As text generated by large language models proliferates, it becomes vital to understand how humans engage with such text, and whether or not they are able to detect when the text they are reading did not originate with a human writer. Prior work on human detection of generated text focuses on the case where an entire passage is either human-written or machine-generated. In this paper, we study a more realistic setting where text begins as human-written and transitions to being generated by state-of-the-art neural language models. We show that, while annotators often struggle at this task, there is substantial variance in annotator skill and that given proper incentives, annotators can improve at this task over time. Furthermore, we conduct a detailed comparison study and analyze how a variety of variables (model size, decoding strategy, fine-tuning, prompt genre, etc.) affect human detection performance. Finally, we collect error annotations from our participants and use them to show that certain textual genres influence models to make different types of errors and that certain sentence-level features correlate highly with annotator selection. We release the RoFT dataset: a collection of over 21,000 human annotations paired with error classifications to encourage future work in human detection and evaluation of generated text.

TL;DR: We give a threat model literature review, propose a categorization and describe a consensus threat model from some of DeepMind's AGI safety team. See our post for the detailed literature review. The DeepMind AGI Safety team has been working to understand the space of threat models for existential risk (X-risk) from misaligned AI. Our aim was to clarify the case for X-risk to enable better research project generation and prioritization. First, we conducted a literature review of existing threat models, discussed their strengths/weaknesses and then formed a categorization based on the technical cause of X-risk and the path that leads to X-risk.

Glasp is a social online highlighter that allows users to annotate and organize web-based quotations and ideas in one place, share their learning, and access the annotated content of others with similar interests. In June of 2021, the Glasp team set out to create a platform that would make it easier for anybody to benefit from the knowledge and expertise of others. Glasp's ultimate goal is to make it possible to quickly and easily access all information from the highlighted parts of the world. You may highlight the original page in one of four colors with the help of the Glasp browser plugin (available for Chrome, Safari, Brave, Edge, Opera, and more). As well as importing notes and highlights from other services like Obsidian, Roam Research, Notion, and Readwise, you can export them to .txt,.md, .html,

ChatGPT is a powerful new tool that is making waves in the world of business and technology. Its advanced natural language processing (NLP) capabilities enable it to deliver a seamless, human-like conversation experience to users. It is based on the GPT-3 (Generative Pretrained Transformer-3) architecture, which is a state-of-the-art language model developed by OpenAI. OpenAI is a research organization that aims to promote and advance artificial intelligence in a responsible and safe manner. It was founded in 2015 by a group of entrepreneurs, researchers, and philanthropists who wanted to advance the field of AI and make it more widely available.

One potential approach would be to use advanced propulsion systems, such as antimatter engines or fusion drives, to reach high speeds and traverse the vast distances of interstellar space. These propulsion systems would require a reliable source of energy, which could potentially be provided through the use of advanced power generation technologies, such as matter-antimatter annihilation or fusion reactions.

Two days ago OpenAI released ChatGPT, a new language model which is an improved version of GPT-3 and, possibly, gives us a peek into what GPT-4 will be capable of when it is released early next year (as is rumoured). With ChatGPT it is possible to have actual conversation with the model, referring back to previous points in the conversation. I wanted to try out if I could use this model as a pair programmer that I can give some instructions and it produces the code for me. I would still double-check those code snippets, of course, but at least I won't have to write them from scratch anymore. So in this blog post I describe how I used ChatGPT to create a simple sentiment analysis application from scratch.

Most stock quote data provided by BATS. US market indices are shown in real time, except for the S&P 500 which is refreshed every two minutes. Chicago Mercantile: Certain market data is the property of Chicago Mercantile Exchange Inc. Dow Jones: The Dow Jones branded indices are proprietary to and are calculated, distributed and marketed by DJI Opco, a subsidiary of S&P Dow Jones Indices LLC and have been licensed for use to S&P Opco, LLC and CNN. Standard & Poor's and S&P are registered trademarks of Standard & Poor's Financial Services LLC and Dow Jones is a registered trademark of Dow Jones Trademark Holdings LLC.

It's difficult to reflect on the past year--or forecast the next--without a sense of wonder regarding the sheer magnitude of innovation taking place across the AI landscape. On a weekly basis, researchers across industry and academia have published work advancing the state-of-the-art in nearly every domain of AI, toppling benchmarking leaderboards and accomplishing feats beyond what we could have imagined even a few years ago. In large part, this progress is due to the rapid advancements we've seen in large AI models. Recent progress in supercomputing techniques and new applications of neural network architectures have allowed us to train massive, centralized models that can accomplish a wide variety of tasks using natural language inputs--from summarizing and generating text with unprecedented levels of sophistication, to even generating complex code for developers. The combination of large language models and coding resulted in two of the most powerful AI developments we witnessed in 2022: the introduction of the OpenAI Codex Model--a large AI model that can translate natural language inputs into more than a dozen programming languages--and the launch of GitHub Copilot, a programming assistant based on Codex.

The recent media coverage of a reasonably intelligent bot able to hold a conversation on just about any subject suggests that this is a major breakthrough, but like many others who've been involved in the development and adoption of machine learning tools for many years, I'd like to point out that there are dozens of companies out there working on easy-to-use software…

Reading this article, one can notice small nuances that show the whole text was written via software. My personal two cents on the topic of automated content creation is that AI can be used as a source of inspiration, or as a basic structure idea. AI can be a great tool for brainstorming, and getting an essential outline of the issues you want to talk or write about, but it will not become the main source of creation. Nothing beats creativity and experience, AI models react from input to create an output. Essentially, AI can be a great and useful tool for overcoming writer's block and helping with content creation.