Author: Lincoln Wang | Founder of MindsLeap | Global Partner at Founders Space | Founder of Founders AI Club
"It Is Less Like Building a Skyscraper and More Like Growing a Plant"
Marc Benioff asked a question almost every entrepreneur is thinking about: what has surprised you most over the last four and a half years?
Dario Amodei did not answer with a product launch, a funding round, or a single technical breakthrough.
He gave a metaphor.
Traditional software, he said, is like building a skyscraper. You draw the blueprint. You design the structure. You write every line of code. AI systems are different. They are closer to biology, more organic. You are "growing" these models. You set a high-level recipe, like growing a plant or baking a cake.
It sounds casual, but the sentence carries a lot of weight.
It means adopting AI is fundamentally different from adopting ERP, CRM, or any traditional enterprise software. You are not buying a deterministic tool. You are introducing an uncertain, evolving system into the company. That combination of uncertainty and massive economic impact is why Amodei treats safety and responsibility not as public relations language, but as core enterprise AI infrastructure.
From Human Metabolism Maps to Deep Neural Networks
Amodei did not begin as a software engineer. He studied neuroscience and later worked as a computational biology postdoc at Stanford.
He was trying to understand cancer and human metabolism. At one point, looking at a map of human metabolic pathways, he realized that the complexity of biology had exceeded the scale that humans could comfortably reason through.
Just as that problem began to feel almost hopeless, early deep neural network papers from Google and the University of Toronto appeared. He reproduced some of the results himself and had a simple reaction: this thing actually works.
So he moved into AI.
The value of this story is not merely that a scientist changed fields. It reveals a pattern. When the complexity of a domain exceeds what individual humans can process, a new kind of tool becomes necessary.
Today, enterprise management, customer service, supply chain coordination, financial compliance, and many other operating systems inside companies are moving toward that same level of complexity.
This is not just a metaphor. It is part of the logic behind Amodei's belief that AI will reshape the economy.
A Natural Experiment in Two Paths
Benioff asked another sharp question: after you left OpenAI, what surprised you most?
Amodei's answer was unusually direct.
Over time, he said, you can see paths diverge through the people you hire and the choices a company makes. OpenAI moved toward consumers. Anthropic moved toward enterprises. In his words, it became a kind of natural experiment: two companies starting from a similar origin and taking very different approaches.
That is an underrated observation.
Most people focus on model capability, parameter counts, and benchmark scores. But the industry may be shaped just as much by this strategic split between two companies with a shared origin. OpenAI chose consumer products, broad adoption, and fast iteration. Anthropic chose enterprise deployment, trust, and regulated industries.
This is not about which path is correct. It is about two very different organizational capabilities.
A consumer path requires growth instincts, product taste, and fast experimentation. An enterprise path requires trust, compliance, integration, and patience. The teams, rhythms, and performance systems needed for each path are not the same.
For Chinese entrepreneurs, the value of this natural experiment is clear: before choosing an AI strategy, you need to know what you are really building around - traffic, or customer relationships.
Sibling Co-Founders and the Cost of Switching Modes
Benioff specifically mentioned Amodei's sister, Daniela Amodei. She and Dario are two of Anthropic's seven co-founders.
Their division of work is revealing. Daniela focuses on company operations: organizational structure, process design, and business architecture. Dario focuses on strategy: research direction, safety philosophy, and company vision.
Amodei made a point that applies to almost every founding team: both people can do both kinds of work, but switching between operating mode and vision mode is hard. Separating them makes both people more effective.
This is not simply one person handling the inside and one handling the outside. It is a sober view of cognitive load.
Operations are concrete. Today's problems need to be solved today. Processes need to be clear. Metrics need to be measurable.
Strategy is uncertain. You are looking several years ahead. You are dealing with ambiguity. You are making choices that may be hard to reverse.
When one person tries to do both at the same time, the risk is obvious: using operational certainty to handle strategic uncertainty, or using strategic ambiguity to weaken operational discipline.
Anthropic's growth has not just been the story of a single heroic founder. It is also the story of a deliberately designed organizational split.
Agents Are Not Just Toys for Programmers
When Amodei was asked what excited him most about the next stage of models, his answer was agentic capability.
