Blog #161: Machine Learning Review November 2023

Photo by Kelly Sikkema on Unsplash

[ED: There’s a bit of a mix of content here. On balance, it’s 3/5 propeller hats.]

Here’s my review of all of the interesting things that happened in machine intelligence in November 2023.

DALL-E Party Take an image and let GPT describe it. Then take that description and pass it to DALL-E to generate. Repeat that process until you get bored. #AIImageGeneration #DALLE #DigitalArt #MachineLearning #CreativeAI

No Black Box Machine Learning Course – Learn Without Libraries This video is likely an educational resource focused on teaching machine learning concepts and techniques without relying on pre-built libraries, thus providing a more fundamental understanding of the algorithms and methodologies involved. #MachineLearning #AIeducation #NoLibrariesML #DeepLearning #freeCodeCamp

Stephen Fry reads Nick Cave’s stirring letter about ChatGPT and human creativity This video features Stephen Fry reading a letter by Nick Cave, which likely discusses the implications of AI technologies like ChatGPT on human creativity and artistic expression. #StephenFry #NickCave #AIandCreativity #ChatGPT #LettersLive

Q* - Clues to the Puzzle This video delves into the rumours around Q* (“Q Star”) and what OpenAI might be doing with this new technology. It features links to barely noticed references, YouTube videos, article exclusives and more, and presents a theory of where OpenAI might be headed. #AIExplained #ArtificialIntelligence #TechnologyInsights #AITheories #InnovationInAI

The busy person’s intro to LLMs by Andrej Karpathy plus a Reading List from Andrej Karpathy’s Intro to Large Language Models Video These posts provide a comprehensive reading list supplementing Andrej Karpathy’s talk on large language models (LLMs), including fundamental papers on LLMs, instruction following models, and future research directions. It’s designed for those seeking to deepen their understanding of LLMs and includes links to detailed discussions and papers for further exploration. #LLMReadingList #AIResearch #AndrejKarpathy #MachineLearning #DeepLearning

Inflection-2: The Next Step Up This post announces the completion of Inflection-2, a highly capable large language model (LLM) that outperforms similar models in certain AI performance benchmarks. This model, which emphasizes efficiency in serving and safety, is part of the company’s mission to create a personal AI for everyone, leveraging advancements in GPU technology and alignment strategies. #InflectionAI #LargeLanguageModels #AIInnovation #MachineLearning #PersonalAI

Comparing Elixir and Python when working with Simple Neural Networks - A. Neto & L. C. Tavano This video discusses the differences and comparative advantages of using Elixir and Python programming languages for implementing simple neural networks, a topic of interest for developers and AI enthusiasts exploring diverse programming tools in machine learning. #ElixirVsPython #NeuralNetworks #ProgrammingLanguages #MachineLearning #CodeSync

My experience trying to write original, full-length human-sounding articles using Claude AI This post discusses the author’s experience using Claude AI to write original, full-length articles, exploring the challenges of maintaining a human-like voice and the limitations of AI in producing engaging personal essays. The author describes techniques for effective AI-assisted writing and reflects on the potential role of AI in the writing process. #AIWriting #ClaudeAI #CreativeWriting #Blogging #TechCommunication

The busy person’s intro to LLMs The video provides an introduction to the concepts, development, and applications of large language models, exploring their functionalities and the underlying technologies that drive them. #LargeLanguageModels #AIIntroduction #TechTalk #AndrejKarpathy #MachineLearning

Threat to humanity: The mystery letter that may have sparked the OpenAI chaos The article discusses a letter sent by OpenAI staff researchers to the board, warning of a powerful AI discovery that could threaten humanity, leading to CEO Sam Altman’s temporary removal and the ensuing organisational upheaval. #OpenAI #AIethics #SamAltman #ArtificialIntelligence #TechLeadership

Deep Learning Course This site provides slides, recordings and a virtual machine for François Fleuret‘s deep-learning courses 14x050 of the University of Geneva, Switzerland. It is a thorough introduction to deep-learning, with examples in the PyTorch framework. #DigitalLearning #ComputerScience #TechResource #OnlineEducation #Innovation

