This article is from Big Technology, a newsletter by Alex Kantrowitz. Life got interesting for Anthropic two weeks ago, when OpenAI nearly lit itself on fire. Anthropic had been operating comfortably in OpenAI's shadow, collecting billions in investment from Amazon, Google, and others as it developed similar technology with an increased focus on safety. Then, as the chaos rolled on, companies that built their products entirely on top of OpenAI's GPT-4 model looked for a hedge. And Anthropic was there, waiting for them.

Wednesday will be a fateful day in Brussels, a faraway city of which post-Brexit Britain knows little and cares less. It's the day on which the EU's AI proposals enter the final stages of a tortuous lawmaking process. The bill is a landmark (first in the world) attempt to seriously regulate artificial intelligence (AI) based on its capacity to cause harm and will soon be in the final phase of the legislative process – so-called "trilogues" – where the EU parliament, commission and council decide what should be in the bill, and therefore become part of EU law. However, the bill is now hanging in the balance because of internal disagreement about some key aspects of the proposed legislation, especially those concerned with regulation of "foundation" AI models that are trained on massive datasets. In EU-speak these are "general-purpose AI" (GPAI) systems – ones capable of a range of general tasks (text synthesis, image manipulation, audio generation and so on) – such as GPT-4, Claude, Llama etc.

Synthetic text plays a vital role in data augmentation, model robustness, privacy preservation and scenario analysis. It is usually formulated as conditional text generation where a given source document is transformed using substitutions, paraphrasing, back translation, mixups etc. [1] to obtain a modified document. We argue that conditioning on the unstructured text limits the ability to mix text fragments coherently and produces transformations that are not confined to essential information, a critical necessity for long-form text. Furthermore, explaining the generated text becomes challenging, particularly detecting hallucinations [2]. We propose here a decompose and expand technique to generate synthetic text, where the semantic frames [3] of a source document are first extracted, and this compact interim form is used to generate the transformed text.

Recent AI-assistant agents, such as ChatGPT, predominantly rely on supervised fine-tuning (SFT) with human annotations and reinforcement learning from human feedback (RLHF) to align the output of large language models (LLMs) with human intentions, ensuring they are helpful, ethical, and reliable. However, this dependence can significantly constrain the true potential of AI-assistant agents due to the high cost of obtaining human supervision and the related issues on quality, reliability, diversity, self-consistency, and undesirable biases. To address these challenges, we propose a novel approach called SELF-ALIGN, which combines principle-driven reasoning and the generative power of LLMs for the self-alignment of AI agents with minimal human supervision. Our approach encompasses four stages: first, we use an LLM to generate synthetic prompts, and a topic-guided method to augment the prompt diversity; second, we use a small set of human-written principles for AI models to follow, and guide the LLM through in-context learning from demonstrations (of principles application) to produce helpful, ethical, and reliable responses to user's queries; third, we fine-tune the original LLM with the high-quality self-aligned responses so that the resulting model can generate desirable responses for each query directly without the principle set and the demonstrations anymore; and finally, we offer a refinement step to address the issues of overly-brief or indirect responses. Applying SELF-ALIGN to the LLaMA-65b base language model, we develop an AI assistant named Dromedary. With fewer than 300 lines of human annotations (including < 200 seed prompts, 16 generic principles, and 5 exemplars for in-context learning). Dromedary significantly surpasses the performance of several state-of-the-art AI systems, including Text-Davinci-003 and Alpaca, on benchmark datasets with various settings.

Chat models, such as ChatGPT, have shown impressive capabilities and have been rapidly adopted across numerous domains. However, these models are only accessible through a restricted API, creating barriers for new research and progress in the field. We propose a pipeline that can automatically generate a high-quality multi-turn chat corpus by leveraging ChatGPT to engage in a conversation with itself. Subsequently, we employ parameter-efficient tuning to enhance LLaMA, an open-source large language model. The resulting model, named Baize, demonstrates good performance in multi-turn dialogues with guardrails that minimize potential risks. Furthermore, we propose a new technique called Self-Distill with Feedback, to further improve the performance of the Baize models with feedback from ChatGPT. The Baize models and data are released for research purposes only at https://github.com/project-baize/baize-chatbot. An online demo is also available at https://huggingface.co/spaces/project-baize/chat-with-baize.

