Silicon Valley's New Mafia: OpenAI and Anthropic Are Mass-Producing Founders

Original Author: David, TechFlow

Silicon Valley hasn't collectively used the word "Mafia" in a long time.

The last time was over twenty years ago. In 2002, eBay bought PayPal for $1.5 billion, and a group of young people who had experienced the company's journey from 0 to 1 suddenly achieved financial freedom overnight, then scattered in all directions.

The rest of the story is well known: Musk built Tesla and SpaceX, Peter Thiel built Palantir, Hoffman built LinkedIn, and Chen Shiyu (Steve Chen) and Karim built YouTube...

Silicon Valley called them the PayPal Mafia.

Mafia isn't a derogatory term here; it's a certification, a certification that you come from a winner and have the ability to create another winner.

The word lay dormant for a long time. Its prerequisites are too demanding: a company that wins big enough, a concentrated distribution of wealth, and a group of people who have seen the world but haven't been worn down by it. Google didn't spawn a mafia, and neither did Meta. Until recently, it started being frequently applied to another group of people.

Those who left OpenAI and Anthropic.

Just over half of 2026 has passed, and the number of people who have left these two AI giants to found new companies is already long enough to fill a list:

Former OpenAI VP of Research Jerry Tworek founded Core Automation. Former Anthropic researcher Behnam Neyshabur and others formed Mirendil. Among those who just left are researchers working on verifiable mathematics, others on truly personal AI, and some who want to reinvent the personal computer from the hardware level...

This path has been successfully traversed before. Anthropic itself was founded five years ago by people who left OpenAI, and is now valued at $380 billion, becoming its former employer's most troublesome competitor.

The list is still growing. These departures are each using their expertise to prune the branches of the AI tree.

When the Mafia Starts Encircling Territory Outside the Megamodels

Let's start with a question. Why is it that almost none of the people who left OpenAI and Anthropic in 2026 are trying to build megamodels again?

The answer is pragmatic: there's no room left on the main trunk. Training a frontier model costs billions of dollars. OpenAI, Anthropic, and Google are already locked in close combat. A startup entering head-on would be committing suicide.

But the stronger the models become, the larger the surrounding open space grows.

Today's models are smart enough that the industry's bottleneck is no longer "can it do this?" Looking at this group of departures collectively, you'll find they are focusing on "getting work done," using their expertise to fill areas where the megamodels' reach hasn't extended yet.

For example, can AI actually get down to doing the work? Once the work is done, can it be trusted? And even if it's trusted, can it be managed?

These layers of trouble are things the big companies can't afford to address, and some aren't suitable for them to answer themselves. The companies on this 2026 list are almost all growing in these open spaces.

The most radical piece of open space is letting AI research AI itself.

Former OpenAI VP of Research Jerry Tworek and several colleagues founded Core Automation, an automated research laboratory that lets models read papers, propose hypotheses, and run experiments on their own.

The judgment behind it is quite cold-blooded: the bottleneck for AI progress is no longer algorithms, but the researchers available to do the work.

Meanwhile, Mirendil, formed by former Anthropic researcher Behnam Neyshabur and others, just secured $200 million. It's another extension of the same logic: self-accelerating systems that let models participate in improving the models themselves. The human role shifts from the subject of research to a supervisor.

Once the work gets done, new troubles follow: how do you confirm it was done correctly?

So, another piece of open space focuses on AI trustworthiness.

In most fields, verifying an AI's answer costs far more than generating one. Math Inc is targeting this most expensive step. Former OpenAI researcher Jesse Han left to found it, with the goal of turning mathematical proofs into forms that machines can verify line by line.

Mathematics is one of the few fields where right and wrong can be thoroughly scrutinized. Getting verification to work here first is the prerequisite for porting the same capability into other industries. Once AI starts making decisions for people, proving it right might be worth more than letting it do the work.

Going one layer deeper is turning intelligence into "the everyday."

The distance between a model being able to answer questions and a model being able to handle tasks for you is filled with a whole set of grunt work: calling tools, breaking down tasks, remembering context, carrying a task through from start to finish... These, too, require supporting tooling.

