DeepSeek Has Gotten OpenAI Fired Up

It’s been just over a week since DeepSeek upended the AI world. The introduction of its open-weight model—apparently trained on a fraction of the specialized computing chips that power industry leaders—set off shock waves inside OpenAI. Not only did employees claim to see hints that DeepSeek had “inappropriately distilled” OpenAI’s models to create its own, but the startup’s success had Wall Street questioning whether companies like OpenAI were wildly overspending on compute.

“DeepSeek R1 is AI’s Sputnik moment,” wrote Marc Andreessen, one of Silicon Valley’s most influential and provocative inventors, on X.

In response, OpenAI is preparing to launch a new model today, ahead of its originally planned schedule. The model, o3-mini, will debut in both API and chat. Sources say it has o1 level reasoning with 4o-level speed. In other words, it’s fast, cheap, smart, and designed to crush DeepSeek. (OpenAI spokesperson Niko Felix says work on o3-mini began long before DeepSeek’s debut and the goal was to launch by the end of January).

The moment has galvanized OpenAI staff. Inside the company, there’s a feeling that—particularly as DeepSeek dominates the conversation—OpenAI must become more efficient or risk falling behind its newest competitor.

Part of the issue stems from OpenAI’s origins as a nonprofit research organization before becoming a profit-seeking powerhouse. An ongoing power struggle between the research and product groups, employees claim, has resulted in a rift between the teams working on advanced reasoning and those working on chat. (OpenAI spokesperson Niko Felix says this is “incorrect” and notes that the leaders of these teams, chief product officer Kevin Weil and chief research officer Mark Chen, “meet every week and work closely to align on product and research priorities.”)

Some inside OpenAI want the company to build a unified chat product, one model that can tell whether a question requires advanced reasoning. So far, that hasn’t happened. Instead, a drop-down menu in ChatGPT prompts users to decide whether they want to use GPT-4o (“great for most questions”) or o1 (“uses advanced reasoning”).

Some staffers claim that while chat brings in the lion’s share of OpenAI’s revenue, o1 gets more attention—and computing resources—from leadership. “Leadership doesn’t care about chat,” says a former employee who worked on (you guessed it) chat. “Everyone wants to work on o1 because it’s sexy, but the code base wasn’t built for experimentation, so there’s no momentum.” The former employee asked to remain anonymous, citing a nondisclosure agreement.

OpenAI spent years experimenting with reinforcement learning to fine-tune the model that eventually became the advanced reasoning system called o1. (Reinforcement learning is a process that trains AI models with a system of penalties and rewards.) DeepSeek built off the reinforcement learning work that OpenAI had pioneered in order to create its advanced reasoning system, called R1. “They benefited from knowing that reinforcement learning, applied to language models, works,” says a former OpenAI researcher who is not authorized to speak publicly about the company.

“The reinforcement learning [DeepSeek] did is similar to what we did at OpenAI,” says another former OpenAI researcher, “but they did it with better data and cleaner stack.”

OpenAI employees say research that went into o1 was done in a code base, called the “berry” stack, built for speed. “There were trade-offs—experimental rigor for throughput,” says a former employee with direct knowledge of the situation.

Those trade-offs made sense for o1, which was essentially an enormous experiment, code base limitations notwithstanding. They did not make as much sense for chat, a product used by millions of users that was built on a different, more reliable stack. When o1 launched and became a product, cracks started to emerge in OpenAI’s internal processes. “It was like, ‘why are we doing this in the experimental codebase, shouldn’t we do this in the main product research codebase?’” the employee explains. “There was major pushback to that internally.”

Last year, the company announced Project Sputnik internally—an effort to go through the code and figure out which parts should merge and which parts should remain separate.

Employees believe the project was not fully realized. Rather than merge the stacks, employees were encouraged to simply prioritize using the “berry” stack, which angered some people working on chat. An OpenAI spokesperson denies this and says Project Sputnik was successfully deployed.

The perceived issues with the codebase had tangible repercussions, sources say. Ideally, after an employee launches a training job, the GPUs involved in that job are freed up for someone else to use. The way the berry code base is structured, that doesn’t always happen. “People would just squat on the GPUs,” one former employee says. “There was gridlock.”

Outside OpenAI, the industry is divided on how to interpret DeepSeek’s success. Earlier this week, shares of Nvidia plunged as investors worried that the industry had wildly overestimated the number of chips needed to work on AI.

But that interpretation is shortsighted, experts say. If DeepSeek discovered a way to do model development more efficiently, as it claims, it might accelerate the model development process, but ultimately the winner will still be the company with the most chips.

“You do need less compute per unit of intelligence, but people are still going to want more units to scale up even more,” says Miles Brundage, an independent AI policy researcher who worked at OpenAI for six years, most recently as a senior adviser for AGI readiness.

Perhaps Stargate, OpenAI’s flashy new infrastructure project, will ease the feeling of scarcity internally. Crusoe, the company building Stargate’s first data centers in Abilene, Texas, has already broken ground on a 998,000 square foot facility, according to Andrew Schmitt, a spokesperson for the company.

While the details of the project are opaque, I’m told it could grow to encompass more data centers, chip manufacturing, and supercomputers. OpenAI plans to appoint a new CEO to lead the project—at least in theory.

Current CEO Sam Altman “is very good at making promises of what will happen in the future. And then in the future those things are just completely unreliable,” says another former staffer.

Time Travel

In 2023, Steven Levy provided a definitive look inside OpenAI in the months before its various, highly publicized upheavals. Even then, the tensions that are boiling over today were already apparent.

It’s not fair to call OpenAI a cult, but when I asked several of the company’s top brass if someone could comfortably work there if they didn’t believe AGI was truly coming—and that its arrival would mark one of the greatest moments in human history—most executives didn’t think so. Why would a nonbeliever want to work here? they wondered. The assumption is that the workforce—now at approximately 500, though it might have grown since you began reading this paragraph—has self-selected to include only the faithful. At the very least, as Altman puts it, once you get hired, it seems inevitable that you’ll be drawn into the spell.

At the same time, OpenAI is not the company it once was. It was founded as a purely nonprofit research operation, but today most of its employees technically work for a profit-making entity that is reportedly valued at almost $30 billion. Altman and his team now face the pressure to deliver a revolution in every product cycle, in a way that satisfies the commercial demands of investors and keeps ahead in a fiercely competitive landscape. All while hewing to a quasi-messianic mission to elevate humanity rather than exterminate it.

That kind of pressure—not to mention the unforgiving attention of the entire world—can be a debilitating force. The Beatles set off colossal waves of cultural change, but they anchored their revolution for only so long: Six years after chiming that unforgettable chord they weren’t even a band anymore. The maelstrom OpenAI has unleashed will almost certainly be far bigger. But the leaders of OpenAI swear they’ll stay the course. All they want to do, they say, is build computers smart enough and safe enough to end history, thrusting humanity into an era of unimaginable bounty.

End Times ChronicleA passenger flight crashed into a military helicopter in Washington, DC, on Wednesday evening.

Last but Not LeastDeepSeek left a core database exposed, leaking 1 million records, including user prompts and API authentication tokens.

Elon Musk told friends he’s been sleeping at the DOGE offices near the White House.

And, no big surprise, Elon Musk lackeys have already started taking over the US Office of Personnel Management.

Update 1/31/25 11:32 ET: This story has been updated to include additional comment from OpenAI about the timing of o3-mini.