Written by: Deep Thought Circle

Have you ever thought that "writing code" could one day become as ordinary as "sending a text message"? Not a metaphor, but a literal idea. Recently, I watched an interview featuring Boris Cherny from Anthropic at the AI Ascent 2026 conference hosted by Sequoia Capital. Boris is the creator of Claude Code and one of the forefront figures in AI programming. He shared many things in the interview that caused me to pause and think repeatedly.

He stated that programming has been solved. He mentioned that by 2026, he had not written a single line of code by hand. He now sends dozens of PRs (Pull Requests) daily from his phone. He even said that one day he set a record by sending 150 PRs in a single day.

When I first heard these words, it felt a bit like reading science fiction. But the more I pondered, the more I felt this isn't just the future; it's reality already unfolding—though most people are yet to realize it.

How Claude Code Came to Be

Boris mentioned that Claude Code was actually born out of "accident." At the end of 2024, he joined a small incubation team within Anthropic called Anthropic Labs, with the goal of transforming the potential of the models into real, usable products. They created various products, including Claude Code, MCP (Model Context Protocol), and a desktop application for Claude. After completing their mission, the team disbanded. Now they are reuniting for a second round.

He said they identified a "product overhang"—the model's capabilities far exceeded what any product on the market could achieve at the time. At the end of 2024, the most advanced product in programming was type-ahead, essentially where you open an IDE, hit Tab, and complete one line of code at a time. This was what Claude Sonnet 3.5 could do at the time.

But what Boris and his team saw was that the model could do much more. No need for line-by-line completion; they could let the AI agent finish all the code directly. So that's what he set out to do.

However, he also confessed that Claude Code was not very useful at first. For the first six months, he only used it to write about 10% of his code. After its release, there was no explosive growth. The real turning point came with the release of Opus 4, after which every model version led to a leap: 4.5, 4.6, 4.7, all the way up.

I think this history is particularly worth reflecting on. Boris and his team started development before the product had PMF (Product-Market Fit), knowing full well that there wouldn’t be PMF within six months because they were preparing for the next generation model. This is a very unique product mindset: they focused not on the current market but on the market that would emerge after the model evolved. It requires strong judgment and great patience, which most startups give up at this stage.

"Programming has been solved," what does this mean?

This was the most shocking assertion for me in the entire interview. Boris said that for him personally, programming has been solved. His code is 100% written by the model; he no longer handwrites a single line of code.

He explained why Claude Code uses TypeScript and React to write its own codebase. The reason is simple: these two are the most widely distributed languages and frameworks in the training data. At the time, the models weren’t as intelligent, and using an "on distribution" tech stack allowed for more stable performance from the model. Today, the model can learn languages and frameworks it hasn’t seen before, but at that time, selecting the right technology stack was a key variable influencing the quality of AI output.

He also added that the statement "programming is solved" does not apply to everyone. For very large codebases, obscure languages, and specialized historical systems, the models still cannot fully handle these cases. But his judgment is: usually just wait for the next model and it will be fine.

My understanding of this statement is as follows: when Boris says "programming has been solved," he is referring specifically to the action of "writing code." It's like saying "typing has been solved"—not that writing has lost its value, but that the physical action of typing is no longer a bottleneck. The technical barrier to writing code is collapsing at a visibly rapid pace. The remaining core issues now become what you want to build, whether you understand the domain, and if you have the judgment to know whether what the AI produces is correct.

Boris's Work Method Redefined "Work" for Me

He shared his current work methods. Most of the time, he works on his phone. He opens the Claude app, where there is a code tab on the left with five to ten sessions running simultaneously, with multiple agents in each session, totaling about several hundred. Every night, there are also thousands of agents doing deeper work.

One feature he has particularly enjoyed using recently is called `/loop`. The principle is simple: it allows Claude to create a scheduled loop task using cron that can run every minute, every five minutes, or once a day. He now has dozens of these loops running continuously. For example, one monitors his PR, automatically fixing CI (Continuous Integration) issues, and automatically rebasing; another ensures CI is healthy, automatically fixing flaky tests; and another grabs user feedback from Twitter every thirty minutes, automatically clustering and organizing it.

He said that now there is no manually written code at Anthropic. All SQL is written by the model, and everything is built by the model. What’s more interesting is that his Claude and his colleagues' Claude communicate with each other via Slack, negotiating to solve unknown issues during the loop's execution.

When I first heard this, I was stunned. This isn’t about "humans using AI tools to increase efficiency"; this is about "humans managing a team of AIs." The nature of work has changed—not just accelerated but fundamentally transformed. Your role has shifted from executor to conductor, from code writer to problem definer, result reviewer, and agent coordinator. This transformation is much deeper than just being "ten times more efficient."

