Deconstructing Anthropic: The Top AI Company, Potentially Also a Form of "Organizational Invention"

Original Title: "Decoding Anthropic: The Best AI Company, Perhaps Also an Organizational Invention"
Original Author: Celia, Overseas Unicorn

Over the past year, Anthropic may have been the most researched company in the entire AI industry.

At the beginning of this year, it achieved the fastest exponential growth in human business history: ARR grew from $9 billion to $45 billion. If the computing power can keep up, it is highly probable that ARR will reach $100 billion by the end of the year, $200-300 billion next year, directly on par with Meta's scale.

In the secondary market, its current valuation has reached $1 trillion, surpassing OpenAI.

We have spent a lot of time researching how Anthropic managed to surpass others.

Ultimately, to understand this company, the core is to understand two points:

One is strategic judgment, and the other is organizational culture.

By now, everyone should have a fragmented understanding of this, but there is still no complete picture. Therefore, this article attempts to provide a more comprehensive review and reconstruction.

The hope is to explain some of the external curious questions from the perspectives of strategy and organization, such as:

• Why could Anthropic realize that coding might be the most important direction in 2021?

• The personality differences between Dario and Sam, and how they shaped two companies on completely different strategic paths?

• Why is Anthropic's talent attrition rate so low?

• Why does almost every person at Anthropic praise its culture? How is this culture maintained during the company's rapid expansion?

01. The Underestimated Importance of Focus

Firstly, from a strategic perspective, OpenAI has always been more like a company that wants everything.

In terms of model capabilities, math, science, coding, reasoning, multimodal, and architectural innovation, OpenAI has been striving.

On the product side, Codex, browsers, robots, enterprise platforms, smart hardware, chips, data centers, and more are all advancing simultaneously. It is said that the number of projects within OpenAI has once reached around 300.

In contrast, Anthropic, as one of the three families, was the only one to give up multimodality very early on. They have never talked about architectural innovation, have not emphasized concepts such as reasoning models, RL, or continual learning, but have focused solely on scaling language models. They have emphasized only one direction - coding - and have first mastered the most crucial capability.

Regarding why coding is so important, the market is now also clear, with the core being threefold:

1. Coding is the gateway to everything. The vast majority of tasks in the digital world can be expressed through code.

2. Coding is the most suitable ability for model learning. Results are highly verifiable, the feedback loop is short, and user data can more significantly impact model training.

3. Coding is the core accelerator for AGI development. The leading AI labs have already entered this accelerated cycle, and this year's progress in models in one quarter is faster than the progress made in the past year.

The final outcome confirms that coding is indeed the most critical direction, exerting its influence across the entirety of the sector.

OpenAI, on the other hand, only woke up in March, cutting off sideline businesses like Sora and prioritizing coding as the company's top priority.

How did Anthropic pinpoint coding?

One thing we have always been curious about is: How did Anthropic manage to pinpoint coding right from the start?

Looking back, it is found that half was foresight, and half was luck.

Anthropic faced great difficulties in early fundraising. Without much money, they had to find a more efficient way to advance toward AGI.

They needed to tell a story in a vertical scenario first, proving that they could form a business loop. So at that time, they seriously studied whether coding might be the best choice if they could only choose one direction: first train a better coding model → provide it to customers for use → obtain customer usage data in a real engineering environment → feed back into model training. This could potentially create a flywheel effect.

The head of Growth at Anthropic once mentioned that he had seen an internal document written by the company's co-founder, explaining why they should focus on coding. The key point was that the document was dated 2021, long before anyone knew what the actual market opportunity in this direction was.

However, as the funding situation improved and the company had more resources, the focus on coding was no longer mentioned. Instead, they decided to first build a more general model foundation.

A turning point came after the ChatGPT became popular. Anthropic realized that the consumer market had already been captured by OpenAI, so it regretfully (but in hindsight, luckily) shifted the battleground, turning its focus to the B2B sector.

This strategic shift was overall cautious and empirical, rather than a decisive gamble.

While training Claude 3, Anthropic intentionally strengthened its coding capability and received very positive market feedback on Sonnet 3.5.

Subsequently, they continued to invest while seeking validation, gradually firming up their judgment on the potential of coding, both in terms of commercial value and research acceleration. The team then began to concentrate on moving forward along this path, completely abandoning the consumer sector and not even diverting attention to multimodality.

