Thread by @Smerity

The way this is written it sounds as if these were recently revealed details of @OpenAI Five, pulled out of secrecy. The post which clarified API use (and debate around it) was literally a year ago (1v1 at TI 2017) and @OpenAI have been impressively transparent since.

https://twitter.com/GaryMarcus/status/1031567959712444416

I complained thoroughly about it at the time. The main issue was that none of the technical details were released with the surprise reveal at TI 2017 which, in combination with @elonmusk overhyping the result in media, lead to speculation then rollback.

https://twitter.com/Smerity/status/897224659036393472

The @OpenAI team seem to have learned from that transparency mistake, first clarifying in the media near immediately and then with a follow-up more technical blog post less than a week after TI 2017 (August 16, 2017). They then hinted at their work on 5v5.
blog.openai.com/more-on-dota-2/

I still find so much value in the discussion of "would the task/domain benefit from vision rather than API use" (h/t @ebetica) but using it as a negative in regards to both @OpenAI 1v1 and @OpenAI Five, given they noted in their mandate a year ago they wouldn't, is odd.

If there were iffy results, fine, but @OpenAI Five has well and truly proven itself. Behaviours seen in released bot v bot play exhibit strategy I'd only associate w/ advanced humans. I genuinely think humans can _learn strategy_ from them. Yet we're still hung up on API use..?

I meant to also +1 @OpenAI on their transparency since 1v1 as it got to 5v5. They released details about the architecture, training, and issues and have had live matches anyone can watch. I wish there was a standard research paper but this isn't a standard research project :)