What Will Future Teams Look Like?

Boris's assessment of future teams is, I believe, the most pragmatic and valuable part of this interview.

He stated that in the future there would be more and more generalists. But he emphasized that the definition of "generalist" is expanding. In the past, a generalist engineer might have meant someone who could write both iOS and backend code, while future generalists will be true cross-disciplinary professionals—strong in engineering while also strong in design, or product and data science.

He gave Claude Code team's example: engineering managers, product managers, designers, data scientists, finance, and user researchers—each of them writes code. Each person is an expert in their field, but now everyone can also program.

This assessment resonates deeply with me. I have always said that in the AI era, what is most scarce isn't "programmers who can use AI," but "full-stack executors who understand the business." In the past, technology acted as a barrier, preventing many idea-driven individuals from entering the field. In the future, the technological barriers will decrease, with the real barriers being: do you understand what you want to achieve, do you have sufficient industry knowledge, and do you have the judgment to evaluate whether what AI produces is correct.

In other words, AI is not making the technical barrier disappear; rather, it is shifting the position of that barrier—from "Can you write code?" to "Do you understand this field?"

Will SaaS Die? No, but Many Moats Will Disappear

Someone asked Boris whether SaaS would face a significant collapse when the cost of coding drops by ten times or even a hundred times.

Boris said he loves this question. His answer referenced Hamilton Helmer's "Seven Powers" theory on business moats. He believes that the advent of AI will weaken some moats while making others more important.

Moats that will weaken include switching costs—previously, if you used a tool for three years, the cost of migrating data and retraining employees was too high; but now the model can help you migrate quickly, drastically reducing switching costs. Also, process power (process barriers)—many SaaS companies' moats actually come from complex business process designs. But Boris said, with Opus 4.7, it can already "hill climb anything." Give it a goal, let it iterate, and it will optimize continuously until it achieves that, which makes companies relying solely on process design to build barriers very fragile.

Moats that won't weaken include network effects, scale economies, and unique resources. These are not significantly affected by AI and will remain effective.

He shared a second judgment, which I find even more important: in the next decade, there will be ten times more startups than now, disrupting various fields. The reason is simple: a small team can now build products level with large companies using AI, while large companies will face internal resistances, needing to change processes, retrain employees, and overcome organizational inertia. Startups do not have these burdens and can build AI-native solutions from day one.

I strongly agree with this judgment. We are in a "window period for small companies." The efficiency advantage of large companies is diminishing, and the disadvantages of startups are shrinking, quickly compressing the gap between the two. This window won’t remain open forever, as large companies will eventually transition, but right now is the best time to attack.

Programming Will Become as Common as Literacy

This is the most historic statement Boris made, which I reflected on for a long time after hearing.

Someone asked him whether programming would become a skill that everyone possesses, like Microsoft Office. He said, not only that, but it would go further—it would become as common as sending a text message.

He then drew a historical analogy that I found very apt. In the 1400s, before the invention of the printing press in Europe, only 10% of Europeans were literate. These individuals were employed by kings and nobles to read and write as this was a specialized skill beyond the reach of ordinary people. After the invention of the printing press, the cost of publishing decreased by a hundredfold, and books became widely available. Over the next few centuries, global literacy rates jumped from 10% to 70%. Now, we all can read and write; no one needs a specific "reading and writing degree" to be literate, even though professional writers still exist.

Boris said that software is undergoing the same transformation. Writing code is transitioning from a skill needing professional training to a basic ability that anyone can acquire. This time, the speed will be much faster than during the printing press era.

He also mentioned a very insightful point: in this new world, the person best suited to write accounting software isn't an engineer, but a highly skilled accountant. Because the challenge lies in understanding accounting, not writing code.

I profoundly resonate with this judgment. Having worked in cross-border fields for many years, I often encounter a situation: individuals doing well in their industry have very clear tool needs, but due to a lack of technical knowledge, these ideas remain just ideas. The advent of AI is liberating these individuals. True productivity comes from deep industry knowledge, not solely from technical skills.

What’s the Gap Between Anthropic Internally and Externally?

Someone asked an excellent question: how far ahead are Anthropic's internal engineers in work methods compared to those outside? Is it three months? Or six months?

Boris's response was candid. He said that at the model level, there is not much difference because Anthropic itself is a platform, and the model versions used internally are similar to those of external users. The real gap lies in the organizational processes and work methods.

