From Vibe Coding to Vibe Researching: The Era of the 'Automated Researcher' Opened by GPT-5

In recent years, with the evolution of artificial intelligence (AI) and large language models (LLMs), a new programming paradigm called "vibe coding" has become a subject of debate. This refers to a style where developers step slightly away from "writing code" and instead generate and modify code by giving instructions to AI. Meanwhile, Mark Chen and Jakub Pachocki of OpenAI have mentioned a concept that takes this trend even further: "vibe researching". In this article, based on the strategies and research stance for the GPT-5 generation discussed by them, we will organize and explain the current status and limitations of vibe coding, as well as the future prospects of vibe researching, incorporating specific examples, papers, and reports.

1. What is Vibe Coding?

1-1. Definition and Context

"Vibe coding" refers to a development style where developers express their intentions in natural language (verbally or in writing) and entrust code generation and modification to AI.

• This concept is said to have been proposed by OpenAI co-founder Andrej Karpathy in February 2025.

• In other words, it is a form where humans do not write all the code by hand, but rather, in a sense, "talk to the AI to make it create the code."

1-2. Academic Considerations and Limitations

Vibe coding is still a new concept, and research such as the following is underway:

• "Vibe Coding: programming through conversation with LLMs" is a paper that analyzes the thought processes, prompt strategies, and behaviors observed in the field during debugging.

• Another paper presents the argument that, from the perspective of "intent mediation," vibe coding may be in the process of restructuring traditional development processes.

• On the other hand, code generated by AI is prone to becoming a black box, and issues such as maintainability, readability, and the division of responsibility (who fixes the mistakes) have also been pointed out.

A point to note is that simply leaving everything to AI is highly likely to lead to fatal bugs or unexpected results, so human oversight and verification capabilities remain indispensable.

2. GPT-5 and OpenAI's Vision

2-1. The Goal of GPT-5: Strengthening Reasoning and Agentic Capabilities

Mark Chen and his colleagues position the release of GPT-5 as a step toward making "reasoning" and "agentic behavior" more standard features, rather than just a "response generation model."

They state that they aimed for a model that integrates the past designs of the GPT series (instant response type) and the "o" series (deliberate thinking type), without forcing users to choose a mode, and instead automatically determining the appropriate "amount of thought."

2-2. From Evaluation Metrics (Evals) to Discovery and Economic Value

Given that traditional evaluation metrics (e.g., accuracy and scores) have reached a saturation point, OpenAI is beginning to emphasize the axes of "whether the model can discover unknown hypotheses" and "whether it can produce economically valuable discoveries" as its next challenge.

In particular, achievements in mathematics and programming competitions are positioned as indicators of "intellectual discovery power" that go beyond traditional benchmarks, and results in events like AtCoder and the IMO (International Mathematical Olympiad) are attracting attention.

2-3. The Concept of Vibe Researching

While there has previously been a trend toward 'AI assisting researchers' or 'becoming a research support tool,' what Chen and his colleagues call 'Vibe Research' goes a step further.
They set their future goal as an 'automated researcher' capable of autonomously discovering new ideas.

In other words, they are looking toward a future where AI can formulate research themes, design experiments, verify hypotheses, and produce output comparable to writing a research paper.

3. Discussion points for the transition from Vibe Coding to Vibe Research

The following are the main points to discuss in connecting Vibe Coding and Vibe Research.

3-1. The trade-off between agency and stability

To realize Vibe Research, it is essential for AI models to have the ability to autonomously perform multi-step planning and tool calling. However, allowing such complex behavior increases the risk of error accumulation and quality degradation at each step.
Chen and his colleagues also state that balancing 'multi-step planning' with 'quality stability' is a critical research challenge.

3-2. Durability and long-term reasoning/memory capabilities

The act of research requires the ability for long-term thinking, memory, and self-monitoring, rather than short-term prompt responses.

They emphasize enabling models to 'maintain thought/retain information over many hours,' and have set extending the long-term horizon as a research objective.

3-3. Verifiability vs. Emergence

While correctness can be determined in fields that can be formalized, such as mathematics and physics, research fields are often ones where 'solutions are not clear/evaluation is difficult.'
To make Vibe Research viable, a methodology is required to overcome the barrier of how to evaluate emergent hypothesis generation and its validity. When asked about this 'expansion into areas with low evaluability,' Chen replied that 'the difference between the two might gradually fade.'

3-4. Resource allocation and roadmap management

Resources such as 'compute,' 'data,' and 'talent' are indispensable for AI research.

OpenAI is constantly mindful of portfolio management regarding how much computing power to allocate to which project, and emphasizes balancing long-term exploratory research with immediate product development. Furthermore, the flexibility to dynamically decide whether to 'allocate to compute or talent' when additional resources are obtained is also considered key.

4. Practical examples and critical perspectives

4-1. Case study: Vibe Coding works by non-engineers

A Business Insider report introduces a case where an author with no programming knowledge used Replit's agent to create a 'Musk vs. Altman racing game' through Vibe Coding.

Although the code was buggy and showed unstable behavior, the fact that a working app was created in a few hours demonstrated both the potential and the limitations of Vibe Coding.

4-2. Critical voices: What will happen to the role of professional engineers?

Former OpenAI research lead Bob McGrew has stated, 'I don't think Vibe Coding will make engineers unnecessary anytime soon.' The view is that while prototype generation is possible, the contributions of traditional engineers are essential to ensure maintainability, safety, and understandability.

5. Future Outlook and Implications for Readers

• The Superiority of Hybrid Systems: Rather than full automation, a research and development structure where AI and humans complement each other is more realistic and powerful.

• Innovation in Evaluation Design: Evaluation criteria that can measure 'discovery capability' and 'long-term planning ability' beyond traditional accuracy metrics will be required.

• Changes in Researcher Skills: In the future, emphasis will likely be placed on idea conceptualization, prompt design, and experimental direction skills, while the relative importance of coding ability may decrease.

• Preparing for Risks: It is essential to have designs and systems in place that ensure verification and accountability, without over-relying on AI-generated insights or code.