One-quarter of YC-backed startups are generating as much as 95% of their code with AI
In recent years, the rapid development of artificial intelligence (AI) has brought automatic code generation into the spotlight. In software development, the use of large language models (LLMs) such as GitHub Copilot and ChatGPT has led to an increasing number of cases where programmers leave part or most of their coding to AI. Notably, at the famous startup accelerator Y Combinator (YC), it has been reported that one-quarter of the startups participating in the latest batch are "generating as much as 95% of their code with AI." This article delves into this phenomenon, its background, comments from key figures, specific risks, and a new development style called "Vibe Coding."
1. The current state of AI-driven code generation
AI-driven code generation has gained significant attention over the past few years. The improved accuracy of large language models (LLMs) has brought revolutionary productivity gains to developers and is being widely adopted by everyone from startups to large corporations.
1-1. Trends shown by YC's latest batch
The fact that one-quarter of startups in the Winter 2025 (W25) batch are generating 95% of their code with AI symbolizes that AI utilization is becoming mainstream. Jared Friedman, a managing partner at YC, spoke about this in a YouTube discussion:
"These startups are run by founders with technical expertise. A year ago, they would have written all the code themselves. But now, we are in a situation where 95% of the product is built by AI."
The 95% figure mentioned here is calculated by comparing the code actually written manually by humans versus the code generated by AI, excluding things like library import code. It is noteworthy that AI is handling the majority of the core logic, rather than just simple code completion.
1-2. The new trend of "Vibe Coding"
In a discussion with YC CEO Garry Tan, managing partner Harj Taggar, and general partner Diana Hu, attention was focused on a new style of creating code using natural language and intuition. This overlaps with the concept of "Vibe Coding," coined by Andrej Karpathy, a former Tesla AI director and OpenAI researcher.
Vibe Coding: A development method where one does not focus on the details of the code, but instead conveys natural language instructions or intuitive images of "how I want this to move" to an AI model to develop software.
For example, if you tell an AI in natural language, "I want a feature where the user logs in, goes to the dashboard, and displays a list of user data," the AI generates that as code and creates a prototype immediately.
2. Benefits and challenges of AI-generated code
While generating code with AI brings significant improvements in productivity and speed, it also points to risks in quality and security.
2-1. Benefits: Speed and innovation
Development tasks that would have taken days or weeks a year ago can now potentially be completed in hours or minutes. Furthermore, by comparing multiple code samples and design proposals presented by AI, it is also attractive to be able to gain innovative ideas that one might not have thought of before.
"This is not a temporary fad; it is becoming the mainstream development method. The option of not using AI will gradually fade away." (Garry Tan)
In fact, hundreds of millions of dollars have been invested in AI-powered coding assistance tools and startups over the past 12 months, indicating high expectations for the "productivity improvement" market.
2-2. Challenges: Bugs and security risks
On the other hand, the fact that AI-generated code is not always accurate is a major concern. The following cases can be considered as risks in terms of code quality.
Insertion of security holes: There are cases where vulnerable code or implementations that are weak against classic attacks like SQL injection are proposed.
Difficulty of debugging: The generation process can easily become a black box, and if the developer does not convey their intent correctly, incorrect code may be output.
Maintainability: Even if AI-generated code appears to work correctly at first glance, it may have a structure that is fragile for future expansion or changes.
In fact, Diana Hu, a general partner at YC, sounds the alarm as follows:
“Even if you leave most of the product building to AI, the ability to find and fix bugs is essential. You need basic programming knowledge and experience to judge whether the LLM is spitting out correct code.”
3. Developer skill sets required when using AI
As automated code generation by AI becomes more common, the required engineering skills are changing. Many investors and entrepreneurs, such as Garry Tan and Harj Taggar, unanimously emphasize the 'importance of classic coding skills.'
3-1. Ability to read code
To judge whether generated code is correct and to fix bugs as needed, you need the literacy to determine whether the code is readable. Skills to analyze error messages and understand the logic and algorithms of the code are indispensable.
“If a startup launches a product early using AI-generated code, when the user base grows from 1 million to 100 million, whether that code can maintain robustness is a major challenge.” (Garry Tan)
3-2. Debugging and problem-solving skills
Although the accuracy of code proposed by AI is improving year by year, it is not perfect. When a malfunction occurs, developers are required to have the ability to identify the root cause and make precise corrections.
Example: Skilled engineers need to compensate for points that AI tends to overlook, such as code that cannot withstand large-scale traffic or implementations that do not consider API call limits.
Friedman states, 'No matter how excellent AI is, if the human using it does not spot bugs and make necessary corrections, the sustainability of the product is at risk.'
4. Future prospects and the potential of 'Vibe Coding'
'Vibe Coding' is still a new concept, but if a style of building systems using only natural language or intuitive instructions becomes widespread, it could lead to a world where not only is the development process streamlined, but non-engineers can also create software.
4-1. Opening the door to non-engineers
In the past, it was considered difficult for people without programming experience to build tech products. However, if AI takes on the majority of code generation, we can see a future where you can shape a product as long as you have the ability to generate ideas.
On the other hand, if you have absolutely no programming knowledge, there is a risk that services with insufficient security or performance will circulate in the world, so minimal technical literacy is essential.
4-2. Impact on emerging startups
Startups that develop and provide AI coding tools are also appearing one after another, raising huge amounts of capital from VCs. For example, Bolt.new, Codeium, and Cursor have been attracting attention in recent years. These services are expected to assist developers in coding efficiently using AI, leading to shorter development periods and reduced labor costs.
“This is not a passing fad. It is certain to become the mainstream coding method, and companies and developers who do not use it risk being left behind.” (Garry Tan)
AI-driven code generation has already become a major trend in the startup world, as demonstrated by YC's latest batch.Vibe Coding A new natural language-based development style has emerged, and it can be said that the roles and skill sets of so-called "programmers" are changing significantly.
Increased productivity and creation of new businesses: With technology now at a practical stage where 95% or more can be left to AI, it is becoming possible to test more ideas in a shorter period.
Risks and challenges: To maintain the quality of AI-generated code, traditional coding knowledge and debugging skills are indispensable. Literacy in security and maintenance is also required.
Future predictions: If generative AI models continue to evolve, there is a possibility that current bugs and security risks can be further reduced. On the other hand, expert supervision and judgment in bug fixing will be key to ensuring the robustness that supports products in the long term.
Finally, I will quote the words of YC CEO Garry Tan.
"This is not a fad, and it is not going away. AI-driven development is becoming the dominant method. If you are not using it, you might already be left behind."
We need to build a hybrid development system where human engineers ensure quality and safety while leveraging the benefits brought by AI-generated code. Such a new development style will become the foundation that supports the future software industry.
Related articles