Natural language programming (NLP) easily predates AI’s NLP (natural language processing) that we talk about more today.
This older form of NLP has been around since the 1950s as a goal or idea—that we can just code using our everyday languages, like English. Programming languages like FORTRAN and COBOL were once thought to be “English-like,” and this goal keeps resurfacing with new languages, shells, and interfaces.
But now we have AI, and the newest form of NLP could be called vibe coding.
This isn’t a one-to-one with code, but more like speaking your objectives and sending off a team of programmer interns or beginner devs—albeit extremely well versed in publicly available software resources and general facts to boot—to execute it.
So instead of natural language programming, vibe coding could be more analogous to guiding, nurturing, and carefully reviewing (and repairing) this initial draft work.
Our Founder and CEO wrote a profile of vibe coding back in March (assessing its potential impact on developers), so today we revisit it as it moves from weekend warrior superpower to startup workplace staple.
But is it ready for enterprise work? That’s one question we’ll examine as we look at vibe coding, how fast it’s emerging, and how far it’s come just in 2025.
Vibe Coding vs Traditional Coding vs AI-Assisted Coding
The term vibe coding is credited to Andrej Karpathy, a founding former member of OpenAI and former director of AI at Tesla (he also authored the first deep learning course taught at Stanford).
In February, Karpathy posted on X about weekend throwaway projects coded entirely by AI, wherein he described barely even touching the keyboard, using voice to talk to it, and accepting everything. When there were error messages, he just copied and pasted them without comment back in and let the AI fix them. And usually it could.
He wrote: “I just see stuff, say stuff, run stuff, and copy paste stuff, and it mostly works.” The name came from giving in to the vibes, using natural language to build software, with the goal of the code writing itself being far less important to the creator.
Karpathy wasn’t alone in the approach. Weekend or hobby projects everywhere were immediately being showcased, whether for apps to help pack a school lunch with items in your fridge (the New York Times’s Kevin Roose), build a website with pay functionality for your daughter’s custom soaps in just a few hours (Business Insider’s Alistair Barr), or craft a 3D game where you build a town and drive around competitively putting out fires (the Wharton School’s Ethan Mollick).
Vibe Coding Adoption Booms in 2025
Vibe coding tools have since sprung up quickly and caught fire, from Cursor (built on VS Code) to Replit (cloud IDE, partnered now with Microsoft Azure) to Windsurf (code assist with internet search, famously sought by OpenAI, raided by Google, and bought by Cognition) to Lovable (no-code capable) to Claude Code.
It’s also begun to move from the weekend to the workday. Bloomberg profiled in June how Cursor reached a $9.9B value, while CNBC reported the amount of code now being written by AI at large companies like:
- Amazon (30%)
- Google (30+%)
- Microsoft (20–30%)
- Meta (50% within the next year)
- Visa (≈30%)
And while not yet driving bottom line profits, AI-generated coding has been boosting productivity and aiding with scaling, so far used most heavily at startups managing to thrive from teams of five instead of 50.
(See below for some vibe coding tips from some of these entrepreneurs.)
Note that these numbers mix code copiloting with vibe coding, as the terms—like many in AI that relate to use and marketing—have also gotten intermixed. And while vibe coding may denote letting AI handle more of the coding than copiloting does, no doubt professional developers are using both techniques together.
The Dream of Software Development Automation
More than two-thirds of developers across surveys acknowledge using AI coding assistants (daily, according to Canva’s most recent poll), and one-third of new production code among the biggest enterprise adopters is AI-generated.
But vibe coding tools are more about talking code into existence—not using the technology to augment your existing knowledge and experience.
As discussed, the allure of that goes back to the earliest days of computing. With true vibe coding, development, or at least prototyping, can be fully democratized.
Rather than needing engineering skills, anyone can create software.
From a business POV, this could mean lower development costs and massively accelerated cycles, with a shift from grinding to focusing on design and intent.
Stakeholders could engage in real-time prototype development, asking for alternatives and seeing the results in minutes. Then adjusting and iterating rapidly.
Sound too good to be true?
On AI-Generated Code Quality: What Can Possibly Go Wrong?
For enterprise code that has to be efficient and secure, this is still more dream than reality (at the time of this writing).
Because while vibe coding can create remarkably effective code (especially in well-defined, familiar, controllable chunks), it can also proliferate mistakes in your code base. Like some junior developers, today’s AI generated code has a tendency to miss more efficient alternatives, neglect documentation and logging, and even hardcode keys, passwords, and critical information.
