Vibecoding and the future of AI

Vibecoding is the idea that hopes and dreams plus artificial intelligence equals computer programs.

program = vibes + ai

I can see why this is appealing.

After using the AI for programming, I have vibes too. I have used ChatGPT, Claude, Llama 4 "Scout", and various Mistral models. The recent AIs are all very similar when it comes to coding. Claude is the most concise. ChatGPT seems to regenerate the entire program after every suggested change (which is actually good because you can "diff" the results against your current working copy (hint: use "meld" https://gitlab.gnome.org/GNOME/meld ) ). Llama 4 runs on my personal computer. It seems to be pretty good at coding (reddit weirdos didn't like it but they are weirdos). I haven't used any of the AIs that are integrated into IDEs, such as Cursor or CoPilot. They might work for you. IDEs just aren't my style. So, keep in mind that the AIs that I am using aren't autocompletes, they are chatbots.

There are two main limitations of the AI: context length and correctness. Both of these things will improve, but as of now, we will have to work around them. I will explain how.

Context length. It is a number. It is a number of tokens. It represents a consumable, depletable thing. In regards to text, a token is about 3/4 of a word. Your conversation with the AI uses up the context space. Each token, each word, uses up this precious resource. When the conversation get so long that the context space has been used up, the AI starts forgetting things and goes schizo.

So keep your programs small. I have found that small programs can be made correctly. You can try things like asking about single functions instead of the entire program. Simply generating code is not enough. You will need to question the AI about what it wrote, and get it to make changes. You need enough context length for code and a conversation about it.

Now for the subject of correctness. The AIs are wrong very often but computer programs need to be correct. You must know that the AI doesn't know anything. It has no source of truth and no way to discern it. It will tell you very incorrect things. It will make up functions that don't exist. I will write dangerous code that may result in loss of data. You need to inspect AI generated code closely. Does it make sense? Once the AI generates code, be sure to reference some real software documentation if you are unsure about what the code does.

Try:

• "Is there a simpler way to do this?"
• "I tried this and it didn't work."
• You can paste in an error message.
• Don't ask for advice. Ask for options. You are the decider.
• Suggest using a specific technique. You will have to steer the conversation.
• Provide snippets of relevant documentation to the AI.
• Contradict the AI. Tell it that something you know to be true is false. Be horrified when the AI agrees with you.

These limitations are related. In the search of correctness, the conversation will become long enough that the context length will become exhausted and the AI will go into full bull-nonsense mode. There is usually no warning for this. Be prepared for your program to get too large for the AI to handle.

 

Example vibecode technique. You need a text editor, a terminal, and a web browser (for the AI).

Make 3 files:

1. myprogram.txt: Use this for notes at first. As you continue vibe coding, it will become your design document.
2. myprogram.py: This is the program. I am using Python in this example.
3. scratch.txt: Paste AI output here so that you can diff it.

Folow this algorithm:

1. Ask the AI to program. Copy the AI generated program to scratch.txt
2. Diff: meld myprogram.py scratch.txt
3. Compare differences, test and edit your program, reason about it, take notes, discover the software requirements and refine your notes into a design document.
4. Clear scratch.txt, go to step 1, and keep going!

Due to the availability of training data, the AI is best for the most widely used computer languages. Requesting Python, Bash, or Javascript will have better results than Prolog or Erlang, for example.

I really need to mention again that the AI will be wrong. Distrust it. You should treat it with disdain and its code with suspicion. It only seems like magic. It isn't. At some point, you will actually need to learn to code in order to code, but why would you do that to yourself? It had better be for a very good reason because people who learn to code are never the same afterward.

For a relatable example of AI wrongness, consider prose. Consider the text of this very blog post. I asked the AI to check my grammar and also things such as "did I use the word 'whom' right?" and it goes and makes all kinds of very wrong suggestions. It went and corrected one sentence by turning it "inside out" (my words)! I was like, why? so I can sound like a robot? Then I was like, I am not going to use the words "among" or amongst" (one of its suggestions). I mean, the words are in my vocabulary, but like, seriously, who talks like that? and then it was like "put a semicolon here" and I was like, uh, no. I'm not programming, why would I use a semicolon? I also happen to agree with Kurt Vonnegut in regards to semicolons. I did use its suggestion that I should use a hyphen between the "kick" and the "ass" of kick-ass because it is a "compound adjective". Ok, that does vibe. But then I checked up on the internet about grammar (because you should always check up on what the AI says), and I found out that the Millennials ruined grammar, so it doesn't matter anyway.

Even with its limitations, the AI has a huge advantage over human programmers. It is much more polite. You can ask it simple questions without judgement. For example, It doesn't make fun of you for not knowing something (yet). This is also the reason why we need to immediately replace programmers with AI whether or not that even works.

While I don't think that the AI will replace programmers, I do think that the AI will help many more people program. Remember all of the people who learned to code Flash/Actionscript? They were mostly self-taught and they were able to make some very interesting things (It is quite a shame that we lost Flash). The AI can also help people who need to program an Excel spreadsheet (It is quite a shame that we still have Excel). Such people are able to teach themselves, they can reason and refute, which is what I think is necessary to successfully use the AI for coding, because the AI is wrong all the time. 

It seems to me that a person needs to have a certain type of personality to program and considering that many could program Flash and many can program spreadsheets, there are many people who have the requisite personality traits. The schooling for programming and the culture of it does further selection of this group, selecting the people whom we must absolutely replace with AI.

The future of AI:

1. Artists and Accountants make a lot of kick-ass awesome stuff using AI.
2. Programmers get jealous and cringe and whinge.
3. Programmers take away the AI.
4. Something must be done about those Programmers because of what they did and because they make fun of people for not knowing things (yet), and other reasons.
5. Programmers are replaced by AI and cron jobs, which completely humiliates them (the cron jobs are what they find humiliating -- they still think that AI is cool).
6. Artificial General Intelligence never happens, hence the cron jobs.

So, yeah, I realized this and it changed my life.