YT YouTube2Text

YouTube2Text → Transcripts → 21 Claude-Max-Accounts, ein (menschlicher) Entwickler. | Alexander Weipprecht #41

21 Claude-Max-Accounts, ein (menschlicher) Entwickler. | Alexander Weipprecht #41 — Transcript

by TokenMade · 5,818 words · 954 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00Welcome to Tokenate.
0:02We speak here with people who have built
0:04something with AI in the last few months
0:05.
0:07And today, we’ve arrived at an episode
0:09that has never happened—it's the fastest
0:10episode in the podcast world I've ever
0:12experienced.
0:13I woke up this morning to a message from
0:16a listener who likes the podcast and
0:19said I should speak with Alexander.
0:22There was a link to his Instagram, I
0:26messaged you 30 minutes ago, and Alex
0:30asked if I was spontaneous.
0:34Of course, I’m skipping my lunch for
0:36this podcast episode.
0:38Some know how hard that is for me, but I
0:41think it’s worth it. Alexander, thank
0:43you very much for your time. Who are you
0:46and what have you built with AI in the
0:48last few months?
0:50Okay, first of all, thank you very much
0:52for the invitation. Thanks for making it
0:53happen on such short notice.
0:55Uh yeah, so who am I? I’ve been
0:57self-employed since I was 14.
0:59I founded a software company at 15,
1:02built websites, started a web hosting
1:05company, and sold it in 2017 with nearly
1:0856,000 customers.
1:11I trained as a software developer,
1:13studied economics, and spent 8 years
1:16doing website consulting for large
1:18corporations, reputation management, and
1:21SEO.
1:22But as I said, I’m a software developer
1:25by trade, and for the last two years,
1:27I’ve been working in the AI field
1:29because I operated several websites in
1:32the affiliate space, and now there are
1:34298 websites that I run completely
1:37agentically, supported by AI agents.
1:40That means I have several PCs with AI
1:42agents that run these websites and
1:43market themselves autonomously.
1:47And in the last 12 months, I’ve founded
1:4912 startups that are all performing
1:51relatively well, with over 50,000 users
1:53in the areas of podcasts, online
1:54learning for children, and other
1:56projects.
1:58And as a software company, we work for
2:00about 40 medium-sized businesses, for
2:02whom we program software agentically
2:04with AI alongside my four employees here
2:06on-site.
2:08And exactly, I run a TikTok channel with
2:10now almost 10,000 followers on the topic
2:12of AI programming.
2:14I have a YouTube channel with nearly
2:1611,000 followers that I’ve had for 3
2:18months in the field of AI programming,
2:19and that is what I do all day long. I
2:21program, I program, I develop, and I
2:23talk about what I’m doing. Exactly.
2:27And maybe in closing, I have specialized
2:28over the last few months primarily in
2:30building a software called Codeguard.
2:33The goal was to close the gap between a
2:35senior developer and a "vibecoder" by
2:37making the code secure and the
2:39structures and processes clean.
2:41That is one of my main focuses at the
2:43moment.
2:44Ah, I think there was just some news
2:47from Lovable, that the Handelsblatt
2:51reported that a few things might not be
2:54entirely secure.
2:57Yes, exactly.
2:59So software development as such is an
3:02area that isn't just a question of code
3:05quality, but also of systematic planning
3:08of infrastructure and processes.
3:12Like, how do I publish software content
3:14to a server?
3:15How must the server be configured? What
3:17do the processes look like after the
3:19first time you publish?
3:22So, things like ongoing deployment,
3:24maintenance, and ensuring—uh,
3:26deprecations as they're called in the
3:28field, meaning phasing out features—uh,
3:31security architecture, backup planning,
3:34customer data protection, GDPR; it’s a
3:36huge area.
3:38And I believe it doesn't just end with
3:40choosing a Large Language Model, but
3:42rather, it's a bird's-eye view that you
3:44have to take, as an entrepreneur as well
3:46as a developer, when you're working with
3:48software today.
3:50The EU regulation also says that product
3:52liability has become stricter.
3:54Uh, the GDPR and the EU AI Act clear up
3:57quite a lot of heavy things.
3:59That's why you have to look at the big
4:01picture overall to see what's important
4:03, and I want to help with that. Yes,
4:06>> how would one do that as a Vibecode?
4:09So, assuming I'm building a small app
4:12here, would I just ask the same AI I
4:15used to build the app if it could check
4:18it now?