In addition to the market-focused direction, it is also worth mentioning the determination in the technological roadmap.

Over the past two years, there have been repeated assertions from prominent researchers that scaling laws hit a wall and the marginal returns of pretraining have peaked. Based on our interactions with various research teams, Anthropic has always been the most steadfast believer in scaling laws, having the most solid foundation in pretraining and data, without diluting focus on new paradigms.

In hindsight, this approach was also proven to be correct. A significant part of Claude's leap in capabilities came from the solid investment in pretraining.

Founder's Character

However, this raises another curiosity: why is Anthropic always able to make decisive trade-offs in key directions and maintain its focus?

Firstly, it is naturally due to resource constraints. Anthropic's historical funding amount is roughly only one-third of OpenAI's. However, looking deeper, the strategic differences between these two companies are also closely related to the founders' characters and backgrounds.

Anthropic has 4 co-founders, all of whom were core authors of the scaling laws paper that year. Dario himself is the most central research lead of GPT-3 and had already been working in the AI field for a decade, possessing firsthand experience of AI technological advancements and being more willing to make judgments.

Furthermore, Dario is a person who does not experience FOMO at all, and has even been described as somewhat narcissistic and stubborn, rarely being swayed by market consensus.

When he was 24 years old, well before Anthropic had achieved explosive growth, he said something that I still believe is a crucial point to understanding this company. The essence of it was:

“The most profound lesson I learned over the past decade is that there will always be a so-called consensus in the market. However, after witnessing the consensus flip overnight several times, I began to focus on my own bet.”

“I don't know if we are always right, but to be honest, even if we are only right 50% of the time, it is still highly valuable, as you are offering something that others do not possess.”

This is very different from Sam Altman. Based on our conversations with some people close to Sam:

1. Sam is considered one of the most ambitious founders in Silicon Valley, wanting everything from the start. Additionally, his past experience in investing at Y Combinator made him very familiar with the "spray and pray, parallel bets" approach, which is why OpenAI grew countless branches.

2. Sam does not come from a technical background, so his judgment on technical direction is not as strong as Anthropic's, relying more on the team to push forward from the bottom up. Sam leverages his resource-gathering abilities and provides ammunition to individual teams.

3. Sam's VC background makes him particularly fond of breakthrough fancy ideas. Therefore, within the OpenAI culture, there is a strong emphasis on 0 to 1 paradigmatic innovation, but less focus on refining from 1 to 10. Many product lines like Sora, Atlas Browser, Voice Mode, and many others do not have continuity; once released, they are left unattended.

4. Both Sam and Mark Chen (Chief Research Officer) have personalities that only say yes and never say no. For side tasks, as long as the team pushes hard, resources will still be allocated from the top.

As OpenAI's forces are continually spread thin across various side projects, Anthropic can gain an edge on the most critical battlefield through a Sun Tzu-style strategy.

The Brilliance of Strategy Lies in "Strategicness"

Anthropic's strategic focus has given us an inspiration, where the importance of focus has been underestimated.

I recall a podcast I listened to last year, where the guest was David Senra, the host of the Founders podcast. For the past 8 years, he has almost exclusively done one thing: researching a great entrepreneur every week.

When asked, if he were to distill all the entrepreneurial experiences from over 400 founder biographies he has read into one thing, what would that be?

He answered: Focus.

Great entrepreneurs are often not well-rounded achievers but rather extreme obsessives. They identify the one or two variables that are most important to them, such as Costco's pricing, Apple's design experience, ByteDance's recommendation algorithm & data flywheel, and then push them to the extreme, even to the point of making competitors look ridiculous.

It is important to clarify here that many people think they are very focused, but they do not truly understand the meaning and cost of focus.

The so-called focus must be broken down into two levels:

First is judgment, knowing what is most critical, and being willing to sacrifice everything else.

Second is intensity, being able to devote overwhelming resources to penetrate the key elements.

The former is a cognitive issue, while the latter is a willpower issue, and both are indispensable.

For example, when Google was founded, the prevailing consensus in the entire internet industry was that the future belonged to "portals." Search giants like Yahoo were stuffing their homepages with more and more features, news, weather, shopping, games, horoscopes... Every feature was seen as a lever to "increase ad value."

However, Google believed that information would only continue to grow, and what users needed was not a larger portal but to immediately find the most relevant answer.

