QFM001: Machine Intelligence Reading List - January 2024

The Machine Learning Video Library by Professor Yaser Abu-Mostafa at Caltech provides a comprehensive collection of video segments on various machine learning topics, including theoretical foundations, algorithms, and practical applications, aimed at enhancing understanding and education in the field.

The second article in this series discusses the integration of Large Language Models (LLMs) with unstructured data and the role of Knowledge Graphs (KGs) in enhancing Retrieval-Augmented Generation (RAG) systems, highlighting design choices in data preparation and retrieval to improve LLM's response accuracy in enterprise settings.

is an AI agent framework designed for simplicity and power aimed at engineers. It emphasises the ease of building sophisticated multi-agent interactions, making it accessible for creating advanced AI solutions. The framework is designed to orchestrate role-playing autonomous AI agents. It focuses on collaborative intelligence, allowing agents to work together seamlessly to handle complex tasks. More here: GitHub: joaomdmoura/crewAI{:target="_blank"}

This article discusses significant advancements and findings in AI research, focusing on large language models and their applications. It covers topics such as efficient training techniques, domain-specific language models, and advancements in model fine-tuning and scaling. The highlighted papers reflect the rapid evolution and diverse applications of AI technology in 2023.

Nitro is a fast, lightweight 3MB inference server designed to enhance applications with local AI capabilities. It is OpenAI-compatible, extremely efficient for app developers, and supports cross-platform operation on both CPU and GPU architectures. GitHub: janhq/nitro{:target="_blank"}:

Meta has launched a comprehensive prompt engineering guide to optimise interactions with AI, such as ChatGPT, by employing techniques like detailed instructions, format specification, restrictions, role assignments, and chain-of-thought prompting, enhancing the quality and relevance of AI-generated responses.

This article explores the concept of Model Autophagy Disorder (MAD) in generative AI, where using synthetic data to train subsequent model generations without sufficient fresh real data leads to a decline in model quality and diversity, analogous to mad cow disease.

This article discusses a study showing that consultants using AI, specifically ChatGPT-4, outperform those who don't in various tasks when working at a management consulting firm. It explores the concept of the "Jagged Frontier" of AI, where AI's capabilities are uneven across different tasks. The article also discusses two approaches to integrating AI in work: Centaur and Cyborg methods, emphasising the need for a strategic combination of human and AI capabilities.

This paper introduces Direct Preference Optimisation (DPO), a new method for training language models to align with human preferences without explicit reward modelling or reinforcement learning. DPO offers a simpler, more stable, and efficient alternative to existing methods, performing equally or better in tasks like sentiment modulation, summarisation, and single-turn dialogue.

This article discusses how AI, rather than eliminating jobs, will transform them by broadening job roles and enhancing capabilities. It argues that AI's ability to digitise and amplify skills will turn workers into creative generalists capable of handling diverse tasks with AI assistance, leading to more dynamic, flexible, and efficient work environments.

This article discusses the emerging threat of generative AI (GenAI) in undermining the effectiveness of Know Your Customer (KYC) processes. The article highlights how GenAI can be used to create realistic deepfake images and videos that can bypass traditional identity verification methods, including advanced liveness checks, posing a significant challenge to financial institutions and security systems.

The article discusses the Eliza program, an early example of a chatbot designed to pass the Turing Test by mimicking a psychotherapist. It delves into Eliza's simplistic yet effective approach to simulating conversation, including its use of pattern matching, knowledge base, and memory. The article also provides insights into how Eliza handles synonyms, transformations, and user inputs, emphasising its success in appearing human-like despite its basic algorithm. I remember stumbling across Eliza on the Apple ][ in the late 80s on a public machine in my local library. I was totally blown away by something so simple yet so seemingly intelligent.

When doing a bit of research on "embodied AI"{:target="_blank"} recently, I came across the concept of "The Frame Problem". This idea in artificial intelligence and cognitive science involves the challenge of representing the effects of actions in a dynamic world using logic, specifically dealing with the difficulty of specifying which aspects of a situation remain unchanged after an action. This problem has led to various solutions in logical reasoning and has broader implications in philosophy, particularly in the context of knowledge representation and common sense reasoning.

This video presented at the EMPEX Conference explores the synergies between Elixir, a programming language known for handling concurrent processes, and the future development of large language models (LLMs). Moriarty discusses how Elixir's capabilities can effectively manage the complex and resource-intensive demands of LLMs, highlighting the potential for more efficient and scalable AI systems.

This article explores the advanced capabilities of ChatGPT's API for structured data output, discussing techniques like prompt engineering, system prompts, and the use of Pydantic for simplifying schema input/output. It provides practical examples and explores the potential of structured data support in ChatGPT, including how to optimise outputs and the utility of function calling with ChatGPT.

Stanford University researchers have developed an innovative AI system called Mobile ALOHA, a robot capable of performing complex household tasks like cooking, cleaning, and doing laundry. This system is an extension of the existing ALOHA system, now enhanced with a wheeled base for mobility, making it a cost-effective alternative to conventional bimanual robots. Mobile ALOHA can operate autonomously or be controlled by a user and is capable of performing a range of tasks demonstrated in a video, including cooking shrimp, handling spills, and interacting with people. The system is designed to learn from human demonstrations, with as few as 50 demonstrations per task, using transformers similar to those in large language models. Although impressive, the system is still in the development phase and requires further optimisation before it's ready for widespread use.