Artificial intelligence is widely expected to transform our lives. Leaders from across the sector gathered for a TIME dinner conversation on Nov. 30, where they emphasized the need to center humans in decisions around incorporating the technology into workflows and advocated for governments and industry leaders to take a responsible approach to managing the risks the technology poses. As part of the TIME100 Talks series in San Francisco, senior correspondent Alice Park spoke with panelists Cynthia Breazeal, a pioneer in social robotics and the Dean for Digital Learning at MIT, James Landay, a computer science professor and vice director of the Institute for Human-Centered AI at Stanford University, and Raquel Urtasun, CEO and founder of self-driving tech startup Waabi, which recently put a fleet of trucks into service on Uber Freight's trucking network. The panelists discussed the ethical considerations of AI and the ways in which leaders can ensure its benefits reach every corner of the world. During the discussion, the three panelists highlighted the transformative journey of AI and delved into its profound implications, emphasizing the need for responsible AI deployment.

The introduction of large public legal datasets has brought about a renaissance in legal NLP. Many of these datasets are comprised of legal judgements - the product of judges deciding cases. This fact, together with the way machine learning works, means that several legal NLP models are models of judges. While some have argued for the automation of judges, in this position piece, we argue that automating the role of the judge raises difficult ethical challenges, in particular for common law legal systems. Our argument follows from the social role of the judge in actively shaping the law, rather than merely applying it. Since current NLP models come nowhere close to having the facilities necessary for this task, they should not be used to automate judges. Furthermore, even in the case the models could achieve human-level capabilities, there would still be remaining ethical concerns inherent in the automation of the legal process.

To train well-performing generalizing neural networks, sufficiently large and diverse datasets are needed. Collecting data while adhering to privacy legislation becomes increasingly difficult and annotating these large datasets is both a resource-heavy and time-consuming task. An approach to overcome these difficulties is to use synthetic data since it is inherently scalable and can be automatically annotated. However, how training on synthetic data affects the layers of a neural network is still unclear. In this paper, we train the YOLOv3 object detector on real and synthetic images from city environments. We perform a similarity analysis using Centered Kernel Alignment (CKA) to explore the effects of training on synthetic data on a layer-wise basis. The analysis captures the architecture of the detector while showing both different and similar patterns between different models. With this similarity analysis we want to give insights on how training synthetic data affects each layer and to give a better understanding of the inner workings of complex neural networks. The results show that the largest similarity between a detector trained on real data and a detector trained on synthetic data was in the early layers, and the largest difference was in the head part. The results also show that no major difference in performance or similarity could be seen between frozen and unfrozen backbone.

The evolution of legal datasets and the advent of large language models (LLMs) have significantly transformed the legal field, particularly in the generation of case judgment summaries. However, a critical concern arises regarding the potential biases embedded within these summaries. This study scrutinizes the biases present in case judgment summaries produced by legal datasets and large language models. The research aims to analyze the impact of biases on legal decision making. By interrogating the accuracy, fairness, and implications of biases in these summaries, this study contributes to a better understanding of the role of technology in legal contexts and the implications for justice systems worldwide. In this study, we investigate biases wrt Gender-related keywords, Race-related keywords, Keywords related to crime against women, Country names and religious keywords. The study shows interesting evidences of biases in the outputs generated by the large language models and pre-trained abstractive summarization models. The reasoning behind these biases needs further studies.

This paper presents the first dataset for Japanese Legal Judgment Prediction (LJP), the Japanese Tort-case Dataset (JTD), which features two tasks: tort prediction and its rationale extraction. The rationale extraction task identifies the court's accepting arguments from alleged arguments by plaintiffs and defendants, which is a novel task in the field. JTD is constructed based on annotated 3,477 Japanese Civil Code judgments by 41 legal experts, resulting in 7,978 instances with 59,697 of their alleged arguments from the involved parties. Our baseline experiments show the feasibility of the proposed two tasks, and our error analysis by legal experts identifies sources of errors and suggests future directions of the LJP research.