From forming an “advisory board” of stuffed animals to becoming an AI giant with thousands of employees, Anthropic has grown from a small AI security startup to an important player in the industry.
Author: Li Jia, Wall Street CN
Today, Anthropic has made "AI safety" its core label, but few know that the company's co-founder and president, Daniela Amodei, once discussed management issues with her husband, Holden Karnofsky, using a group of stuffed toys.
These toys were given different personalities: some were specifically responsible for "having fewer meetings and causing less trouble for themselves," some emphasized rules, details, and execution, and others were responsible for reminding decision-makers to remain compassionate. When facing a thorny workplace issue, Daniela and her husband would have these virtual characters each "express opinions," thereby simulating the choices different types of managers might make.
This rather private, even somewhat childlike, episode took place during the period when Daniela was still working at OpenAI. It was also during this period that she and her brother, Dario Amodei, were gradually diverging from OpenAI's management: Daniela had friction with multiple senior executives at work, while Dario began to grow increasingly wary of the company's leadership and the direction of advanced AI development.
In 2020, the siblings ultimately decided to leave OpenAI. This split, later called a "divorce" by insiders, also became the starting point for the birth of Anthropic.
The Atlantic recently cited journalist Kevin Roose's forthcoming book, The AGI Chronicles, in disclosing this little-known experience. In a 2023 interview with Roose, Daniela said this method was "very useful for analyzing different personality types."
From assembling an "advisory board" out of stuffed toys back then to growing into an AI giant with thousands of employees today, Anthropic has evolved from a small AI safety startup into a major industry player. According to media reports, the company's potential IPO valuation may approach $2 trillion.
How did a group of stuffed toys become an "advisory board"?
This informal "advisory group" was not simply a joke, but a way of thinking that simulated different decision-making styles.
Among them, Jinji was an easygoing cat who tended to reduce Amodei's workload, such as skipping unnecessary meetings; the pink bear Maura represented the completely opposite style—organized, demanding, and detail-oriented; and the panda named Beary Bonds was designed to add more compassion and generosity to discussions.
When facing the same workplace issue, the Amodeis would have the "personalities" represented by different dolls each respond separately.
The real function of this method was not in the toys, but in forcing decision-makers to temporarily step away from their own position and re-examine the issue through another personality.
In essence, this was a low-cost thought experiment: without real external advisers, it could still break the same issue into different decision paths through the simulation of different roles.
And this method was almost unknown inside OpenAI.
Beyond the stuffed toys, divisions inside OpenAI were widening
At the same time, the relationship between the Amodei siblings and OpenAI was also changing.
Daniela Amodei joined OpenAI in 2018, later moving into a management role on the technical team, and had disagreements with multiple senior executives, including Greg Brockman and Mira Murati.
She and Brockman were once close. Brockman had spoken at her 2017 wedding to Holden Karnofsky, but afterward, their working relationship gradually deteriorated.
Dario Amodei became increasingly focused on OpenAI's leadership and the direction of advanced AI development, and grew increasingly wary of Sam Altman and Brockman.
At the same time, he began privately organizing events with a group of researchers who wanted to place greater emphasis on the risks of powerful AI systems. Outside the company's normal structure, these people communicated and held meetings, and some colleagues called this faction "Dario's little circle."
Internal conflicts eventually reached a stage that could no longer be avoided.
In 2020, Dario began assembling a team willing to leave OpenAI and pursue advanced AI research under a different safety philosophy. Daniela became one of the central figures in this departure and was responsible for negotiating with OpenAI while the team prepared to leave.
This split was later privately called a "divorce" by those familiar with the matter.
One departure gave birth to an AI giant
In December 2020, the Amodei siblings formally left OpenAI and co-founded Anthropic.
Jack Clark, Sam McCandlish, Tom Brown, Chris Olah, and Jared Kaplan also participated in founding the company, and Ben Mann was later identified as another co-founder.
Anthropic was initially just a small research organization centered on AI safety. Since then, the company has gradually expanded to thousands of employees and grown into one of OpenAI's most important competitors.
Daniela Amodei now serves as Anthropic's president and board chair, with authority covering most of research, engineering, product, and business operations. Holden Karnofsky, who was once a member of OpenAI's nonprofit board, later also joined Anthropic.
From using stuffed toys to simulate managers to now growing into an AI giant with thousands of employees, this episode also bears witness to Anthropic's journey from a fledgling startup to maturity.