So, while others aimed to keep users on the page longer, Google wanted users to leave faster. Google's homepage at that time was exceptionally clean, with nothing but a search box.

The same goes for the business model. Yahoo had dozens of monetization methods, while Google focused all its energy on the "search keyword auction" mechanism, taking nearly a decade to seriously pursue a second line of business.

Even today, one of Google's Ten Things is "It's best to do one thing really, really well."

The core of strategy is not only about deciding what to pursue but also about deciding what to give up. I believe most people don't say "no" enough.

02. Culture is the Biggest Secret Sauce

Perhaps the most unique aspect of Anthropic is not its strategy but its organizational culture.

Over the past six months, amidst the intense AI talent war, Anthropic has experienced a much lower talent turnover rate compared to other AI labs.

The following two graphs summarize the talent flow data from '21 to '23.

The first graph shows the proportion of job hops between various AI labs, revealing:

• For every 10.6 people who move from DeepMind to Anthropic, only 1 moves back to DeepMind.

• For every 8.2 people who move from OpenAI to Anthropic, only 1 moves back to OpenAI.

The second graph illustrates the proportion of employees who stay at a company after 2 years of joining.

Anthropic has an 80% talent retention rate, the highest among the top AI labs at that time, slightly surpassing DeepMind's 78%.

For a younger, rapidly evolving company like Anthropic to achieve a higher retention rate than the established DeepMind is quite remarkable.

In contrast, OpenAI is at only 67%.

It is worth noting that this data was compiled before OpenAI's heyday, when Anthropic had not yet emerged.

Looking at the news over the past two years, Anthropic's talent attraction and stability become even more apparent.

For example, in a recent Twitter post that gained significant traction, several celebrity CTOs were willing to jump ship to Anthropic to become regular technical employees (i.e., MTS, member of technical staff):

The primary reason for this is often attributed to Anthropic's organizational culture.

Listening to podcasts featuring Anthropic team members, almost everyone mentions Anthropic's culture. Some even consider this cult-like culture as Anthropic's biggest secret sauce.

“I truly believe that culture is Anthropic's secret weapon, our strongest defense, something that others cannot replicate. It's not something that happened naturally; the leadership has invested a lot in this area.”

— Amol Avasare, Head of Growth at Anthropic

If one doesn't specifically pay attention to this issue, they might not notice it because discussions about culture or values often feel abstract, as if it were just a slogan. However, when all the first-hand information and public interviews are viewed together, it is genuinely inspiring.

Three Key Traits of Anthropic

If we break it down further, three key traits that set Anthropic apart from other AI labs are:

1. Mission-oriented

Anthropic's mission is to “ensure the world safely transitions to transformative AI,” placing safety above all else.

While many companies claim to be mission-driven, the level of seriousness with which Anthropic approaches this is almost religious.

It is a frontier lab with a strong moral imagination: it genuinely believes AGI can save the world, but also genuinely believes AGI could destroy the world. It aims to lead everyone across that narrow tightrope between the two.

The head of Claude Code, Boris Cherny, once said: "At Anthropic, if you randomly ask someone in the hallway 'Why are you here,' the answer will always be safety."

He and product manager Cat Wu had both left Anthropic last year to join Cursor, but returned within two weeks because they found themselves deeply missing the internal cultural atmosphere of Anthropic. The feeling of everyone being purely dedicated to a greater mission.

Some people were skeptical before joining Anthropic, but upon entering, they found that "Wow, the atmosphere inside is even more serious than what was described outside."

There are even early employees here who have said in all-hands meetings—if Anthropic ultimately achieves its mission but the company itself fails, it would still be a good outcome.

This statement explains a lot about Anthropic.

In the logic of most companies, commercial success always comes first, and the mission is just a facade. But what makes Anthropic unique is that internally there is indeed a group of people who prioritize the mission over the survival of the company.

Examining what Anthropic actually does, it is also a manifestation of their words. For example, their governance structure designed around a nonprofit trust, research on interpretability, various investments in security, including sacrificing a $200 million order from the U.S. Department of Defense due to values misalignment, and so on. I will not go into detail on these aspects.

2. High trust, low ego

When we interact with other cutting-edge labs, we often hear about internal politics and turf wars. But not at Anthropic. On the contrary, everyone is very united and willing to selflessly help others.

The most magical thing here is that Frontier AI is a place where star culture and resource struggles easily grow. AI researchers are arguably the smartest, highest ego group of people in the world, naturally striving to propose a different solution, make a name for themselves, but resources are very limited, so departmental conflicts always occur.

