
42:22trust that's working. What if we want to
42:24turn this from a oneperson app or a team
42:26app, what if we really want to scale it?
42:27Like what are the things that you start
42:29thinking about? Um, and we don't have to
42:30get super super technical, but what are
42:32the high level things you start thinking
42:33about around
42:35infrastructure, databases, security,
42:38that sort of things. And then if you
42:40have any like examples of maybe
42:41landmines that you've stepped on when
42:43trying to scale something,
42:44>> that'd be super interesting.
42:46>> Yeah, let's dive into it. This is my uh
42:48this is my fun zone pretty much.
42:50[laughter] I like it. So um let's indeed
42:53make it a little bit more like tact
42:55tactical and therefore also technical
42:58for the audience because up until this
43:00point also my um recommendations for
43:04people may have been quite obvious. Just
43:06use cloud code, right? Just ask the
43:08model, bro. And the funny thing is um
43:11that's true. That's what you should do.
43:13But there is still a part when you
43:16really want to take things to the next
43:18level where um it is very worthwhile to
43:22like train yourself and to get a better
43:24understanding of how software works. And
43:29I keep coming down to describing that as
43:33just the highlevel architecture
43:36of your application and the software
43:37that you're building.
43:39And there are two distinctions between
43:41that. And this is very worthwhile to
43:44like research a little bit what I'm
43:45about to say. So this could even be you
43:47with like codeex with with cloud like
43:50spend some time asking the the the
43:53following terms that I'm going to
43:54mention. So architecture first of all we
43:57have it at uh what I call the system
43:59level. And this means how do all of the
44:04individual components like your back
44:06end, your front end, and your database
44:09and potentially even other services, how
44:11do they talk together? Because most
44:14software products
44:16are a combination of those things,
44:18right? And it's very it's very important
44:21to understand what the role is of each
44:25and every one of those how they
44:27communicate to each other and what are
44:30some of like the traits the unique
44:32things and also the things that you need
44:34to care about when it comes to those
44:37individual things. Right? So this
44:39depends heavily on first of all what are
44:41you building? Are you building a web
44:44application, a desktop application, a
44:47mobile application? uh something that
44:50combines all of those. So it depends. It
44:52also depends on what language do you
44:55use. Do you for example use languages
45:00where you have a very clear separation
45:02between your back end let's say that is
45:04built in Python for example or in Rust
45:07or in C and do you have a front-end
45:10application so the visual layer that is
45:12maybe in JavaScript you're using Nex.js
45:14JS. So then you have usually like
45:18completely different code bases. Now you
45:20can also have projects that for example
45:23just work in one stack. So you do
45:25everything in TypeScript for example. So
45:27TypeScript that whole project it is your
45:29back end and it is also your um like
45:31your front end your visual component to
45:33that. So those are all things that
45:35matter based on what you're building. So
45:38now that you get a little bit more
45:39serious, that is a good like first
45:42question to ask your AI agents. Hey, I'm
45:45building XYC.
45:47What would be the best stack? And then
45:51stack is the important word here. What
45:52would be the best stack in order to
45:54build this? Do I need a separate back
45:56end and a front end? What kind of
45:58database should I use? Can I put
46:01everything together? So those are then
46:03questions that you can go through and it
46:05will give you examples and you can just
46:06ask like hey what does that mean? What's
46:08the simplest solution? what fits the
46:10best for kind of like what I'm building
46:12and because that is a really important
46:14starting range because once you log in
46:16on a on a programming language like
46:19within the language itself you can do a
46:20lot right you can scale up you can scale
46:22down you can refactor you can build but
46:24switching a project to a different
46:26language even though with coding agent
46:29it becomes easier and easier and easier
46:32that is usually not a trade that you
46:34that you want to make right we had to do
46:36it with glido at some point um it was a
46:38lot of like we started out on a
46:40different stack than what we use right
46:42now because in the beginning we didn't
46:44do our research properly. It was also a
46:46little bit because of team structure. We
46:48had someone working on that and the
46:49other one working on that and then we
46:51decided to go in a direction and then
46:52later we figured out
46:54>> um we need to do this differently. So
46:57that's that's a landmine that uh that I
47:00that that we stepped on. it was choosing
47:02the kind of like things we're we're
47:05familiar with rather than picking uh the
47:09tool or the stack that was really like
47:11most optimized for what we eventually
47:13wanted to build. And in the past
47:17um this would pretty much have not been
47:19possible. So if you are comfortable with
47:21a with a stack, you've been developing
47:22that for 10 years, you most likely want
47:24to use that, right? But I think that now
47:26changes. You could now very well say,
47:28"I'm confident with that stack, but
47:30there is there's evidence there's
47:33actually evidence that this stack is way
47:35better for what we're trying to do.
