A former OpenAI employee is convinced that Dendi’s victory over the bot in 2017 would have slowed the development of AI

You can laugh at this, but I’m almost certain that if Dendi had won, OpenAI would have raised far less money afterward, and the development of artificial intelligence would have proceeded more slowly.

At the time, OpenAI was a relatively unknown research lab (with about 50 employees), and this was the first time since the company’s initial launch that it had made headlines in major media outlets.

There were several reasons:

Overall, as far as I know, this decision was made almost by chance.

A few random facts about the development:

A few days before the match, the bot still wasn’t good enough to beat professionals. We continued making last-minute changes right during The International. Specifically, I worked extensively on feature extraction from game data, while Jakub and Szymon ported the old weights to the new model. We didn’t have time to test the changes, so we literally sent everything straight to production, since we only had time for a single training run. And it worked.

We knew of several ways to exploit the bot’s vulnerabilities (our training environment wasn’t versatile enough, so it was possible to do a lot of strange things, such as luring the bot toward other towers). At the same time, I was certain that defeating it in a direct head-to-head match based on micro-control was practically impossible. As long as Dendi stuck to his usual 1v1 playstyle, I estimated our chances of winning at 98% or higher.

Many professional players thought we had “solved” the question of which items heroes should buy in 1v1 duels. Specifically, the bot would buy a huge number of consumables. As far as I know, pros were subsequently inspired by his play, and the mid-lane meta shifted toward buying more consumables. The truth was that we simply hard-coded the item purchase order and never had enough time to explore other options.