While much research has explored enhancing the reasoning capabilities of large language models (LLMs) in the last few years, there is a gap in understanding the alignment of these models with social values and norms. We introduce the task of judging social acceptance. Social acceptance requires models to judge and rationalize the acceptability of people's actions in social situations. For example, is it socially acceptable for a neighbor to ask others in the community to keep their pets indoors at night? We find that LLMs' understanding of social acceptance is often misaligned with human consensus. To alleviate this, we introduce SocialGaze, a multi-step prompting framework, in which a language model verbalizes a social situation from multiple perspectives before forming a judgment. Our experiments demonstrate that the SocialGaze approach improves the alignment with human judgments by up to 11 F1 points with the GPT-3.5 model. We also identify biases and correlations in LLMs in assigning blame that is related to features such as the gender (males are significantly more likely to be judged unfairly) and age (LLMs are more aligned with humans for older narrators).

A large body of work in psycholinguistics has focused on the idea that online language comprehension can be shallow or `good enough': given constraints on time or available computation, comprehenders may form interpretations of their input that are plausible but inaccurate. However, this idea has not yet been linked with formal theories of computation under resource constraints. Here we use information theory to formulate a model of language comprehension as an optimal trade-off between accuracy and processing depth, formalized as bits of information extracted from the input, which increases with processing time. The model provides a measure of processing effort as the change in processing depth, which we link to EEG signals and reading times. We validate our theory against a large-scale dataset of garden path sentence reading times, and EEG experiments featuring N400, P600 and biphasic ERP effects. By quantifying the timecourse of language processing as it proceeds from shallow to deep, our model provides a unified framework to explain behavioral and neural signatures of language comprehension.

Communicating science and technology is essential for the public to understand and engage in a rapidly changing world. Tweetorials are an emerging phenomenon where experts explain STEM topics on social media in creative and engaging ways. However, STEM experts struggle to write an engaging "hook" in the first tweet that captures the reader's attention. We propose methods to use large language models (LLMs) to help users scaffold their process of writing a relatable hook for complex scientific topics. We demonstrate that LLMs can help writers find everyday experiences that are relatable and interesting to the public, avoid jargon, and spark curiosity. Our evaluation shows that the system reduces cognitive load and helps people write better hooks. Lastly, we discuss the importance of interactivity with LLMs to preserve the correctness, effectiveness, and authenticity of the writing.

In a giant warehouse in Reading, Massachusetts, I meet a pair of robots that look like goofy green footstools from the future. Their round eyes and satisfied grins are rendered with light emitting diodes. They sport small lidar sensors like tiny hats that scan nearby objects and people in 3D. Suddenly, one of them plays a chipper little tune, its mouth starts flashing, and its eyes morph into heart shapes. This means, I am told, that the robot is happy.

In "A Lot of Things Have Happened," your story in this week's issue, the narrator remembers an old girlfriend through a series of events and coincidences--her fear of palmetto bugs is recalled by way of the narrator's new house in Florida; her congratulations to him and his new wife are recalled by way of parallel stories about rodents; her sister's death is recalled by way of an apology the narrator once extorted from a student. In a story without a linear, driving narrative, how do you go about parsing out the inciting events? The shortest, most honest answer here is: accidentally. The slightly fancier-sounding version of that answer is: through a process of trial and error. For whatever reason, I've been drawn to ellipsis and anecdote lately and become more impatient with artful transition.

In his talk yesterday at COSM 2021, venture capitalist and philanthropist Peter Thiel -- the ultimate Silicon Valley insider, prophet, and sometimes needed gadfly -- offered a cold shower for transhumanism, The Singularity, the computers we will supposedly merge with by 2030, and all that. Those things, he thinks, are uncertain. We should worry about what's happening now in everyday time, to which, in his view, too few are paying heed: The growth of total AI-based surveillance and the disappearance of privacy. Thiel considers arguments about whether computers that think like people will ever be developed to be "above his pay grade." Given that he is reputed to be worth $3.7B dollars, that's a polite way of saying that such arguments are a pleasant waste of time.

I recently wrote the book that I wish existed when I was introduced to machine learning: Human-in-the-Loop Machine Learning: Active Learning and Annotation for Human-Centered AI. Most machine learning models are guided by human-annotated data, but most machine learning books and courses focus on algorithms. You can often get state-of-the-art results with good data and simple algorithms, but you rarely get state-of-the-art results from the best algorithm with bad data. So if you need to go deep in one area of machine learning first, you could argue that the data side is more important. In addition to the technical focus of the book, it features anecdotes from 11 machine learning experts. Each shared an anecdote about data-related problems they encountered building and evaluating machine learning models in real-world situations. Their stories tell us something important about machine learning leadership more broadly, with each anecdote tying into a lesson about running successful data science projects.

One billion hours, veteran game designer Sid Meier notes in this light and enjoyable memoir, is an unfathomable length of time. And yet it took just six years for players to spend a billion cumulative hours on the fifth iteration of Meier's engrossing Civilization series, a nation-building game that has seen them shepherding their peoples from the foundation of their first city in 4000BC to an eventual victory through military, cultural or scientific might, millions of times over. What makes the Civilzation games so compelling? The fans are a cut above; I'd know as I am one of them, having racked up 530 hours on Civilization VI since I bought it at the beginning of lockdown. For Meier, good game design comes down, at its core, to a series of "interesting decisions".

It's common knowledge that Barack Obama met the woman who eventually became his wife, Michelle Robinson, when he came to work at her law firm as a summer associate. George W. Bush met the future Mrs. Bush, who was Laura Welch back then, at a barbecue and took her mini-golfing the next day. And we all remember that Bill and Hillary Clinton were law school sweethearts. The historical record is full of these president-and-first-lady origin stories: Harry Truman was just 6 when he met the woman he would go on to marry, in church. So it's only natural to ask how the current crop of presidential candidates' how-they-met stories stack up.

This was a large applied research effort that presented many opportunities for unusual experiences. In one such experience, I was called in, at the last minute, to help improve our ALV proposal. The proposal was a 300-page document that segued smoothly from problem description to corporate capabilities and managerial plan, omitting any mention of technical approach. This taught me a rule of thumb I have seen validated many times: the larger the project (in dollars and scope), the poorer the technical proposal. In a second experience, I was demonstrating a dynamic programming algorithm at a quarterly review.