The goal is to let models complete tasks end to end. This is most visible in coding, but coding is only the leading indicator. In financial services, healthcare, insurance, manufacturing, and other regulated industries, agents can be valuable in the same way.
He mentioned sales and customer management workflows: connecting multiple steps, pulling in context, learning information, and executing actions.
This is where the word "agent" is often misunderstood. An agent is not simply a magical AI assistant that can do anything. In business terms, an agent is the ability to complete a defined workflow from end to end.
Programmers are early adopters because code is structured and verifiable. But the larger commercial value may sit inside unstructured, cross-system enterprise processes that require judgment.
For traditional companies, the starting question should not be "what can AI do?" It should be: which workflow is painful, cross-system, and in need of end-to-end orchestration?
The Enterprise Itself Is a Superintelligence
One of the most imaginative lines in the conversation came when Amodei reframed what a company actually is.
In a sense, he said, an enterprise is already a superintelligence. It acts in the world. It has strategy, knowledge, and influence far beyond any individual human. Even if models become smarter than any one person, there is still a higher ceiling above them: what enterprises are trying to do in the world.
That changes the relationship between AI and the enterprise.
AI is not simply here to replace the company. AI is an engine that can make this existing "superintelligence" operate faster. The more complex the company, the deeper its strategy, and the more global its resource coordination, the higher the ceiling for AI becomes.
This is why Amodei's big ambition is to put AI at the center of every enterprise. Not because AI can do everything, but because the enterprise itself is an almost bottomless demand pool.
Data Center Mania and What Actually Matters
Benioff also asked about the data center buildout. Amodei's answer is worth reading carefully.
He said he worries a little that the media has become too focused on data center construction and data center deals. Some deals, he suggested, may look questionable or involve double counting. But then he pulled the conversation back to something more basic: all of those data centers are spending. The most important thing is earning.
If demand exists and customers are willing to pay, then you can pay your compute suppliers and eventually get the compute you need.
This is a very pragmatic view.
While everyone watches who signed the largest data center agreement, who bought the most GPUs, and who invested in the most energy infrastructure, Amodei brings the question back to revenue: where are the customers?
Data centers sit on the cost side. Revenue sits on the income side. A dramatic cost-side story cannot replace demand-side validation.
For enterprise decision makers exhausted by AI infrastructure headlines, this is a useful correction.
Trust Is the Moat
Near the end, Amodei returned again and again to one word: trustworthy.
Anthropic wants to be known as a company that is dependable, responsible, and able to do what it says it will do. If a company builds on its products, it should not encounter strange surprises or models suddenly saying something wildly inappropriate.
This is not just brand positioning. It is the real moat for enterprise AI.
Consumers can tolerate occasional nonsense from an AI system. Enterprises cannot. One unpredictable output can become a compliance risk, a customer problem, or a brand crisis.
Anthropic has spent four and a half years turning "trustworthy" from a value statement into a core part of its business model. That is much harder to copy than any single technical metric.
There is a clear line running through Amodei's story: from biological complexity, to the uncertainty of growing models, to the enterprise as a higher-order intelligence, to trust as the final moat.
This is not just a technical roadmap. It is a business argument about how AI enters the operating system of the enterprise.
For Chinese companies, several coordinates on this map are worth marking. Choosing a consumer path or an enterprise path is really a choice between two organizational capabilities. The value of agents is not only in code, but in workflows. Infrastructure narratives cannot replace revenue validation. Trust is not a slogan. It is a product specification.
Growing a plant does not require only stronger execution. It requires a better recipe, more patience, and clearer judgment.
Source Note
This article is Lincoln's interpretation of the Salesforce Events video Dario Amodei & Marc Benioff: Future of AI, published on May 12, 2026.
About MindsLeap
MindsLeap is the China partner of Founders Space, a leading Silicon Valley incubator. We connect global frontier innovation with the real transformation needs of Chinese entrepreneurs and enterprises. Through AI strategy, founder communities, innovation study tours, and executive training, MindsLeap helps organizations build stronger cognition, methods, and execution capabilities for the AI era.
This article was translated and adapted from the Chinese original with AI assistance.