47:37Let's let's go with that stack." So,
47:40architecture at the system level,
47:42backend, front end, database, how does
47:44everything talk to each other? Then you
47:46also have architecture at like the
47:48project level. So, this is when you go
47:50inside, let's say, your GitHub
47:52repository and how you structure your
47:54projects in there. This is less
47:56important than the higher level system
47:59architecture, but it's still important
48:00to have like a rough understanding of
48:03how do you structure projects in terms
48:05of files and folders. So, what do you
48:07put in there? And how do you structure
48:11your code in terms of functions,
48:15classes, and modules that you create? So
48:19when you're creating code and these
48:21agents get better at it every month, but
48:24you can still have these like spaghetti
48:26code bases. You know, you build
48:27something here, then you build something
48:28there, then you build something there.
48:30And this uh can happen quite quickly
48:33when you are vibe coding and you go from
48:35idea to idea to idea and you experiment
48:37a lot, which is a new way of building,
48:38right? You just ask it to build
48:40something and it might build something.
48:42Now you have this artifact in there.
48:44>> Then you decide, let's leave it there. I
48:47don't really like it. let's come back to
48:48it later. You build something else. You
48:50forget about it, you know. So now you
48:52have this mess. You have this junk
48:54sitting in there and that might have
48:57>> some logic in there which you still
48:58really use and which you still really
49:00need, but it's got all of this bloat
49:03around it that was just a test feature
49:05that you wanted to build. Now you start
49:07to build on top of that and you create
49:08these new features. It's all great, but
49:10they still depend on this kind of like
49:11messy little part that's really deep way
49:14into your codebase. So this is how you
49:15get spaghetti code codebase, right? So
49:18over time it gets harder for your AI
49:20agents to manage and to find where where
49:23everything is. And this is where you
49:25you'll run into bugs at scale. So all of
49:27a sudden things stop working, right? It
49:29it worked fine for 3 months and then all
49:31of a sudden
49:33>> things don't work anymore. And it's
49:34because you removed the feature that was
49:37>> not no longer needed, but it contained a
49:39little part that you actually still
49:41needed. So
49:43>> understanding how that works at a high
49:45level is
49:47>> very valuable to do. And I'm going to
49:49give people like one tip to look into.
49:51There's a really good skill for this and
49:54it's called um I think it was recently
49:57changed but it's from Matt BCO. So this
50:00is like he's famous for his skills,
50:01right?
50:02>> Great engineering skills.
50:03>> Yeah, great great engineering skills.
50:04So, it used to be called improve
50:06codebased architecture, but I think he
50:08refactored that a little bit to I think
50:11codebased design or something like that.
50:13But if you
50:13>> we'll find it, we'll link it in the
50:14description.
50:15>> Yeah, if you search for Matt PCO and
50:16then find some of his skills where he
50:18talks about uh improving architecture
50:21about how to create deep modules and
50:24locality and how to work with with the
50:26seams. That is a skill that I very often
50:29uh use to review my code bases and that
50:34uh he even like mentioned that he
50:36created that skill to fight AI slop and
50:40make code bases that are easier to
50:42navigate by AI
50:44>> and imagine how cool that really is
50:47where I just said pretty much there
50:49there's two components that I still are
50:52that are still I think very high
50:54leverage to learn more about
50:55architecture at like the system level
50:57and the codebase level. For the codebase
50:59level, there's literally just a skill
51:01that you can use. And if you just read
51:04that skill and use it in your coding
51:06agents, you'll get better at it.
51:08>> To me, that's just amazing that that's
51:10how you get how you get better at it.
51:12And then the and then the the like the
51:14front end database back end,
51:17>> spend a day re researching that.
51:19>> Like if you're serious serious about
51:21building stuff, just like ask questions.
51:24like even do it on like lunch break, go
51:25out, take a walk, like start to learn a
51:29little bit about what that means, how
51:30they communicate to each other and also
51:32what the different um security aspects
51:36are of those different levels because
51:38security applies to like your entire
51:41application. But there are specific
51:42things that your database handles u your
51:46back end handles, your front end handles
51:49and the communication layers between
51:51those.