Internally at Anthropic, Claude is used for everything. Engineers' Claude runs code in loops while communicating via Slack with other engineers' Claude to negotiate and solve unknown problems within the code. The entire company has no manually written code; all SQL is written by the model. He mentioned that this organizational-level process transformation is the real leading edge, not the technology itself.

This brings to mind a more macro question: the release of AI capabilities is often not constrained by technology but by organizational structures and work processes. The tools are there, but most companies are still using outdated work methods with new tools, akin to the printing press invention while everyone still publishes using the methods of hand-written books.

For startups, this is a massive advantage. Without historical burdens, they can structure their organizations and processes AI-natively from day one. In contrast, large companies aiming to undergo the same transition must persuade every managerial layer, redesign every process, and face exponential resistance.

MCP is the Infrastructure Connecting AI and the Tool World

Boris mentioned MCP several times during the interview, and I believe it is important to explain why this is significant.

MCP (Model Context Protocol) is a standard protocol proposed by Anthropic that allows AI models to connect various external tools and data sources in a uniform manner. Just as USB enables various devices to connect to computers, MCP allows AI to access tools like Salesforce, Google Docs, GitHub, Slack, etc.

Boris says that for the agentification of knowledge work, MCP is the answer. Whether it’s Claude Code, Claude CLI, or Co-work, once connected with various MCP connectors, AI can access the data and capabilities within these tools. For tools that lack MCP, computer vision can serve as a fallback plan—though slower, it works.

My understanding is that MCP is building a bridge between AI and the real working environment. Previous AI tools were often isolated, requiring manual copying and pasting of information in and out. MCP automates and smooths this process. This is not a minor improvement; it’s a critical step for AI's evolution from "an assistive tool" to "the main body of work."

Moreover, he made a statement that I feel captures the essence: for the model, whether it's MCP, API, or other interfaces, they are just tokens. The model doesn't care how it connects; it only needs access to information and execution capabilities. This means that as long as information can be structurally transmitted to the model, it can utilize it. This perspective is very useful for thinking about how to build AI products.

My Deep Reflection: What Does All This Mean for Us?

After listening to this interview, the question I pondered most was: in this new world, what is truly valuable?

Boris stated that programming has been solved. I agree. But I want to add: the fact that programming has been solved does not mean software products have become unimportant. On the contrary, software will become more numerous, cheaper, and more specialized. Many valuable software products that were never created due to high development costs are now being rapidly developed. The cost is declining sharply.

There will be a large number of "specialized tools in vertical scenarios" created that may only have a few thousand users, but for those few thousand, they are immensely important.

In the overseas market, I often encounter a pain point: cross-border e-commerce, SaaS expansion, and content distribution all have many repetitive workflows, yet there have been few good tools specifically targeting these scenarios. The reason is simple: the market has been too small, the development cost too high, and the ROI unviable. But now, with a 100-fold decrease in development cost, this logic changes. Many previously commercially unviable tools can now be created.

This leads me to another deeper question: where will competition between future products lie?

Boris mentioned the network effects, economies of scale, and unique resources that won't be weakened by AI. I believe there is another underlying factor he did not explicitly state, but which was implied throughout the conversation: the depth of domain knowledge.

When code generation can be automated and processes optimized automatically, the hardest quality to replace is a profound understanding of a specific field. You know where the users' real pain points are, you understand the industry’s unspoken rules, you recognize which features seem important but no one actually uses, and you know that the final 20% of details determine whether a product is usable or not. These are things the model won't learn in the short term.

Therefore, I conclude that in the AI era, product competition will superficially seem to rely on who can better use AI. However, it fundamentally rests on who understands their users more deeply. The tools have changed, but this underlying logic hasn't.

Another point I strongly resonate with from Boris is regarding the opportunities for startups. He said that in the next decade, the number of startups would increase tenfold. The resistance to transformation faced by large companies is a natural advantage for startups. I have seen many entrepreneurs in China give up good ideas due to technological barriers. In the future, this reason will become increasingly untenable. The quality of ideas themselves will more and more determine whether a company can succeed.

Finally, I want to extend what Boris said about the printing press analogy. The printing press popularized writing and gave rise to the Renaissance, the Enlightenment, and the Industrial Revolution. It wasn't because the printing press directly propelled these movements, but because knowledge began to flow widely, allowing more people to build upon the shoulders of their predecessors to think and create. The popularization of programming by AI may also bring about a similar effect: when more people with deep industry insights can directly build tools and products, the speed of innovation across society will accelerate. True breakthroughs often come from those who deeply understand the problem and possess the ability to implement solutions. AI is making such individuals move from being a minority to a majority.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。