4:19So do I conduct a DD, a security due
4:22diligence with the AI on my AI product,
4:24or what is the best way to do that?
4:28>> Well, there are actually a few good
4:30standard systems, a security review from
4:32Opus, meaning from Anthropic themselves
4:35.
4:36But there are also very good
4:38repositories from, for example,
4:40Cloudflare, which provide a very, very
4:42good security repository nowadays.
4:45Um, but actually it's not as hard as you
4:48might think.
4:50It's much more a question of whether I
4:51am asking the right questions to the LLM
4:53.
4:54So, if you have a modern LLM, like Opus
4:563.5 or something from Anthropic, or a
4:58GPT-4o, then it’s more a question of
5:01whether you're approaching the problem
5:03from the right vector?
5:05So, are you asking the right questions
5:07about your software?
5:08And I believe that anyone who does this
5:11right can bring very, very good
5:13deployable software online today that
5:15also meets very high security standards
5:17. Only, as I said, most people lack the
5:19question of what they should ask, and
5:21that is where the problem somewhat lies
5:23.
5:25>> All right, and uh, why don't you show us
5:28what you're working on?
5:31I, uh, would like to understand and see
5:33what you have set up for yourself.
5:37With pleasure. So, I can, if you like,
5:39briefly provide some insight into how I
5:41build my setup and why I have it
5:43structured that way. Um, in addition, I
5:46should mention that a lot of my work
5:47runs through the Code Guard. I'll also
5:49talk a bit about how it works and what
5:51it does, and then you can get a little
5:52insight into it yourselves.
5:54Um, okay, I'll just share my screen. Uh
5:57, I must say, in fairness, I'm working
6:00locally here with Hurder.
6:02Hurder is a kind of multi-terminal
6:04management system.
6:06I have that running right now, and there
6:08are 16 client projects running in
6:10parallel as we speak. I have a
6:11six-monitor setup here; people know that
6:13from my YouTube channel.
6:16Um, that's why I've opened an emergency
6:18session in Visual Studio Code here,
6:20where I can also show how I do my setup
6:22.
6:24Um, just so you're not confused, as I
6:25said, I work with Hurder.
6:27It's very, very cool. Um, exactly.
6:30I, I hope you can see my, um, terminal
6:33setup here now; I've put together a
6:35terminal, a setup, and I'll briefly show
6:38you how I do my basic setup and what my
6:40strategy is when I'm planning software
6:43development.
6:44And I think, because I was asked to show
6:46something practical. I hope that's okay
6:48. Um, exactly. So, first of all,
6:51regarding the setup, how do I build it?
6:53I have two terminals up here and I'll
6:56show you briefly how I do it. I have two
6:59terminals up here. Um, one moment,
7:02Terminal, New Terminal. Uh, bam, I'll
7:04pull one terminal to the top.
7:07So, and this terminal up here, I'll
7:09split into two halves.
7:11And the terminals down here, I'll split
7:13into four individual sessions.
7:16And I'll start Cloud Code briefly for
7:18each session.
7:20Um, just so you can see how I proceed
7:22there.
7:23Um, so.
7:28So,
7:32exactly. Now you can sort of see briefly
7:34what something like that looks like.
7:36I'll briefly go into, uh, bypass
7:38permission mode.
7:40I work with it because my Code Guard, my
7:43system that I have, enables me to work
7:46securely because I have so-called guards
7:49in place.
7:51I have developed a parser in Codeguard
7:53that parses bash commands—specifically,
7:57any destructive commands related to SQL
8:00or the operating system, such as bash
8:03shell commands.
8:05This means that every command the LLM
8:07tries to send out is run through a
8:09parser and is blocked if it is
8:11destructive, such as delete commands,
8:14destructive commands that recursively
8:16delete data, break databases, and so on
8:18.
8:20Um, exactly.
8:21Um, now I am here in my base session,
8:24and before I start building software,
8:27all of my sessions are assigned a very
8:30clear defined role.
8:32I will start by renaming the
8:34Orchestrator, which is my, is my, is my
8:38orchestrator, my main session.
8:42Um, my second session here on the right
8:44, and I will explain that in a moment,
8:46is the Task Manager
8:49and then come the individual workers:
8:52Worker 1,
8:55Worker 2,
8:59Worker 3. The setup is done relatively
9:01quickly; you just have to, um, Worker 4
9:04.
