Skip to content
/ LLM-Learning Public

A repository to store helpful information and emerging insights in regard to LLMs

License

MIT license
20 stars 0 forks Branches Tags Activity
Star
Notifications You must be signed in to change notification settings

j-webtek/LLM-Learning

BranchesTags

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

38 Commits
Data to LLM Processes.pdf
Data to LLM Processes.pdf
 
 
Foundation Models.pdf
Foundation Models.pdf
 
 
LICENSE
LICENSE
 
 
LLM Stack.pdf
LLM Stack.pdf
 
 
Multi_Models.pdf
Multi_Models.pdf
 
 
README.md
README.md
 
 

Repository files navigation

📘 LLM-Learning

Welcome to LLM-Learning! This repository is dedicated to offering a curated collection of intriguing research papers, innovative projects, and insightful documents from the world of Large Language Models (LLMs) and beyond. LLMs have transformed the AI landscape with their ability to understand and generate human-like text. Here, we gather some of the most notable works that provide insights into their capabilities, applications, and potential impact. Whether you're a seasoned researcher, a developer looking to harness the power of LLMs, or just a curious enthusiast, there's something for everyone. Dive into the transformative capabilities of LLMs and discover their potential across diverse applications.

📈 Repository Statistics

Research Papers: 14

Projects: 11

Links & Articles: 79.

🆕 ## Latest Additions 10/27/23

Welcome to this week's AI news summary, highlighting key developments, debates, and breakthroughs in the AI and ML field.

AI & ML Weekly Summary: week

A curated list of the most interesting AI and Machine Learning stories from the past week. Dive in to stay updated!

Stories

  • New data poisoning tool lets artists fight back against generative AI:
    • Analysis: Nightshade is a new tool developed by a team led by Ben Zhao from the University of Chicago that allows artists to fight against AI companies using their work without permission. The tool adds invisible changes to the pixels in an artist's work, causing AI models trained on scraped data to produce incorrect and unpredictable outputs. This can damage image-generating AI models and disrupt their ability to generate accurate results. The tool, currently submitted for peer review, aims to tip the power balance in favor of artists and protect their copyright and intellectual property. The team also developed Glaze, a tool that allows artists to mask their personal style to prevent it from being scraped by AI companies. Nightshade will be integrated into Glaze, and both tools will be open source, allowing others to use and modify them.
  • Show HN: Autolicious – AI-powered bookmark cataloging Chrome extension:
    • Analysis: Autolicious is a Google Chrome extension that automatically organizes bookmarks using ChatGPT and the OpenAI API. With just one button press, users can add bookmarks and the extension will automatically categorize them with a category, sub-category, and description. However, it's important to note that users will be charged for requests made to the GPT API by OpenAI, with prices starting from $0.0015 per bookmark added.

To use the extension, users need to provide their OpenAI API key, which can be obtained by registering with OpenAI and creating an API key. The extension's popup window allows users to enter their API key and save the settings.

It's important to consider privacy and security concerns when using this extension, as it sends the page content to the OpenAI API for categorization. Although OpenAI claims to handle user data securely, users who have concerns about privacy and the security of their browsing history may want to avoid using this extension.

This version of Autolicious is still in its early stages and only supports viewing saved bookmarks, without the ability to manually organize them.