Daniel Freeman, who joined from Google to Anthropic, said that other modeling companies internally feel like separate, secretly competitive fiefdoms, but he "has never felt this way at Anthropic."

After joining Anthropic last fall, former Stripe CTO Rahul Patil mentioned that the culture here was the most surprising to him. It's hard to imagine that a group of such intelligent people could also be so humble at the same time.

He gave an example: If the company told you tomorrow that the best position for you is not to continue as an executive but to become an IC (individual contributor) because that would be your greatest contribution to the mission, would you be willing to do it? He believes that 100% of Anthropic's people would do it without any ego.

3. A Strong Humanistic Background

The writer for The New Yorker had spent several months doing an in-depth embedded visit at Anthropic, and two interesting descriptors stuck with him about the people here:

• Bookish misfits

• A disproportionate number of Anthropic employees seem to be the children of novelists or poets.

In other words, the people here are not quite like typical Silicon Valley elites or the traditional image of tech nerds; they are somewhat bookish, somewhat nerdy, and somewhat idealistic. Many give the impression of having grown up in families of writers and poets.

To some extent, you can already see this from the Claude model naming: Haiku, Sonnet, Opus, each corresponding to the concise haiku, Shakespearean sonnet, and heavyweight classical works.

As a comparison, OpenAI's GPT-4 / 4o / o1 follows an engineering numbering naming convention, and Google's Gemini Ultra / Pro / Flash uses classic product line names. This contrast can shed some light on certain issues.

The head of Claude Code, Boris, once shared an interesting detail in a podcast:

During his first lunch at Anthropic, he casually mentioned a very niche book by hard sci-fi writer Greg Egan.

How niche was this book? He had never met anyone who had read it before.

At the dinner table, he casually mentioned a reference from a book, and to his surprise, everyone at the table caught on.

This incident both shocked him and made him feel like he truly belonged.

Bookworms who enjoy science fiction often have a grand sense of humanism and historical responsibility, as well as a better ability to reason about the butterfly effect.

This consensus based on reading interests reassured him that this might be the best place to push the boundaries of AI.

How Culture Becomes Institutionalized

The next question is, how is this pure, almost cult-like culture maintained?

After all, Anthropic is no longer a small AI lab; it is a large company with 3000 people, and it has maintained its cultural intensity as it expands at the fastest pace in history.

To this, Dario directly said he probably spends 1/3 to 40% of his time ensuring that Anthropic's culture is strong.

Even though there are countless things to do technically, product-wise, in fundraising, and in government-industry relations. But he believes that his leveraged work is to make Anthropic a highly cohesive place where top talent enjoys working.

In terms of specific practices, there are a few key points:

1. Unique Hiring Standards

Anthropic's approach to hiring is different from many AI labs.

On one hand, in terms of talent preferences, unlike most companies vying for big names, Anthropic prefers to hire underdogs. Instead of external labels, they value direct evidence of ability, such as "Have you done independent research, written truly insightful blogs, or made substantive contributions to the open-source community."

On the other hand, Anthropic has very strict cultural screening. During interviews, they have a specific Cultural Interview round, where candidates are asked 15-20 scenario questions in an hour.

Based on leaked interview questions online, the focus is on three main points:

（1）Do you really put the safety mission first?

One of the most typical screening questions is: If Anthropic decides not to release a model because it cannot guarantee safety, are you willing to accept your stock going to zero?

（2）Are you a nice, humble person?

This includes kindness, empathy, people skills, and the ability to admit ignorance and mistakes.

（3）Can you handle complexity?

Many of the issues that Anthropic deals with internally are highly complex and ever-changing. They greatly value whether a person has a systems mindset, can deeply reason about the second-order effects of things, and can consider how a decision will impact other parts of the system.

They have spent a lot of time in recruiting doing "reverse screening," and have truly turned down many top 10x developers for this. Rahul Patil, former CTO of Stripe, mentioned that before joining Anthropic, he had a long conversation with the then CTO of Anthropic.

Instead of persuading him to come over, the CTO of Anthropic spent two to three weeks specifically discussing with him repeatedly why he should not join Anthropic, kindly dissuading him, stating that it's not worth it unless you are truly aligned with the culture and mission.