The OpenAI Cookbook is a comprehensive resource for developers, providing detailed guidance on using various OpenAI APIs, including examples and best practices for building applications. It covers a range of topics, from basic API usage to advanced techniques, aimed at helping developers integrate AI into their projects effectively.

This paper provides a comprehensive guide on deep learning algorithms, focusing on the mathematical foundations and techniques of artificial neural networks, including their types, calculus, approximation abilities, optimisation methods, generalisation errors, and applications in solving partial differential equations.

Coca-Cola's use of an AI-powered image generator for Christmas cards led to unexpected and often bizarre results. The AI, trained on Coca-Cola's Christmas imagery, was meant to create charming holiday-themed images but instead produced a variety of odd and sometimes unsettling creations, highlighting the unpredictable nature of AI in branding and marketing.

This article explains the mathematics behind a transformer model, specifically focusing on a simplified end-to-end example. It covers the basics of tokenisation, embedding, positional encoding, and the self-attention mechanism, providing insights into how transformers work during inference, especially in language translation. The article is designed to demystify the complex steps and numerous parameters involved in transformer models, making it accessible for those with basic linear algebra and machine learning knowledge. And don't forget The Illustrated Transformer{:target="_blank"}, which is required reading for this article.

AlphaCodium is an advanced open-source tool for code generation, surpassing most human competitors in code contests. It utilises a novel, iterative approach with Large Language Models (LLMs), focusing on test-based, multi-stage processes to handle complex coding challenges and edge cases effectively. This method has significantly improved the performance of LLMs in code problems, demonstrating a notable increase in accuracy and efficiency over previous models like AlphaCode​.

This GitHub repo offers comprehensive resources and guidance on Machine Learning Engineering, covering practical aspects like scalability, the use of transformers, and MLOps, with a focus on Python and PyTorch.

This article discusses the development of Persuasive Adversarial Prompts (PAPs) that effectively persuade Large Language Models (LLMs) to perform actions outside their intended use, with a 92% success rate in jailbreaking aligned LLMs like GPT-3.5 and GPT-4. It highlights the increased vulnerability of advanced models to PAPs and explores defence strategies to mitigate these risks.

Linus Torvalds discusses the profound impact of Artificial Intelligence on programming, highlighting the evolution of programming languages, the improvement of development workflows with machine learning, and future predictions for AI's integration in software development.

This site is a non-profit educational initiative featuring a vast collection of Midjourney styles and guides, run by enthusiasts and volunteers, dedicated to aiding the Midjourney community with over 284,950 users and 52,505 images stored in its database.

This article discusses how users interact with AI bots like ChatGPT, Bing Chat, and Bard in six distinct conversation styles. These styles vary based on the users' skill levels and information needs, highlighting the importance of diverse user interface designs to accommodate these different conversation types.

Journalist Karen Hao joins Big Think’s Editor-in-Chief, Robert Chapman-Smith, to discuss the recent events at OpenAI, including the ousting and reinstatement of CEO Sam Altman, as well as the ideological clashes regarding the development and release of powerful AI models like ChatGPT. The video discusses the complexities and potential existential risks associated with large language models like ChatGPT from OpenAI.

This article delves into various methods of compromising AI models, including adversarial attacks, data poisoning, backdoor attacks, and model extraction. It highlights the vulnerabilities of neural networks and the importance of considering security in AI development, offering insights into different types of attacks and their implications.

The article discusses the current state and future predictions of AI for 2024, emphasising generative AI's rise, including advancements in text, image, video, and audio models. It forecasts a significant increase in AI's role in consumer assistants, code generation, and unstructured data processing. The article predicts a shift towards commoditised software and the increasing importance of product design, along with a trend towards AI models moving to edge devices and the integration of AI in major companies like Apple and Amazon.

DocLLM is a novel extension of large language models that enhances document understanding by incorporating both textual and spatial layout elements without the need for heavy image encoders, outperforming state-of-the-art models in several document intelligence tasks.

This article examines how the length of reasoning steps in Chain of Thought (CoT) prompts affects the reasoning abilities of large language models (LLMs). The study finds that longer reasoning steps significantly enhance LLMs’ problem-solving abilities, even without adding new information. Surprisingly, even incorrect rationales in the CoT prompts lead to better outcomes if they maintain sufficient length, indicating that the quantity of reasoning steps is more critical than their factual accuracy. As the Xitter{:target="_blank"} thread discussing the paper and its implications suggests, AI is weird.

This article identifies key challenges in implementing RAG systems, which combine document retrieval with language models like ChatGPT. It highlights seven potential failure points, including missing content, missed top-ranked documents, context issues, extraction errors, format problems, incorrect specificity, and incomplete answers. The paper emphasises the need for thorough testing and refinement of RAG systems, citing case studies that illustrate how larger contexts and metadata improve performance and the parity of open-source embedding models with closed-source ones in certain contexts.