How OpenAI lost Musk and took aim at “something magical”

In 2018, Reid Hoffman received a panicky call from Sam Altman. Things were not going well inside OpenAI. “We knew what we wanted to do,” Altman said. “We knew why we wanted to do it. But we had no idea how.” Its people tried applying AI to video games, as DeepMind had already done, and devoted too much time experimenting with a robotic hand they had built that could unscramble Rubik’s Cube. They were floundering. But the real problem was Musk. “Elon’s not happy,” Altman told Hoffman.

The race to build the first powerful AI model had always been personal for Musk. In the summer of 2015, he and Larry Page had gotten into a bitter argument about AI. Where Page saw artificial intelligence as an accelerant that could elevate humanity, Musk argued the technology was more likely to lead to our doom. Reportedly, the two stopped speaking because of it. A few weeks later, Musk met with Altman and the others at the Rosewood Hotel, where the idea for OpenAI was hatched. Yet despite OpenAI’s efforts, Google remained the undisputed leader in artificial intelligence. In 2016, a DeepMind model called AlphaGo had wowed the world by beating an eighteen-time world champion at Go, an ancient game that is considered more complex than chess and one more heavily based on human intuition. That same year, DeepMind released WaveNet, a neural network that learned to emulate human speech. Musk complained when talking with Altman that they had fallen hopelessly behind Google. The large slug of money Musk was scheduled to deposit in their bank account was in doubt.

Musk’s solution, as invariably it is, had him taking over the effort. He proposed that he either take a majority stake in OpenAI and operate the company alongside his other companies, including Tesla and SpaceX, or he would fold the startup inside Tesla, which was already working on self-driving cars. Three years into their effort, he was ready to scrap the idea of OpenAI as an independent lab.

Altman had no desire to work for Musk, a famously mercurial boss. He imagined that most of the people he recruited to OpenAI wouldn’t either. Altman rejected the offer. Musk walked away from the company, leaving Altman to worry about covering salaries and other expenses.

“Elon is cutting off his support,” Altman told Hoffman. “What do we do?” Hoffman committed to investing another $10 million in OpenAI and promised to do what he could to help them raise more. Within the year, Altman, then thirty-three years old, stepped down as president of Y Combinator and took over as OpenAI’s CEO. Hoffman was added to the OpenAI’s board of directors, and Greg Brockman, whose San Francisco apartment had served as OpenAI’s first office, assumed the role of board chairman. Publicly, Musk said he severed ties with OpenAI because of a conflict of interest with Tesla. Privately, he declared that the company had no chance at success.

Try Big Think+ for your business

Engaging content on the skills that matter, taught by world-class experts.

Request a Demo

Money continued to be an issue for OpenAI, despite Hoffman’s largesse. When I first started writing about tech in the 1990s, the cost of building a company was prohibitive. A dot-com needed to hire hordes of programmers to build its site and invest in expensive equipment to host it. Yet by the time I left the Silicon Valley beat in the mid-2000s, the economics of the startup ecosystem had been transformed. The globalization of tech talent allowed startups to tap into an international pool of skilled technologists. Programmers in India or Eastern Europe could be enlisted at a fraction of the cost of U.S.-based teams. And cloud computing eliminated the need to buy their own expensive hardware. By renting computer power as they needed it, a young company could scale its infrastructure costs in line with their growth. The barrier to entry for a startup had been lowered to the point that a few people with a laptop could challenge a giant.

AI flipped that equation back to the old days. Google was trying to hoard as much AI talent as it could. So was Facebook. As a result, top researchers in the field were commanding a salary of $1 million or more. OpenAI, for instance, had dangled $1.9 million a year plus stock to lure Ilya Sutskever from Google. The annual pay for anyone with any AI experience was reaching into the many hundreds of thousands. The labor costs for any AI startup would be enormous. Plenty of money had gone up in flames in previous tech booms. But the cost of building AI systems shocked old-timers.

Even greater was the cost of “compute”—the computer power companies needed to train and run their models. AI startups could still rely on the cloud but training large neural nets can require weeks if not months of nonstop computer time. And it seemed foreordained that those costs would continue to soar for the foreseeable future. Around the time of Musk’s ultimatum, OpenAI had made a breakthrough that would require even more computer power. In 2017, a group of researchers inside Google published what colloquially became known as the “Transformer” paper. Until that point, OpenAI had been experimenting with large language models (LLMs) that learn to chat conversationally by digesting Reddit posts, Amazon reviews, and other publicly available data sets. The Transformer paper offered an entirely new model for teaching a neural net both to better deduce a human’s meaning and to respond in a more natural-sounding way. The authors suggested that AI mimic our own brains and weigh words based on their importance. Rather than analyzing individual words, OpenAI’s large language model, or LLM, would evaluate chunks of words and use context to come up with the next word, as a human would do.

Using Transformer architecture to power its large language models, an OpenAI computer scientist told Wired, “I made more progress in two weeks than I did over the past two years.” The Transformer model proved a better way to train an LLM but it also meant creating vast, expensive-to-run models. “The amount of money we needed to be successful in the mission is much more gigantic than I originally thought,” Altman said in a 2019 interview.

Altman’s solution was to create a for-profit subsidiary that was answerable to its nonprofit board. OpenAI would seek new investors but make clear that theirs was not the typical startup. Stamped across the top of the funding agreement any investor would sign was a warning: “the principles advanced in the OpenAI Inc Charter take precedence over any obligation to generate a profit.” The new entity would be what OpenAI described as a “capped profit”company, though only a venture capitalist could consider the conditions that were imposed a cap. Anyone investing in this first commercial round could make no more than 100 times their original investment.

The company explained the change in a short post on its website in 2019. “The most dramatic AI systems use the most computational power,” it read. “We’ll need to invest billions of dollars in upcoming years into large-scale cloud compute, attracting and retaining talented people, and building AI supercomputers.”

Again, Hoffman proved pivotal. At Altman’s behest, he led this first commercial round by writing another big check to OpenAI. “Sam said it’d be really helpful if I took the lead because they didn’t have a business plan, or a product plan, or any of those things that an investor typically likes to see before putting money into a business,” he said. “It really was a bet on them being able to do something magical with AI.”