Introducing autoresearch for arXiv papers
Change 'arxiv' to 'autoarxiv' in any paper URL
An agent deploys to resolve setup issues on the codebase, run a minimal reproduction, and estimate full replication cost. Read more below
Wrote some thoughts on the deployment gap between scaling robot foundation models and the inference serving infrastructure that .. doesn't exist yet!
https://t.co/RTfFpGyxZd
ML interview question:
Here are the weights for Llama 3.1 70B. Generate a token by executing the forward pass manually using pen and paper. You have 30 minutes.
The optimal PPO setup from our training also generalizes to more difficult tasks like Walker2D, HalfCheetah, and Ant (with slight changes to account for task-specific constraints).
Notes: PPO as the algorithm is held constant. In the future, the agent should be able to iterate on the choice of algo as well. Many environments require longer rollouts to achieve optimal performance. 12-minute rollouts suffer from short-horizon bias and may not generalize to longer training runs.
I adapted @karpathy's autoresearch for RL.
Given a locomotion RL environment, the agent continuously optimizes a PPO algorithm to solve the task.
I ran it for 6 hours, 29 experiments overnight, nearly solved BipedalWalker-v3 on a single MacBook CPU.
https://t.co/o2k8jN8En5
Very excited to share the news I’ll be joining @xai, focusing on talent.
I grew up chasing the American Dream the only way I knew how - watching my immigrant mom work nonstop so I could have a chance to try. From Chinatown to SF, the path hasn’t been linear, but it’s been life changing.
I wrote more about that journey below.
Grateful to @TheGregYang and @barisakis for the opportunity and trust.
As my mom always says: be kind, be honest, take care of people.
We’re building something special here.
Would love to meet thoughtful builders and kind humans.
Time to make @grok 🚀🚀🚀
in college you meet people who remind you of yourself, and your interactions with them make you realize you're becoming more similar to your friends that you used to disagree with