51:52>> Mh. [clears throat]
51:54Yeah, I think it's really cool that you
51:55brought up the skills of Matt PCO
51:57because one of the things that I was
51:59thinking about as you were going off
52:00there was what you mentioned at the
52:03beginning which which was that the fact
52:04that really under the hood software
52:09engineering is still the same but the
52:11human role has changed
52:13>> because of the fact that these things
52:14are so smart and can move so fast but
52:16still the principles of building good
52:18code are still the principles of
52:20building good code. And for someone like
52:22me or for a lot of people probably
52:24watching this, maybe when you started
52:26saying like Python, C, Rust, they were
52:28like, "What are those things?" You know,
52:30but that's been around for so long. And
52:32people like Matt PCO, people like you
52:35know what it looks like to have a good
52:37codebase. And that means that you can
52:39leverage the subject matter expertise
52:41and experience of other people if you
52:42ask the right questions, if you have it
52:44do research, if you leverage skills like
52:46that. And I think that that part is
52:48really cool and hopefully gives people a
52:50lot of a lot more comfort in in the fact
52:52that they're not alone in building this
52:54app because there's so much stuff out
52:56there from people that understand what
52:58what does good look like when maybe you
53:00can't define good as well as someone
53:01else can, you know?
53:02>> Yeah.
53:03>> So, the last piece I wanted to touch on
53:06then is I know you hit on a little bit,
53:08but
53:10what should people be thinking about
53:13when it comes to the security stuff like
53:16not only from like a data privacy but
53:19also how do you think about
53:21what's the potential of your app getting
53:22hacked you know that sort of thing too
53:25how do you think about that especially
53:26if you have no background in any sort of
53:28like cyber security
53:30>> sort of stuff
53:32>> so the cool thing about this is that
53:36with
53:38um very little uh like ju with just the
53:42the basics you can go really are. So, of
53:47course, security is this big domain and
53:50there's different levels to it, but I
53:52also don't come from like a security
53:53background. I wouldn't even consider
53:56myself a security expert. Uh we do have
53:59yours on our team who knows way more
54:01about that and I've picked up most of
54:03the things from him uh over the years.
54:05But like I've said, the cool thing is
54:08there are actually like if you do a hand
54:10uh a handful few things, you already are
54:14it's almost already impossible for you
54:16to get hacked. So let me give you a
54:18couple of those things to look into and
54:22there is a little bit of overlap between
54:26like what we would call infrastructure.
54:28So that is the whole concept of okay we
54:31talked about architecture at like the
54:32system level backend front end database
54:36right infrastructure pretty much means
54:38like where do all of those services live
54:41where are they deployed and how do they
54:43talk to each other and what are for
54:45example the firewalls around them
54:47there's a little bit of overlap in there
54:49and again um very um high lever thing
54:53that you can look into even if it's just
54:56like from a research curiosity
54:58perspective ive to learn more about
54:59this. So let me walk you through some
55:01examples to make this a little bit more
55:03tangible. So
55:05let's start with like a database. For
55:07most people building something, you want
55:10a database, right? Most applications
55:12need a database. You need to store user
55:14accounts, passwords, any information
55:16that you might need in your database.
55:18And right now the most popular option
55:21for that is Superbase. Superbase is a
55:24really popular database. And Superbase
55:26is amazing. If I would say to anyone, if
55:30you don't have a very strong reason to
55:33not use uh Superbase, use Superbase.
55:37Meaning that if you're nontechnical, if
55:39you're just starting out, just use
55:40Superbase. It is flexible enough that
55:42you can do anything with it. Don't be
55:45like uh bothered by people saying, "Oh,
55:47you really need like an unstructured
55:49database for this use case," or like,
55:51"Oh, you need a dedicated factor
55:52database." Like Superbase can do all of
55:54that. Just use that. And then like I've
55:56said, if you have a very strong argument
55:58as to why you don't like Superbase, then
56:00you're probably technical enough to make
56:02the decision on your if you get what I
56:03mean, right? So
56:04>> that makes sense. Yeah.
56:05>> Yeah. So database, use Superbase.
