The paper titled "CaBRNet, an open-source library for developing and evaluating Case-Based Reasoning Models" presents a new framework aimed at enhancing the field of explainable artificial intelligence (AI). The authors, Romain Xu-Darme and his colleagues from LSL, highlight the growing interest in self-explainable models, which serve as a more principled alternative to traditional post-hoc methods that attempt to clarify decisions made by opaque models after the fact. Despite the advancements in self-explainable models, the authors point out several challenges that persist in this area. These include issues related to reproducibility, difficulties in making fair comparisons between different models, and the lack of standardized practices across the field. To address these challenges, the authors introduce CaBRNet, a modular and backward-compatible framework specifically designed for Case-Based Reasoning Networks. This framework aims to provide a structured approach to developing and evaluating models, thereby facilitating better reproducibility and comparison. The paper was submitted on September 25, 2024, and is set to be presented at the 2nd World Conference on eXplainable Artificial Intelligence in July 2024 in La Valette, Malta. The authors encourage the use of their open-source library to foster collaboration and innovation in the development of explainable AI systems. By providing a robust platform for researchers and practitioners, CaBRNet aims to contribute significantly to the advancement of self-explainable models in artificial intelligence.

xAI announced its next model, with 128k context length and improved reasoning capabilities. It excels at retrieval and programming.

Cohere For AI has created a 30B+ parameter model that is quite adept at reasoning, summarization, and question answering in 10 languages.

This work introduces a framework to evaluate the trustworthiness of Retrieval-Augmented Generation (RAG) systems across six key areas: factuality, robustness, fairness, transparency, accountability, and privacy.

Anthropic researchers have unveiled a method to interpret the inner workings of its large language model, Claude Sonnet, by mapping out millions of features corresponding to a diverse array of concepts. This interpretability could lead to safer AI by allowing specific manipulations of these features to steer model behaviors. The study demonstrates a significant step in understanding and improving the safety mechanisms of AI language models.

Casper Labs has introduced Prove AI, developed with IBM, to bring transparency and auditability to enterprise AI applications.

Beth Barnes' nonprofit METR is partnering with major AI companies like OpenAI and Anthropic to develop safety tests for advanced AI systems, a move echoed by government initiatives. The focus is on assessing risks such as AI autonomy and self-replication, though there's acknowledgment that safety evaluations are still in early stages and cannot guarantee AI safety. METR's work is seen as pragmatic, despite concerns that current tests may not be sufficiently reliable to justify the rapid advancement of AI technologies.

Baldur Bjarnason, a web developer from Hveragerði, Iceland, recently shared insights on the evolving discourse surrounding fair use in the context of generative AI models. He referenced a paper by Jacqueline Charlesworth, a former general counsel of the U.S. Copyright Office, which critically examines the claims of fair use made by proponents of generative AI. The paper highlights a significant shift in legal scholarship regarding the applicability of fair use to the training of generative models, particularly as a clearer understanding of the technology has emerged. Charlesworth argues that the four factors outlined in Section 107 of the Copyright Act generally weigh against the fair use claims of AI, especially in light of a rapidly changing market for licensed training materials. A key point made in the analysis is that the argument for fair use often relies on a misunderstanding of how AI systems operate. Contrary to the belief that works used for training are discarded post-training, these works are actually integrated into the model and continue to influence its outputs. The process of converting works into tokens and incorporating them into a model does not align with the principles of fair use, as it represents a form of exploitation rather than a transformative use. Charlesworth draws a distinction between the copying of expressive works for functional purposes—such as searching or indexing—and the mass appropriation of creative content for commercial gain. The latter, she argues, lacks precedent in fair use cases and cannot be justified by existing legal frameworks. The paper emphasizes that the act of encoding copyrighted works into a more usable format does not exempt it from being considered infringement. Furthermore, the notion that generative AI's copying should be deemed transformative because it enables generative capabilities is critiqued as a broad and unfounded assertion. This argument essentially posits that the rights of copyright owners should be overridden by the perceived societal benefits of generative AI, which does not hold up as a legal defense in copyright disputes. The narrative pushed by AI companies—that licensing content for training is unfeasible—faces scrutiny, as these companies have shown they can engage in licensing when it serves their interests. This undermines their claims that copyright owners are not losing revenue from the works being appropriated. Overall, Bjarnason encourages readers to explore Charlesworth's paper, noting its accessible language and the importance of understanding the legal implications of generative AI in relation to copyright law.

This growing library will help you understand the emerging patterns of interaction, affordances, and heuristics in AI.

Anthropic has published the system prompts used to guide its Claude AI models and plans to continue being transparent moving forward.

This blog post outlines common themes in building generative AI systems. It covers many of the building blocks a company should consider when deploying its models to production.