  • UK officials use AI to decide on issues from benefits to marriage licences:
    • Analysis: A Guardian investigation has revealed that UK government officials are using artificial intelligence (AI) and complex algorithms in various areas, including welfare, immigration, and criminal justice. However, the investigation has also found evidence of potential discrimination, such as an algorithm used by the Department for Work and Pensions leading to the removal of benefits for dozens of individuals and a facial recognition tool used by the Metropolitan police making more mistakes in recognizing black faces than white ones. The use of AI in decision-making processes highlights the need for better regulation and understanding of the potential biases and implications of these technologies.
  • AMD wants to know if you would like Ryzen AI support for Linux:
    • Analysis: The content is a GitHub issue discussing a link to an article about AMD releasing open-source demo code for Ryzen AI. The issue highlights that the code is only available for Windows and that many libraries are binary only, meaning they are not fully open-source.
  • Cortical Labs: "Human neural networks raised in a simulation":
    • Analysis: The content is about Cortical Labs, a company that is developing a new form of intelligence by growing human neurons into silicon. They have created a simulation called the Biological Intelligence Operating System (biOS), which runs the neurons and sends them information about their environment. The neurons are cultivated on a silicon chip and are directly connected to the digital world, creating a high-bandwidth connection between organic and digital networks. The ultimate goal is to create a new form of learning machine that surpasses the limitations of traditional silicon-based computers. Cortical Labs is inviting developers to join their program to gain early access to their software development kit (SDK) and train the minds they are developing. The focus is not on creating smarter computers or data centers, but on exploring the potential of organic-digital intelligence.
  • Riley Reid on AI: 'I Don't Want Porn to Get Left Behind':
    • Analysis: Riley Reid, a porn performer, discusses her involvement with Clona.ai, a virtual companion platform that allows users to chat with AI versions of real people. Clona.ai uses Meta's open-source language model, LLaMa, to create virtual versions that reflect the personalities of the individuals behind them. Reid co-founded Clona.ai to ensure that she has control over the use of her likeness and to be a part of technological advances in society. She believes that AI has a place in the future of pornography and online sex.
  • Animated AI:
    • Analysis: The content being reviewed is about neural networks and includes instructional videos on various topics related to AI. The topics covered include the basic algorithm of convolution, padding, stride, groups in convolution, depthwise and depthwise-separable convolution, and pixel shuffle for changing resolution. The content is available on the creator's Patreon and YouTube channel and is licensed under the MIT License.
  • We cannot allow AI to make Big Tech even bigger:
    • Analysis: In this op-ed, Steve Case, co-founder of AOL and venture capitalist, expresses his concern about the potential expansion of Big Tech through the use of artificial intelligence (AI). He argues that if AI is allowed to dominate tech industry, it could lead to further consolidation of power and limit competition. Case emphasizes the importance of keeping the tech industry diverse and innovative by encouraging competition and preventing excessive concentration of power.
  • What AI won't replace in your programming:
    • Analysis: The article discusses how generative AI can be used to augment the developer experience, but it emphasizes that AI will not replace the expertise and intuition of programmers. It highlights the importance of programmers having a deep understanding of coding and being able to effectively prompt and evaluate AI tools. AI assistants can be helpful for new or inexperienced developers by automating tedious tasks, but they should not be relied upon to do all the work. The article encourages developers to start small and automate simple tasks to get started with generative AI and software development.
  • Lawrence Lessig on why AI and social media are causing a free speech crisis:
    • Analysis: The Verge article features an interview with Harvard law professor Lawrence Lessig, who discusses the free speech crisis on the internet caused by artificial intelligence (AI) and social media platforms like TikTok. Lessig, an expert in internet policy, expresses concerns about the abundance of disinformation and misinformation online, which is harmful to individuals and democracy. While the US First Amendment protects speech, Lessig mentions various strategies, such as copyright law, that have emerged to regulate content on the internet. The article highlights that there are different approaches to address these speech regulations.
  • Deciphering Latest Restrictions on AI and Semiconductor Manufacturing:
    • Analysis: The US government has introduced new restrictions on AI chips and wafer fabrication equipment in its ongoing trade war with China. The restrictions specifically target AI semiconductors and close most of the loopholes that allowed companies to release modified chips with slightly lower bandwidth while maintaining high performance. The new regulations also include a performance density threshold to prevent shipments of chips with smaller die sizes but still efficient computing power. The ruling has caught companies like Nvidia and GPUs such as the L40S off guard, as they expected to ship their products to China. It is believed that the 4090 GPU may be granted a license due to its datacenter focus. The article also mentions China's efforts to work around these restrictions and the impact on various companies and earnings going forward.
  • AI 'breakthrough': neural net has human-like ability to generalize language:
    • Analysis: A neural network has been created with the ability to generalize language, similar to humans. The artificial intelligence (AI) system can quickly incorporate newly learned words into its vocabulary and use them in various contexts, showcasing systematic generalization. In comparison, the popular chatbot ChatGPT performs poorly at this task despite its human-like conversational abilities. This breakthrough could lead to more natural interactions between machines and humans. Neural networks have traditionally struggled with systematic generalization, while humans effortlessly use new words in different settings. The study involved testing 25 individuals on their ability to deploy newly learned words in different situations, laying the foundation for training neural networks to be more systematic.
  • Show HN: Spice.ai – Building blocks for data and time-series AI applications:
    • Analysis: Spice.ai, an application platform for building intelligent data and time-series AI applications, has announced its general availability. The platform provides developers with composable data and AI building blocks, including pre-trained machine learning models and a petabyte-scale cloud data platform. The platform also includes a new community-centric developer hub where developers can create, fork, and share datasets and ML models. Additionally, Spice.ai has introduced new tools such as Spice Firecache, a real-time, in-memory SQL service, and Spice Functions, a hosted function compute service for data transformation and processing. Developers can also customize their data with custom datasets and views.
  • Show HN: Prompt-Engineering Tool: AI-to-AI Testing for LLM:
    • Analysis: The content is a GitHub repository belonging to the user "artas728" titled "spelltest". It is described as an AI-to-AI Testing simulation framework for LLM-based applications. The repository contains code files, examples, and documentation related to the spelltest framework. It is licensed under the MIT license and has gained 52 stars and 2 forks. The content provides information on how to use the framework, including describing simulations, running simulations, and analyzing results. The repository also includes important notices for users, such as project maturity warning and cost expectations. Overall, the content serves as a resource for implementing AI-to-AI testing using the spelltest framework.
