The Florida Street Incident#

While the race to catch up with GPT-4 continued to accelerate, an unexpected event would shake OpenAI to its core and expose the internal tensions that had been quietly building inside the leading organization.

Let’s move to November 2023. On Friday the 17th, OpenAI’s board of directors did the unthinkable: they removed Sam Altman as CEO, under the premise that they had “lost confidence in his leadership” and that Sam had not been “consistently honest in his communications with the board”.¹ The implications, motives, and consequences are enough to make an entire movie. In fact, one is already being made.²

During a long time (2023-2024), one of the most repeated explanations of the incident was that there was a conflict between Sam Altman, interested in accelerating the development and commercialization of OpenAI’s products, and Ilya Sutskever, increasingly concerned about the safety of advanced systems.³⁴

The news spread around the world and caught the entire community off guard, spawning a wide range of opinions, speculation, and memes—my personal favorite being: “What did Ilya see?” However, later documents and testimonies would reveal a more complex story.

In a 2025 court deposition, Ilya acknowledged that he had been considering for at least a year the possibility of replacing Sam Altman. At the request of the independent directors, he ended up preparing a 52-page document whose central argument stated that: Sam exhibited a consistent pattern of lying, undermining his executives, and pitting them against each other.⁵

A significant part of the material used to build this case came from Mira Murati.⁵ She would later, under oath (2026), make similar statements: she described Sam as someone capable of telling one thing to one person and the opposite to another, creating chaos among the executives and making her own work as CTO difficult.⁶

But to summarize what were days of absolute chaos: after the threat of a mass resignation by hundreds of OpenAI employees and Microsoft’s offer to build Sam an entire organization/lab inside Microsoft itself—along with his entire entourage,⁷ the board was left with no real choice but to reinstate Sam just five days after his dismissal, only to see all of its members replaced shortly after, except Adam D’Angelo.⁸

This episode demonstrated something many of us had long suspected but were not yet ready to accept: OpenAI was no longer that non-profit organization devoted to developing AI for the benefit of humanity, governed by a board free of conflicting interests and guided by what was best for the world. Instead, it had become a company like any other, subject to power dynamics among founders and the pressure of investors.

SSI (Safe Superintelligence): Ilya’s Lab#

The consequences of the November incident did not take long to materialize. Several months later, in May 2024, the first of two major splinterings within OpenAI occurred: the resignation—or more precisely, the indirect dismissal of Ilya Sutskever, co-founder and chief scientist of the company.⁹ We speak of an “indirect dismissal” because, according to multiple sources, after the failed attempt to remove Sam Altman, Ilya was effectively isolated within the company¹⁰, with reduced access to resources and virtually no influence.

Ilya’s response would come in June 2024, with the creation of his own startup: Safe Superintelligence Inc., or simply SSI. With a single stated objective: to build superintelligence safely, without any intermediate products.¹¹ SSI is arguably the most secretive lab in the industry to date.

Although the specific details of their work remain undisclosed, some directions can be inferred from interviews Ilya has given over the past couple of years. Unlike the industry’s dominant approach of ever-larger, increasingly general models, SSI appears to be pursuing something fundamentally different: not absolute intelligence, nor a hyper-ambivalent model like today’s LLMs, but a system (model and learning algorithm) that behaves more like a human teenager. That is, a system that, without being an expert in everything from the start, can be deployed in different environments to learn how to perform specific tasks in the real world. In this vision, Ilya has repeatedly emphasized the importance of a powerful value function—the mechanism that allows a model to evaluate which actions bring it closer to its goals—as the key element underlying human learning capability.¹²

SSI’s opacity is striking. Other details about the startup remain tightly guarded: its stage of development, the exact size of the team, the specific techniques being explored. The few data points that have surfaced are fragmentary: the team is small (estimated at a couple dozen researchers), much of it comes from Israel, they are using Google TPUs for their compute infrastructure, and their valuation, according to 2025 reports, has already reached 32 billion dollars.¹³¹⁴

But in July 2026, Ilya Sutskever announced that they have research worth scaling, along with a partnership with NVIDIA to access an order of magnitude more compute.¹⁵

Some say—rightly so—that if the history of AI were a movie, Ilya Sutskever would be its protagonist.

Thinking Machines Lab: Mira’s Project#

The second major split would arrive just over a year after the Sam Altman incident. In February 2025, Mira Murati, who had left her position as OpenAI’s CTO five months earlier, decided to embark on her own venture: Thinking Machines Lab.¹⁶

Unlike Ilya’s quiet departure, Mira’s exit was a sizable exodus. She took with her a significant number of key OpenAI researchers, as well as select talent from other competing labs.¹⁶

Without having launched a single product, Thinking Machines Lab secured an estimated initial valuation of 12 billion dollars.¹⁷ In the age of AI, pedigree and promises are worth almost as much as results.

But Mira and her team would not remain as mysterious as Ilya for long. Months after raising their initial capital, they began publishing a series of blog posts sharing genuinely interesting research: they explored the deep causes of non-determinism in LLMs (that seemingly random behavior that persists even at temperature 0), introduced a novel optimization algorithm variant called Manifold Muon, developed advanced techniques using LoRA (Low-Rank Adaptation, a method for efficiently adapting large models) for LLM fine-tuning, and shared new approaches to model distillation.¹⁸

Who would have thought? They ended up being more transparent and more “open” than today’s OpenAI.

However, Thinking Machines Lab’s most significant launch was neither a paper nor a model, but a commercial service: Tinker. This is an API-accessible platform that allows companies and developers to perform customized fine-tuning of LLMs, still in private beta as of late 2025. You control the algorithm and the training data; they provide all the computational infrastructure and scalability.¹⁹

And in the AI world of 2025, a handful of blog posts and a working API are enough to open the doors to the next funding round, which was estimating valuations around 50 billion dollars,²⁰ but it ended up stagnating, and becoming an investment from NVIDIA for an undisclosed amount together with an agreement to provide 1GW of compute systems.²¹

The development of their own models would come later.

Be that as it may, this would not be the first time that a seed fallen from OpenAI’s tree germinates and flourishes spectacularly on its own soil.

Right, Dario?

Author’s Note:

1. The final question refers to Dario Amodei, co-founder of Anthropic and former Vice President of Research at OpenAI, who left the company in late 2020 along with other key researchers to found what today is one of OpenAI’s main competitors.²² Anthropic, creator of the Claude family of models, was born precisely from those same tensions between speed of development and safety considerations that years later would trigger the Florida Street incident.²³ History, as it often does, tends to rhyme.

References#