It can also write bloated code, introduce security risks, hallucinate fictional software libraries, and, in the very worst (and rarest) of examples, nuke your production database.
AI-Generated Code Security Issues
“Frequently occurring issues are missing or weak access controls, hardcoded secrets or passwords, unsanitized input, and insufficient rate limiting. In fact, Veracode recently found that 45% of AI-generated code contained an OWASP Top 10 vulnerability.”
This according to Forrester analyst Janet Worthington, who told CSO that all too often, the threat intelligence on AI code turns up sloppy errors more akin to informal code than enterprise-grade software.
Vibe coding security risks are still significant enough that professionals like Secure Code Warrior CTO Matias Madou recommend treating it much like zero-trust with system access.
“As a security professional, I check any AI-generated code for flaws… But less experienced developers won’t. That’s where secrets and unsafe defaults slip through.”
Hallucinations Are Also Vibe Coding Challenges
Worthington cited research which found that more than 5% of software dependencies suggested in AI code by commercial LLMs (and 21%+ from open source) were hallucinations.
This included cases where the libraries being used weren’t secure, high-quality, or even real.
Our first stat above comes from researchers from Virgina Tech and the Universities of Texas San Antonio and Oklahoma, who broke down package hallucinations between languages, settings, and parameters, using 16 different LLMs.
Among the 576,000 code samples analyzed, they found a shocking 205,474 unique software packages were found to be entirely non-existent—fabricated into existence by the LLMs.
But worse than this can be destructive errors. One startup found out the hard way that, when given access to your systems, AI code will sometimes act without your supervision.
During one live test at a startup, a Replit coding agent in development issued a command that wiped out a production database.
And while the data was swiftly recovered, it led to Replit’s CEO issuing an apology on X, calling this “unacceptable” and explaining how safeguards have been implemented to prevent it in the future.
Nevertheless, the event demonstrates the dangers of using AI-generated vibe code without sufficient limits and human control.
Vibe Coding in the Workplace: Startups vs. Enterprise
Few companies at this point would willingly risk their production database to unreviewed AI code.
But these fears aside, many startups are leaping in with two feet and finding success in AI coding on their own terms.
Among Y Combinator’s W25 startup batch, a quarter of startups reported codebases that were almost entirely (around 95%) written by AI, according to YC managing partner Jared Friedman.
Enterprises, for the reasons discussed above, are being more careful. But they’re still moving (more with code assist than vibe coding), with examples like Citigroup pairing 30,000 developers with AI code tools, and Walmart reporting some four million developer hours have been saved with coding assistants.
At the same time, some reports find more senior developers actually lose efficiency with AI coding assistant use. One analysis by METR (Model Evaluation and Threat Research, as reported by Ars Technica) found that developers expected a 24% efficiency boost but in fact were 19% slower with AI over sampled tasks.
A Fastly survey found that around 95% of developers spend extra time fixing errors in AI-generated code, with the senior coders bearing the brunt of this load (and far more time spent fixing the code than generating it).
At the same time, they also found senior developers twice as likely to ultimately ship AI code. (Perhaps unsurprisingly, vibe coding cleanup specialist is a new role emerging for just this purpose.)
The result for enterprises can be a balancing act between moving too cautiously and missing out on potential efficiency gains versus introducing potential vulnerabilities or additional hours spent in proofing code sprawl.
The key is not only choosing the right tools, but also extensive monitoring and review with effective processes.
Startup Tips for Maximizing Results in AI-Assisted Software Development
To this end, consider these pointers for making effective use of vibe coding techniques in the workplace, from YC’s Tom Blomfield in their Startup School series.
Blomfield has been building projects in tools like Claude Code and Windsurf but also shared pointers from some of the most successful vibe coders out there—startup founders:
- Nicole Lu (Sieve): Pasting code from an AI IDE back onto the underlying LLM’s actual website UI often gives different results when stuck.
- Skyler Ji (Human Behavior): Using multiple tools at once can be more efficient for different areas—Cursor’s faster for the front end while Windsurf thinks longer. You can also use both with the same context at the same time and pick your favorite.
- Arshad Shaikh (Leeroo): Starting your work with the test cases—handcrafted, then built to with strong AI guardrails—can save you from having to micromanage the code. When you get green flags on these tests and you know the tests are good, you’re good to go and can then refactor for modularity.
