Markus Oberlehner
@markus.oberlehner.net
Working on modernizing Austria's digital tax services. Vue / Nuxt, React / Next.js expert, TDD practitioner, Node.js enthusiast, ML padawan.
No matter what I do with AI, I end up with the realization that the "I" in AI is an exaggeration. The "Intelligence" in LLMs is nothing like human intelligence.
November 18, 2025 at 10:58 AM
No matter what I do with AI, I end up with the realization that the "I" in AI is an exaggeration. The "Intelligence" in LLMs is nothing like human intelligence.
Yet I've had great success using agents with prompts like, "Here is how I've implemented X, please implement Y in the same way," or for refactoring tasks that can't be 100% automated by the IDE.
November 18, 2025 at 10:58 AM
Yet I've had great success using agents with prompts like, "Here is how I've implemented X, please implement Y in the same way," or for refactoring tasks that can't be 100% automated by the IDE.
Ultimately, every single time I've tried to let an AI build something I'm planning to maintain for a long time, I've ended up throwing the LLM-generated code away and writing it myself, much simpler.
November 18, 2025 at 10:58 AM
Ultimately, every single time I've tried to let an AI build something I'm planning to maintain for a long time, I've ended up throwing the LLM-generated code away and writing it myself, much simpler.
But what I find is that AI-generated code is anything but simple. This not only makes the code very hard for a human to debug, but the AI itself will also struggle to iterate on the code it has written.
November 18, 2025 at 10:58 AM
But what I find is that AI-generated code is anything but simple. This not only makes the code very hard for a human to debug, but the AI itself will also struggle to iterate on the code it has written.
I mean, coming from Google, it reads like a threat anyway.
November 18, 2025 at 7:00 AM
I mean, coming from Google, it reads like a threat anyway.
Yeah but only if you’re very strict about scope of feature branches.
I’m a big fan of leaving the codebase a better place so I like to refactor code as I go and always having to create a separate branch for that is no fun..
I’m a big fan of leaving the codebase a better place so I like to refactor code as I go and always having to create a separate branch for that is no fun..
November 13, 2025 at 9:54 AM
Yeah but only if you’re very strict about scope of feature branches.
I’m a big fan of leaving the codebase a better place so I like to refactor code as I go and always having to create a separate branch for that is no fun..
I’m a big fan of leaving the codebase a better place so I like to refactor code as I go and always having to create a separate branch for that is no fun..
Ideally, Git would automatically rebase quicksaves into end-of-level commits.
November 13, 2025 at 9:36 AM
Ideally, Git would automatically rebase quicksaves into end-of-level commits.
The original post said, "What does OpenAI offer the world?" implying that it is not.
November 10, 2025 at 2:42 PM
The original post said, "What does OpenAI offer the world?" implying that it is not.
I want to ban all cars from cities, and I don't drive either. 95% of the time. Yesterday, a mother in Austria rolled over her own infant. Tragic. Still, in 5% of the time, I acknowledge that driving is useful. And others even more so. This tragic incident and hundreds more daily do not change that.
November 10, 2025 at 2:39 PM
I want to ban all cars from cities, and I don't drive either. 95% of the time. Yesterday, a mother in Austria rolled over her own infant. Tragic. Still, in 5% of the time, I acknowledge that driving is useful. And others even more so. This tragic incident and hundreds more daily do not change that.
But look, I don't even want to defend AI; I want to critique it, too. But in a way that makes the average Joe listen. Stating that it's useless when their experience is clearly the opposite will get us nowhere.
November 10, 2025 at 2:33 PM
But look, I don't even want to defend AI; I want to critique it, too. But in a way that makes the average Joe listen. Stating that it's useless when their experience is clearly the opposite will get us nowhere.
Okay, so you don't Google as well? Your only source of information is primary, peer-reviewed sources from libraries and trusted journals?
November 10, 2025 at 2:33 PM
Okay, so you don't Google as well? Your only source of information is primary, peer-reviewed sources from libraries and trusted journals?
Okay, do you drive a car or take an Uber sometimes?
November 10, 2025 at 2:29 PM
Okay, do you drive a car or take an Uber sometimes?
The same is true for Google results and StackOverflow answers. And oftentimes the information is self-validating: e.g., when it solves a programming problem for me, it either works or it doesn't.
November 9, 2025 at 7:19 PM
The same is true for Google results and StackOverflow answers. And oftentimes the information is self-validating: e.g., when it solves a programming problem for me, it either works or it doesn't.
Especially when it comes to learning the basics, LLMs are correct in the vast majority of cases. And just an hour ago an LLM helped me to solve a pretty complex programming problem, too.
But yes, hallucinations are a problem! Yet that doesn’t mean they can’t be useful.
But yes, hallucinations are a problem! Yet that doesn’t mean they can’t be useful.
November 9, 2025 at 5:34 PM
Especially when it comes to learning the basics, LLMs are correct in the vast majority of cases. And just an hour ago an LLM helped me to solve a pretty complex programming problem, too.
But yes, hallucinations are a problem! Yet that doesn’t mean they can’t be useful.
But yes, hallucinations are a problem! Yet that doesn’t mean they can’t be useful.
How often did that happen?
November 9, 2025 at 5:13 PM
How often did that happen?
Okay. Do you drive or use an Uber sometimes?
November 9, 2025 at 12:18 PM
Okay. Do you drive or use an Uber sometimes?
How many people are using ChatGPT and how many suicides can be connected to its usage?
November 9, 2025 at 11:58 AM
How many people are using ChatGPT and how many suicides can be connected to its usage?
I think this is a vast oversimplification. I’ve seen people using it learning new skills much more effectively than they could with Google alone.
There is a lot to criticise, yes. But how do you expect people taking your arguments seriously when their experience is that it _is_ useful to them?
There is a lot to criticise, yes. But how do you expect people taking your arguments seriously when their experience is that it _is_ useful to them?
November 9, 2025 at 8:06 AM
I think this is a vast oversimplification. I’ve seen people using it learning new skills much more effectively than they could with Google alone.
There is a lot to criticise, yes. But how do you expect people taking your arguments seriously when their experience is that it _is_ useful to them?
There is a lot to criticise, yes. But how do you expect people taking your arguments seriously when their experience is that it _is_ useful to them?
In Österreich: Duzen im alltäglichen beruflichen Kontext. Siezen beim Knüpfen neuer beruflicher Kontakte oder im privaten bei deutlich älteren oder Respektspersonen (Amt, Krankenhaus, Justiz usw.).
Nie Du (veraltet). Email / Snailmail immer Sie wenn man sich nicht kennt.
Nie Du (veraltet). Email / Snailmail immer Sie wenn man sich nicht kennt.
November 1, 2025 at 4:08 PM
In Österreich: Duzen im alltäglichen beruflichen Kontext. Siezen beim Knüpfen neuer beruflicher Kontakte oder im privaten bei deutlich älteren oder Respektspersonen (Amt, Krankenhaus, Justiz usw.).
Nie Du (veraltet). Email / Snailmail immer Sie wenn man sich nicht kennt.
Nie Du (veraltet). Email / Snailmail immer Sie wenn man sich nicht kennt.