  • Cosine: AI Co-Developer:
    • Analysis: Cosine is an AI knowledge engine that helps developers understand and work with their codebase. It offers features such as answering coding questions, explaining code features, and providing suggestions for new code. The tool can be installed either through the website or as a Visual Studio Code extension. Unlike other AI copilots, Cosine reads every code snippet in the codebase, providing a higher level of contextual semantic understanding. It utilizes OpenAI to create semantic embeddings for code, which are stored locally on the user's system for privacy. Cosine does not store code on its servers and the OpenAI API is configured to not submit user inputs for training. Testimonials from users highlight the tool's usability and effectiveness in improving productivity and understanding of code.
  • Dante: Advanced custom AI chatbots trained on your data:
    • Analysis: Dante AI is an AI chatbot builder that allows users to create custom AI chatbots without any coding. The platform offers features such as the ability to upload various file types, share website links, images, and videos to train intelligent chatbots. Dante AI can be integrated with websites and other platforms using an iframe or a chat bubble. The platform also provides advanced insights, analytics, and customization options. Users can store their data securely on encrypted AWS servers, and the platform supports over 100 languages. Dante AI offers a 2-day free trial and integrates with Zapier for compatibility with over 5,000 applications.
  • Nightshade, a tool allowing artists to 'poison' AI models:
    • Analysis: Nightshade is an open-source tool that allows artists to "poison" AI models with corrupted training data, making the models generate incorrect and distorted output. The tool, developed by researchers at the University of Chicago, can be added to digital art before uploading it to the web, altering pixels in a way that is invisible to the human eye but confuses AI models. For example, an image of a dog can be poisoned to make an AI model perceive it as a cat. The tool aims to give artists a way to fight back against AI models that use their work without permission.
  • Lexis+ AI Is Generally Available for US Customers:
    • Analysis: LexisNexis has released Lexis+ AI, a product that uses large language models to answer legal research questions, summarize legal issues, and generate legal document drafts. It is now available for all US customers after being tested and refined with various law firms, legal departments, and courts. Lexis+ AI claims to deliver trusted results with "hallucination-free" linked legal citations, offering accurate and verifiable information. The product is said to be "multiple times faster" than other legal generative AI solutions and allows users to provide instant feedback for continual improvement. It also features conversational search capabilities. However, there is a dispute regarding LexisNexis' claim of being the only product with inline linked citations, as vLex's Vincent AI also offers this feature.
  • Puncia – Subdomain and Exploit Hunter Powered by AI:
    • Analysis: The content is a GitHub repository for a tool called Puncia, which is described as the "Panthera(P.)uncia of Cybersecurity - Subdomain & Exploit Hunter powered by AI." It is used to automate the process of finding subdomains and exploits in cybersecurity. The tool utilizes two AI-powered APIs - Subdomain Center and Exploit Observer - to gather results. Installation instructions are provided for installing Puncia from PyPi or from the source code.
  • AI titans throw a (tiny) bone to AI safety researchers:
    • Analysis: The Frontier Model Forum, an industry body consisting of Anthropic, Google, Microsoft, and OpenAI, has announced a $10 million fund to support research on tools for testing and evaluating advanced AI models. The fund will be administered by the Meridian Institute and will support researchers from academic institutions, research institutions, and startups. The initial funding will come from the Frontier Model Forum and its philanthropic partners. The fund's primary focus will be on developing evaluation techniques for potentially dangerous capabilities of frontier AI systems. While $10 million is a significant amount, it is comparatively conservative given the resources of the fund's backers. Other organizations, such as Open Philanthropy and The Survival and Flourishing Fund, have provided larger grants for AI safety research. The Frontier Model Forum alludes to the possibility of a larger fund in the future.
  • US orders immediate halt to some AI chip exports to China, says Nvidia:
    • Analysis: The US government has instructed chip designer Nvidia to immediately halt the export of some of its high-end artificial intelligence (AI) chips to China. The restrictions were originally scheduled to come into effect 30 days after October 17, when the Biden administration announced measures to prevent countries like China, Iran, and Russia from receiving advanced AI chips. Nvidia did not provide a reason for the accelerated timing of the restrictions and stated that it does not expect an immediate impact on its earnings. Fellow semiconductor company Advanced Micro Devices is also affected by the export restrictions, but has not commented on the matter. Intel, which sells its Gaudi 2 chips in China, stated that it is reviewing the regulations and assessing the potential impact. The US Department of Commerce declined to comment on the situation. The export ban includes Nvidia's A800 and H800 modified advanced AI chips, as well as the A100, H100, and L40S chips.
  • Jina AI launches open-source 8k text embedding:
    • Analysis: Jina AI, a Berlin-based artificial intelligence company, has launched jina-embeddings-v2, an open-source text embedding model with an 8K (8192 tokens) context length. This makes it comparable to OpenAI's proprietary model, text-embedding-ada-002, in terms of capabilities and performance. Jina-embeddings-v2 outperforms OpenAI's model in multiple benchmarks such as classification, reranking, retrieval, and summarization. The model was built from scratch over the course of three months, offering extended context potential for applications such as legal document analysis and medical research. Jina AI aims to provide innovative solutions in the field of AI.
  • MetaCLIP – Meta AI Research:
    • Analysis: The content is a GitHub repository for MetaCLIP by facebookresearch. It includes curation/training code, metadata, distribution, and pre-trained models. The repository contains various files and folders related to the project.
  • Show HN: Polyfire – Javascript SDK to build AI apps without a backend:
    • Analysis: The content is a GitHub repository for a project called polyfire-js, which is described as a managed backend for AI apps. It allows for calling AI models from the frontend. The repository includes various files and folders related to the project, such as code, examples, and configuration files. There have been several commits to the repository, with the latest one made on October 26, 2023.
  • The Beatles: 'final' song Now and Then to be released thanks to AI technology:
    • Analysis: The long-awaited "final" Beatles song, titled "Now and Then," is set to be released thanks to AI technology used to enhance the audio in Peter Jackson's documentary, "Get Back." The software was used to separate John Lennon's vocals from a demo tape he left to Paul McCartney. The song was written and sung by Lennon in the late 1970s and was given to McCartney by Yoko Ono in 1994. Technological limitations at the time prevented the vocals and piano from being separated and added to new parts recorded by the other three Beatles. However, using AI-assisted software, the original audio from footage of the Beatles recording their final album, "Let It Be," was de-mixed and isolated for the documentary. This technology was then used to isolate Lennon's vocals and piano for "Now and Then," and McCartney and Ringo Starr have contributed new parts to the song, with George Harrison's guitar parts from 1995 also included.

