In recent weeks, the internet has been flooded with experiments from various people showcasing what is now being called ‘vibe coding’, a term coined by OpenAI co-founder Andrej Karpathy. In this, having an ‘idea’ is enough.
“There’s a new kind of coding I call ‘vibe coding’, where you fully give in to the vibes, embrace exponentials, and forget that the code even exists,” wrote Karpathy on X. “I just see stuff, say stuff, run stuff, and copy paste stuff, and it mostly works.”
While AI coding tools such as GitHub Copilot have been available for years assisting professionals by autocompleting code, newer tools like Cursor, Replit, Bolt, and Lovable are now making it easier for beginners to join this growing trend with better features.
VC firm Andreessen Horowitz outlined some limitations for text-to-web app tools, despite their wide potential.
“They excel at simple builds. And if you can’t code otherwise, they can feel like magic. But there’s a limit to what they can reliably generate. Integrations are difficult, bugs persist, and code can get “too big” quickly,” said Justine Moore, partner at a16z, on X.
Another user on X said: “Vibe coding is all fun and games until you have to vibe debug.” One possible solution to this is prompting effectively.
On a larger level, how would the rise of ‘vibe coding’ impact the job of a junior or entry-level developer?
On NYT’s Hard Fork podcast, Anthropic CEO Dario Amodei spoke about the effect of AI on software development and coding. “I don’t think our hiring plans have changed yet, but I certainly could imagine over the next year or so that we might be able to do more with less [developers].”
Anthropic recently released a new version of Claude, the 3.7 Sonnet, touted as the best coding model yet, with some calling it the best model to ‘vibe code’ with. AIM has covered in depth how to build apps easily using simple, yet effective prompts.
Rise of ‘Weekend’ Projects
Vibe coding could prove to be one of the most consequential activities yet. In a lecture at UC Berkeley, Karpathy discussed the scalability of weekend projects to bigger companies.
For instance, the GPT series began as a Reddit chatbot project and evolved into GPT-4. GitHub Copilot started as an internal developer tool and became a global AI code completion platform, reshaping software development.
Likewise, Midjourney, initially a research project on AI image generation, is now a leading platform rivaling DALL-E. Hugging Face, once a chatbot app, is now a $2 billion hub for open-source AI models and datasets, widely used by AI researchers.
Learning How to Prompt is the New Moat
But what’s the secret to successful vibe coding? It all comes down to effective prompting. As AI models become more advanced, the ability to steer them correctly is turning into an essential skill.
Recently, OpenAI president Greg Brockman shared a guide on how to prompt effectively. “o1 is a different kind of model. Great performance requires using it in a new way relative to standard chat models,” he said. This was in the context of their o-series of models, which is different from the usual models, and requires different, contextual prompting.
“Think of prompting as the new user interface (UX) in the age of AI,” Unscript CEO Ritwika Chowdhury told AIM, emphasising that while just three years ago, one had to learn programming languages to instruct computers, today, prompting has become the essential skill for everyone to learn.
“AI models like GPT-4 or other large language models are trained on vast amounts of data but rely heavily on how they are prompted to produce meaningful, accurate, and contextually appropriate responses,” added Sudipta Biswas, co-founder at Floworks, part of Y Combinator’s Winter-2023 batch.
Prompt engineering is a real, growing profession. “Prompt engineering is a job today—that’s not something we could have predicted,” said Thinking Machines Lab CEO Mira Murati in an interview.
Biswas added that in its early stages, prompt engineering was more of an experimental practice, with users trying to figure out how to phrase their inputs to generate the most useful outputs. However, it has now matured into a profession owing to factors like complexity of AI models, industry demand, and optimisation for use cases.
“This reflects a broader trend where interdisciplinary skills, including those from the humanities, are increasingly valuable in the AI domain, highlighting the importance of understanding both technology and human communication,” concluded Chowdhury.