9:08And what I have now is a division: one
9:11Session Orchestrator, one Task Session
9:14Orchestrator—which is special, I’ll get
9:17to that in a moment—and four worker
9:21sessions that work in their own Git work
9:24trees. This means I can always work with
9:26four workers simultaneously.
9:28And the cool thing is, this is where
9:30Codeguard comes into play, so to speak.
9:32I will first introduce my terminals to
9:36their tasks.
9:39To the Orchestrator, I say, you act as
9:41Orchestrator.
9:46Orchestrator,
9:50act
9:52as Task Orchestrator.
9:55These are fixed roles that my Codeguard
9:57has.
9:58You can also see that the skill "
9:59Session Orchestrator" is being loaded.
10:03Here, the skill "Session Orchestrator"
10:05is loaded with the special instruction "
10:07Task Orchestrator."
10:09That is, Senior Dev now starts the Task
10:13Orchestrator, and now I say, introduce
10:16yourself to the workers,
10:20introduce yourself to the worker
10:22sessions, and I tell the Orchestrator
10:24that, too. Introduce yourself to all
10:26sessions.
10:29So, what I am doing now is having them
10:31introduce themselves to each other and,
10:34so to speak, write into their memory so
10:36that they know who has which task.
10:40And what the Orchestrator does first is
10:43create a Git repository and assign the
10:45instructions to the individual workers
10:48here to work with so-called Git
10:50worktrees.
10:53I hope this isn't too technical for your
10:55channel's target audience, but I assume
10:56since you talk about AI programming,
10:58this is all still trivial, so to speak.
11:02Um, while that's happening, I'll explain
11:05an important fundamental strategy of
11:08this approach: the orchestrator I'm
11:11currently in—its tokens and context
11:14should remain as small as possible, and
11:17in the best case, it's never reset, only
11:20compacted.
11:22This means the orchestrator always keeps
11:24an overview of everything that has run.
11:27I should perhaps mention that I
11:29specialize in software applications that
11:31run for weeks. So, the tasks I build
11:34here run for weeks. I have one that's
11:36been running for over a month, using
11:37this exact setup. This means the
11:39orchestrator isn't reset for a whole
11:41month. That means the context must be
11:43kept as small as possible.
11:46Um, and that's why it's important that
11:48the task orchestrator is the one I talk
11:50to, and it distributes the tasks to the
11:52corresponding—um, to the corresponding
11:54sessions, and you can see it down here,
11:56it's currently logging into them.
11:59You can also see now that the messages
12:01are coming in. That means the
12:02orchestrator is now saying "git init" is
12:04recommended. Bam.
12:06Yes, that means it just said directly, "
12:08As the orchestrator, I now have to take
12:10care of the versioning of the Git
12:11repository."
12:13And the role of the task orchestrator is
12:15super exciting because it does the
12:17following: it doesn't just look at, "
12:20Hey, what to-dos do I have open?"
12:23There is a separate to-do management in
12:25my codeguard, but it looks at the Git
12:28trees and the commits of the individual
12:30worker sessions.
12:32Which task is already finished? How much
12:35have we already accomplished there? Are
12:37there merge conflicts? Are they in
12:39collision with each other?
12:41So, this means the task orchestrator
12:43ensures that the Git structure we are in
12:46is completely clean.
12:48Yes, and it is extremely important that
12:51one session always takes care of this
12:53matter.
12:54And as I said, this is such a process;
12:56you have to imagine, I usually have not
12:59just four worker sessions, but 15 to 20
13:01worker sessions.
13:02This means I am recreating software; for
13:05example, we have clients. For them, we
13:07are rebuilding an ERP system.
13:09It was developed over 10 years by seven
13:12developers, so 2.5 million lines of code
13:15, and we are reverse engineering and
13:18completely rebuilding it on a platform
13:21in 3 months.
13:24This means, um, we are first building a
13:27tool that can handle the entire
13:29mirroring.
13:31That means we look at database
13:33structures, infra, business logic, build
13:36entire mappings via a graph, so to speak
13:39, clone the logic into an actual format
13:42, and then completely rebuild it on a
13:45Laravel framework, for example, as a web
13:47application with a full security concept
13:50and an MCP connection.
13:53This means the entire ERP solution will
13:55be controlled via MCP in the future. Yes
13:57, and for that to work, it needs open
14:00endpoints with token authentication,
14:02rights management systems, and so on.
