Previously @OpenAI (Reasoning, Superalignment)
Joined July 2009
- Tweets60
- Following220
- Followers316
- Likes363
Harri Edwards retweeted
Final superalignment paper! We define a multi-agent game; get LLMs to play it and show that it makes their reasoning more “legible” to humans 1/n
Harri Edwards retweeted
AI proves formal math theorems🧮This paper uses language models for theorem proving, and expert iteration to automatically build a curriculum towards ever harder statements.
The final system solves two(!) IMO problems. Watch the paper review here:
youtu.be/lvYVuOmUVs8
ALT Expert Iteration for formal mathematics
Harri Edwards retweeted
In September 2020 @QuantaMagazine wrote about researchers trying to build an AI system that can achieve a gold-medal score at the IMO. New work by @spolu and co. at @OpenAI takes another step in that direction. openai.com/blog/formal-math/
Harri Edwards retweeted
When I started this project 2 years ago I couldn't have dreamt of us getting that far. But this is also only the beginning💥
Some thoughts on what we achieved so far 🧵
We trained a neural network that solved two problems from the International Math Olympiad. openai.com/blog/formal-math
I'm in awe of the great work my teammates have done.
We trained a neural network that solved two problems from the International Math Olympiad. openai.com/blog/formal-math
Harri Edwards retweeted
📔 New MiniF2F paper! arxiv.org/abs/2109.00110
Introduces MiniF2F a benchmark of Olympiad-level problem statements formalized in Lean/Metamath/Isabelle.
GPT-f applied to MiniF2F/Metamath ~ 2% 🥶
GPT-f applied to MiniF2F/Lean ~ 29% 🔥
Code: github.com/openai/miniF2F
👇
Harri Edwards retweeted
Posted my first paper on arXiv💥🙌
GPT-f is a Transformer-based automated theorem prover. We show that Transformer + Search is suitable to formal reasoning and continuous self-improvement 🦾
arxiv.org/abs/2009.03393
Harri Edwards retweeted
RL agents get specific to tasks they are trained on. What if we remove the task itself during training?
Turns out, a self-supervised planning agent can both explore efficiently & achieve SOTA on test tasks w/ zero or few samples in DMControl from images!
ramanans1.github.io/plan2exp…
Harri Edwards retweeted
My group enjoy talking about prunes: arxiv.org/abs/1810.04622
Harri Edwards retweeted
Excited to share the release of Obstacle Tower! Inspired by Montezuma's Revenge, we've built it to act as a benchmark for hard problems in DeepRL: requiring vision, control, planning, and (importantly) generalization in order for agents to perform well. github.com/Unity-Technologie…
Harri Edwards retweeted
We’re releasing CoinRun, an environment generator that provides a metric for an agent’s ability to generalize across new environments - blog.openai.com/quantifying-…
Harri Edwards retweeted
My latest blog post on meta-learning in general and "How to train your MAML" in particular is now out. bayeswatch.com/2018/11/30/HT…
The post thoroughly explains MAML, some of its problems, and proposes some solutions. In addition visualizes the learned per-step per layer learning rate
Harri Edwards retweeted
We did a thing!!
Releasing Spinning Up in Deep RL, a new educational package to let anyone become a deep reinforcement learning practitioner: blog.openai.com/spinning-up-…
Next year, we'll teach workshops on Spinning Up at CHAI, OpenAI, and other venues.
Harri Edwards retweeted
We’ve built an energy-based model that can quickly recognize, generate, and transfer simple concepts after only 5 training demos: blog.openai.com/learning-con…
Harri Edwards retweeted
"Dilated DenseNets for Relational Reasoning," Antoniou et al.: arxiv.org/abs/1811.00410
Harri Edwards retweeted
Nice article by @jjvincent on OpenAI's new exploration + RL results.
theverge.com/2018/11/1/18051…
Harri Edwards retweeted
Random Network Distillation: A prediction-based method that achieves state-of-the-art performance on Montezuma’s Revenge -blog.openai.com/reinforcemen…
Harri Edwards retweeted
A a very p(r)uny blogpost on Prunning Neural Networks. bayeswatch.com/2018/10/26/pr…
Harri Edwards retweeted
Interested in stabilizing the training your MAML? Do you want to substantially improve the generalization of MAML whilst cutting down hyperparameter tuning by learning most of the hyperparms automatically? Look no furthr
Paper arxiv.org/abs/1810.09502
Code: github.com/AntreasAntoniou/H…