“Now is the Time to Start a Business” — The Speed, Product, and Human Element Required in the AI Era

As CTO of Microsoft, Kevin Scott is a leading figure who oversees everything from AI research and implementation to business development. In recent years, as the progress of Large Language Models (LLMs) and agentic AI has accelerated, he has been actively sharing his insights in interviews and lectures. This article focuses on Kevin's remarks, providing a detailed explanation of his assessment of the Chinese-developed Large Language Model "Deepseek," AI scaling laws, the future of agents, and more. While these are highly technical topics, we will focus on clarity and approach their essence using concrete examples and quotes.

1. Why now is the best time for entrepreneurs

1-1. Kevin Scott on the intersection of AI and entrepreneurship

Kevin emphasizes, "This is the best time to be alive if you have an entrepreneurial spirit." He states that because new technologies like AI are emerging one after another and existing common sense is changing drastically, there are more business opportunities being created than ever before.

The background to this is that the evolution of large language models and AI infrastructure has created an environment where it is faster and easier than ever to build prototypes and develop iteratively while checking user reactions. It can be said that this is an era where both existing companies and emerging startups can explore the "creation of unknown products" on equal footing.

1-2. It is the "product," not the model, that creates value

Kevin repeatedly states, "Models aren’t products." AI models are merely technical foundations, and what ultimately solves user problems is the UI/UX—that is, the form of the product.
Therefore, he emphasizes that even if the accuracy and performance of an "AI model" are excellent, how it is integrated and what kind of experience it provides determines its business and social value. This is also an important perspective when launching a new business.

2. How far will scaling laws continue?

2-1. The potential of scaling: "The end is not yet in sight"

Regarding large models, there is a debate that "even if we increase the size of the model, we will soon see the limits." However, Kevin states, "I can very clearly see what we’re doing now and what we’re doing next, and I don’t see the limit to the scaling laws."

Of course, we will eventually face problems with physical laws, costs, and data volume, but at this moment, the recognition is that "the situation where performance continues to improve as we scale up is continuing."

2-2. There are rational constraints, but no sudden stagnation

Since scaling requires massive computational resources, power, and data, there may come a time in the future when the cost-effectiveness no longer makes sense. Kevin himself, while comparing it to the number of neurons in the human brain, acknowledges that a certain upper limit will eventually arrive. However, he notes that this is in the distant future, stating, "We’re not at that asymptote yet," and expresses the view that significant performance improvements will continue in the short term.

3. Data, compute, and algorithms: What will be the bottleneck?

3-1. The importance of data quality and evaluation

Kevin points out that the biggest problem in model training is not just the amount of data, but "quality and evaluation." He argues that we have not yet established evaluation methods to determine not just "collecting" large amounts of text, but the accuracy of the information contained within it and whether it directly leads to improvements in model performance.

He also states, "No one has good science about the incremental value of a single token of data," and expresses the view that efficient learning through "high-quality synthetic data" and "expert feedback" will be the key in the future.

3-2. Separating advanced reasoning from "facts"

If we only increase the amount of knowledge through data accumulation, AI will only become a "giant dictionary" or a "vast database." Kevin emphasizes that "the essential value of a model is not in memorizing data, but in its ability to reason."
This is a clear difference from search or databases, suggesting the possibility that AI models will function as "thinking engines" that perform complex tasks rather than just information indices.

4. Open source vs. closed: A future where both coexist

4-1. The Evolution and Convenience of Open Source

In the recent large language model scene, the open-sourcing of large models, starting with the LLaMA series released by Meta and others, has become a major topic. However, Kevin points out, using search engines as an example, that “even if there are open-source search projects, it is the search engines equipped with massive infrastructure (Google and Bing) that are creating actual market value.”

Ultimately, he expresses the view that it is important that “open source exists as an option,” and that it is not the case that everything will become open source, nor that closed systems alone will hold hegemony.

4-2. Differentiation at the Product Layer

The number of companies and individuals who can build their own infrastructure using open-source large language models is increasing. However, the “process of translating this into a product used by people” still requires specialized knowledge and large-scale investment.
Kevin envisions a future where not only the binary opposition of “open source vs. closed” exists, but diverse forms coexist depending on the delivery method and use case.

5. The Era of Agents: Beyond Conversational UI

5-1. The Strengths and Limitations of Conversational UI (Chat)

Since the advent of ChatGPT, conversational UI has tended to be expected to become the new gateway to computing. However, Kevin states, “I don’t believe in one agent for everything,” and envisions a future where multiple agents exist tailored to the user’s tasks and domains.

On the other hand, since conversational UI itself has significant advantages such as high versatility and low learning costs, the trend of “various services introducing chat interfaces” will accelerate in the near future.

