Seven Important Judgments by Claude Code's Founder at the Sequoia Conference

Author: AI Product Aying

Original Video: 《Anthropic's Boris Cherny: Why Coding Is Solved, and What Comes Next》

Boris Cherny, founder of Claude Code, shared a wealth of information at the Sequoia Conference, with many viewpoints I heard for the first time. This guy really has a solid understanding of AI.

I’ll share my own summary.

#01 Code is No Longer Scarce

For many mainstream development scenarios, writing code by hand has begun to become an inefficient task.

In the past, to deliver a feature, an engineer would sit down, think through how to implement it, and then type out the code line by line. In this process, the engineer's greatest value was: whether they could write, how well they could write, and how quickly they could write.

Now, the working method is different.

For the same feature, what engineers do now is more like: first clarify the requirements, break the task into pieces to assign to an Agent, set an acceptance standard, and then check if the results produced by the Agent are correct; if not, adjust the prompts and let it run again.

AI can already handle most coding tasks. Of course, it’s not 100%, as there are still many large and complex codebases, niche languages, or special environments where today’s models still fall short.

Overall, the value of engineers has shifted from knowing how to write code to knowing how to break down tasks, articulate goals clearly, assess results, and manage Agents.

This change is very similar to the Industrial Revolution.

Before the Industrial Revolution, a blacksmith did everything from forging, shaping, polishing to assembling all by themselves. A skilled blacksmith was naturally valuable.

Then the assembly line appeared. Each worker was responsible for one process, but the overall output was dozens or even hundreds of times higher than in the handmade era.

At this point, the valuable role in the factory was no longer the craftsman who excelled at a specific process, but the one who could design, manage, and streamline the assembly line.

Workers did not disappear, but their roles changed.

Software engineering is now experiencing a similar shift. Code itself is no longer a scarce resource. Knowing how to write code is becoming a basic skill, much like knowing how to use PowerPoint.

What is truly scarce is the ability to break down vague requirements into clear tasks, to choose the best option from several provided by the Agent, and to coordinate a group of AIs to accomplish a task.

Many older engineers initially find this hard to accept. The act of writing code by hand has been the reason many people have loved this field for decades.

Handing this over to machines represents not just a change in work methods for many, but a reshaping of identity.

But trends are trends.

#02 Like the Gutenberg Printing Press

Coding is transitioning from a specialized skill to a basic ability. This can be likened to the invention of the printing press in 15th century Europe.

Before the printing press, only about 10% of people in Europe were literate. These individuals were often employed by illiterate nobles to read and write for them.

Then the printing press was invented. In just 50 years, the number of books published in Europe exceeded the total from the previous thousand years, and book prices dropped by about 100 times. It took several hundred years for supporting systems (education systems, economic structures) to catch up, leading to a global literacy rate of 70% today.

Boris believes that AI's impact on software is an accelerated version of the printing press revolution. Software will be completely democratized within decades, becoming something anyone can master.

Ultimately, being able to create software will become as natural as sending a text message.

#03 What Abilities Are Most Important?

When the barrier to writing code is lowered to an extremely low level by AI, what truly distinguishes a person's abilities is their product sense and genuine understanding of a specific field.

For example, two people are tasked with creating a product for doctors. One is a fast-coding engineer, and the other has worked in a hospital's information department for several years.

In the past, the engineer was more likely to produce a viable product because they could implement the idea.

Now it has reversed. Anyone can implement an idea. At this point, the person who truly understands the daily workflow of the hospital becomes more valuable. They know which features doctors will actually use and which ones only sound reasonable.

In other words, once AI levels the execution barrier, the gap in judgment becomes magnified.

This directly rewrites the meaning of the term generalist.

In the past, when we referred to a generalist, we typically meant an engineer who could write iOS, Web, and backend code. This type of generalist was essentially a full-stack engineer within the engineering domain.

The future generalist will be a cross-disciplinary full-stack.

Some will understand product, design, and engineering simultaneously. Others will understand product, data science, and engineering. Such combinations were nearly impossible in the past because each area required long periods of specialized training.

But now that AI has lowered the execution barriers in each area, a person can span several fields while still maintaining depth of expertise.

The Claude Code team exemplifies this. Engineering managers, PMs, designers, data scientists, finance, and user researchers are all writing code.

Designers can run interaction prototypes themselves for the team to see, no longer just providing designs for engineers to implement.

Finance can build an analysis tool to run complex financial models without waiting in line for BI. User research colleagues have started running their own data, taking over tasks that previously required collaboration with the data team.

Everyone's depth of expertise remains intact. But with AI assistance, writing code has become a shared language for everyone.

#04 The Moat of SaaS is Crumbling

In the past decade, there have been several almost axiomatic consensus points in the SaaS industry.

The first is the switching cost. Once a company uses your system, it accumulates years or even decades of data, configurations, fields, and permission relationships.