14:05And to manage such large projects on the
14:07fly in the enterprise sector, you have
14:10to imagine, if I only had to make a
14:12decision myself for every fourth problem
14:14in this development, it would take me a
14:17year just to make decisions for software
14:19like this.
14:21So, I don't decide anything anymore. My
14:23entire system has an autopilot.
14:27It has an LLM Council behind it, which
14:29you might know.
14:31I don't know if you are familiar with
14:33LLM Councils, but my LLM Council has a
14:35special user, called the Pareto Engineer
14:38.
14:40That means he has to find pragmatic
14:42solutions that don't necessarily have to
14:44meet pure efficiency, or the perfect
14:47solution, but rather address the
14:49question of how we can reach a result
14:51quickly.
14:53Because one of the core problems of
14:55large AI engineering is infinite looping
14:58, which you can't get out of because
15:00tasks keep increasing and the LLMs keep
15:03finding more errors. Um, exactly. Then
15:06perhaps it's important that within this
15:08system, I work with static code analysis
15:10.
15:10That means I use PHPStan, linting
15:12methods, and I use large frameworks that
15:14already have a clean security concept,
15:17like Laravel.
15:19For example, this means I don't work
15:21with vanilla CSS, JavaScript, or legacy
15:24code, but rather use large existing
15:27frameworks to have a very, very clean
15:30structure here.
15:32And now comes perhaps one of the most
15:34exciting points, which I can
15:35unfortunately only show halfway because
15:36there are too many client projects open
15:38in parallel here.
15:40Specifically, I have developed a
15:42so-called acceptance criteria
15:44contracting procedure.
15:46I’ll explain that briefly.
15:48There is a skill I have called Goal
15:50Engine, and Goal Engine describes the
15:52target definition of my to-dos.
15:56So, if I were to say now, "build me a
15:59tool"—I can show you— "build me a tool
16:02where I can manage my tasks."
16:08So, let's plan that out.
16:11Then Goal Engine steps in and does the
16:14following.
16:16It interviews me now in a classic
16:18brainstorming session using superpowers
16:21and the Grillmi skill, so I have the
16:23Grillmi skill integrated here.
16:26That means it first asks me for all the
16:28important aspects necessary to build
16:30something like a task manager.
16:34But next, it is important to say at this
16:37point, once it has asked these tasks, uh
16:40, once it has asked these questions,
16:42then the question is, how do I formulate
16:45my tasks?
16:47Therein lies a big difference in
16:49software engineering today, because here
16:50, now it starts: what is it about?
16:53Who is working with the tool? This
16:54determines whether it needs login,
16:56assignment, or permissions. Now I will
16:58say just me alone. Then what it is about
17:01, what should it do better later than
17:02today? It should stop forgetting things
17:04, know what is next on the agenda. I
17:06want to track deadlines and see progress
17:08. Bam.
17:10Um, what it is about, you need to access
17:12your task, let's say via a browser. Yes
17:14, bam.
17:16What it does now is it takes these
17:18answers and builds acceptance criteria
17:21out of them. I'll give you an example.
17:24If you program a login somehow and you
17:26just say build a login, then an
17:28acceptance criterion would be to say a
17:30login has a username, a password, but
17:32isn't "forgot password" perhaps missing
17:35there?
17:36What about, uh, saving password login?
17:39What about, uh, five failed login
17:41attempts? How often are you allowed to
17:43do that anyway?
17:44That means, for a login screen alone,
17:47there are certainly 50 acceptance
17:49criteria that must be cleanly defined to
17:51say this is a good login and this is a
17:53bad login.
17:56And for this, I have my own engine that
17:58tries to find the most important and
18:00best acceptance criteria as best
18:03practice with an online search for a
18:05single task and then creates a catalog
18:07of acceptance criteria from it, and
18:09those are in turn defined into
18:11milestones and sub-acceptance criteria.
18:16And the concept behind it is that you
18:19have a measurable end result.
18:23And what comes out of it is a SlashGPT
18:25prompt.
18:27That means it is now creating a contract
18:29with precise acceptance criteria,
18:30milestones, and goals.
18:33And at the end of the day here, now it
18:35makes the goals, goal 1, goal 2, goal 3
18:37, and at the end I get a goal statement
18:39with a clear SlashGPT command, and when
18:41that is started, the software develops
18:43completely autonomously.