56:07>> So Superbase
56:09um it's you can also self-host that by
56:12the way, but for most people I would
56:14recommend just use it in the cloud. That
56:16means go to superbase.com, create an
56:19account, create a database. There you
56:21can get on a uh a plan and you have your
56:24database. Now then what I recommend to
56:27do is again spend a couple of hours
56:30talking about the security aspects and
56:34best practices that you can set up
56:35around Superbase. Superbase is great out
56:38of the box. But there are some things
56:40you need to be aware of. This mostly has
56:43to do with rowle security. Again I'm
56:45just putting out these terms in here not
56:47to go into them but for people to like
56:49research into that. So rowle security is
56:52important. like look up what that means,
56:55how you can work with that. Uh make sure
56:58that your database has like a strong
57:00password, like common sense, strong
57:02password. Make sure that your user
57:03account that you can log into on
57:05Superbase, strong password and also
57:08potentially two-factor authentication.
57:10This is all very simple stuff, right?
57:12But this is just avoiding like people
57:14and random bots being able to like
57:16randomly guess a password and can get
57:18into something. Now then also, so that's
57:22your database. So now your data lives in
57:23there. If you use Superbase in the
57:25cloud, you put on rowle security and you
57:29have strong passwords and you make sure
57:31that all your services that connect to
57:34it use the credentials in like
57:36environment variables without like
57:38actually putting it into into code,
57:39you're already in a good spot. There's
57:41more that we can do, but this is already
57:43from a database perspective a good
57:45start. Now then when it comes to your uh
57:49back end which is the second most
57:52important thing because for your back
57:53end usually you have data flowing right
57:56so your front end application
57:58um may load data directly from your your
58:01database or it can also go uh through
58:03your back end right so your back end
58:05talks to your database and then your
58:06back end like uh uh puts that uh or like
58:10flows that data through to your front
58:11end. So the back end is another part
58:14where you need to be aware of first of
58:16all like what language do you use and
58:18then second where do you deploy that? So
58:21where does that code live? So when you
58:22start building something it's on your
58:24laptop it's local host. You're running
58:26things you're testing things but when
58:27you deploy something you actually put it
58:29in the cloud and this is where people
58:31can get access to it. So there are of
58:34course various a lot of different
58:36services that you can use. You have
58:38tools like Railway and Render that make
58:41it really easy one-click deployment. Um,
58:44you can rent a virtual private server.
58:47You can use Azure, AWS, GCP. So,
58:50different types of ways that you can set
58:51this up.
58:53Um but the most important thing is that
58:55when you deploy your back end that you
58:59set up ideally some type of firewall
59:02rule around this where you restrict
59:05certain traffic to it and this is again
59:09something that you just need to do some
59:10research in depending on which platform
59:12do you use which language do you also
59:14use uh and then use your AI coding
59:17agents to talk about okay what is like a
59:19firewall how can I set that up now this
59:21also applies applies to your your
59:24database and you can do this in the
59:26superbase uh dashboard. But why this is
59:29important is usually there is a part of
59:32your application that is userfacing.
59:35That's typically your front end meaning
59:37that's the that's what people actually
59:39see. They go into the browser or in a
59:40desktop application and they use that
59:43and then your back end and your database
59:46no one should be able to touch that.
59:49only your user application is you should
59:52be able to pull information from that
59:55and uh your front- end application is
59:58then also deployed somewhere again can
1:00:01be any of the services that you use
1:00:03maybe you use it on for sale uh the
1:00:06thing is that deployment there uses an
1:00:10IP address so that is a specific address
1:00:13from which it reaches out to either your
1:00:15database or to your back end and I know
1:00:17it gets a little bit technical but
1:00:19Firewall pretty much means you set up a
1:00:21rule where you say in your back end and
1:00:23in your database
1:00:25only open the door if the request is
1:00:29coming from the IP address that I
1:00:31whitelisted which is my known front end
1:00:34or desktop application. And now there is
1:00:37very different rules depending on how
1:00:39you set up your like your infrastructure
1:00:41and how you set up your architecture and
1:00:43your applications. But that is the like
1:00:46general uh best practices around
1:00:50security. Setting up the proper
1:00:52firewalls and then making sure secrets,
1:00:55API keys, credentials, all of that are
1:00:58handled correctly. So you cannot read
1:01:00them in your application code. people
1:01:02cannot read them uh by for example in a
1:01:05web application inspecting the code in
1:01:06the browser right so if you protect your
1:01:08secrets and your credentials you have
1:01:12rowle security on and you put proper
1:01:15firewalls on your database your back end
1:01:17and whitelist it for your like front end
1:01:20application that is security in a
1:01:22nutshell and again it's technical but
1:01:24it's more so for people like to take
1:01:26this snippet summarize it put it
1:01:27[clears throat] into like a chat and
1:01:29then reason through this what that means
1:01:31for your application
1:01:32Yeah, I love that. That was a master
1:01:33class. And I think one of the things
1:01:35that
1:01:36>> have become very clear throughout this
1:01:38conversation with different technical
1:01:39concepts that you've brought up is that
1:01:41a lot of this sounds like it's just
1:01:43about awareness. If you can bring up
1:01:45this conversation to your agent and let
1:01:47it go do the thinking and the
1:01:48understanding of
1:01:50>> implementing,
1:01:51>> but you were the one who had to say,
1:01:52"Hey, think about this and do this."