Moving to another system means that just migrating these things as they are can be enough to make people reluctant to act.

The second is workflow lock-in. Employees' daily operations, cross-department collaboration, and approval nodes all grow around this SaaS.

Switching systems is not just about moving data; it means tearing down the muscle memory built up over the past few years in the entire company.

These two factors combined formed the deepest moat in the SaaS industry. But with sufficiently strong models, the logic of the situation begins to change.

First, looking at the switching cost side. In the past, switching from one SaaS to another meant that just aligning fields and replicating data structures could keep engineering teams working overtime for several months.

Now, you can directly hand over the interfaces and data structures from both sides to the model, letting it clarify the mapping relationships itself, gradually climbing toward the optimal solution. What used to take months can potentially yield a usable version in just a few days.

Now, looking at the workflow lock-in side, it gets even more interesting. In the past, workflows could lock in customers because these processes were complex, implicit, and reliant on people.

The tacit understanding of who needs to approve what and when in employees' minds cannot be directly transferred.

However, models like Opus 4.7 excel at understanding, breaking down, and reconstructing complex processes in a new environment. The newly reconstructed version may even run smoother than the original.

Thus, the moat built on data accumulation and process entrenchment is crumbling.

For those in the SaaS business, this may be bad news. But for all customers using SaaS and teams preparing to create the next generation of SaaS, this is a genuine opportunity window.

#05 The Best Era for Entrepreneurs

In the next 10 years, the startups that truly disrupt industries may outnumber those of the past decade by tenfold.

The reason is not complicated.

Small teams can use AI to create products that are on par with or even better than those of large companies. Conversely, large companies wanting to truly leverage AI find themselves at a disadvantage.

How so?

A company with a history of over a decade has developed a complete set of business processes, role divisions, collaboration habits, training systems, and KPI assessments. These things were assets and barriers in the past.

But to truly embed AI means that all of this must be re-evaluated: business processes need to be restructured, all employees need retraining, and every step forward will encounter significant internal resistance, requiring coordination among multiple departments and layers of approval.

In contrast, a startup team of three can treat AI as a default foundation from day one. They have no historical baggage to dismantle, no habits to change, and no one to persuade. They can clarify discussions today, produce a demo tomorrow, and launch it for users the day after.

This speed difference existed even before AI. Startups already had a speed advantage over large companies. But AI has amplified this gap many times over.

Why?

Because the stronger the AI, the greater the leverage one person can exert in a unit of time. A small team that truly knows how to leverage AI can produce outputs equivalent to what ten people could do in the past today, and potentially what thirty people could do tomorrow.

However, the organizational weight of large companies has not lightened; in fact, it has become heavier due to the need to digest AI. The stronger the AI, the greater the gap between the acceleration of small teams and the drag of large companies.

This is what Boris refers to as negative assets. It’s not that large companies lack money, people, or willingness; rather, the muscle that used to generate profits is now stuck in the way of AI truly realizing its value.

#06 MCP Will Not Die

MCP will not die.

After Skill became popular, many people felt that MCP was no longer needed. The founder of OpenClaw shares a similar viewpoint.

But Boris disagrees. He believes that MCP will become the software connection layer in the AI era.

In the past, the way software connected on the internet was through APIs.

But the core issue with APIs is that they are designed for engineers. To use an API, one must first read the documentation, request a token, write code, align fields, and handle exceptions. In short, APIs are written for human developers.

MCP is different. It allows models to connect directly and use them; the model can understand and adjust without needing a programmer to translate.

Thus, Boris refers to APIs as Human Developer Interfaces and MCP as Model Interface Protocols. One is for human use, and the other is for models.

This is quite similar to the past. In the mobile internet era, it was assumed that all services needed to be API-based. In the AI era, it is assumed that all services need to be MCP-based.

#07 Computer Use Remains Important

Many people now discuss Computer Use and feel that this direction may not work.

The reasoning is quite reasonable: it consumes too many tokens, runs slowly, and is unstable. It seems more like a flashy demo, still far from being truly usable.

But Boris sees it from a completely different perspective.

What he truly values is that Computer Use addresses the biggest pain point in AI implementation: in the real world, there are numerous systems that have neither APIs nor MCPs.

Especially in the corporate world.

Anyone who has worked in a company knows that many core systems are quite old. ERP, OA, financial systems, internal approvals, supply chain backends, various custom systems. Many do not have open interfaces, documentation, or automation capabilities. They are just there, manually operated by countless employees every day.

So why not just create APIs for them?

Because it’s not feasible. The vendors that developed these systems may no longer exist. IT departments lack the motivation and budget to restructure.

Business departments are even less likely to stop and wait for six months to a year. These systems will never wait for a perfect API to save them.

In the short term, major models should still focus on enhancing their Computer Use capabilities.