18:47All decisions that need to be made are
18:49made via the LLM Council, via a reviewer
18:51.
18:52That means the LLM Councils are allowed
18:55to go online and search for a best
18:56practice solution for a task and also
18:58document it.
19:00When you have to make technological, uh
19:02, technological decisions, you can have
19:05a Senior Developer Skill, which already
19:07has a ton of data, go online and develop
19:09the best solutions based on its
19:11experience and pass them to the LM
19:13Council as information so that the
19:15Council has a better basis for
19:17decision-making.
19:19And, uh, you might think this is all
19:22totally spaced out and exaggerated, but
19:24I have 14 different software tools here
19:27, three Linux computers, and 300
19:29websites, and they are optimized every
19:31day by automatically cross-referencing
19:34Matomo statistics, Google Ads campaigns
19:36, and visitor statistics.
19:39This means every website has about 100
19:41measuring points, and every day an agent
19:43goes in and looks: Did the users–how did
19:45they behave there? Did they click on
19:47affiliate links? Did it generate revenue
19:49?
19:50Did the Google Ads campaign work? What
19:52needs to be improved in that regard?
19:54That’s how I manage nearly 300 projects
19:57completely autonomously with several AI
19:59agents that run here all day long, and
20:01this only works in an absolutely perfect
20:03autonomous setup that only informs me
20:06about the most important things.
20:08Via email, where I only intervene when I
20:10really have to do something, and I
20:12basically just watch, uh, and then I
20:14have to check: Is it legally secure?
20:16There is a separate legal skill that
20:18must check every text it writes to see
20:20if the text is legally correct, if I am
20:21violating third-party rights, or if I
20:23might run into data protection issues
20:25with it.
20:26So, in a sense, it must build everything
20:29to be legally compliant with the EU AI
20:31Act.
20:32If I incorporate an image that is
20:33AI-generated, it must have the AI label
20:35.
20:37That means my system not only builds
20:39software autonomously and builds
20:41websites, but also develops itself
20:44independently via a so-called recursive
20:46self-improvement process.
20:49This means that after every process, a
20:51separate procedure checks whether the
20:53process could have been improved and
20:54thus independently improves itself.
20:57And it does this for 300 processes
20:59multiple times every day.
21:02This means we have a constantly evolving
21:04process, and now you see, now it has
21:06defined acceptance criteria, goals,
21:08acceptance criteria for the app, so a
21:10complete structure.
21:12And now I can say, I am approving this
21:14acceptance criteria catalog. This is
21:16version 001.
21:18Uh, that means there is actually a
21:19graphical interface for me now. I can
21:21now see all contracts there and all
21:23processes, they are basically included.
21:25I just can't show that right now because
21:26there are too many other customer
21:27projects in there.
21:29Um, and now you'll see what comes out,
21:32which is a complete goal prompt that
21:34allows me to have this contract
21:36processed completely autonomously until
21:39it's finished.
21:42Here comes the slashgoal right now.
21:46The contract is under personal task tool
21:49contract V001 Jason with all acceptance
21:52criteria.
21:54I hope you can read it well. You just
21:56have to let me know. Yes.
21:59And um and then it runs exactly now it
22:01is finished. Now here comes my slashgoal
22:03.
22:04That is essentially the contract. Now I
22:06say slashgoal. Boom.
22:10And now I have entered the contract and
22:12now this contract is being completely
22:15implemented via a slashgoal instruction
22:17and it's only finished when the
22:19application is completely done, working
22:21, end-to-end tested, programmed with
22:24unit tests, and fully checked by the
22:26security architecture via my skills.
22:29That means the Code Guard skills are now
22:32running over it.
22:34Um, the security architecture is being
22:36checked, the end-to-end tests are being
22:38checked, a static code analysis is being
22:40run over it, and the database is being
22:42checked for clean indexes and clean
22:44normalization of the fourth normal form
22:45, as you would do in a proper software
22:47architecture.
22:49Um, exactly. And that means the thing is
22:51now building the app completely like a
22:53senior developer.
22:55So that was my task to build an app and
22:56it's now running almost completely
22:58autonomously and now we are at 23
22:59minutes.
23:00Maybe you still have a few questions.
23:02Exactly. I'll switch back again.
23:04>> Yes, incredible. Would you like to give
23:07a little overview?
23:08So, you said you have projects running
23:11there.
23:12Uh, can you categorize a bit what that
23:14brings in terms of revenue on one side
23:17and what it means in terms of costs on
23:19the other?
