It was only August of last year (2018) when OpenAI barely lost to some good human players, and now here we are. It's starting to feel like AlphaGo all over again. If it does go that route, OpenAI Five will continue to get better as it keeps playing itself, till it leaves humanity behind.
Didn't D. Lenat's 'Automated Mathematician' ('AM') software (from the 'Eurisko' line) perform similarly (let's say 'learning new heuristics by discovery, through training') when it won at a championship built upon a game named 'Classic Traveller', in 1981 and 1982?
I can't find any occurrence of the words 'eurisko' nor 'lenat' in the openai.com website.
It OpenAI's software ran autonomously, starting with a completely generic version (not taylored for games, especially for Dota) then interacting with humans exclusively through the Dota game (no source code nor parameters modification) it may be a breakthrough.
I believe that rather than gathering information via pixels OpenAI gets it through the API. As a result it knows exactly where its teammates are at all times, their conditions etc. without any delay time (to my knowledge).
Sure, but that's not the same as not playing by the rules.
Previous incarnations didn't have invisible units, drafting, wards and more. Hence humans had to adapt to an essentially new game. This version of OpenAI Five appears to play standard DOTA.
This. ^ And as a piggyback what makes OpenAI five radically different from when they were trounced last year? They scaled up the computing and trained more hours. Butdid anything change? People are still saying it was its micro not it's long term strategy that won the day. Furthermore people who know OG are suspicious that they threw the match because they said "that's not how they play, they made to many amateur mistakes". I'm highly suspicious about OpenAI especially with how they handled GPT-2. Talking to journalist before releasing anything else. Propaganda is definitely what it sounds like.
Chart of appx AI MMR over time: https://openai.com/five/