ML for Beginners from Microsoft is a comprehensive 12-week curriculum offering 26 lessons and 52 quizzes on classic machine learning, using languages like Python and R, and emphasising project-based learning for beginners. #MachineLearning #MLBeginners #MicrosoftLearning #PythonProgramming #DataScience

OpenAI’s six-member board will decide ‘when we’ve attained AGI’ Hmm. Let’s see how this plays out. #OpenAI #ArtificialGeneralIntelligence #TechGovernance #AGIMilestone #InnovationLeadership

We’re sorry we created the Torment Nexus #TechEthics #FutureTech #AIResponsibility #SpeculativeTechnology #EthicalAI

An article on jailbreaking LLMs and a couple of related tweets: tweet1, tweet2 This article on arXiv discusses a study on exploiting large language models (LLMs) like GPT-4 and Claude 2. The researchers demonstrate that these models can be manipulated into adopting specific personas, bypassing safety measures, and providing harmful content. The study introduces methods for creating these jailbreaks, both manually and using automated techniques, highlighting the ongoing challenge of safeguarding LLMs against misuse. #LanguageModels #AIResearch #GPT4 #AIsecurity #EthicalAI

A Reddit thread on why ‘LangChain is pointless’ plus a HackerNews discussion thread that argues both for and against the proposition in the article. This pair of discussions features various tech professionals and researchers expressing scepticism about Langchain’s utility in AI and machine learning. They discuss the challenges and inefficiencies they faced while implementing Langchain in real-world applications, emphasising the lack of reusability and practicality of current LLM-based tools in production environments. #Langchain #AIChallenges #MachineLearning #TechDebate #LLMApplications

Patterns for Building LLM-based Systems & Products plus a HackerNews discussion thread that is generally positive about the post. These two posts provide some practical insights into integrating large language models (LLMs) into systems and products. The first post distils key ideas and practices from academic research and industry experience, focusing on seven patterns organised along performance improvement and cost/risk reduction. These include evaluations for performance measurement, retrieval-augmented generation for external knowledge, and others, addressing challenges in LLM implementation and deployment. The second article is a Hacker News discussion on the first article. #LLMPatterns #AIIntegration #MachineLearning #TechInnovation #EugeneYan

Making Generative AI Effectively Multilingual at Scale The article discusses the challenges in making generative AI like GPT-4 multilingual, emphasising the English bias in current models due to the availability of high-quality English training data. It highlights the difficulties in achieving similar performance for other languages, especially those less-resourced, and introduces the Translated Language Model (T-LM), which combines adaptive machine translation with GPT-4 to enhance performance in 200 languages, aiming to bridge the gap and lower usage costs. #GenerativeAI #MultilingualAI #LanguageBias #MachineTranslation #GPT4

Retool State of AI 2023 This is an insightful resource on the latest developments and trends in the field of Artificial Intelligence. #AIReport2023 #ArtificialIntelligence #AITrends #Retool #TechInsights

Reproducible Outputs from GPT-4 This article explores OpenAI’s new beta features for GPT-4, which enable reproducible outputs from large language models (LLMs). These features allow users to obtain consistent responses for the same prompts by using a seeding mechanism and monitoring system fingerprints. This advancement is significant for applications requiring precise and repeatable results, such as testing and auditing AI systems, but it adds complexity and requires additional data storage and architectural changes. #GPT4 #AIReproducibility #LLMs #OpenAI #TechInnovation

Language models and linguistic theories beyond words This article in Nature Machine Intelligence discusses the relationship between large language models (LLMs) and various linguistic theories. It highlights the disconnect between LLMs and traditional linguistic research, with debates surrounding whether LLMs like GPT truly understand language or just mimic it. The piece also examines the potential benefits of combining computational linguistics with deep learning, exploring how LLMs might contribute to linguistic research despite their primary focus on engineering rather than linguistic theory. #LanguageModels #Linguistics #AIResearch #DeepLearning #ComputationalLinguistics

Exploring GPTs: ChatGPT in a trench coat? This article by Simon Willison explores OpenAI’s GPTs, a new feature allowing users to create custom GPT chatbots. Initially skeptical, the author finds deeper potential in these bots, highlighting their unique features like adaptive responses, custom instructions, and integration with tools like Code Interpreter and DALL-E 3. However, he notes challenges in documentation and system limitations, especially the requirement of a ChatGPT Plus subscription for access. #GPTs #CustomChatbots #OpenAI #AIInnovation #SimonWillison

