How ChatGPT and DALL-E affect children's minds. Virtual talk
Plus: Video lecture on recent research in AI alignment
Hello, fellow human! As usual, I’m sharing some of the virtual talks we’ve lined up to give you a glimpse into cutting-edge research.
Next Thursday, we’re hosting a talk with Eliza Kosoy, who will present her research from UC Berkeley and Google DeepMind on how children aged 5 to 12 perceive and interact with GenAI models like ChatGPT and DALL-E.
Her findings reveal that children generally view AI positively, often associating it with beneficial traits and showing a knack for imaginative queries, particularly with visual-based AI.
Join the virtual talk to learn more.
Just a reminder that this Thursday, I’ll be having a discussion with Alex Irpan, who spent the last 8 years at Google DeepMind working on robotics research. We’ll talk about the SOTA of robotics and why he believes it has reached an inflection point. We’ll also touch on his recent shift in research focus to AI safety.
The talk is virtual – join the conversation.
How can we align LLMs to the preferences of hundreds of millions of people who interact with them every day? With different conversational norms, value systems, and political beliefs, it's a complex challenge. The researchers behind the PRISM Alignment Dataset made an attempt to give AI a better understanding of human differences.
To give a sense of the dataset’s scale: 1,500 participants from 75 countries interacted with 11 open-source LLMs and 10 private LLMs, generating 68,371 utterance-feedback pairs. This is a great effort to align AI with diverse groups. The caveat — although participants came from different countries, they were all English-speaking and interacted with the LLMs in English, so non-English-speaking groups are not represented in the dataset.