Vibe Coding: It's Four Months Old. What's Up?

Forbes

5 hours ago

Female freelance developer coding and programming. Coding on two with screens with code language and ... More application.
What is vibe coding, anyway?
The term, coined by Andrej Karpathy a few months ago, is now shorthand for an entire shift in the way that we view software engineering. It's the idea that AI is 'hands-off' in terms of code generation: the machine just takes the human's inputs, and comes up with source code on its own.
Although LLMs are not 100% autonomous with this yet, and there's often a need for some debugging, vibe coding essentially asks the programmer to back off and let the AI do its thing. In contemporary coverage of the phenomenon, Forbes Council member Shubham Nigam quotes Rhiannon Williams of the MIT Technology Review:
'Not all AI-assisted coding is vibe coding. To truly vibe-code, you have to be prepared to let the AI fully take control and refrain from checking and directly tweaking the code it generates as you go along—surrendering to the vibes.'
So while it's a colloquial term (good vibes, man!) it's also a request for the human in the loop to take distance – not to stand over the LLM's shoulder as it creates.
Here's more from a recent panel at Imagination in Action in April. Nikolay Vyahhi of Hyperskill interviews Artem Lukoianov, Heena Purohit, and Aldo Pareja about this trend.
'I guess the whole beauty of this term is because it so accurately presents what's actually happening,' Lukoianov said. 'You don't even have to read the code that it produces. You … just basically teleprompt to the system, it generates a part of the code for you, and then, quite often, because developers are notorious for being lazy, you don't even read the codes.'
'You don't even try to understand what the code is doing,' added Pareja, theorizing about what will happen when IDEs and other tools start to incorporate unit testing. 'You don't even read the code. You're just feeling it.'
Panelist Zach Lloyd talked about the realities of source code management.
'(Developers will) get into trouble, and get and they'll try to vibe their way out of it, or they'll get into trouble with their production system and try to vibe their way out of it,' he said. 'So in the terminal, we see it goes beyond producing code, to this whole feeling of: 'let me see if the AI can just fix this thing for me, and maybe I won't have to understand exactly what it's doing.'
Lloyd described this power as a kind of double-edged sword: on the one hand, it's, as he said, a 'magical' thing for a developer who feels stuck. On the other hand, he suggested, it can be dangerous for the human coder not to know what the system is doing at all.
Panelist Heena Purohit pointed to some challenges with letting the AI have its own project.
She argued the systems are not typically good at 'distance thinking,' or how various components of a system interact.
'Sure, you can have millions of lines of code be generated in minutes or seconds, but you still need to understand what the code is actually doing, so that you can troubleshoot it and debug it when you need to,' she said, suggesting that in many cases, scaling might be a problem.
By contrast, Lukoianov gave a sort of qualified opinion that we are mostly there, and will get there soon.
'(Vibe coding capability) is already good enough for us to stop coding anything,' he said. 'To me, it's more the question of how we engineer the system around this… how do we … provide the correct information to the LLM, how do we summarize our code base, and how do we provide the right tools to the LLM to actually perform … better? in my personal opinion, I feel that like the LLM is already there. It's all about, how do we properly provide this information about your code base, about what you want, about any regulations, security issues that are around there. So it's all just about the correct information, correct inputs, correct tools, to the LLMs, and eventually we'll get there.'
In the 1980s, he pointed out, we had to be very close to the hardware – now, it's different.
'You don't think that much about operating systems,' he said. 'You don't think that much about hardware, unless you're working in something very specialized.'
Regardless of the change, Pareja argued that full stack developers are still valuable.
'If you're using (tools) to synchronize different processes, and you have a million processes, your system is going to break,' he said. 'You need to understand these constraints.'
If I was going to boil down some of the biggest ideas in this panel, I think most of them would be around the need for coding knowledge to manage the detail and periphery of systems development. In other words, the AI can do everything, but it might not do it 100% the way that you need it to be done. And there's that old adage: if you want it done right, you have to do it yourself.
Maybe LLMs get us 80% of the way there without any oversight, but skipping the context and being completely ignorant of what the machine is doing is typically not a good idea, partly for the reasons that the panelists laid out. So the human in the loop is still relevant for now. But the bottom line is that vibe coding is something so new and fundamental that we'll probably be spending a lot of time figuring out how to do it best.