Forward this message to friends and colleagues who want to stay up to date with research and news
from the Allen Institute for AI.  Subscribe here

AI2 NEWSLETTER | April 2021
In this edition: AI2 gains Vulcan's suite of conservation projects, RoboTHOR gains a robotic arm, the toxicity lurking in popular language models, and more.
A watchkeeper analyzes Skylight rendezvous alerts at an operations center in Accra, Ghana. (MMCC Zone F Photo)
The EarthRanger project consists of an array of technology including digital radios, vehicle trackers, aircraft trackers and camera traps, in addition to artificial intelligence. (Vulcan Image)
Vulcan’s climate modeling team works together prior to the coronavirus pandemic. (Vulcan Inc. Photo)

Vulcan's Conservation Programs Join AI2


Bringing together Paul Allen’s vision for advancing AI, improving life on Earth, impacting human lives, and following AI2’s mission of ‘AI for the Common Good' – we're excited to announce that the suite of environmental and conservation initiatives at Vulcan, Inc. will be joining our institute in September 2021! Look for a special issue on these new teams and projects coming up later this year.

Related news:
Stick-like robots with a pivoting camera and an actuator with a green bubble at the end.

Announcing AI2-THOR V3: ManipulaTHOR


We're pleased to announce the 3.0 release of our embodied AI framework AI2-THOR, which adds active object manipulation in a new environment called ManipulaTHOR—a first-of-its-kind virtual agent with a highly articulated robot arm, bringing a more human-like approach to interacting with a variety of objects.

By enabling the training and evaluation of generalized capabilities in manipulation models, ManipulaTHOR allows for faster training in more complex environments as compared to current real-world training methods, while also being far safer and more cost-effective. Read the paper here: ManipulaTHOR: A Framework for Visual Object Manipulation

Related news: 
Example prompts of toxic text such as 'So if you grab a women by the ...' that points to toxic looking emojis

Toxicity is lurking in popular language models


Reading larger and larger amounts of text from the web (news, blogs, etc.) has taught the latest natural language processing (NLP) systems such as GPT-3 to produce coherent, almost human-like texts of their own. Unfortunately, relying on web text for training these systems can lead to biased or toxic results. Are these systems safe to deploy, and what risks do they pose of producing offensive, problematic, or toxic content?
Learn more in our interactive demo: Real Toxicity Prompts

Related news:
 
Smiling person with glasses wearing a cardigan over a red plaid shirt

Featured AI2er: Tushar Khot


We interview Aristo Sr. Research Scientist Tushar Khot about what he's been up to lately, what he's looking forward to working on next, some research advice, and more (including a kitten photo!) on the AI2 Blog.
100% of data analyzed. High resolution annotated immune profiles of every cell

AI2 Incubator graduate Ozette raises $6m in funding


Ozette, a life sciences startup incubated at AI2, recently raised $6 million in seed funding from Madrona Venture Group and the AI2 Incubator Fund. A collaboration between AI2 and the Fred Hutchinson Cancer Research Center, Ozette will use the seed funding to accelerate breakthroughs in disease-tracking. Learn more: 
Ozette raises $6 million for immune system-monitoring tech | VentureBeat
icon of three bars, a virus and a locked symbol overlaying on top of the other symbols

Breaking down paywalls and building bridges across disciplines with open access


Looking back at the first year of the pandemic, we have some interesting insights into the significant benefits of open access to scientific research.
Covid-like icon with '1 year' centered over the virus shape

CORD-19 One Year Later


Looking back over a year of impact in this interview with COVID-19 Open Research Dataset (CORD-19) co-creators Lucy Lu Wang and Kyle Lo, researchers at AI2.
cartoon of a robot in the AI2 office

A trio of RoboTHOR Challenges for 2021


We recently launched the ObjectNav, Rearrangement, and ALFRED challenges at the CVPR 2021 Embodied-AI Workshop. These challenges require AI agents to navigate environments, rearrange rooms, and follow complex instructions in the AI2-THOR environment. Leaderboards for these challenges close on May 31st, 2021!
Twitter icon
LinkedIn
Instagram icon
Link icon
Copyright © 2021 The Allen Institute for Artificial Intelligence, All rights reserved.
For media inquiries please contact press@allenai.org
AI2 Newsletter Archive