dailynewsnblog
OpenAI is now one of the world’s most talked-about companies, and the ongoing legal spat between co-founders Elon Musk and Sam Altman has resulted in some truly juicy revelations around exactly what was going on in the firm’s earliest years. One example: it turns out Gabe Newell gave the company over $20 million, and was the sole member of its “informal advisory board” in 2018.
One of the first times OpenAI attracted wider attention was when one of its bots defeated pro Dota 2 player Danil “Dendi” Ishutin in a 1v1 match at the 2017 The International (TI) tournament. This “much simpler” variant was then built out into an AI called the OpenAI Five, a bot that taught itself Valve’s MOBA through the simple technique of playing 180 years’ worth of games against itself every single day. The goal was to beat a human team in a 5v5 matchup, and it did.
Article continues below
OpenAI said at the time that the OpenAI Five’s training process required 256 GPUs and 128,000 CPU cores—that’s a lot of compute, to use the parlance. And thanks to the ongoing Musk v. Altman trial, which has seen the release of a slew of internal emails and documentation, we now know precisely where it was all coming from.
“In 2017, OpenAI needed even more compute to power its ‘Dota’ project, in which OpenAI’s AI bot would attempt to defeat a human Dota champion,” says the document. “Musk personally called Nadella to request additional discounted compute. Nadella agreed.”
“The day OpenAI won the Dota challenge on August 11, 2017 [sic], Nadella congratulated Musk, Altman, Brockman, and Sutskever by email. Musk replied: ‘Indeed, much appreciated. Will make sure people know about Microsoft’s help.'”
Another email from earlier in the year gave Musk a progress report, with some eyebrow-raising claims that were not borne out (as of yet). It’s from Shivon Zilis to Musk: Zilis is a tech venture capitalist who was on OpenAI’s board until 2023, and is also the mother of four of Musk’s children. At this time Zilis was both working for OpenAI and providing Musk with the inside track on what was happening.
In the email to Musk, sent on April 23 2018, she lists some facts under various headings. Under “Tech” she discusses how OpenAI’s attempts to have its AI beat a five-player team in Dota 2 is progressing, and what she’s been told by Sam Altman:
- Says Dota 5v5 looking better than anticipated.
- The sharp rise in Dota bot performance is apparently causing people internally to worry that the timeline to AGI is sooner than they’d thought before.
- Thinks they are on track to beat Montezuma’s Revenge shortly.
Well it’s 2026 and we still don’t have AGI, so that correlation to Dota skills didn’t pan out. The more amusing element here however is the reference to Montezuma’s Revenge, a 1981 Atari 2600 game that is weirdly resistant to typical AI game-playing strategies. In a seminal 2015 paper titled “self-taught AI agent masters Atari arcade games” Montezuma’s Revenge was the big outlier, being the only game where the AI bot got 0% of the average human score.
This clearly made Montezuma’s Revenge something of a white whale for OpenAI, which made a huge effort to produce an AI bot that could master the game and achieve a high score. This OpenAI blogpost from 2018 goes into considerable detail on why the game’s structure baffles certain AI techniques, and explains how the company built a model with elements of “curiosity” in order to better play the game:
“For an agent to achieve a desired goal it must first explore what is possible in its environment and what constitutes progress towards the goal… simple exploration strategies are highly unlikely to gather any rewards, or see more than a few of the 24 rooms in the level. Since then advances in Montezuma’s Revenge have been seen by many as synonymous with advances in exploration.”
OpenAI did eventually achieve a breakthrough with Montezuma’s Revenge, though it was 2021 before a different team of researchers claimed to have cracked it with their own AI model based on multiple algorithms, collectively called Go-Explore. I guess us children of the ’80s were just built different.
