Really thoughtful piece — especially the framing around why vibe coding still breaks down for most users.
One angle I’ve been running into a lot in practice is that some of this frustration isn’t just UX or user skill, but the way platforms operate under the hood. Many route or switch models mid-session for cost or load reasons, which can materially change agent behavior halfway through a build.
From the user’s perspective it feels like you suddenly have a different model — same codebase, same prompts, but degraded reliability.
Curious how you think about this from a business-practice / economic perspective, not just tooling.
I have a masters in computer engineering. Started coding in the 70s.
I now run a retail bakery and do all the coding for the web site and our back end - i.e. payroll and accounting.
Here's the deal on using AI for coding from someone who has been coding since the start.
In the end, the basics have not really changed with AI. Its just the next level of abstraction that has been going on since I started coding. Coding started as assembler, went to basic, then C/C++, scripting like Awk/Python, web created HTML, PHP, Javascript, etc, DBs have SQL... You get the point. All those made it more efficient for me as the "programmer" to create an "app".
AI is a layer on top of all these. It can generate any of these. But here's the catch, you need to be very, very, very clear on what you are trying to build. You need to know how to state requirements in a way that the AI create the code that aligns with what you want. You need to be able to translate requirements into tests. That debugs both your requirements and the code generated by the AI. When it does not work, you need to be able to figure out where the problem is. AI can help but only if you give it information to help on.
Vibe coding certainly works for creating "toy" applications for your personal fun.
AI is not magical. It's just an abstraction layer on top of all the current coding languages.
Like all the AI chatbots, it comes down to how good you are at the prompts. Those prompts are the requirements. Also how good you are at checking the results - i.e.hallucinations - by both you (you really did not specify the prompt precisely) or by the AI.
Junior coders just implemented requirements. They ideally created test plans for the requirements. AI makes junior coders less relevant. For AI to work, the people who defined the requirements and the senior engineers that understand the whole system are the key.
Those people were once junior coders. The question is how does someone new to doing software get sufficient experience to define requirements and understand the system. Even more importantly, how to test and debug the requirements.
There is no magic in AI. Its just another abstraction compiler. A very costly one in this case. It's not clear that the cost of running and training the AI engine really is less than just having Junior programmers create code from requirements. The business case is not proven. Copilot and Claude are subsidized today by investors. If they were not subsidized, would they really be lower cost than just hiring a human junior programmer.
That remains to be seen.
For me and my business. Copilot has made it so I can do much more in my less time. But I don't even pay for Copilot. I can use a free version. I don't use the copilot embedded in Visual Studio really. Auto code line recommendations can be disruptive.
That's my take on doing this for a year now. Happy vibe coding out there.
Writing a series of articles ‘AI Primer for Normies’ and this is exactly what I’m focusing on next. It is a major pain for anyone without a strong tech background to understand how to actually create something that functions.
So nice to se a16z sharing here. Some free guide on coding in the AI era
from the 🇧🇷 community ✨
https://ocoelhobranco.substack.com/p/o-coelho-branco-026-extra-mastering
Really thoughtful piece — especially the framing around why vibe coding still breaks down for most users.
One angle I’ve been running into a lot in practice is that some of this frustration isn’t just UX or user skill, but the way platforms operate under the hood. Many route or switch models mid-session for cost or load reasons, which can materially change agent behavior halfway through a build.
From the user’s perspective it feels like you suddenly have a different model — same codebase, same prompts, but degraded reliability.
Curious how you think about this from a business-practice / economic perspective, not just tooling.
Is it really that hard? I mean, many from the product space, technical, but not coders, have picked it up easily enough.
Thx for posting this. I have a CS degree and am building with AI IDEs and still end up debugging hallucinations at 3AM - :}
Nice. Thank you.
I have a masters in computer engineering. Started coding in the 70s.
I now run a retail bakery and do all the coding for the web site and our back end - i.e. payroll and accounting.
Here's the deal on using AI for coding from someone who has been coding since the start.
In the end, the basics have not really changed with AI. Its just the next level of abstraction that has been going on since I started coding. Coding started as assembler, went to basic, then C/C++, scripting like Awk/Python, web created HTML, PHP, Javascript, etc, DBs have SQL... You get the point. All those made it more efficient for me as the "programmer" to create an "app".
AI is a layer on top of all these. It can generate any of these. But here's the catch, you need to be very, very, very clear on what you are trying to build. You need to know how to state requirements in a way that the AI create the code that aligns with what you want. You need to be able to translate requirements into tests. That debugs both your requirements and the code generated by the AI. When it does not work, you need to be able to figure out where the problem is. AI can help but only if you give it information to help on.
Vibe coding certainly works for creating "toy" applications for your personal fun.
AI is not magical. It's just an abstraction layer on top of all the current coding languages.
Like all the AI chatbots, it comes down to how good you are at the prompts. Those prompts are the requirements. Also how good you are at checking the results - i.e.hallucinations - by both you (you really did not specify the prompt precisely) or by the AI.
Junior coders just implemented requirements. They ideally created test plans for the requirements. AI makes junior coders less relevant. For AI to work, the people who defined the requirements and the senior engineers that understand the whole system are the key.
Those people were once junior coders. The question is how does someone new to doing software get sufficient experience to define requirements and understand the system. Even more importantly, how to test and debug the requirements.
There is no magic in AI. Its just another abstraction compiler. A very costly one in this case. It's not clear that the cost of running and training the AI engine really is less than just having Junior programmers create code from requirements. The business case is not proven. Copilot and Claude are subsidized today by investors. If they were not subsidized, would they really be lower cost than just hiring a human junior programmer.
That remains to be seen.
For me and my business. Copilot has made it so I can do much more in my less time. But I don't even pay for Copilot. I can use a free version. I don't use the copilot embedded in Visual Studio really. Auto code line recommendations can be disruptive.
That's my take on doing this for a year now. Happy vibe coding out there.
Writing a series of articles ‘AI Primer for Normies’ and this is exactly what I’m focusing on next. It is a major pain for anyone without a strong tech background to understand how to actually create something that functions.
100% accurate. Now what?
I am thankful you have validated that this is still a real thing. I will keep building.