23:20>> Well, I have learncastaai.com for
23:22example. That's about 6,000 users now.
23:25Um, I have wemolino.com, I have Bankreif
23:29as a project, I have um I can gladly
23:32show a few projects, but I have Market
23:36Update AI, Ranchen AI, that's a new
23:39company, it's about AI visibility, we
23:42have findexpert.com.
23:45We have My Startup 24, we have State by
23:47Steep a fitness app, we have Website
23:50aient.com, that is a chatbot for
23:52e-commerce websites that is agentic.
23:55That means it's a bot, it's not just a
23:58simple RAG agent chatbot, but an agentic
24:00loop system.
24:02That means it's like a cloud coder only
24:03for online shops, which corrects its own
24:05results in the chatbot itself and helps
24:07the website operator to improve their
24:09website.
24:10Um and um as I said, regarding the total
24:13volume, we are at about
24:19AI costs between 11 and 15,000 dollars a
24:21month. I have 21 Cloud Max accounts
24:24myself, um, that we operate, um, and
24:27these must, of course, be used in a
24:30GDPR-compliant way. That means we
24:32clearly separate customer projects from
24:35staging.
24:36Um, so we develop on staging and then
24:38deploy live with the other plans, with
24:40the GDPR-compliant team plans.
24:43Um, and we currently make about 1.1
24:46million in revenue per year. That is
24:48roughly our, our thing here.
24:52So, we do have a good margin in the
24:53range of 45 to 60 percent, depending on
24:55how much.
24:57I just bought a 9, 8,000 euro Mac Studio
24:59, last week 11,000 euros in new hardware
25:02for my employees. Um, and sometimes we
25:04also have to buy additional tokens for 3
25:06, 4,000 dollars. So, depending on that,
25:08um, the revenue is sometimes more,
25:10sometimes less. That's about our scale,
25:12and we've been operating at that scale
25:14for a year and a half now. Exactly.
25:16>> And where is the journey headed, also?
25:20In terms of SaaS. There are always these
25:23waves where people say SaaS is dead,
25:25long live SaaS.
25:27How, how do you see it? Especially when
25:29you say, hey, we just rebuilt an ERP
25:31here and everything.
25:34>> Is the future heading there, or is it
25:37more like everyone is building their own
25:39little solution, but none of these AI
25:41products really get traction?
25:44>> I don't think the solution will be that
25:46everyone builds their own little
25:47solution. Um, as, as the company itself
25:50. I believe there will be specialized
25:52firms like us that build such software
25:55solutions, because the, um, requirements
25:57—I notice this with our customers—they
25:59are growing exponentially with the
26:01possibilities. That means, um, if we
26:03develop software today—I mean, I
26:05developed software before, but if I, if
26:07I get an order for software today, for
26:09which I would have said earlier that we
26:12need a year, that is now expected to be
26:14delivered in about two and a half to
26:16three weeks. That means the speed of
26:19what needs to be delivered, in what time
26:21, is changing, and the question is who
26:23can get this stuff done efficiently,
26:26cost-effectively, and at a high level?
26:29Um, at the same time, the software tools
26:30themselves are no longer that simple. So
26:33, people suddenly want to have authentic
26:35AI everywhere. That means they want all
26:37of their processes to be automated. This
26:39means we are not just simply rebuilding
26:41an ERP, but it has to be completely
26:43authentic afterward. It has to be fast,
26:45it has to be connectable, uh, and for
26:46that, you definitely need architecture
26:48experience.
26:50You need to know where the pitfalls are
26:52, keyword: fallback LMs.
26:54I’m not going to build an entire
26:55application on one cloud just so that if
26:57the cloud has downtime tomorrow, the
26:59whole company comes to a standstill.
27:00That means we need multi-large language
27:03models, fallback solutions, and we need
27:05monitoring. For example, if complex
27:07pipelines are running for invoicing or
27:09whatever, and they run via LMs, what
27:11happens if they have downtime or fail,
27:14yeah, uh, when do you detect that in
27:16time?
27:17Or we’ve often had the issue where large
27:19language models sometimes have error
27:21rates, where they suddenly just start
27:23spitting out garbage.
27:26Yeah, what happens if I suddenly get
27:27errors during a highly sensitive process
27:29?
27:31That means I have to monitor their
27:32status page and, if in doubt,
27:34automatically shut down the client’s
27:36software to prevent errors and so on.