Therefore, Anthropic's recruitment logic has never been about bringing in as many top talents as possible, but rather about screening out unsuitable candidates as early as possible. "We are very good at filtering out those who come for money and fame."

In contrast, after OpenAI grew larger as a company, they stopped conducting specific cultural interviews, reportedly leading to some management issues.

This contrast was very apparent in the round of talent poaching by Meta last year. Faced with Meta's sky-high package offers, OpenAI's response was more in line with market practice: counteroffers, offering retention bonuses, removing the new employee's vesting cliff to accelerate stock ownership.

Anthropic's response, on the other hand, was very Anthropic. They told employees that coming to Anthropic is primarily for the mission, not to continually drive up their own price in external bidding wars.

We will not offer you a salary ten times higher than your equally excellent colleagues just because Mark Zuckerberg happened to pick you. That would be unfair. If you want to leave, feel free to go.

The outcome of this situation also speaks volumes. OpenAI reportedly lost dozens of people, while Anthropic only lost 2, and these two individuals had already worked at Meta for 6 and 11 years, respectively.

2. Culture of Context Sharing

Anthropic fosters a culture of high information transparency.

First and foremost, Dario actively, frequently, and repeatedly provides meaning. He often holds all-hands meetings to share with the entire company, sometimes as frequently as every two weeks, aptly named Dario Vision Quest (even Dario himself joked that the name's evangelistic nature is too obvious, sounding like he went to the mountains to gain some enlightenment).

He stands in front of the whole company and speaks for an hour, usually accompanied by a three to four-page document covering everything from the company's direction and product strategy to industry changes. He then directly answers questions on the spot.

Many internal employees have mentioned that he speaks very directly and candidly. "Dario is the most straightforward person I've ever met. His words are not calculated but spoken as he truly thinks."

In addition to all-hands meetings, he regularly writes a lot in his own Slack channel, unadorned, jotting down his stream of consciousness: recent company events, his concerns, and how he views issues of interest to everyone.

This culture ensures that everyone in the company understands how decisions are made and which issues should take top priority. Consequently, in a complex and ever-changing environment, each individual can make relatively consistent distributed decisions.

Moreover, this transparency is not unidirectional but open to challenge. If someone disagrees after hearing Dario's sharing at an All Hands meeting, they can directly go to Dario's notebook channel and publicly say, "I disagree with your assessment," initiating an immediate debate. Challenging the leadership openly is encouraged.

Furthermore, this writing culture is not exclusive to Dario but rather a collective thinking mechanism involving all employees.

Many individuals in Anthropic have their own notebook channels, akin to personal Twitter feeds, where they constantly document their thoughts, activities, and progress. Others can subscribe, observe, join the discussion, or challenge viewpoints.

Many employees have praised the company's writing culture, with Slack being a huge repository where much unfolds.

Therefore, Anthropic seems to have cultivated a strong alignment soil within the company, where everyone's projects, viewpoints, and ideas are transparent enough and fluid enough, to the point that someone once exclaimed even financial data is transparent.

(However, on the contrary, secrecy is strictly maintained in technology, and it is rumored that some teams are even intentionally isolated from each other and cannot dine together.

The result is that researchers from other companies regretfully note that all the key know-how here is scattered across different people's minds, and it is impossible to piece together a complete picture by poaching a few individuals.)

3. 7 Founders with Equal Voting Rights, the Founding Structure Itself is a Cultural Mechanism

Anthropic's founding structure has a very unconventional design: it has 7 founders, and Dario boldly decided at the time to give each person the same amount of equity, rather than taking a larger share for himself.

At the time, everyone advised him that this would be a disaster, warning that it would blur leadership, cause misaligned incentives, and the company would easily fall apart due to internal conflicts.

However, Dario believed that the company does not revolve around a single founder, but around the mission, and equal voting rights are the most unmistakable evidence of this concept.

They had worked together for many years and had a high level of trust in each other. Equal voting rights are not essentially a governance design, but rather a proof of commitment, a mechanism for cultural diffusion.

The 7 co-founders, like 7 cultural replication nodes, can project values to a broader audience on different fronts. As a result, even if the company expands, the original culture is not easily diluted.

In comparison, OpenAI's executive team has actually been very volatile, with 11 founding team members leaving in succession, leaving only Sam Altman, Greg Brockman, and Wojciech Zaremba.

