The ongoing legal battle between co-founders Elon Musk and Sam Altman is shedding light on some of the most fascinating secrets from OpenAI’s early years. Recent revelations have highlighted everything from Gabe Newell donating over $20 million to acting as the sole member of an informal advisory board in 2018, to how OpenAI's Dota 2 ambitions were fueled by high-level executive connections.

The Legend of OpenAI Five and the Battle for Dota 2 Supremacy

One of the first moments that brought global attention to the company was when an AI bot defeated pro player Danil "Dendi" Ishutin in a 1v1 match at the 2017 The International (TI) tournament. This evolved into the OpenAI Five, a sophisticated bot that mastered Valve's MOBA by playing roughly 180 years' worth of games against itself every single day.

The goal was ambitious: defeat a professional human team in a 5v5 matchup. In 2018, the OpenAI Five played three exhibition matches against human pros with conclusive results. While the AI dominated most of the series, it did lose one game after Twitch chat influenced a poor team composition.

The sheer precision of the AI left professionals stunned. Player MoonMeander recalled a moment during the second match where the bot saved a teammate: "There was one time when I was about to fissure kill a Lion and the courier came at the frame-perfect moment, delivered a salve, and it instantly used it. No way a human could have done that. No way."

How Elon Musk Secured Microsoft Compute for OpenAI's Dota 2 Project

To achieve such high-level performance, the training process required massive resources, specifically 256 GPUs and 128,000 CPU cores. Internal documents released during the current trial—specifically the "proposed findings of fact" from Musk’s lawyers—reveal exactly how this hardware was obtained.

In 2017, as the company worked to power its OpenAI's Dota 2 project, Elon Musk personally called Microsoft CEO Satya Nadella to request additional discounted computing power. Nadella agreed to the arrangement well before Microsoft made its massive official investment in the firm.

The close relationship between the two leaders was evident in their correspondence. On August 11, 2017, the day OpenAI won its Dota challenge, Nadella emailed Musk, Altman, Brockman, and Sutskever to offer congratulations. Musk replied, "Indeed, much appreciated. Will make sure people know about Microsoft’s help."

Leaked Emails: A Glimpse into the Early AGI Timeline

Another revealing piece of evidence comes from an April 23, 2018, email sent by Shivon Zilis to Musk. At the time, Zilis—a tech venture capitalist and mother of four of Musk's children—was working for OpenAI and providing Musk with internal progress reports.

Under a "Tech" heading, Zilis provided several updates regarding the state of AI development:

  • The OpenAI 5v5 Dota project was progressing better than anticipated.
  • The rapid rise in bot performance caused internal concerns that the timeline for AGI might arrive sooner than expected.
  • The team believed they were on track to master Montezuma's Revenge shortly.

The Quest to Master Montezuma's Revenge

While the Dota 2 project saw success, the reference to Montezuma's Revenge highlighted a significant hurdle. This 1981 Atari 2600 classic is notoriously difficult for AI because it resists standard game-playing strategies. In a famous 2015 paper, researchers noted that the game was an outlier where AI bots achieved 0% of the average human score.

OpenAI treated this game as a "white whale," developing models centered on "curiosity" to encourage better exploration within the game's complex environment. While OpenAI worked extensively on these techniques, it wasn't until 2021 that a different team of researchers finally claimed to have cracked the game using the Go-Explore algorithm.