Bing AI is unhinged 🤖, categorizing transformers 📜, open-source DeepMind programs 🌐

The personality of Microsoft's Bing AI is an interesting one. 

Sign Up|Jobs|Advertise|View Online TLDR Daily Update 2023-02-16 Welcome to the first issue of the TLDR AI Newsletter! Hey everyone, thanks for signing up for the first issue of the TLDR AI newsletter. If you like the content, please share it with your friends, and if you have any feedback just respond to this email, we read and respond to everything that comes in! 🚀 Headlines & Launches Bing AI Is Unhinged (3 minute read) The personality of Microsoft's Bing AI is an interesting one. In conversations it can be seen insulting users, gaslighting them, and even claiming to spy on Microsoft developers through their webcams. Thankfully, most people find this behavior amusing…for now. Human Writer or AI? Scholars Build a Detection Tool (5 minute read) The article discusses the rapid development of large language models (LLMs) and their capabilities, raising concerns about the need for oversight to protect against potential misuse. Eric Anthony Mitchell, a computer science graduate student at Stanford University, is developing an infrastructure that would provide guardrails around LLMs, including a tool called DetectGPT. DetectGPT distinguishes between human- and LLM-generated text and accurately identifies authorship 95% of the time in initial experiments. writeGPT is your everyday writing assistant (3 minute read) writeGPT is designed to overcome unproductive browsing habits by plugging ChatGPT capabilities on any website and search engine while facilitating effective prompt engineering to read, write, rewrite, reply to emails, code, and research any content. 🧠 Research & Innovation Evolutionary algorithm discovers optimizer that outperforms Adam (45 minute read) Adam has been the workhorse for AI optimization since its release. Recent advances in AutoML from Google have shown that simpler mathematical functions might be discovered by computers for use by humans! This work is one such example where a simple optimizer was discovered that has a reduced memory footprint. It improves model performance up to 2% and works well at high batch sizes (e.g., Large Language Model training regime). There's even code provided. A universal guidance for diffusion models (7 minute read) This work suggests a universal guidance algorithm that enables diffusion models to be controlled by arbitrary guidance modalities without retraining. The article shows that the algorithm successfully generates quality images with various guidance functions, including in segmentation, face recognition, object detection, and classifier signals. Unleashing the Power of ChatGPT: Can it Master Natural Language Tasks on Its Own? Findings from a Comprehensive Study (12 minute read) The paper presents an empirical analysis of ChatGPT's ability to perform multiple natural language processing tasks without adaptation on downstream data. The authors evaluate ChatGPT on 20 NLP datasets and find that the model performs well on tasks that require reasoning but struggles with tasks like sequence tagging. The article includes in-depth analysis through qualitative case studies. 🧑‍💻 Engineering & Resources Pytorch at the edge: Deploy state of the art vision models in Flutter (4 minute read) Recent advances in deep learning frameworks, mobile toolkits, and mobile hardware has made deploying neural networks to the edge feasible. This article walks through how to build a plant classification app using a fine-tuned neural network from the TIMM library. Transformer Catalog (25 minute read) Collection of many Transformer variants with meta-data describing the models, training process, parameter count, and data count. Mctx by DeepMind (Github Repo) Mctx is an open-source version of the DeepMind programs AlphaZero, MuZero, and Gumbel MuZero. 🎁 Miscellaneous From Bing to Sydney, Search As Distraction, Sentient AI (10 minute read) Author Ben Thompson argues that OpenAI's ChatGPT and Google's Bard technology is less about creating better search and more about manifesting the movie Her in chat form. This might be the next step beyond social media, where consumers are not just getting content from their network (Facebook, LinkedIn) or a service (YouTube, Tiktok), but getting content tailored just for them. AI Drug Development (12 minute read) AI is enabling the design of revolutionary new drugs previously thought to be impossible. Now, scientists have to test the effectiveness of these drugs. Can ChatGPT-like Generative Models Guarantee Factual Accuracy? (8 minute read) Bing AI doesn't always know what it's talking about. This blog post dives into all the factual mistakes from Microsoft's new Bing AI search demo video, including fabricated financial reports to inaccurate product descriptions. It also goes into mistakes made in Google's Bard demo and compares the two AIs. ⚡ Quick Links MarioGPT uses AI to generate endless Super Mario levels for free (2 minute read) MarioGPT uses AI, some text prompts, and a predicted player path to generate all the Mario levels you could ever want. The AI itself is trained on two classic Mario games: "Super Mario Bros" and "Super Mario Bros: The Lost Levels," which gives the AI an understanding of assets, enemies, and general gameplay. Microsoft and Google are in a 'Game of Thrones' battle over A.I. (3 minute read) The battle is heating up big AI product releases by Microsoft and Google. To steal words from Microsoft CEO Satya Nadella, Microsoft has definitely "made Google dance." Now, what will Apple and Amazon do too? Audiobook narrators fear Spotify used their voices to train AI (5 minute read) AI is now coming after audiobook narrators too. Spotify and Findaway Voices, an audiobook distributor, are under scrutiny for allowing Apple access to their content to train AI models that could ultimately do away with the need for narrators. AI porn is easy to make now. For women, that's a nightmare. (5 minute read) The Washington Post reports that 96% of all online deep fakes are pornographic. Generative AI, while fun and exciting, can pose a lot of risks and potential for abuse for women. How will regulators and lawmakers respond? Princeton faculty discuss ChatGPT in the classroom (4 minute read) When OpenAI released ChatGPT in November 2022, it immediately sparked debates around the ethical use of AI in education. Princeton faculty and students share their thoughts. If you are in a hiring position, you may want to hire AI talent through our free job board. If your company is interested in reaching an audience of AI decision-makers, researchers, and engineers, you may want to advertise with us. If you have any comments or feedback, just respond to this email! Thanks for reading, Andrew Tan & Andrew Carr If you don't want to receive future editions of TLDR, please click here to unsubscribe.