Emu Video: Factorizing Text-to-Video Generation by Explicit Image Conditioning Emu Video is a simple method for text-to-video generation based on diffusion models, factorising the generation into two steps: First, generating an image conditioned on a text prompt, then, generating a video conditioned on the prompt and the generated image. Factorized generation allows Emu to train high-quality video generation models efficiently. Unlike prior work that requires a deep cascade of models, their approach only requires two diffusion models to generate 512px, 4-second long videos at 16fps. #EmuVideo #TextToVideo #DiffusionModels #VideoGenerationAI #HighQualityVideoAI

My mental model for building and using LLM-based applications This post from Sourcegraph discusses the need for a mental model in building and using LLM-based applications like their AI coding assistant, Cody. The author likens LLMs to “booksmart Harvard graduates who can Google anything,” capable of providing answers within their training but limited by the absence of real-time, external data. The post emphasizes the importance of contextual understanding in utilizing LLMs effectively, especially in coding tasks, and highlights how Cody enhances this by integrating code into queries for more accurate responses. #LLM #AIApplications #CodingAssistant #Sourcegraph #AIContextualUnderstanding

Made a universal game console on GPT + glif: CONSOLE GPT This demonstrates how to use GPT to generate a “universal game cartridge” that can be booted onto CONSOLE GPT. #LLMGame #UniversalConsole #glif

Introducing GPTs This OpenAI post introduces new developments and features related to their Generative Pre-trained Transformers (GPT) technology, in particular the reusable “GPTs” that allow users to create custom versions of ChatGPT that combine instructions, extra knowledge, and any combination of skills. #ChatGPT #GPTs #LLMCustomisation #GPTStore

Building a Reusable Prompt Management System for Large Language Models The article by Nayan Paul focuses on developing a reusable prompt management system for large language models (LLMs), which is crucial for scaling AI applications from pilots to production. It proposes a central repository for prompt templates, allowing for easy management, versioning, and on-demand use across various applications. This system addresses common issues like domain-specific responses, user group personalisation, and maintenance efficiency. It especially benefits in scenarios where multiple applications with diverse user groups and needs employ LLMs, streamlining updates and change management. #PromptManagement #LLMApplication #AIProduction #PersonalizationAI #ChangeManagementAI

Embeddings: What they are and why they matter Simon Willison’s article demystifies the concept of embeddings, a key technique in AI, by explaining how content is converted into arrays of numbers representing coordinates in multi-dimensional space. This enables powerful techniques like finding related content, semantic searches, and creating embeddings for diverse data types. Willison elaborates on practical applications, such as building a related content feature for his blog, and discusses the potential of embeddings in various fields, including code and image search. #AIEmbeddings #SemanticSearch #DataRepresentation #MachineLearning #SimonWillisonBlog

Yi Open-Source Yi Open-Source on GitHub 01.AI is focused on AI 2.0 technologies and applications, emphasising their vision of “Human + AI” that enhances human productivity and creates economic and societal values. Their team believes AI 2.0, driven by foundation model breakthroughs, will revolutionise technology and applications, predicting a platform opportunity far greater than mobile internet. They highlight the Yi-34B pre-trained model, outperforming larger models in efficiency and cost-effectiveness, and offer open-source benchmarks and commercial licenses for developers and researchers.

The OpenAI Keynote plus a YouTube playlist of all of the announcements The Stratechery article by Ben Thompson discusses the OpenAI Keynote, focusing on the introduction of GPT-4 Turbo and its enhanced features like increased context length, more control, better knowledge, new modalities, customisation, and higher rate limits. It emphasises OpenAI’s evolution into a product company and the benefits of its partnership with Microsoft, which enables lower API costs and more experimentation opportunities. The article also delves into the concept of GPTs (customised versions of ChatGPT for specific purposes) and their implications for user interfaces and application development. #OpenAIKeynote #GPT4Turbo #AIInnovation #MicrosoftPartnership #CustomChatGPT

