Don't get involved in the large model business anymore, unless you're Elon Musk? A 90-minute in-depth interview with the chairman of OpenAI.

Is the chairman of OpenAI's board actually advising people to stop developing large models?

According to a report from Zhidongxi on August 2nd, recently, the well - known overseas tech blog Lenny's Podcast released a high - information - content conversation with Bret Taylor, the chairman of OpenAI's board and the founder of the startup Sierra, an Agent company. In this 90 - minute conversation with a transcript of nearly 20,000 words, Bret Taylor shared in depth his insights on the future development pattern of the AI industry and high - value entrepreneurial directions.

Before officially reading the interview content, it's necessary for us to first understand Bret Taylor's legendary career.

In 2003, after graduating with a master's degree from Stanford University, he became Google's youngest product manager and created Google Maps, which received tens of millions of visitors on its first day of launch, completely reshaping the local life industry.

After leaving Google in 2007, Bret Taylor founded the social media company FriendFeed and invented core elements in modern social media such as the Newsfeed and the like button. He eventually sold FriendFeed to Facebook for nearly $50 million and joined Facebook as CTO.

In 2012, Bret Taylor left Facebook and founded the collaborative document enterprise Quip. He finally sold Quip to Salesforce for a sky - high price of $750 million, and he joined Salesforce and eventually served as co - CEO.

At the beginning of 2023, when generative AI technology was just emerging, Bret Taylor resolutely left Salesforce, founded the startup Sierra, an Agent company. In the same year, he also joined OpenAI and served as the chairman of the board.

It can be said that in Bret Taylor's career over the past 20 years, every step has been precisely on the cutting - edge. His successful experiences spanning all different levels and positions such as CEO, CTO, COO, CPO, product manager, engineer, and board member are still of reference significance to all entrepreneurs in the AI era.

The core information of this interview is as follows:

1. The AI market will have three major segments - models, tools, and applications. There is no future for new startups in the model market, unless they can raise billions of dollars like Elon Musk. The tool market is facing the impact of leading model companies, and the value will be concentratedly released at the application layer.

2. Google Maps evolved from the failed product Google Yellow Pages. This shows that when new technologies emerge, instead of directly digitizing past experiences, it's better to create a brand - new experience and answer the new customers' question of 'Why should I use it?'

3. Agents are the applications of the new era and will be one of the main forms of AI products. Due to their highly automated and measurable results, Agents enable enterprises to directly see the productivity improvement brought by AI, thus promoting purchases. Their model is similar to that of SaaS companies, with higher profit margins.

4. AI should not be billed by tokens but should be billed by results. In the result - oriented billing model, the goals of both the supply and demand sides are unified: enterprises hope to get satisfactory work results from AI products, and AI companies can only make money by creating products that can deliver good results.

5. As AI replaces human programming, the cost of writing code continues to decrease. Languages like Python, which are designed to make human programming easier, are actually not suitable for the characteristics of AI. We urgently need to reconstruct a programming system for AI that is suitable for building complex systems and is easy to adjust flexibly.

6. Due to insufficient product capabilities, existing AI programming tools often backfire in production scenarios. Most of these problems are due to the lack of sufficient context in the models, which can be solved by MCP.

7. When developing AI applications, one cannot just wait for the underlying models to improve on their own. The improvement of models will eventually happen, but if you want to realize the potential of AI in advance, you have to rely on the engineering design in the application. This is also the fundamental reason for the existence of application - oriented AI companies.

8. There are three models for the marketization of AI products - developer - led, product - driven, and traditional direct sales. Although the first two models are popular among entrepreneurs, it is difficult to scale them up, or their scope of application is limited. Many AI companies choose the wrong strategy and should consider more direct sales.

9. Learning computer science is not equal to writing code. Programming may be replaced by AI, but understanding the principles and mastering systematic thinking are still the foundation of software engineering in the AI era.

10. Entrepreneurs need to be flexible about their identities, be willing to change constantly, and be able to accurately judge: What work should I do now to have the greatest impact?

11. Encourage children to integrate AI into their lives. ChatGPT can promote educational equality, but it will also lead to a polarization between students with strong and weak subjective initiative.