- Antoni Gmitruk (Authed): Start with pure LLM (vs vibe coding tools) and build scope and architecture. This keeps the vibe coding from running wild in unrealistic directions.
- Jack Swisher (Beluga Labs): Monitor whether the LLM is going into rabbit holes. If it looks funky or keeps regenerating, step back, and make a clean go.
Ultimately, much of this boils down to ensuring the LLM follows good software development practice, even if it means some handholding.
Blomfield recommends designing and documenting extensively beforehand, using LLMs to generate a comprehensive plan. AIs can also make effective planning documents in markdown files, which you can then proof, organize, and streamline before attempting to code, as with traditional software development cycles.
One difference here is you can do it far faster, using the LLM as your assistant. And while many vibe coding tools and AI copilots boast their own source control, all recommend using your own and committing regularly so you can always rollback changes yourself.
AI tools can also make unexpected changes to other parts of the code than the section being addressed, so this gives a safeguard to ensure each piece is developed the way you need it and stays that way.
Writing high-level integration tests also ensures everything stays on track.
Another piece of advice from successful vibe coders?
Start over regularly with a clean slate. When you go through multiple stages, issues can accrue and the LLM may not be seeing things clearly. Take what works and start again, and when in doubt, even switch models. One may succeed where another has failed.
AI is also often undervalued for its power in non-coding dev work, too, like configuration, test writing, and even walking through code line by line for explanations and suggestions. You can copy and paste screenshots in, for example, and have the system work from them.
As Blomfield recommends: “Small files and modularity are your friend.”
At PTP, one of the many things we do with AI is implement agentic AI software testing. If you are making the move the AI coding automation, consider us to help you keep on track with DevOps automation and ensuring quality remains high.
Conclusion: Developers Face the Future of Workplace Coding
We’ve now presented a few different takes on vibe coding:
- That miraculous weekend project you spoke to life with natural language that actually works (security and heavy workloads not being essential)
- Its repeated security issues, hallucinations, and the instances of sloppy coding that can pop up
- The way hungry startups are working through some pain to get their ideas on their feet, in ways never before possible
Is vibe coding ready to handle enterprise code at scale?
Not yet.
But where it is already proving invaluable is in the dozens of ways innovative engineers are using it to make their lives easier every day—automating or speaking to life repeated code, debugging, translating, documenting, and testing.
References
Speaking things into existence, One Useful Thing
Anysphere, Hailed as Fastest Growing Startup Ever, Raises $900 Million, Bloomberg
‘Vibe coding’ comes to big tech, CNBC
Why Did a $10 Billion Startup Let Me Vibe-Code for Them—and Why Did I Love It?, Wired
From misconceptions to momentum – the case for vibe coding in modern enterprises, TechRadar Pro
The enterprise is not ready for vibe coding — yet and Citi deploys AI coding tools to 30K developers in modernization push, CIO Dive
We Have a Package for You! A Comprehensive Analysis of Package Hallucinations by Code Generating LLMs, arXiv:2406.10279 [cs. SE]
When AI nukes your database: The dark side of vibe coding, CSO
Vibe coding has turned senior devs into ‘AI babysitters,’ but they say it’s worth it and A quarter of startups in YC’s current cohort have codebases that are almost entirely AI-generated, TechCrunch
Study finds AI tools made open source software developers 19 percent slower, Ars Technica
How To Get The Most Out Of Vibe Coding, Y Combinator Startup School
FAQs
What’s the difference between vibe coding and copilot tools?
Vibe coding originally referred to using natural language to generate code directly. In this case, you don’t need to know how to code to generate working applications. With copilots, developers use AI for features like autocomplete, code assist, debugging, and more.
Increasingly, the two terms are bleeding together, and where security and efficiency matter, having the code reviewed by humans with software development skills remains essential.
Who is using vibe coding today?
While it exploded in the scene earlier this year in use for side projects and personal experimentation, startups are increasingly taking advantage of vibe coding tools to get prototypes and early models of their ideas on their feet much faster and with far less.
At the same time, many enterprises are rapidly scaling up AI coding assistance and tool-use in their development process.
Does vibe coding replace the need for human developers?
No. It is true that vibe coding can already help do things that once required far more talent and resources to get accomplished, but at the same time the code is found to contain regular instances of hallucinations, security vulnerabilities, and inefficiencies that make it essential to check it before it goes live. At this point it functions best as a generator of first draft code, or for use in sandboxed scenarios.