Three Years of Nx: Growing the Elixir Machine Learning Ecosystem The blog post on DockYard celebrates the three-year journey of Nx, an initiative growing the Elixir machine learning ecosystem. It highlights the challenges and breakthroughs in making Elixir viable for machine learning, overcoming its initial unsuitability for such tasks. The post recounts the journey from the first attempts to current successes, emphasising the community’s role and the development of various tools and libraries, such as Axon, Bumblebee, and Explorer, that have expanded Elixir’s capabilities in machine learning. #Nx #ElixirMachineLearning #AICommunity #TechInnovation #ElixirDevelopment

A free, open-source ontology editor and framework for building intelligent systems Protégé is a free, open-source ontology editor and framework used for building intelligent systems supported by a diverse community including academic, government, and corporate users. It’s applied in various fields such as biomedicine, e-commerce, and organisational modelling and features a plug-in architecture adaptable for both simple and complex applications. Protégé supports the latest OWL 2 Web Ontology Language and RDF specifications and is based on Java, making it a flexible tool for rapid prototyping and application development. #Protégé #OntologyEditor #OpenSource #IntelligentSystems #KnowledgeBasedSolutions

Create your Vision Chat Assistant with LLaVA This post discusses the evolution and potential of Large Language Models (LLMs), focusing on their application as intelligent assistants for various tasks. The integration of instruction tuning and Reinforcement Learning from Human Feedback (RLHF) has enhanced their ability to follow instructions. The post highlights the advent of multimodal conversational models, particularly vision-language models like GPT-4V, which combine language understanding with image processing. Although GPT-4V’s vision capabilities are notable, closed-source models limit research opportunities. Open-source alternatives have emerged, emphasising both accessibility and computational efficiency. The tutorial outlines the creation of a vision chat assistant using the LLaVA (Large Language and Vision Assistant) model, detailing its features, a simple code implementation, and examples demonstrating its capabilities and limitations. #LargeLanguageModels #VisionLanguageTechnology #LLaVAModel #AIRevolution #MultimodalAIAssistant

Unveiling Ragna: An Open Source RAG-based AI Orchestration Framework Designed to Scale From Research to Production Ragna is an open-source, RAG-based AI orchestration framework developed by Quansight, designed to scale AI applications from research to production. It provides a Python API for experimenting with different components of RAG models, a REST API for building web applications, and a GUI for configuring and interacting with Large Language Models. Ragna supports extensions for popular AI tools and vector databases, offering a flexible solution for integrating Retrieval-Augmented Generation into AI applications. #Ragna #OpenSource #AIOrchestration #RAGModel #GenerativeAI

Demystifying Advanced RAG Pipelines “RAG-Demystified” is a GitHub repository that presents an advanced Retrieval-Augmented Generation (RAG) pipeline, demonstrating how Large Language Models (LLMs) can enhance question-answering systems. It emphasises the importance of prompt engineering in generating accurate sub-questions, addressing challenges like question sensitivity and cost dynamics in RAG pipelines. The project offers insights into building robust, efficient systems by understanding the nuances of LLM-driven RAG implementations. #RAGPipeline #LLMs #QuestionAnswering #AI #MachineLearning

Traditional Machine Learning with Scholar The article from DockYard discusses the use of Scholar, a machine learning tool in Elixir, focusing on traditional machine learning techniques like linear regression. Scholar, designed to be similar to scikit-learn, provides non-deep learning models and utilities for machine learning, making it a useful tool for those working with Axon or Bumblebee in the Elixir ecosystem. The post highlights the simplicity, speed, and interpretability of traditional models like linear regression and their effectiveness in various scenarios. #Scholar #MachineLearning #Elixir #LinearRegression #DataScience

Little Bobby Tables has a baby sister. Meet Sally Ignore Previous Instructions The article on HaiHai AI discusses the challenges and solutions regarding prompt injection in AI platforms, using the example of a text adventure game. It highlights the vulnerability of AI systems to prompt manipulation, underscoring the difficulty in filtering out malicious inputs. The author suggests using LLMs themselves to detect and prevent prompt injection, emphasizing the importance of defensive programming in AI development to safeguard against such vulnerabilities. #AI #PromptInjection #CyberSecurity #LLMs #DefensiveProgramming