Rational and Zavify, founded by former OpenAI and Anthropic employees respectively, are both focused on agent workflows, helping enterprises hand their business processes over to intelligent agents.

Similarly, former xAI co-founder Igor Babuschkin founded River AI, also aiming to create AI that is truly personal and shaped by the individual.

Among this group there's an outlier: Daniel Edrisian, a former OpenAI Codex engineer, founded Blackstar, which is taking a hardware-first approach, aiming to build a personal computer redesigned for the AI era.

After all the work gets done, someone has to watch over it.

There's a subtle opportunity here. When AI companies themselves declare, "Our models are safe," no one believes them. An athlete can't also be the referee. So safety has become an independent business, and those running it are mostly people who have personally researched AI safety within these two labs.

In plain terms, they do two things. One is finding faults: Syntony, founded by a former Anthropic team, deliberately tries to induce AI to make mistakes and trick it into crossing boundaries, testing the system's weaknesses before the bad guys can strike.

The other is scoring: Resolution, founded by a former OpenAI researcher, studies how to confirm whether an AI is truly acting according to human intent, and aims to calibrate the degree of that certainty. And then there's setting rules: Guidelight, also founded by former OpenAI employees, works to define what counts as safe practices across the industry and pushes everyone to follow them...

These companies don't touch the ceiling of AI capabilities; they guard the baseline. The more capable the models become, the better their business might be.

If you ask what distinguishes this "AI Mafia" from the PayPal Mafia, the answer might be divergence versus convergence.

The PayPal departures went into everything—payments, social media, aerospace, intelligence analysis—because that was an era of opportunity everywhere;

The new Mafia members are far more concentrated in their directions, because there are only a few open spaces in AI. And the people who understand models best in the world are voting with their feet on where the next bottleneck lies.

Why Now?

The PayPal Mafia playbook had a critical prop: a single, concentrated liquidity event. The eBay acquisition gave everyone money and freedom at the same time, which is what allowed the Mafia to take shape.

In the AI industry of 2026, the prop is in place.

Last autumn, OpenAI arranged a round of secondary share sales, with employees cashing out a total of $6.6 billion as the company's valuation reached $500 billion. An even bigger event is still to come: both OpenAI and Anthropic are preparing for IPOs, potentially within the year. For early employees, the eve of an IPO is the perfect time to leave—their paper wealth is about to turn into real money, and leaving any later means being bound by golden handcuffs for several more years.

With the money in place, the talent poachers were already there.

Silicon Valley VCs have already formed a routine of waiting by the exit doors of the two labs. Aliisa Rosenthal, OpenAI's first head of sales, even switched careers to become an investor, publicly stating she plans to source deals through her network of former colleagues. Former consumer products head Peter Deng has also joined the venture firm Felicis.

Mira Murati, mentioned in the previous chapter, secured $2 billion without even having a product. Mirendil had $200 million upon its debut... Money is chasing people to an almost exaggerated degree.

The risk calculation for the leavers is also quite simple: in the worst case, they can always go back to a big tech company for a million-dollar salary.

The Cost of Crowding

Currently, dozens of companies founded by the AI Mafia are crowded into the same few patches of open ground, with competitors standing on every path who are equally smart, equally credentialed, and equally well-funded.

Another way to say crowded is: most will lose.

The PayPal Mafia story is compelling precisely because we only remember Tesla and LinkedIn, forgetting the dozens of companies that died along the way. This list will likely be no different. Looking back a few years from now, perhaps no more than five names will still be recognized.

There's also a subtler problem. The businesses of these new companies either make AI do work or watch over AI doing work. Their customers and prospects are all tied to a single premise: that model capabilities will continue to advance rapidly. If that premise slows down, many of these open spaces will disappear simultaneously.

But even factoring all that in, this list is still worth keeping.

The true legacy of the PayPal Mafia isn't any particular company. It's a principle: when the most important talent of an era starts moving out from the same place, following them is usually not a mistake.

Twenty years ago, that group defined the second half of the internet. Today, this group is gathered around the AI tree.

The ones who planted the tree are the first to know which direction it will grow.