R. Alex Hofer @rhofour

SWE by day, pretend ML researcher by night. he/him Pittsburgh, PA Joined June 2014

Tweets

12
Followers

16
Following

90
Likes

55

Google DeepMind @GoogleDeepMind

2 years ago

Meet Gemini Robotics: our latest AI models designed for a new generation of helpful robots. 🤖 Based on Gemini 2.0, they bring capabilities such as better reasoning, interactivity, dexterity and generalization into the physical world. 🧵 goo.gle/gemini2-roboti…

170 436 2K 643K 515

View Details

Alex Reinhart @capnrefsmmat

2 years ago

LLMs like ChatGPT have a specific writing style: big words, lots of "delve", always happy and agreeable. But it's hard to measure or identify what is specifically so different in their writing. With our new corpus of human and AI writing, now we can. 🧵

1 2 6 269 1

View Details

🌟Cheshire Cat ᓚᘏᗢ, @autismsupsoc

2 years ago

CW: Body Horror? This AI video attempt to show gymnastics is one of the best examples I have seen that AI doesn’t actually understand the human body and it’s motion but is just regurgitating available data. (Which appears to be minimal for gymnastics) x.com/Przemek8739456…

374 906 5K 2.1M 1K

View Details

R. Alex Hofer @rhofour

9 years ago

@texttheater @nsaphra I'd love to know how that works. I recently switched to zsh primarily for it's better handling of history, but I don't have things setup quite how I'd like yet.

1 0 0 0 0

View Details

R. Alex Hofer @rhofour

9 years ago

@beenwrekt From what I've seen policy gradient methods tend to be substantially more sample-efficient than genetic algorithms. In OpenAI's ES paper they claimed a wall-clock improvement, but at the cost of using 10x more compute. The paper you link looks like the exception, not the norm.

0 0 0 0 0

View Details

R. Alex Hofer @rhofour

9 years ago

@beenwrekt Are these problems mostly restricted to just vanilla policy gradients? It seems quite a few of the issues you raise can be fixed by doing things like DDPG, using entropy regularization, and using advantage estimators.