27:40Um, then GDPR, we work a lot with
27:42Mistral, but also with Microsoft Azure
27:45or Ionos as an AI Cloud Hub, which are
27:47GDPR-compliant.
27:50That means multiple APIs, you have to
27:52put a credit card everywhere, you have
27:53to manage all of that, and many
27:55companies don't want to deal with that
27:57kind of thing, which I can understand.
28:01Yeah.
28:01>> And how do you view the European AI
28:04players? Like Ionos, Mistral, and things
28:08like that?
28:09Are there still chances?
28:11Do we have any possibility of keeping up
28:14with the American models using
28:15open-weight models?
28:18Yes, because I believe the marginal
28:20utility of LMs decreases after a certain
28:23point.
28:24I mean, to exaggerate: if we had
28:26software in Europe at the level of Opus
28:293.5, at least on the day it came out, I
28:32mean LMs, then you could really build 90
28:35%of the software you need to build with
28:38that.
28:39Yes, so um, the question is what an Opus
28:426.6 could actually do much better, um,
28:44besides finding security vulnerabilities
28:46that no one really wanted to find anyway
28:49.
28:50Yes. Um, that means the question will
28:52ultimately boil down to the end result:
28:55how much smarter does it really need to
28:57get for you to build something a human
28:59can actually use? Because I can already
29:02see that I’m—I’m building.
29:03For some clients, for instance—the
29:05client works with their own cloud, and I
29:07just enable an agent on my end that the
29:09client talks to.
29:11So we have software projects that the
29:13customer builds themselves by giving
29:15instructions to Sclaud, our Cloud, and.
29:18And so, I don't manage these projects as
29:20a human anymore; instead, the customer
29:21just provides their requirements and
29:23discusses what they need with their
29:24Cloud.
29:25It connects via a server-wide MCP to my
29:28Cloud, and they then discuss the best
29:30solution with each other and build it
29:32without me having to approve it.
29:35That means there's a framework where my
29:37agents build software for customers that
29:39the customers trust, but the advantage
29:41is that the customer doesn't have to
29:43build this setup, they don't need to
29:44have this guardrail, this backup
29:46infrastructure, the staging environment
29:48, and so on.
29:49They don't have to conceive any of that
29:51; I can do all of that for them. So,
29:52they get a finished software solution
29:54but can give me instructions through
29:55their Cloud on what they want. Yes, and
29:57I can intervene in case of doubt. Yes,
30:02>> That's insane.
30:04Uh, in closing, everyone is talking
30:06about the Anthropic IPO right now.
30:09You are with, what did you say, 21
30:13Claude Max accounts, probably one of the
30:15power users.
30:17Are you looking forward to the IPO? Will
30:19you be subscribing? Are you perhaps
30:21already stocked up on secondaries? Um,
30:23or do you think the future will look
30:26different?
30:28>> Uh, I have, uh, I have been watching
30:30these numbers at OpenAI very closely.
30:33which are now down by, like, half a
30:35billion dollars in valuation.
30:38Uh, I fear that these AI companies are
30:42extremely overvalued.
30:44Right now, not necessarily in the long
30:47run. I think in the long run it will
30:49settle down.
30:50There's this study, I once heard, that
30:530.3%of people use AI professionally.
30:56I believe that the demand for it will
30:58increase, but I also believe there is
31:00currently a shortage, specifically of
31:01computing power. They are building data
31:03centers like crazy right now, and I
31:05believe that at some point, there will
31:06be a tipping point. I have to imagine,
31:08what happens if they set the Max
31:10accounts to $ 1,000 tomorrow?
31:12Yes, then there will be a few crazy
31:14people like me who will probably keep
31:15their 21 accounts anyway and continue to
31:17use it.
31:18But there are probably very, very, very
31:20many who will no longer do so and will
31:22switch to local models instead, and I
31:24believe that is the biggest danger for
31:25such big players like Anthropic or
31:27OpenAI.
31:28Are these really these local models from
31:31Nvidia, Spark, the new um systems, they
31:33can already deliver extremely good
31:36performance, and I just believe in the
31:38long run um there will be some kind of
31:41mix and there will be a normal, normal
31:43valuation.
31:44I'd say maybe based on what is valued
31:47today, but whether it can necessarily
31:49keep going up like that. I'm a bit out
31:52of the speculative bubble.