The following is the complete compilation of Bret Taylor's in - depth interview (to improve readability, Zhidongxi adjusted the order of some Q&As and made some additions, deletions, and modifications without violating the original meaning):

01.

The AI market will have three major segments

The value will be concentratedly released at the application layer

Host: Let's start by talking about the business strategy of AI. One of the most concerned questions for many entrepreneurs now is: What should I do? Will the foundation model companies directly swallow up what I'm doing? On one hand, you're starting an AI business, and on the other hand, you're a member of OpenAI's board. How do you think the AI market will evolve? Where should entrepreneurs focus their energy, and which directions should they avoid?

Bret Taylor: I think the AI market can be divided into three main segments, and they will all eventually be quite significant markets. Let me share my overall judgment.

The first type is the 'leading model market' or the 'foundation model market'.

I think this market will ultimately be controlled by only a few super - large companies and large - scale laboratories, just like the current cloud infrastructure service market. The reason is simple. Building a leading model is a capital - intensive task. You must have huge capital expenditure capabilities to train these models.

All startups that have attempted to do this have either been acquired or are about to be acquired, such as Inflection, Adept, Character AI, etc. I don't think there is a viable business model for startups in this market at present because the capital investment required for model training is too large, and startups don't have enough financing space to reach the 'escape velocity'. At the same time, the value of foundation models as assets decays very quickly, so you must have a scale advantage to get a reasonable return from them.

I think almost no entrepreneurs should build foundation models.

Host: Unless you're Elon Musk.

Bret Taylor: Yes, Elon Musk is really different. He has the ability to raise billions of dollars in capital. I guess most of your listeners don't have this ability. And he is one of the greatest entrepreneurs of this era. He is different. Don't compare yourself with him.

Another part of the market is the 'tool - layer market'.

Many people are selling shovels in the 'gold rush', including data annotation services, data platforms, evaluation tools, and some specialized models. For example, ElevenLabs' voice model is very excellent and is used by many companies. The core question in this market is: What tools and services are needed to succeed in the AI era?

But there are also risks in this market. It's too close to the'sun'. Just like in the cloud service market, various cloud platform providers (AWS, Azure, etc.) are also moving upstream and providing various tools. Companies close to the infrastructure layer are easily replaced directly by platform providers.

Of course, there are truly valuable companies, such as Snowflake, Databricks, Confluent, etc. But there are also many companies that have been eliminated because infrastructure platforms have launched similar functions. For such companies, the biggest risk is: What if a foundation model company suddenly launches the tools provided by startups at a developer conference?

There may be many people who need your tools, but the problem is, if a foundation model company launches similar tools, why would people choose yours? The tool - layer market has great potential, but it's also quite dangerous.

The third part is the 'application - oriented AI market'.

I think this part will be dominated by companies that build Agents. Agents are the applications of the new era, and this will be one of the main forms of AI products. My company, Sierra, helps enterprises build Agents that can answer calls and handle customer service chats to improve customer experience and service efficiency.

There are also companies like Harvey, which build Agents for the legal and quasi - legal fields, such as anti - monopoly reviews, contract reviews, etc. There are also companies in content marketing and those specializing in supply - chain analysis.

I think these companies are more like the 'Software as a Service (SaaS)' model, and their profit margins may be higher because the products they sell are business results, not by - products of models. Of course, they also have to 'pay taxes' to foundation model companies and pay for using the underlying models. This is why these model providers will eventually develop into extremely large - scale companies, but their profit margins may be slightly lower.

I think the application market will become less technical. If you think about the purest form of Software as a Service, people don't ask what database you're using, but care about your functions and features. I think the future of Agents will be the same. As time goes by, products will be more important than technology itself. Just like when Salesforce was founded in 1998, running a database in the cloud was a technological achievement. Now, no one will ask you this question because you can just start a database on AWS or Azure. It's effortless.

I think today, organizing an Agent process on top of a model sounds very complex, advanced, and difficult, but I'm sure that in three to four years, with technological progress, all of this will become easy. Gradually, when people talk about an Agent company, it will look a bit like a SaaS company. You'll talk less about how to handle models, just like in modern SaaS products, few people will ask what database you're using. Instead, you'll be asked more about your work processes and what business results you bring. Are you creating potential customers for the sales team? Or are you reducing procurement costs? No matter what value you provide, this direction will gradually develop.

