They couldn't see it continuously, but they can see Five's chats about the win prob when it hits certain thresholds. Was a bit of an experiment to see how humans would react :).
Currently training at our largest-ever scale (thanks to heroic efforts from the team), which means even we can't predict how much progress Five will make this week.
Released post about our progress and goals with Five: openai.com/five/. My personal favorite: this chart of progress vs humans to date. When creating it, was stunned by how linear our MMR (Dota skill) growth has been. Means roughly exponential progress. Same as 1v1 bot.
OpenAI Five will be playing against five of the world's top Dota 2 professionals in front of nearly 20,000 audience members at The International on Wednesday, Thursday, and Friday: openai.com/five/
Intern Open House tonight at OpenAI! Talks will be recorded for those who can't attend in person. Good luck to our interns Alex, Fish, Sadhika, Will, and full-timer @machinaut! Will also contain a bonus Q&A with @ilyasut. Let me know if you have questions for him :).
We have been training the same parameters dating back to June 9th. We have constantly make significant changes to the game restrictions or model architecture, and the model has adapted its existing knowledge quickly.
Just finished debugging a last issue with our Dota 7.19 build (latest Dota version, which we'll play at The International). Will start training on it tomorrow. Less fun than nitter.vloup.ch/gdb/status/10257…, but some things are just necessary to stay up late for :).
Stayed up late last night integrating @benbarry's awesome new OpenAI Five logo into the Dota UI. Some features are just worth staying up till 2:30a for.
Looking for Dota teams with average MMR 6k+ to test the latest version of OpenAI Five. This week & weekend only! Please ping brooke@openai.com if you're interested.
Need all the help we can get before The International! Wish us luck :).
Solution: delete 3 copies of Dota from source machine, and then transfer will complete overnight. Pretty sure issue was number of files rather than number of bytes.
Thanks everyone for the suggestions and sympathies :)