As artificial intelligence systems grow more powerful, questions of ethics and safety grow as well. How can we ensure AI aligns with human values as it becomes deeply integrated into society? Advancing AI Responsibly argues for a holistic approach. Technical safety measures like verification and robustness must be paired with ethical training data and diverse development teams. Ongoing policy conversations are needed to shape norms and regulations around emerging capabilities.
Education has a huge role to play in creating realistic expectations. Below I have selected four models of academic centered AIs’. At this publication they are all open-sourced models.
Semantic Scholar is an AI-powered academic search engine developed by the Allen Institute for Artificial Intelligence. It aims to organize published scientific literature and make it more accessible. The system uses natural language processing and machine learning algorithms to analyze millions of research papers and extract key information such as topics, methods, datasets, metrics, and authors. It structures this information into knowledge graphs that capture relationships between various research concepts. Users can search academic papers by keywords or phrases and quickly get to the most relevant results. Key capabilities include automatically generating summaries of papers, finding connections between disparate research areas, and recommending papers based on a user's areas of interest. By enhancing discovery and understanding of scientific literature, AI Semantic Scholar seeks to accelerate research and advance science. The tool is free to use, and the database encompasses over 175 million papers across a wide range of disciplines including computer science, biomedicine, social sciences and more.
SciSpace aims to accelerate scientific breakthroughs by creating an AI-powered workspace for discovering, understanding, and collaborating on research. Its database contains metadata on over 200 million papers and 50 million open access PDFs across scientific domains. For any paper, SciSpace provides explanations, answers, and connections to relevant research using natural language processing. This automates repetitive tasks and simplifies discovery so researchers can work efficiently. SciSpace was founded in 2015 as Typeset, initially focusing on research formatting before expanding into a full research platform. Adopting the new SciSpace brand in 2022 reflected a commitment to making scientific knowledge freely accessible and collaborative. By designing features tailored to research workflow needs, SciSpace seeks to remove friction from the research process so scientists can uncover answers to humanity's challenges faster.
Elicit is an artificial intelligence system developed by Anthropic to conduct natural language conversations. It is designed to be helpful, harmless, and honest through self-supervision techniques based on constitutional AI principles. The system utilizes transformers and memory networks to contextually respond to open-ended prompts with relevant knowledge. A core technique involves conversational elicitation where the AI assistant asks clarifying questions to better understand the user's intent before responding. This enables more coherent and meaningful dialogue. The self-supervised learning approach allows AI Elicit to dynamically expand its knowledge by consuming diverse unlabeled text across the internet. A censorship module filters out inappropriate content to align with human values. By optimizing for constitutional objectives, transparency, and robustness during training, the goal is to create an AI that is universally helpful to all people. The system is currently available in a research preview as Anthropic plans to make it more widely accessible.
Consensus is a search engine that surfaces relevant scientific claims from research papers to answer user queries. It utilizes a database of over 200 million academic papers across all scientific domains from Semantic Scholar. A proprietary claim extraction model identifies salient statements of findings from these papers. User queries undergo keyword matching to narrow down potentially relevant claims. Vector search and metadata filters further refine results. A final question answering model ranks claims based on how well they address the query. Consensus combines neural information retrieval with custom fine-tuned language models optimized for surfacing evidence-based answers from research. By synthesizing claims from across papers, Consensus aims to enhance discovery and understanding of scientific knowledge.