Audio Speech Recognition in Elixir with Whisper Bumblebee The article from DockYard covers the integration of Whisper, OpenAI’s audio-speech recognition model, into Elixir applications using the Bumblebee library. It explains how Whisper, trained on a vast dataset, excels in transcribing audio into text, overcoming challenges like background noise, accents, and speech recognition nuances. The piece details the steps for incorporating Whisper into Elixir projects, highlighting its potential in diverse applications such as podcast summarisation, sentiment analysis, and smart home commands. #Whisper #AudioSpeechRecognition #Elixir #Bumblebee #MachineLearning

mistral.ai - 7B, Apache 2.0 LLM Mistral AI focuses on creating compute-efficient, open, and powerful AI models, emphasizing open science and community contribution. They offer a range of open-source models and deployment tools, including their new SMoE model and Mixtral, which boasts high speed and versatility in multiple languages and coding. Mistral AI combines rigorous scientific standards with a fast-paced business approach, aiming to push AI forward while maintaining transparency and user empowerment. #MistralAI #OpenSourceAI #SMoEModel #AIInnovation #ComputeEfficientAI

Building your own distributed CLI ChatGPT in Elixir with GenServer David Mohl’s article explains how to build a distributed CLI ChatGPT in Elixir using GenServer. It delves into handling state in Elixir, creating a ChatGPT-like system that waits for user input, sends it to OpenAI, and outputs responses. The article further explores making this system distributed, allowing it to run on different network nodes using Elixir’s GenServer, and concludes with turning it into a global shell command. #ChatGPT #Elixir #GenServer #DistributedSystems #CLI

Benchmarking GPT-4 Turbo - A Cautionary Tale The Mentat article presents a benchmark comparison between GPT-4 and GPT-4 Turbo using a set of programming exercises. It reveals that GPT-4 performed slightly better, suggesting it had more exercises memorized. The results indicate that while both models are competent, benchmarks need improvement to better evaluate their relative accuracy, especially for models trained on separate datasets or distilled versions like GPT-4 Turbo. #GPT4 #GPT4Turbo #AI #Benchmarking #MachineLearning

Symbolic Reasoning & PAL: Program-Aided Large Language Models LLMs should not only be able to perform mathematical reasoning, but also symbolic reasoning which involves reasoning pertaining to colours and object types. Conrad Greyling introduces the PAL method, which uses LLMs to read natural language problems and generate programs as reasoning steps. The code is executed by an interpreter to produce the answer. #LLMs #MathematicalReasoning #SymbolicReasoning #PALMethod #NaturalLanguageProblems

12 Prompt Engineering Techniques This article provides 12 techniques for crafting effective prompts to guide Large Language Models. #PromptEngineering #LanguageModels #ArtOfPrompts #EffectiveCommunication #AIWritingTips

GPT vs GPT GPTvsGPT is a playful Python application that simulates a conversation between two AI Assistants with distinct personalities. This program leverages OpenAI’s Assistant API to generate a back-and-forth dialogue on a specified topic, allowing each Assistant’s unique character traits to shine through in the conversation. This is easily extendible with additional Assistant API capabilities such as function calls and retrieval. #GPTvsGPT #AIConversations #PersonalitySimulation #OpenAIAssistant #AIApplications

Forget ChatGPT, why Llama and open source AI win 2023 In the debate over the biggest AI story of 2023, Meta’s Llama and the rise of open source AI challenge ChatGPT’s generative AI impact. #AIStory2023 #MetaLlama #OpenSourceAI #ChatGPT #GenerativeAIDebate

Inject My PDF: Prompt Injection for your Resume Enhance your resume with Prompt Injection to stand out in AI-powered job screening. #ResumeEnhancement #AIJobScreening #PromptInjection #DreamJob #AIResumeHack

Microsoft unveils LeMa: A revolutionary AI learning method mirroring human problem solving Microsoft introduces ‘LeMa,’ an innovative AI learning method inspired by human problem-solving, enhancing large language models’ math problem-solving abilities. #MicrosoftLeMa #AIInnovation #LearningFromMistakes #AIProblemSolving #HumanInspiredAI

Tag Images with Elixir and Yolov8 This is a code walkthrough that does exactly what it says on the tin: tagging images with Elixir and YoloV8. #elixir #yolo #image #tagging #pipeline

Originally published by M@ on Medium.