Thank you for reading this week's summary! Stay tuned for more exciting updates from the world of AI and Machine Learning!

🚀 Quick Navigation

  1. Visual Diagrams
  2. Research Papers
  3. Projects
  4. Links & Articles
  5. License

📊 Visual Diagrams

For a pictorial insight:

📄 Research Papers

Dive deep into the world of research with these enlightening papers:

💼 Applications and Deployment

  • Think before you speak: Training Language Models with Pause Tokens

    • 📜 Read the Paper
    • 📝 Language models traditionally generate responses token-by-token but introducing a learnable "pause token" allows the model to process additional computation before finalizing an answer.

  • MotionLM: Multi-Agent Motion Forecasting as Language Modeling

    • 📜 Read the Paper
    • 📝 The study introduces MotionLM, a model that predicts multi-agent motion by treating trajectories as sequences of discrete motion tokens, akin to a language modeling task.

  • Large Language Models for Compiler Optimization

    • 📜 Read the Paper
    • 📝 A study on the use of Large Language Models for code optimization.

  • PROMPT2MODEL: Generating Deployable Models from Natural Language Instructions

    • 📜 Read the Paper
    • 📝 A method that transforms natural language task descriptions into deployable models.

📊 Data Handling and Annotation

  • Language Models Represent Space and Time

    • 📜 Read the Paper
    • 📝 Large language models (LLMs) like Llama-2 exhibit capabilities that suggest they learn coherent world models rather than just superficial statistics.

  • AnnoLLM: Making Large Language Models to Be Better

    • 📜 Read the Paper
    • 📝 Large language models (LLMs) like GPT-3.5 can serve as effective crowdsourced annotators when given sufficient guidance and example demonstrations.

  • Want To Reduce Labeling Cost? GPT-3 Can Help

    • 📜 Read the Paper
    • 📝 The study explores using GPT-3 as a cost-effective data labeler for training other models in NLP tasks.