1:01:54Right. I think that's very cool.
1:01:56>> Yeah. It's it's those unknowns unknowns
1:01:59that you in the beginning as a new
1:02:02builder developer need to become aware
1:02:04of so that you can ask the right
1:02:06questions to your AI agents. It's not
1:02:09about the skills or even the
1:02:10intelligence or uh the years of
1:02:13training. It's just the awareness being
1:02:15able to like ask the right questions and
1:02:18then also with the base level
1:02:20understanding that you have being able
1:02:22to somewhat reason over those answers
1:02:24right where you can say yeah that is
1:02:26important and let's not get into this
1:02:28right now because those are not kind of
1:02:30like the basics that goes way deeper
1:02:32that's not important for my application
1:02:33right now at this stage
1:02:35>> I love it let's let's wrap up here with
1:02:38with one final sort of like
1:02:41big statement from you what I want to
1:02:43hear from you is
1:02:46basically I want you to start the
1:02:47sentence like this. If I was starting
1:02:49today with no technical experience, I
1:02:53would be most excited about this. Like
1:02:55what is the opportunity and what gets
1:02:57you excited like thinking about the
1:02:59future of where AI is headed and what
1:03:01you're able to do with it in this, you
1:03:03know, software engineering or building
1:03:05apps sort of space.
1:03:08>> Cool. That's the I really like this. So,
1:03:12um, yeah, here's I would answer. So, if
1:03:15I were just starting out and I see what
1:03:18what's going on around me with all of
1:03:20this technology,
1:03:22what I would be most excited about is
1:03:25the opportunities
1:03:27it can create for each and every one of
1:03:30us. So I out of university
1:03:34quite randomly happened to land on a
1:03:38freelance gig and worked self-employed
1:03:42pretty much ever since then. Started to
1:03:44build businesses off of that and that
1:03:48has given me insane amounts of freedom
1:03:52and getting to work on exciting things
1:03:55that I want to build. And I think that
1:03:58is now within reach for more and more
1:04:02people because of this gigantic
1:04:06transformation that we as the world will
1:04:09be going through in this AI
1:04:11transformation.
1:04:13And you might be thinking, yeah be when
1:04:15the tools get easier, everyone can do
1:04:18this. True, but we still need
1:04:21implementers. We still need builders.
1:04:23The fact that it is possible and that we
1:04:25can do it doesn't mean that everyone and
1:04:29every business will actually take the
1:04:31initiative to start to start doing this,
1:04:33right? So there are just going to be so
1:04:37much opportunities
1:04:40to for like pretty much um
1:04:44entrepreneurial things to build things
1:04:46on your own to to sell that to make that
1:04:49available to make content to help local
1:04:51businesses or maybe not even local but
1:04:53to help larger businesses to be able to
1:04:55be excited about the technology about
1:04:58the automation about the transformation
1:05:00and then getting better at that craft as
1:05:03you do it and then being able to like
1:05:05monetize off of that. And if that is
1:05:09exciting to people watching, which I
1:05:11know it probably is because I know your
1:05:12audience, right? Then I would say like
1:05:14go go all in on that. And if you even if
1:05:18you're not just starting out, but you
1:05:19may already have a job, like figure out
1:05:21how you can do this on the side, right?
1:05:24Um, I for example have an entire
1:05:26community of freelance developers and
1:05:28data professionals and I also share that
1:05:31with people because the opportunities
1:05:33are real. You can do this next to a
1:05:35full-time job. You can pick up a project
1:05:38and while you're doing your day job, you
1:05:41can just kick off another agent and let
1:05:43it do the work, right? You can even use
1:05:44your phone to talk to Claude on lunch
1:05:47break. Give it a master prompt and when
1:05:49you're back home like it has built the
1:05:52automation for your client. So I would
1:05:54say that's the most exciting thing. Um I
1:05:58recently watched a podcast uh it was on
1:06:01Lex Fritman and we had DHH the I don't
1:06:05even like it's David blah blah blah. DHH
1:06:09is is is what he goes by online
1:06:12typically. And he described it as having
1:06:14a genie in a bottle. Like we as
1:06:17developers now have a genie in a bottle.
