research @AnthropicAI, prev @GoogleDeepMind @MIT
NYC
- Tweets 241
- Following 526
- Followers 1,954
- Likes 9,498
Thrilled to see Gemma released today, loved working on post-training with the team!
Introducing Gemma - a family of lightweight, state-of-the-art open models for their class, built from the same research & technology used to create the Gemini models.
Blog post:
blog.google/technology/devel…
Tech report:
goo.gle/GemmaReport
This thread explores some of the performance characteristics of these models.
8
8
51
20,793
Thanks to all who came to our workshop in Exploration in RL! The videos, slides, and papers are now available: sites.google.com/corp/view/e….
Thanks again to our speakers and panelists @pabbeel Doina Precup, @white_martha, Emo Todorov, @RaiaHadsell, @pulkitology, and @jeffclune! :)
1
11
53
Check some new work I got to work on with @catherineols and others!
Our paper "Skill Rating for Generative Models" is now up! arxiv.org/abs/1808.04888
tl;dr: A new idea & proof-of-concept for evaluating generative models. Train a bunch of GANs. Have the generators "play against" all the discriminator snapshots. Rate them like chess players. 1/n
21
Gemma v1.1 Instruct 2B and “7B” :) are out! See @robdadashi’s thread for details, featuring improvements in multi-turn, fixing some overly-chatty features, and new RL. Even more to come soon!
I am very happy to announce that Gemma 1.1 Instruct 2B and “7B” are out! Here are a few details about the new models:
1/11
1
2
19
2,792
Check out our newest paper! Perhaps biased policy gradients really are the future of RL...
We looked at the sources of variance in policy gradient estimators for some common continuous control tasks, and I was surprised by the results: arxiv.org/abs/1802.10031.
3
14
We've launched our first DFL Fellowship to help support more people in creating high-quality ML curricula -- please apply and share! :)
We’re thrilled to announce the DFL fellowship, generously funded by @JaneStreetGroup.
Have curriculum ideas? We are offering 4 fellows a $4000 grant each to build a 6 week curriculum and run weekly on-line discussions.
Learn more and apply at fellowship.depthfirstlearnin…! (1/3)
5
13
Ben Eysenbach and I organized a workshop at ICML 2018 and decided to write about our experience, what we learned, and what we would've tried differently -- check it out! We hope it helps anyone looking to try organizing a workshop: medium.com/@erl.leads/hitchh… (also go vote!)
13
Check out our guide about TRPO that I helped co-write for @DepthFirstLearn! Feedback and suggestions are welcome :)
We just released our newest study guide! Learn all about TRPO from professors @kumarkagrawal and @suryabhupa → depthfirstlearning.com/2018/….
1
9
Just watched an AI named Ousia absolutely CRUSH a team of very qualified Quiz Bowl veterans at Quiz Bowl in the Human-Computer Question Answering competition track! #NIPS2017
4
12
Related: how different probability metrics are related cf "On Choosing and Bounding Probability Metrics" (2002):
10
@anishathalye presenting adversarial turtles at ICML 2018 with @logan_engstrom @andrew_ilyas @antimatter15!
2
10
Our workshop, Exploration in RL, is starting momentarily in Room T1 at #ICML — come by and hear some amazing speakers talk about solving exploration!
9
The new Google AI Residency has been announced! It's been simply fantastic so far -- please consider applying! research.google.com/teams/br…
6
Huge thanks to the other co-organizers, including Ben Eysenbach, @shaneguML, @HarriLEdwards, @white_martha, @pyoudeyer, @EmmaBrunskill, @kenneth0stanley and sponsors @DeepMindAI and @GoogleAI!
2
6
Check out our new educational effort!
Announcing DepthFirstLearning.com! We are building a repository of study guides targeting consequential papers. Check it out, learn something in-depth, and help us build the next one. @avitaloliver @suryabhupa @kumarkagrawal @cinjoncin
5
Gemma 1.1 improves on lmsys!
Exciting news - the latest Arena result are out!
@cohere's Command R+ has climbed to the 6th spot, matching GPT-4-0314 level by 13K+ human votes! It's undoubtedly the **best** open model on the leaderboard now🔥
Big congrats to @cohere's incredible work & valuable contribution to the open community!
More exciting updates:
- Qwen1.5-32B-Chat almost top-10
- Gemma-1.1-7B-it shows great improvement (1044 -> 1088, on par with Llama-2-70b)
- Starling-7B-Beta still the best 7B with over 13K votes!
6
611
this is really thoughtfully done, thanks for surfacing all of these so clearly! we're on it :)
3
1,292
I recently gave a deep learning talk for prefrosh at MIT and I spent way too long on this slide (the rest was serious, I promise):
1
3
this is remarkable stuff, i can't imagine how much progress this team will continue to make!
Today we're excited to introduce Devin, the first AI software engineer.
Devin is the new state-of-the-art on the SWE-Bench coding benchmark, has successfully passed practical engineering interviews from leading AI companies, and has even completed real jobs on Upwork.
Devin is an autonomous agent that solves engineering tasks through the use of its own shell, code editor, and web browser.
When evaluated on the SWE-Bench benchmark, which asks an AI to resolve GitHub issues found in real-world open-source projects, Devin correctly resolves 13.86% of the issues unassisted, far exceeding the previous state-of-the-art model performance of 1.96% unassisted and 4.80% assisted.
Check out what Devin can do in the thread below.
4
476
A plethora of wonderful ideas/tricks/resources related to self-learning: metacademy.org/roadmaps/rgro….
4
Come see our poster at ICLR 2018! :)
We decompose the variance of a policy gradient estimator with an action dependent baseline which provides insights into previous methods and new opportunities for improvements. Workshop poster #2 at 11am today. @suryabhupa @shanegu @svlevine #ICLR2018
4
Such a wonderful read rec'd by @josephwandile: paulgraham.com/hs.html by @paulg -- would've been lovely to have read this years ago.
2
2
Press release of some of the work in neural program synthesis I did with researchers at @MSFTResearch: microsoft.com/en-us/research….
2
#Caffe2's finally been released! My comments and TODOs from my last summer on AML are still there haha. Wonderful job to Yangqing et. al.!
1
over time the body responds by building muscle in that area to compensate for its usage, very akin to hebbian-like learning
1
Congrats @cloudera for the IPO! Looking forward to a productive future. :)
Today is an important day in the life of Cloudera j.mp/2qf22BB via @MikeOlson on the VISION blog
1
Check out some of new neural network quizzes out on @brilliantorg that I helped write! brilliant.org/explorations/a…
1
Some from his YT channel: piped.video/channel/UCzpDWSO…
1
1