This AI newsletter is all you need (#36)

Author(s): Towards AI Editorial Team Originally published on Towards AI. What happened this week in AI by Louis This week we were pleased to note an acceleration in progress toward open-source alternatives to ChatGPT as well as signs of increased flexibility in access to these models. While many major tech companies are building their own alternative to ChatGPT, we are particularly excited to see open-source alternatives that can make next-generation LLM models more accessible, flexible, and affordable for the machine learning community. We are seeing progress in the release of more open-source foundational LLM models as well as building out of open-source data sets and workflows, together with new innovations to reduce the cost of fine-tuning these models with human feedback. The strategic partnership between Hugging Face and Amazon Web Services (AWS) looks like a positive step in this direction and should increase the availability of open-source data sets and models hosted on Hugging Face. We were also pleased to see the release of Meta’s LLaMA, 4 foundation models ranging from 7B to 65B parameters. Another example is Laion’s project, which involves crowdsourcing annotations for its OpenAssistant ChatGPT replication project. Carper has also developed open-source RLHF workflows, which range from human annotation with CHEESE to RLHF training with trlX. We are also seeing new approaches that can reduce the human feedback requirements and barriers to entry to developing a ChatGPT-like product — such as Anthropic AI’s constitutional AI approach (behind its Claude model) requiring minimal human labels. We hope these open-source models and competition can also put pressure on OpenAI to keep costs affordable and to increase the flexibility of interaction with their models. We see signs of this with a glimpse leaked of its new OpenAI Foundry product — a platform for running OpenAI models on a dedicated capacity. This platform will also allow more robust fine-tuning options for its latest models. Also of note in this release is that the future models (GPT-4?) look to offer 32,000 max context length (up from 4,000 today) which will likely bring a lot of new capabilities. Hottest News OpenAI Foundry will let customers buy dedicated compute to run GPT3 and their other models OpenAI is launching a new developer platform that lets customers run the company’s newer machine learning models, like GPT-3.5, on a dedicated capacity. OpenAI describes the forthcoming offering, called Foundry, as “designed for cutting-edge customers running larger workloads.” 2. What ChatGPT And Generative AI Mean For Your Business? Generative AI could potentially become a powerful tool for businesses, providing a new basis for competitive advantage. Enterprises should consider experimenting with generative AI by identifying existing processes that can be enhanced with this technology. 3. What are ‘robot rights,’ and should AI chatbots have them? This article features an interview with Professor David Gunkel, discussing the issue of what rights robots, including AI chatbots, should have. The discussion centers on the concept of robot rights, including the background and articulation of rights for AI. 4. Hugging Face and AWS partner to make AI more accessible The strategic partnership between Hugging Face and Amazon Web Services (AWS) is expected to make AI open and accessible to everyone. Together, the two companies aim to accelerate the availability of next-generation machine learning models by making them more accessible, efficient, and affordable for the machine learning community. 5. How AI Can Help Create and Optimize Drugs To Treat Opioid Addiction The use of artificial intelligence for drug discovery has shown promise in the development of potential treatments for opioid addiction. Preclinical studies suggest that blocking kappa-opioid receptors may be an effective approach to treating opioid dependence. AI can be used to optimize and create new drugs that can block the activity of the protein responsible for kappa-opioid receptors, making the drug discovery process more cost-effective and efficient. Three 5-minute reads/videos to keep you learning Text-to-Image Diffusion Models: A Guide for Non-Technical Readers The guide provides a simple explanation of text-to-image models and their use of diffusion to create images from natural language. It also introduces various tools for controlling and improving image generation processes, including ControlNET, ControlNET Pose, and ControlNET LORA. 2. The technology behind GitHub’s new code search This post provides a high-level explanation of the inner workings of GitHub’s new code search and offers a glimpse into the system architecture and technical underpinnings of the product. It also discusses how the search functionality allows users to find, read, and navigate code more efficiently. 3. MIT course on Introduction to Data-Centric AI This is a practical course on Data-Centric AI, focusing on the impactful aspects of real-world ML applications. The class covers algorithms for finding and fixing common issues in ML data, as well as constructing better datasets, with a concentration on data used in supervised learning tasks such as classification. 4. Lessons learned while using ChatGPT in education This guide shares an experience of using AI in education to complete tasks such as generating ideas, producing written material, creating apps, and generating images. The article details the success, changes, and process of using AI in education. 5. Writing Essays With AI: A Guide This guide explores the use of AI as a creative tool for essay writing. It discusses how to incorporate AI into writing practices by leveraging it to organize thoughts, capture a voice, summarize complex ideas, assist with idea generation, and evaluate writing quality. Papers & Repositories LLaMA: A repository for Open and Efficient Foundation Language Models A collection of foundation language models ranging from 7B to 65B parameters by Meta. 2. Aligning Text-to-Image Models using Human Feedback This paper proposes a fine-tuning method for aligning text-to-image models using human feedback by collecting human feedback, training a reward function that predicts human feedback, and fine-tuning the model by maximizing the reward-weighted likelihood. 3. The Wisdom of Hindsight Makes Language Models Better Instruction Followers This paper considers an alternative approach to RLHF: converting feedback to instruction by relabeling the original one and training the model for better alignment in a supervised manner. 4. Zero-Shot Information Extraction via Chatting with ChatGPT This work aims to investigate whether […]