And the newly appointed executive team is even more unstable: since the beginning of the year, the Head of Product has taken a leave of absence, the Head of Marketing has resigned for health reasons, the Head of Communications has exited, the Head of Operations has been reassigned, and the Head of Finance has also been sidelined...

4. Strong Emphasis on One Team, Avoiding Silo Mentalities

The Anthropic CTO once mentioned in a podcast that AI labs, compared to traditional companies, are very bottom-up as they follow a sort of inverted pyramid organizational structure where power and creativity flow from the bottom up.

The most critical work happens on the front lines because those individuals are closest to AI's emergent behavior. They run experiments every day and have the most intuitive understanding of what the models are capable of. The majority of product innovations come from the front lines rather than being driven by executive-level roadmaps.

However, there is a challenge that emerges when decision-making authority is delegated – each team tends to cling to its own problem awareness and value system, leading to the formation of isolated silos.

What sets Anthropic apart is that it early on recognized that since decision-making must be decentralized, it is even more crucial to proactively foster unity. Dario does not want the safety team only talking about safety being paramount, or the product team solely focusing on the product's importance, and then passing all conflicts up the chain of command for resolution.

One core management principle he adheres to is to distribute trade-offs to each individual, allowing everyone to have a founder's perspective. Everyone is simply engaging in the same massive trade-off processing from their respective positions.

Thus, they strongly emphasize being one team and employ various institutional designs to blur the boundaries between roles. For example, below the executive level, there is no differentiation in titles; everyone is simply referred to as a member of the technical staff. They intentionally downplay identity distinctions such as "researcher vs. engineer," "senior vs. junior," or "architect vs. implementer."

This stands in stark contrast to OpenAI, which has always had a stronger researcher culture characterized by an internal hierarchy: Researcher > Research Engineer > Software Engineer.

As a result, the product team is often overshadowed by research and lacks significant decision-making power. In times of conflict, researchers are also unwilling to cooperate with the product team.

In terms of product innovation, OpenAI is heavily researcher-driven – new discoveries often come from the research team, with the product team receiving last-minute emails and then scrambling to find applications.

At Anthropic, the product and model teams are more closely aligned, allowing the product to more effectively influence and define the model's capabilities.

This is actually one of the reasons why OpenAI's product capabilities are not as strong as Anthropic's.

The Dual Origins of Culture

The next question is, why has Anthropic developed such a unique organizational culture?

Perhaps it can be viewed from two aspects:

1. The nature of the business itself

I remember hearing a sharing from a senior HR leader at a top company two years ago, which left a deep impression on me and made me think deeply for the first time about what organizational culture really means.

The essence of organizational culture is that employees' behavioral patterns are a key element that can help a company succeed.

So the first principle of organizational culture is actually that the nature of the business determines the culture.

For example, ByteDance and Huawei are both companies with strong organizational capabilities, but if you were to swap the organizational systems of the two, both companies would soon go out of business. This is because they are at two extremes of the same spectrum: ByteDance emphasizes "dare to be the first," while Huawei emphasizes "dare to be the second." One values innovation more, and the other values efficiency more.

This is not about value judgment but about the nature of the business. When both companies are developing a new product, Huawei focuses on things like base stations and chips. If an issue arises, the cost of a recall could wipe out a year's profit. ByteDance, on the other hand, operates differently. It is a typical short-cycle, short-chain business where dozens of versions can be released in a week, allowing for quick corrections and releases.

So ByteDance can encourage innovation and choose "Context, not Control," but Huawei cannot. For Huawei, premature innovation could be a burden. Huawei excels at gradually surpassing the competition step by step with its organizational capabilities and resources once the market achieves product-market fit.

Now, back to Anthropic.

In the AI competition, a core moat is the ability to have smart people do dirty work, especially in the areas of coding and agency. While it may seem like a competition of model capabilities on the surface, at a deeper level, it is actually a competition of engineering capabilities. It is not the kind of problem that can be solved by a few geniuses having a sudden insight. Instead, it requires a large amount of dirty, fragmented, and detailed system engineering.

One of the most crucial barriers is data.

Historical chat data was just simple text data, but Coding and Agentic data are more complex. They are not just chat logs, but also include the task itself, environment setup, execution traces, and finally, a complete evaluation and verification system.

All of this is grunt work. Doing it well is crucial, but it's not like publishing a paper or launching a new product, where it can be a personal shining moment.

