President & Co-Founder @OpenAI
Joined July 2010
- Tweets 4,815
- Following 65
- Followers 498,999
- Likes 1,386
Welcome to the team!
I'm thrilled to let everyone know that I've joined OpenAI as a research scientist on AI policy. I'm excited to work with such a great team of researchers towards the goal of making sure that AI is beneficial for all. blog.openai.com/openai-chart…
2
57
Congrats to my co-founder @ilyasut. He is a true visionary, has an unbridled creativity, and inspires everyone who works with him. Well deserved!
1
7
87
Particularly good HN comment on the talk: news.ycombinator.com/item?id…
3
21
My talk from today entitled "Can we rule out near-term AGI?": youtube.com/watch?v=YHCSNsLK…
6
36
8
187
Boredom = getting into such states. Avoiding boredom = not getting into such states. Sadly English parses ambiguously!
1
1
63
An agent which learned to play Mario without rewards. Instead, it was incentivized to avoid "boredom" (that is, getting into states where it can predict what will happen next). Discovered warp levels, how to defeat bosses, etc. More details: blog.openai.com/reinforcemen…
88
2,851
340
7,736
0
Performance on Montezuma's Revenge, on which the 2015 DQN algorithm famously achieved a score of 0. We've now developed an algorithm which exceeds human performance: blog.openai.com/reinforcemen…
5
59
4
242
That feeling when you point your agent at a totally fresh environment, and it starts learning without any tuning. Never gets old.
5
12
1
195
Congrats to my co-founder @ilyasut for winning the Market for Intelligence award for ideas. Ilya is a visionary and working with him is one of the best parts of my job:
This year’s CDL #MKT4INTEL awards were presented by @Cmdr_Hadfield. Congratulations to this year’s recipients!
Labour: @melsariffodeen of @learningcode
Capital: Jim Fleck
Ideas: @ilyasut of @OpenAI , Parvin Mousavi of @queensu
#BuildSomethingMassive
1
13
1
74
A lot of OpenAI's AGI safety research focus on specifying goals without reward functions — since it's clear that we can't hand-program objective functions for complex real-world tasks. One potential paradigm for avoiding reward functions:
Iterated Amplification: An AI safety technique that has the potential to let us specify behaviors and goals that are beyond human scale: blog.openai.com/amplifying-a…
36
1
169
Now accepting applications for our second OpenAI Scholars program. Diversity is crucial for ensuring that AI benefits everyone: nitter.vloup.ch/openai/status/10….
Apply for our Winter 2019 OpenAI Scholars Program, open to individuals from underrepresented groups in STEM interested in becoming deep learning practitioners -blog.openai.com/openai-schol…
5
30
1
121
Want to get into machine learning, but don't have pre-existing background? The OpenAI Fellows program is a great way for people from adjacent fields (software engineering, science, etc) to become machine learning practitioners:
Come join the OpenAI team as a Fellow or Intern in 2019 — applications now open: blog.openai.com/openai-fello…!
8
25
144
.@vicenews segment on OpenAI Five: news.vice.com/en_us/article/…
4
7
53
Have been making serious progress on OpenAI Five. Looking for Dota teams with average MMR 6.5k+ or Tier 3 and higher to test the latest version. Ping brooke@openai.com if you’d like to try it out!
31
80
8
503
Lots of cool projects being presented tonight at OpenAI Scholars Demo Day (mobile.twitter.com/openai/st…). Really impressed with the progress everyone made in 3 months of study.
Neural waltzes, semantic trees, CycleGANs, and more - details on OpenAI Scholars' 2018 Final Projects: blog.openai.com/openai-schol…
Demo day in San Francisco on September 20th!
8
66
That feeling when your agent learns a task so well it crashes your code:
...
Mean reward: 0.9
Mean reward: 0.94
Mean reward: 1.0
./learn.py:122: RuntimeWarning: invalid value encountered in true_divide
A = (R- np.mean(R)) / np.std(R)
Mean reward: -0.02
...
1
13
231
Weights are shared. The maxpool happens before any recurrence so it’s just about sharing the inputs, which differ slightly per hero.
4
Added a bibtex here: blog.openai.com/openai-five/…. LMK if you have any issues. Looking forward to reading your paper!
1
1