31:54I um I'm sticking a bit to Warren
31:55Buffett at the moment, who is holding a
31:57lot of liquidity.
31:59Uh I unfortunately have the fear that we
32:01are still facing a big bubble and if
32:03things escalate worldwide due to
32:05political and military conflicts, I
32:07don't want to be invested in the stock
32:09market at the moment. That's too hot to
32:11handle for me right now, but that's
32:14another story.
32:15>> Yes. And um you're also a content
32:18creator, so you said TikTok um and as I
32:21said, I was sent the Instagram account.
32:24Uh have you already fully automated that
32:26too? So does Claude edit your your
32:29shorts, your your videos? Um are there
32:32any hot tips for that?
32:34>> I actually have two videos on YouTube,
32:36no, three, that are completely
32:38AI-generated for testing. My employee
32:40made them. Uh I'm completely an avatar
32:42in them. They are really hard to
32:44distinguish from real videos now. Uh
32:46they are also labeled logically. But my
32:48TikToks and Reels I still do completely
32:50myself with the phone in my hand. Uh I
32:53basically film my screen. Nothing AI
32:55related.
32:57Um because I'm someone who says it's raw
32:59material.
33:00I just talk about everyday life.
33:02I have no desire to put that into video
33:04editing. I just chat away and my viewers
33:06quite like that. I sit in the car in the
33:09evening on my way home, then I record it
33:10and upload it directly from the car. So
33:12, when I turn the car keys, I quickly
33:15upload the TikTok.
33:17Um, for me it's just about taking people
33:18along on my life's journey and what I do
33:20and getting them excited about
33:22technology.
33:24Um yes, all this, so we naturally use AI
33:27to analyze the comments under my videos
33:29and I get an email once a day containing
33:31all the important information about the
33:34comments.
33:36Um um and through that I get inspiration
33:38for new videos. That runs fully
33:40automatically. So we evaluate it. Sure.
33:42>> Exactly.
33:43>> Do you have three final hot tips that
33:46everyone should do with AI?
33:52Look into the processes you do every day
33:55that don't involve people directly but
33:58require high precision, and test whether
34:01you can automate them.
34:04Um, anything involving people, I would
34:06be careful with—or rather, not do at all
34:08. So I would continue to handle phone
34:10and email correspondence personally.
34:12Um, also a note: don't forget that with
34:14all processes running automatically, you
34:16still need to know what's happening in
34:18your own shop at the end of the day.
34:21Um, and, very important: maybe don't let
34:24these negative people get to you.
34:27Be careful about publishing stuff live
34:29and launching apps when you don't know
34:31what you're doing, but at the same time
34:33, don't let all those senior developers
34:36ruin your fun by running around saying
34:38that private individuals who aren't
34:40developers can't build anything.
34:43Don't let them spoil your fun, because
34:45there’s nothing wrong with everyone
34:47building their own software and enjoying
34:49it.
34:52The truth is, as I said yesterday in a
34:54longer YouTube TikTok, I think I
34:56recognize after a certain time,
34:58especially noticing it myself, that the
35:01joy of building eventually fades because
35:03you actually have to use it.
35:06I mean, there are many who build ten
35:08tools and use only one or none of them.
35:10And then at some point I ask myself,
35:12should I build an eleventh, and is it
35:14worth the $ 20 a month cloud
35:16subscription, or the Netcup server, or
35:18the Hetzner server for € 30 a month?
35:21Does it make sense for me to pay for it
35:23every month if I don't really have
35:25anything on it that is running or
35:27working?
35:28So, I think the euphoria, the "tinkerer
35:30" euphoria, will eventually die down a
35:32bit and people will go back to buying
35:34their SaaS software subscriptions
35:36because they'll say: "I can afford the €
35:3819.90, I can just buy Alex's cool
35:40software and don't have to build
35:41everything myself." But who knows, I'm
35:44not someone who has a crystal ball; who
35:45knows, maybe it will turn out completely
35:47differently. Yeah,
35:48>> Alexander, thank you very much for
35:50sharing. It was a great episode.
35:52The next one will be back on Tuesday or
35:54Thursday. See you soon. Peace.

About this transcript

This page contains the full transcript of 21 Claude-Max-Accounts, ein (menschlicher) Entwickler. | Alexander Weipprecht #41 by TokenMade, generated from the public captions YouTube serves with the video. The transcript has 5,818 words across 954 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.