5-2. User Experience “Memory” and Asynchronicity

What Kevin emphasizes is that current chat-based agents are “transaction-based (one-off interactions).” He points out that “memory functions” and “mechanisms for delegating tasks asynchronously” will advance from here on.

The image is of an agent that learns the user’s preferences and past interactions and carries out tasks in the background. This has the potential to create a new UI/UX that goes beyond the mere framework of chat.

6. How Will Software Development Change?

6-1. The Future Where “95% of Code is AI-Generated”

Kevin boldly predicts, “95% of net new code will be AI-generated.” Developers who already use code completion in programming support tools are unanimous in saying they “can no longer let go of it,” and the evolution of automatic code generation will further accelerate this trend.

6-2. The Role of Engineers and the Expertise of Product Managers

Even if AI comes to write the majority of code, it does not mean that engineers will become unnecessary. Rather, the “ability to design what you want to build at a high level of abstraction” and the “ability to debug and verify code generated by AI” will be in greater demand.

Furthermore, Kevin emphasizes the importance of “domain experts as product managers.” His outlook is that rather than a single general-purpose agent, there will be a need for talent who can link deep expertise in specific fields, such as medicine or investment, with AI models to design superior user experiences.

7. Are We Underestimating China’s AI Technical Capabilities?

7-1. Evaluation of Deepseek

Regarding "Deepseek R1," which became a hot topic as a large language model from China, Kevin commented, "We've had models more interesting than Deepseek R1 that we didn’t even launch," drawing attention to its technical maturity and the impact of its release method.

However, from the perspective of the US development community, it seems there was surprise at the reaction itself, specifically "how surprised everyone seemed to be like, oh my God, this is coming from China."

7-2. "We should not underestimate China"

Kevin emphasized, "We should really respect the capability of Chinese entrepreneurs, scientists, and engineers," sounding an alarm that we must not underestimate China's AI technical prowess.
This global structure of competition and cooperation will likely be a factor that further accelerates the development of AI.

8. Leadership and organizational theory at Microsoft

8-1. Satya Nadella's leadership

Microsoft CEO Satya Nadella is known as a model of leadership. Kevin describes Nadella's characteristics as "the ability to always energize the team while simultaneously pointing out a clear direction."

He says these two axes—creating energy and clarifying priorities—work effectively in driving organizational transformation.

8-2. Fighting against tech debt

A problem that constantly troubles Kevin as he leads engineering teams is "tech debt." If code is implemented with a priority on release, the possibility of rework later increases. He states, "Tech debt is like financial debt," and expresses hope, saying, "We can turn the zero-sum problem of tech debt accumulation into something non-zero-sum with AI."
Microsoft Research is advancing a large-scale project to "resolve tech debt with AI," aiming to build an efficient and sustainable software development process.

9. The future envisioned by Kevin Scott

9-1. Applications in education and healthcare

Kevin cites "healthcare" and "education" as the fields where AI social implementation should be addressed first. The "diagnostic support" function of large language models is particularly promising for resolving medical disparities, and he even mentions, "The frontier models are probably better diagnosticians than your average GP."

At the same time, he argues that it is important to create an educational environment where children around the world recognize that "AI is a tool for them" and can use it themselves.

9-2. "Positive transformation" accelerated by AI

Many people tend to focus on the disruptive impact of AI, but Kevin argues that "essentially, it is a technology for bringing about positive transformation in society." Of course, there are various practical challenges such as costs, infrastructure constraints, and regulations, but he believes it is well worth exploring ways to overcome them.

As these words indicate, the attitude of continuing to pursue possibilities is Kevin's philosophy.

10. Conclusion

10-1. The Next-Generation Paradigm Brought by Large Language Models

Synthesizing Kevin Scott's views, the impact brought by large language models and agent technology is set to significantly rewrite the common sense of existing application development and data utilization. In particular, the way data and inference are handled will change fundamentally, and the barrier to entry for utilizing AI will drop dramatically. On the other hand, there are still many issues to be solved, such as technical debt, infrastructure construction, and data quality assessment.

10-2. A Future Where Agents and Humans Co-create

While chat-based agents are a powerful UI at present, they will likely evolve into agents that can continuously learn from users and execute complex tasks asynchronously in the future. In that landscape, a collection of domain-specific specialized agents is expected to play a more important role than a single, all-purpose AI.
By multiplying human expertise (domain knowledge) with the vast inference capabilities of AI, it is expected that new industries and business models will be born one after another. It can truly be said that the “best time for everyone with an entrepreneurial spirit” that Kevin emphasizes has arrived.

Related Articles