1:06:18I like that.
1:06:19>> Meaning that you can just think about it
1:06:21>> and you can prompt it into existence.
1:06:24>> And you're not even limited to three
1:06:25wishes. You can just keep asking.
1:06:27>> You're not you're just limited
1:06:28[laughter] to your tokens.
1:06:30>> Yeah. That's that's what you got to be
1:06:31careful
1:06:31>> to your weekly token budget. Yeah. Yeah.
1:06:34But I think that is that that is a great
1:06:37like description and you know this as
1:06:39well Nate because I see all of these
1:06:40videos from you when whenever every time
1:06:43a new model releases like the stuff that
1:06:45they can do just gets crazier and
1:06:47crazier and crazier and it used to be on
1:06:49like this yearly timeline where if you
1:06:52looked one year back like the models
1:06:53were like extremely way way better. I
1:06:57think now we have it already like on a
1:06:59quarterly basis where you have these
1:07:01like really holy this is so much
1:07:04better than what we had right and that
1:07:05will that will continue I I don't see a
1:07:08world where that is going to stop
1:07:10>> so you just need to learn how to like
1:07:14use these tools and then build stuff
1:07:16with it and then use that to ultimately
1:07:21create the life that you want because
1:07:23like building stuff just for the sake of
1:07:25it is fun but I It's a we have a great
1:07:28opportunity to like help yourself, help
1:07:30your family, and to create a life around
1:07:32these tools that was just not possible
1:07:36years ago.
1:07:37>> Mhm. Yeah. I think the the only way we
1:07:39see the the progress slow down is if
1:07:42this whole like pacing the frontier
1:07:44discussion that's going on right now, if
1:07:45there's an actual standard in place for
1:07:47regulation and for
1:07:49>> I guess kind of like a a national and
1:07:51global coordination. So, we'll see what
1:07:53happens there. But
1:07:55>> there go f They go f steam hats. There's
1:07:59no stuff.
1:07:59>> Oh man, we could do we could probably do
1:08:01a whole another hour talking about this
1:08:02stuff. I think it's really interesting
1:08:03as well. But
1:08:05>> um this space has been, you know, so
1:08:07exciting, so fun. I'm so glad that I've
1:08:09been able to connect with people like
1:08:10you and hang out in person and just, you
1:08:12know, you're super smart guy. So I
1:08:13appreciate you coming on. This has been
1:08:14a super informational
1:08:17episode. So really appreciate you coming
1:08:19on. Dave, where can people get in touch
1:08:21with you or learn more from you if
1:08:22they're interested?
1:08:24>> Cool. Appreciate it, Nate. Uh, just go
1:08:26to my YouTube channel, Dave Ablar.
1:08:29Mostly
1:08:31videos for a more technical audience.
1:08:34But that said, what does that matter
1:08:36nowadays? Like I make sure that my
1:08:38tutorials, I also create these kind of
1:08:41like handbook guides for them nowadays
1:08:43because it's so easy. I used to just
1:08:45create tutorials and teach them. Now I
1:08:47create a tutorial and I just say to AI
1:08:49like create a whole entire documentation
1:08:51page for it so people can walk through
1:08:52it step by step. So the cool thing is I
1:08:56mostly do kind of like big builds how to
1:08:58build this. For example, I just launched
1:09:00a video how to build an entire company
1:09:02knowledge base and then actually do it
1:09:04properly like something not just like
1:09:05the toy project but actually something
1:09:07that you could like offer to a company
1:09:10and it's for a technical audience but
1:09:12you could also just use your AI agent
1:09:14and point it at the GitHub repository in
1:09:16the handbook and it can build it for
1:09:17you. So yeah, YouTube Dave Ablar that's
1:09:20where you find me.
1:09:22>> Sweet. I love it. Well, thanks so much
1:09:23for hopping on today Dave and maybe we
1:09:26can do it again sometime.
1:09:28My pleasure, Nate. Talk soon.
1:09:29>> Awesome. All right. See you, Dave.

About this transcript

This page contains the full transcript of How to Actually Build & Sell Software with AI as a Non-Techie by Nate Herk | AI Automation, generated from the public captions YouTube serves with the video. The transcript has 13,311 words across 1,890 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.