Our new collaborative LM initiative, toxicity in ChatGPT, enhanced citations in Semantic Scholar, and more from AI2

Subscribe

May 2023

AI2 Newsletter

Top Story

Announcing AI2 OLMo, an open language model made by scientists, for scientists 

AI2 OLMo - Open Language Model

Today, the Allen Institute for AI is excited to announce that we are embarking on the creation of an open, state-of-the-art generative language model: AI2 OLMo (Open Language Model).  

 

We believe that millions of people want to better understand and engage with language models, and we aim to create the environment where they actually can, leading to faster and safer progress for everyone. Our goal is to collaboratively build the best open language model in the world.

Learn more ➞

Toxicity In ChatGPT

Large language models and chatbots are everywhere, but are they safe? In a large-scale toxicity analysis of ChatGPT, AI2 Aristo team members find that assigning the bot a “persona” significantly increases its tendency to produce toxic output (up to 6X!). This study highlights the importance of AI literacy and model transparency as more people encounter chatbots day-to-day.

ChatGPTToxicity Cropped
Read the blog ➞

Why AI is Incredibly Smart—and Shockingly Stupid

Artificial intelligence has come a long way, but popular understanding of AI often overestimates just how capable models are. AI2 Senior Director Yejin Choi joined the TED2023 conference to highlight where AI is, and where she believes it should go.

Yejin Choi: Why AI is incredibly smart and shockingly stupid

Making Citations Come Alive with CiteSee

Inline citations help scholars explore and discover prior literature, but it can be challenging to make sense of them and prioritize them during reading. That’s why researchers from Semantic Scholar created CiteSee, a tool to better interface with citations as you encounter them by providing useful, personalized context. The CiteSee paper received a Best Paper Award at CHI 2023!

Learn more ➞
Screenshot 2023-04-26 at 11.12.36 AM

An Analysis of Chatbot Data

What’s really in the huge datasets used to train the most popular chatbots? The Washington Post
 teamed up with AI2 team members Jesse Dodge, Yanai Elazar, Dirk Groeneveld, and Nicole DeCario to analyze the contents of Google’s C4 dataset.

Learn more ➞
ReViz

Research Visualization at AI2

The research visualization team at AI2 creates beautiful, informative, interactive applications and demos that help people to understand, share, and build upon our institute’s research. Check out the new website for this important cross-functional team.

Learn more ➞

More from us

➞ What’s AGI, and Why Are AI Experts Skeptical?

➞ The astounding new era of AI: Notes on Session 2 of TED2023

➞ The Semantic Reader Project: AI-Powered Interactive Reading

➞ Inspired by the “pure magic” of AI, Allen School undergraduate Matt Deitke begins his next act

learn more about the allen institute of artificial intelligence

Work with us

AI2 Newsletter archive

Twitter
LinkedIn
YouTube
Website