I'm very excited about this. I don't think startups should build foundation models. You can certainly try. If you have a vision for the future, then go for it, but I think this is a market that has become concentrated and highly competitive. I'm very interested in the other two markets, especially as it becomes easier to build Agents, we'll see a large number of 'long - tail' Agent companies emerging.

I recently browsed a website listing the top 50 software companies by market value. Undoubtedly, the top five are giants like Microsoft, Amazon, and Google, but the next 50 are all SaaS companies. Some of these companies are exciting, and some are very boring, but this is the development trajectory of the software market. I think we'll see a similar trend in the Agent market. It won't just cover a few huge markets like customer service or software engineering, but will also cover many areas where people currently invest a lot of time and resources, and these areas can be completely solved by an Agent. But this requires entrepreneurs to really understand a certain business problem deeply. I think the great value of the AI market will be released in this area.

02.

Why bet all on Agents?

Process automation and measurable results

Host: This is very inspiring. It reminds me of an interview I had with Mark Benioff on my podcast. You were the co - CEOs of Salesforce together. He is extremely obsessed with Agents and talked about Agent Force throughout the interview. Obviously, you've seen some key trends and think you must fully invest in Agents. This is the future direction. So, what do you think people are ignoring? Why is this the key to a major transformation in the way software works?

Bret Taylor: If you talk to economists, such as Larry Summers (who is on the OpenAI board with me), they'll look at the value of technology from an economic perspective: technology drives the productivity growth of the economy. Looking back at history, a significant leap in productivity occurred in the 1990s. Many people I've talked to believe that this leap was the first wave of the informatization era - enterprises started using ERP systems and moved accounting and other content into computers, databases, and even mainframe systems. We're not talking about the PC era here, but the earlier enterprise informatization. That change was truly revolutionary. Imagine how much manpower was needed to complete the accounts of a multinational enterprise before computerization. Informatization completely changed the entire department.

Let me give an example to illustrate this productivity improvement. My father just retired. He is a mechanical engineer. He said that when he entered the engineering industry in the late 1970s, most people in the company were draftsmen. That is, you had a design plan, but you needed to draw views for each floor and each angle and then provide them to contractors.

▲The technology center of GM, showing engineering design in the pre - CAD era (Source: Rare Historical Photo)

Now, there is no longer a draftsman position in his company. Design drawings are directly completed on AutoCAD, and now even 3D drawings are directly generated using Revit. The drawing process has completely disappeared. This work no longer needs to be done. The division between design and drawing no longer exists, only design itself remains. This is a real productivity improvement. The responsibility of a mechanical engineering company is to do design. Drawing is just an intermediate output for delivery to the construction side. It doesn't add actual value itself, but is just a part of the supply chain.

Looking back at the history of the software industry from the PC era, we have indeed achieved productivity improvement, but it's far less significant than the initial leap. I'm not sure why, but it's interesting. The productivity improvement brought by technology has not been fully realized as many people expected.

I think Agents will pull this curve up again, just like in the early days of computers. Because software is moving from helping individuals improve efficiency to autonomously completing complete tasks. So, just like there is no longer a need for draftsmen in mechanical engineering companies, in the future, there will be no need for people to do certain jobs manually. This means these people can shift to more productive work. A small number of people can do more things, driving the overall productivity growth of the economy. The sales of enterprise software will involve discussions about 'value' with customers. You need to design a somewhat complex logic. For example, if you're selling a sales - related software, if each salesperson's performance improves by 5%, then how much more money your company should earn, so you should pay us one million dollars. Such discussions are often difficult to accurately attribute.

This is also why it's so difficult to sell productivity software. I learned this from personal experience. It's difficult to evaluate how much value it is to improve everyone's efficiency by 10%. Did you really improve their efficiency by 10%, or are other factors at play? You're actually not sure. But now, when an Agent really completes a task, it not only drives productivity improvement in a very real way, but also this improvement is measurable. These factors combined make me think this is a leap - forward change in the way we view software because it can complete work autonomously, and this autonomy itself more directly reflects its promotion