President & Co-Founder @OpenAI
Joined July 2010
- Tweets 4,815
- Following 65
- Followers 495,334
- Likes 1,386
I'll be available at NeurIPS on Friday to meet up with anyone who would like to talk about OpenAI, deep learning, or the future of the field. Particularly looking to talk with people whose perspective is different from mine. Ping if you'd like to meet up — gdb@openai.com!
3
6
2
89
A creative instance of human-AI collaboration.
In AI Dungeon 2, a pretty effective strategy is to essentially Jedi mind trick the AI.
Present it with a reality that seems plausible to it, and it'll often work. I once got rid of a marauding ogre by reminding it that it was going to be leaving for Antarctica tomorrow.
1
3
24
We’re starting a new team to apply deep learning to the deep learning iteration cycle. Looking at our own workflows, especially on large projects like Dota, it feels like there’s a lot of room for improvements. Join us:
At OpenAI, we're excited by work in learned optimizers such as Learning To Learn By Gradient Descent By Gradient Descent. These methods will improve the efficiency of our models, and allow for higher quality science. Looking for a founding team member! tinyurl.com/tjz8ct3
3
7
1
97
GPT-2 is a poet and it didn’t even know it.
I’m excited to announce a side project I’ve been working on the past few weeks!
“Transformer Poetry” is a book of famous poetry reimagined by @OpenAI’s GPT-2 language model, and is available as PDF or hard copy at papergains.co 🥳🥳🥳
2
2
19
We're hiring a Head of Security. High-impact role — we have giant infrastructure relative to company size, and we think hard about safety and security in everything we do. Mix of hands-on and leading a focused team:
jobs.lever.co/openai/917ad50…
Questions? Ping me: gdb@openai.com
1
12
2
57
Statistics and deep learning give conflicting intuitions about how models should behave.
Turns out that many deep learning models transition from acting how statisticians think they should, to how deep learning practitioners think they should:
A surprising deep learning mystery:
Contrary to conventional wisdom, performance of unregularized CNNs, ResNets, and transformers is non-monotonic: improves, then gets worse, then improves again with increasing model size, data size, or training time.
openai.com/blog/deep-double-…
1
15
57
That's the first citation in the blog post. Note also that the first author, Mikhail Belkin, provided helpful discussions and feedback throughout this work, as mentioned at the bottom.
1
1
34
You make what you measure. Procgen Benchmark lets you directly measure how well and how quickly an RL agent learns generalizable skills.
We've found that with fewer than ~500-1000 levels, today's algorithms memorize rather than learn something general.
We're releasing Procgen Benchmark, 16 procedurally-generated environments for measuring how quickly a reinforcement learning agent learns generalizable skills.
This has become the standard research platform used by the OpenAI RL team: openai.com/blog/procgen-benc…
3
27
2
129
Enjoyed this article on GPT-2, the last in a six-article series by @scarschwartz on the cultural history of natural language processing.
Earlier this year, OpenAI said its GPT-2 tool could produce long strings of coherent text. Now, there's a reddit group populated entirely with text produced by GPT-2-powered bots. buff.ly/2sFvnYV
1
5
27
"Predictably, the humans made about 50 good cards and the AI one million mediocre ones, but the top 30, which make it into the packs people will receive, were comparable in quality."
Writeup on the results of the experiment: techspot.com/news/82977-card…
1
9
42
Humans are currently performing what looks like a fun and/or painful brainstorming exercise.
Whatever they're doing, it's working! Humans remain in the lead: $49,335 worth of AI-created cards sold vs $50,835 worth of human-created cards!
1
21
The website cardsagainsthumanityaichalle… is very well-done (includes livestream, leaderboard, and a great explainer of how they're using GPT-2), worth taking a look.
One particularly interesting tidbit: the humans get $5,000 bonuses if they win, so there are real stakes on this!
4
12
1
45
Left: Humans generating Cards Against Humanity ideas.
Right: GPT-2 generating Cards Against Humanity ideas.
A human is then curating the best ideas from GPT-2 and posting them here: cardsagainsthumanityaichalle…, where you can upvote your favorites.
As a Black Friday challenge, Cards Against Humanity's writers are creating new cards alongside GPT-2, each trying to create the most popular pack.
Neck and neck so far: have sold $25,115 in GPT-2-authored cards vs $25,955 in human-authored cards.
2
19
3
81
As a Black Friday challenge, Cards Against Humanity's writers are creating new cards alongside GPT-2, each trying to create the most popular pack.
Neck and neck so far: have sold $25,115 in GPT-2-authored cards vs $25,955 in human-authored cards.
For Black Friday, we taught a computer how to write Cards Against Humanity cards. Now we put it to the test. Over the next 16 hours, our writers will battle this powerful card-writing algorithm to see who can write the most popular new pack of cards. cardsagainsthumanityaichalle…
4
59
10
216
‘Twas the neural net before Christmas:
I made the @OpenAI Neural Network write me a new version of The Night Before Christmas using talktotransformer.com
The bold is what I put in, the rest is what the computer wrote. THIS IS NOT A BIT
2
2
1
19
Safety Gym: new environments and tools to evaluate how RL agents obey safety constraints during the *training process*. Normally people measure only about final performance, not mistakes along the way. Fine for simulation, not so much for the real world!
openai.com/blog/safety-gym/
26
6
134
0