According to feedback we received from some researchers, OpenAI's most critical issue today is the difficulty of organizing hundreds of top talents to diligently work on data and do the dirty work.

OpenAI hires only the top talents in the prestige hierarchy—people with good backgrounds and high ambitions. Naturally, everyone is more inclined to make their own bets, to go from 0 to 1. As for cleaning up the mess, filling in the data, few are willing to take on that task.

OpenAI has been successful in the past by relying on some key paradigm breakthroughs to gain a significant competitive advantage. However, as Yao Shunyu recently said in an interview, "The era of individual heroism is over," and "AI doesn't really need intelligence... The most important trait is to be reliable and meticulous in your work."

At this point, you will find that in an environment like Anthropic, which is low ego, highly cohesive, and mission-driven, these advantages will be amplified very clearly.

It is said that Anthropic's co-founder Jared Kaplan personally leads the team every day to work on the data, with extremely careful data cleaning. No other company can do this.

(This also explains a phenomenon: OpenAI's models are the strongest in competition-level coding challenges because these tasks are more of a research problem. However, in agentic tasks in day-to-day work, they often lag behind Anthropic because the latter is more of an engineering problem, testing data, systems, and execution details.)

II. Founding Team Background

Company values can be said to be part of the founder's values, such as Jack Ma's chivalrous spirit, Pony Ma's gentle openness, Steve Jobs's aesthetic orientation, and Ren Zhengfei's military discipline.

More precisely, the values of the founders often come from two things: what the founders originally believed in and what they once deeply despised.

The former determines what you want to become, while the latter determines what you do not want to become under any circumstances.

The Anthropic clearly has both, with the formative power of the latter possibly even greater than that of the former. A quick look at Dario's experience provides insight:

Dario's first encounter with AI was at Baidu's AI lab, where he first observed scaling laws and gradually became a staunch believer in them. However, after Baidu made a breakthrough, internal struggles over control and resources quickly erupted, leading to the eventual disbandment of the team.

Later, Dario joined OpenAI, where he was deeply involved in advancing the GPT series. OpenAI once entrusted him with 50%-60% of the company's total computing power, allowing him to lead the GPT-3 project.

Due to Dario's strong values and personal opinions, disagreements with other OpenAI members on organizational principles began to surface.

For example, Greg Brockman once proposed a stunning idea: selling AGI to the nuclear powers in the United Nations Security Council. Dario almost resigned on the spot after hearing this, as he saw it not as a mere business disagreement but as a fundamental values issue.

Greg and Dario had been at odds for several years, with Sam Altman trying to mediate between them. At this point, Sam used one of his greatest skills — making both sides feel like he was actually on their side. In the short term, this was a balancing act; in the long run, it was a trust overdraft. Eventually, after a reconciliation, everyone realized that what Sam had promised Dario was completely different from what he had promised Greg.

Gradually, Dario formed a close-knit alliance within the company, and some people, simply because he liked pandas, referred to this small group as "the pandas." Their differences with the OpenAI leadership on strategic direction and organizational governance grew, eventually escalating into a severe political struggle.

A serious face-to-face confrontation even erupted among the top executives. Sam accused Dario and Daniela (Dario's sister, who later became one of the co-founders of Anthropic) of organizing negative feedback against him behind his back; both denied it and immediately brought in the alleged sources of these accusations to confront Sam. The sources claimed they had no knowledge of such activities, prompting Sam to retract his accusations.

error

• Coding has already shown its hand, and OpenAI is likely to catch up. A clear trend now is developers migrating from Claude Code to Codex;

• The demand has far exceeded everyone's expectations, and computing power is becoming the new ace in the hole. OpenAI early on secured computing resources far beyond Anthropic's;

• OpenAI's culture of open exploration has its own huge advantage. At the same time, OpenAI has always been more aggressively exploring and betting on new paradigms, and the next leap could potentially turn the tables.

One can only say that looking back over the past three years in 2026, Anthropic has indeed left a memorable mark on the entire industry:

In the AI era, winning doesn't necessarily depend on greater ambition, more exploration, and stronger talent.

Sometimes, winning can also come from the opposite: fewer bets, lower ego, and a naive mission.

P.S. We are also curious about what kind of organizational culture and best practices other cutting-edge AI companies are forming. We welcome friends with first-hand observations and thoughts to contact us through the contact information below!

Perhaps the next generation of great AI companies will first be a new kind of organizational invention.

Original Article Link