This will be the first time an AI has attempted to play the world champions in an esports game. Excited and nervous to see what happens!

Let's find out :)

We'll continue to work on Dota 2 as an environment for AI development!

Yes: docs.google.com/forms/u/1/d/…

Super excited to see OpenAI Five play against the Dota world champions, @OGesports. Will be a unique event showing a taste of what's to come with AI + humans. If you're in the Bay Area (or wanna visit!), definitely request an invite to attend in person:

We don't do co-ops per se, but the Fellows program may be of interest: jobs.lever.co/openai/f5c8d70… (openai.com/blog/openai-fello…)!

Love working with data at massive scale and having end-to-end ownership over your work? Apply to become a data engineer at OpenAI: jobs.lever.co/openai/3558cf3… With results like GPT-2, unsupervised learning is now possible — and curated datasets are key for continued progress.

Energy-based models have many appealing properties (e.g. spend more compute to improve samples — reminiscent of the human creative process), but people have largely given up on training them. We've made some exciting progress:

Been working with a codebase that uses Unicode variable names in Python 3. It's great for when I'm reading (policies are named `π` rather than `pi`!), but makes writing and debugging more difficult (kinda difficult to type `π`...). On the whole, pretty worthwhile tradeoff!

Also interesting: after playing with the small model, people viscerally feel that GPT-2 could be used to generate fake news —old.reddit.com/r/MachineLearning…

My older brother! He spent a few years trying to get me to start, and I finally agreed to do his workout routine over the holidays in 2017 if he'd drop it afterwards. I loved it and have been hooked since.

That's a metal Husafell! It's a hollow metal box which you can fill up with weights. In my event, it was loaded up to a total of 200 pounds.

Thanks, will do my very best!!

Weighed in at exactly 200 (my target weight)! Looking forward to competing tomorrow.

Agree!

Agreed!

Both are important! Look at GPT-2 for instance — that's a general-purpose architectural improvement (i.e. the Transformer) run at massive scale. One interesting point from the essay is that scale gets a bad rap — doing the reverse isn't a good way of fixing the problem!