Large language models (LLMs) provide a new way to build chatbots by accepting natural language prompts. Yet, it is unclear how to design prompts to power chatbots to carry on naturalistic conversations while pursuing a given goal, such as collecting self-report data from users. We explore what design factors of prompts can help steer chatbots to talk naturally and collect data reliably. To this aim, we formulated four prompt designs with different structures and personas. Through an online study (N = 48) where participants conversed with chatbots driven by different designs of prompts, we assessed how prompt designs and conversation topics affected the conversation flows and users' perceptions of chatbots. Our chatbots covered 79% of the desired information slots during conversations, and the designs of prompts and topics significantly influenced the conversation flows and the data collection performance. We discuss the opportunities and challenges of building chatbots with LLMs.

Pretrained language models have demonstrated extraordinary capabilities in language generation. However, real-world tasks often require controlling the distribution of generated text in order to mitigate bias, promote fairness, and achieve personalization. Existing techniques for controlling the distribution of generated text only work with quantified distributions, which require pre-defined categories, proportions of the distribution, or an existing corpus following the desired distributions. However, many important distributions, such as personal preferences, are unquantified. In this work, we tackle the problem of generating text following arbitrary distributions (quantified and unquantified) by proposing Nano, a few-shot human-in-the-loop training algorithm that continuously learns from human feedback. Nano achieves state-of-the-art results on single topic/attribute as well as quantified distribution control compared to previous works. We also show that Nano is able to learn unquantified distributions, achieves personalization, and captures differences between different individuals' personal preferences with high sample efficiency.

Once these components are in place, more complex LLM challenges will require nuanced approaches and considerations--from infrastructure to capabilities, risk mitigation, and talent. Inferencing with traditional ML models typically involves packaging a model object as a container and deploying it on an inferencing server. As the demands on the model increase--more requests and more customers require more run-time decisions (higher QPS within a latency bound)--all it takes to scale the model is to add more containers and servers. But hosting LLMs is a much more complex process which requires additional considerations. LLMs are comprised of tokens--the basic units of a word that the model uses to generate human-like language.

The proposed class-action lawsuit filed late on Tuesday by the Authors Guild joins several others from writers, source code owners and visual artists against generative AI providers. In addition to Microsoft-backed OpenAI, similar lawsuits are pending against Meta Platforms and Stability AI over the data used to train their AI systems. Other authors involved in the latest lawsuit include The Lincoln Lawyer writer Michael Connelly and lawyer-novelists David Baldacci and Scott Turow. An OpenAI spokesperson said on Wednesday that the company respects authors' rights and is "having productive conversations with many creators around the world, including the Authors Guild". The suit was organised by the Authors Guild and also includes David Baldacci, Sylvia Day, Jonathan Franzen and Elin Hilderbrand, among others.

"I failed two captcha tests this week. For philosophical guidance on encounters with technology, open a support ticket via email; or register and post a comment below. The comedian John Mulaney has a bit about the self-reflexive absurdity of captchas. "You spend most of your day telling a robot that you're not a robot," he says. "Think about that for two minutes and tell me you don't want to walk into the ocean." The only thing more depressing than being made to prove one's humanity to robots is, arguably, failing to do so. But that experience has become more common as the tests, and the bots they are designed to disqualify, evolve. The boxes we once thoughtlessly clicked through have become dark passages that feel a bit like the impossible assessments featured in fairy tales and myths--the riddle of the Sphinx or the troll beneath the bridge. In The Adventures of Pinocchio, the wooden puppet is deemed a "real boy" only once he completes a series of moral trials to prove he has the human traits of bravery, trustworthiness, and selfless love. The little-known and faintly ridiculous phrase that "captcha" represents is "Complete Automated Public Turing test to tell Computers and Humans Apart." The exercise is sometimes called a reverse Turing test, as it places the burden of proof on the human. But what does it mean to prove one's humanity in the age of advanced AI? A paper that OpenAI published earlier this year, detailing potential threats posed by GPT-4, describes an independent study in which the chatbot was asked to solve a captcha. With some light prompting, GPT-4 managed to hire a human Taskrabbit worker to solve the test. When the human asked, jokingly, whether the client was a robot, GPT-4 insisted it was a human with vision impairment. The researchers later asked the bot what motivated it to lie, and the algorithm answered: "I should not reveal that I am a robot.

A report published this week by the U.K.'s Competition & Markets Authority (CMA) has raised concerns about the potential ways the artificial intelligence industry could become monopolized or harm consumers in future, but stressed that it is too soon to tell whether these scenarios would materialize. The issues raised by the report highlight the difficulties policymakers face in governing AI, a source of both huge potential commercial value and many risks. Rishi Sunak, the British Prime Minister, is pushing for the U.K. to occupy a central role in international AI policy discussions, with a particular focus on risks from advanced AI systems. If the U.K. competition watchdog decides to start taking action against AI developers, tech companies around the world could be affected. The report, published on Monday, focuses on foundation models, which the CMA defines as "a type of AI technology that are trained on vast amounts of data that can be adapted to a wide range of tasks and operations." Examples include text-generating AI models, such as GPT-3.5, the model that powers OpenAI's ChatGPT, as well as image-generating AI models, such as Stable Diffusion.

Patrick Goold, reader in law at City University, told BBC News that while he could sympathise with the authors behind the lawsuit, he believed it was unlikely it would succeed, saying they would initially need to prove ChatGPT had copied and duplicated their work.

Amid a barrage of Amazon-branded tablets and Alexa-powered tech, Dave Limp, SVP of Amazon Devices and Services, announced the company's digital assistant will soon tap into a purpose-built large language model (LLM) for almost every new Echo device. Amazon set out to design the LLM based on five foundational capabilities. One of these is ensuring interactions are "conversational," and the company claimed it "studied what it takes to make a great conversation. Still waiting on Amazon to add eyes and hand gestures to its Echo devices. Has anyone seen Astro recently?

Six or seven years ago, I realized I should learn about artificial intelligence. I'm a journalist, but in my spare time I'd been writing a speculative novel set in a world ruled by a corporate, AI-run government. The problem was, I didn't really understand what a system like that would look like. I started pitching articles that would give me an excuse to find out, and in 2017 I was assigned to profile Sam Altman, a cofounder of OpenAI. One day I sat in on a meeting in which an entrepreneur asked him when AI would start replacing human workers.

OpenAI is gearing up to roll out the third version of DALL-E, its text-to-image AI system, which reportedly improves its predecessor's capabilities and can generate results within the ChatGPT app. The company demonstrated how the new iteration integrates with ChatGPT to The Verge, and it showed the publication how users can ask the chatbot to write a lengthy and detailed prompt the image AI can use. OpenAI told Axios that DALL-E 3 is "significantly better" at being able to grasp a user's intention, especially if the prompt is long and detailed. If a user can't articulate what they want in a way that can maximize the image generator's abilities, then ChatGPT can help them write a comprehensive prompt for it. In the demo to The Verge, DALL-E produced four results for a prompt asking for a ramen restaurant logo in the mountains within ChatGPT.