OpenAI has released two new "chain-of-thought" models, o1-preview and o1-mini, which prioritize reasoning over speed and cost. These models are trained to think step-by-step, enabling them to handle more complex prompts requiring backtracking and deeper analysis. While the reasoning process is hidden from users due to safety and competitive advantage concerns, it allows for improved results in tasks like generating Bash scripts, solving crossword puzzles, and validating data.

This guide was missed in the excitement of OpenAI's new reasoning models. It shows how prompting this new model is different and requires simpler prompts and a more structured input context.

Google DeepMind introduced the Frontier Safety Framework to address risks posed by future advanced AI models. This framework identifies critical capability levels (CCLs) for potentially harmful AI capabilities, evaluates models against these CCLs, and applies mitigation strategies when thresholds are reached.

Anthropic's Responsible Scaling Policy aims to prevent catastrophic AI safety failures by identifying high-risk capabilities, testing models regularly, and implementing strict safety standards, with a focus on continuous improvement and collaboration with industry and government.

Researchers have developed an AI network where one AI can teach another to perform tasks using natural language processing, a capability not previously demonstrated. The system uses a model called S-Bert that allows AI to perform tasks given via instructions and then communicate that knowledge to another AI. This breakthrough has potential applications in robotics and could further understanding of human cognitive functions.

OpenAI has released a set of code for its rules based rewards for language model safety project. It includes some data they used for training.

A new research paper details the mapping of AI model Claude 3 Sonnet's inner workings, revealing "features" activated by concepts like the Golden Gate Bridge. By adjusting these features' strengths, researchers can direct Claude's responses to incorporate specific elements, demonstrating a novel method of modifying large language models. The research aims to enhance AI safety by precisely adjusting model behaviors related to potential risks.

Anthropic recently published a public research paper explaining why its AI chatbot chooses to generate content about certain subjects over others. Its researchers deciphered what parts of the chatbot's neural network mapped to specific concepts using a process known as 'dictionary learning'. The research showed how neurons associated with a topic fired together when the model was thinking about something associated with the topic - similar sets of neurons firing can evoke adjacent subjects. A link to the paper is available at the end of the article.

Building useful scalable AI applications requires developers to have good data preparation (data cleansing and management) and use retrieval-augmented generation. Models used should be pre-trained or fine-tuned. Custom models can be developed in-house, but usually will require a large amount of capital. Developers should be mindful of latency, memory, compute, caching, and other factors to make sure the user experience is good.

A new benchmark to assist in determining a model's agent-like abilities.

The last week of March 2024 marked a significant moment in open-source large language models (LLMs) with multiple notable releases, including DBRX by Databricks, Jamba by A21 Labs, and Samba-CoE by SambaNova Systems. These launches signify a pivotal moment in the diversification and proliferation of accessible and decentralized AI models. The trend reflects a narrowing performance gap between open-source LLMs and their closed-source counterparts, indicating a vibrant future for AI innovation and enterprise adoption.

Apple CoreNet is a deep neural network toolkit that allows researchers and engineers to train standard and novel small and large-scale models for a variety of tasks, like object classification, object detection, and semantic segmentation.

This article clarifies key AI terms amidst growing confusion due to marketing jargon, highlighting concepts such as Artificial General Intelligence (AGI), Generative AI, and machine learning. It addresses AI challenges like bias and hallucinations and elaborates on how AI models are trained, referencing various models, algorithms, and architecture, including transformers and retrieval-augmented generation (RAG). The piece also mentions leading AI companies and their products, such as OpenAI's ChatGPT, and hardware used for AI, like NVIDIA's H100 chip.

California's SB1047 bill proposes regulations for AI models with computational capacities over 10^26 FLOPs. It focuses on ensuring these models are used safely by requiring secure environments, quick deactivation capabilities, and rigorous misuse potential testing. The bill targets only high-risk scenarios, aiming to balance innovation with safeguards against misuse in response to concerns about AI's potential impact on society.

CFExplainer is a new tool that improves how AI models, specifically Graph Neural Networks, understand and identify security vulnerabilities in software.

Covariant has introduced RFM-1, aiming to revolutionize robotics with a large language model for robot language that enhances robots' decision-making and interaction capabilities across various industries by utilizing a massive data collection from its Brain AI platform.

Covariant has introduced RFM-1, aiming to revolutionize robotics with a large language model for robot language that enhances robots' decision-making and interaction capabilities across various industries by utilizing a massive data collection from its Brain AI platform.

This is a great paper that discusses how brittle modern agent systems are and what the path forward could be to design learned systems. Its authors use programming languages as a test bed where agents can be designed and run unsupervised.

This article discusses the evolution and growing complexity of generative pre-trained transformer models. It touches upon how AI development and use are influenced by the regulatory landscape, with examples stretching from cryptographic software to AI-specific executive orders. The piece highlights several steps in AI model creation, from data collection to inference. It also highlights the potential of utilizing crypto and decentralized technology to make AI more user-aligned, verifiable, and privacy-conscious. Despite the progress, AI democratization remains a challenge.