  • DataFinder: Scientific Dataset Recommendation from Natural Language Descriptions

    • 📜 Read the Paper
    • 📝 The study introduces the task of recommending datasets based on a short natural language description of a research idea.

🎨 Graphics and Imaging

  • Traveling Words: A Geometric Interpretation of Transformers

    • 📜 Read the Paper
    • 📝 This paper offers a novel geometric perspective to understand the inner workings of transformers, particularly how layer normalization confines features to a hyper-sphere, aiding attention in shaping word semantics.

  • CoRF: Colorizing Radiance Fields using Knowledge Distillation

    • 📜 Read the Paper
    • 📝 This research introduces a method to generate colorized novel views from input grayscale multi-view images using Neural Radiance Fields (NeRF). The proposed distillation-based technique effectively transfers color knowledge from 2D colorization methods to the radiance field network, ensuring 3D consistency and producing superior results in both indoor and outdoor scenes compared to other methods.

🚀 Retrieval Techniques

  • Security Weaknesses of Copilot Generated Code in GitHub

    • 📜 Read the Paper
    • 📝 Modern code generation tools, especially those utilizing Large Language Models (LLMs) like GitHub Copilot, pose security challenges, leading to potential vulnerabilities in the generated code.

  • Towards Self-Assembling Artificial Neural Networks through Neural Developmental Programs

    • 📜 Read the Paper
    • 📝 This paper explores the creation of neural networks inspired by the self-organizing developmental processes seen in biological nervous systems.

  • Precise Zero-Shot Dense Retrieval without Relevance Labels

    • 📜 Read the Paper
    • 📝 Introduction of Hypothetical Document Embeddings (HyDE) for zero-shot dense retrieval.

🛠️ Projects

💡 Model Optimization & Fine-Tuning

Data Cleaning and Labeling

  • 🌟 Cleanlab - An aid for cleaning data and labels in ML datasets. Check it out.

Model Adaptation and Fine-Tuning

  • 🌟 LLaMa-Adapter - A lightweight adaptation method for fine-tuning Instruction-following and Multi-modal LLaMA models. Learn more.
  • 🌟 axolotl - A tool designed to streamline the fine-tuning of AI models. Check it out.

🌐 Multimodal Models & Applications

Curated Lists and Collections

  • 🌟 Awesome-Multimodal-Large-Language-Models - A curated list of Multimodal Large Language Models. Explore here.

Frameworks and Systems

  • 🌟 DreamLLM - A learning framework focusing on synergy between multimodal comprehension and creation. Dive in.
  • 🌟 NExT-GPT - The first end-to-end MM-LLM for text, image, video, and audio. Discover more.

🛍️ Collections & Repositories

LLM Collections and Demonstrations

  • 🌟 LargeLanguageModelsProjects - A collection of llama models in different configurations. Explore.

Toolkits and Utilities

  • 🌟 JVector - A Java-based vector search engine with no dependencies. Check it out.
  • 🌟 AutoGen - AutoGen is a tool designed for building next-gen LLM applications centered on multi-agent conversations, streamlining the orchestration, automation, and optimization of complex LLM workflows. Check Out
  • 🌟 LiteLLM - Manages inputs to the provider's completion and embedding endpoints. Discover.
  • 🌟 AutoGPT - A modular toolkit for AI agents. Explore on GitHub.
  • 🌟 localGPT - Interact with documents locally ensuring data privacy. Check it out on GitHub.
  • 🌟 LLM-Finetuning-Hub - Resources for finetuning LLMs tailored to specific use cases. Learn more on GitHub.
  • 🌟 chatgpt-history-export-to-md - Convert your ChatGPT history and data export into neatly formatted Markdown files. It includes YAML headers and a Code interpreter for Advanced Data Analysis. Check it out on GitHub.

🔗 Links & Articles

📈 Research & Innovations

🌌 General AI Insights

AI Model Philosophies and Debates

AI Technologies and Capabilities

  • Microsoft introduces AI meddling to your files with Copilot in OneDrive:Microsoft is set to revamp OneDrive with AI-driven features, with some rolling out in December and others in early 2024.
  • Picking a vector database: a comparison and guide for 2023: The landscape of vector databases in 2023 is diverse, with each offering unique features catering to different needs. While Milvus stands out in performance and community strength, Pinecone shines in developer experience and hosting; the ideal choice varies based on project specifics, budget, and preferences, with the author leaning towards Pinecone and Milvus for their performance and flexibility.
  • Graph-mining: Google's tools for tasks with inherent graph structures, used in products like Search, YouTube, and Maps.
  • OnnxStream: A specialized inference library designed to minimize memory consumption.
  • Released L2E OS v0.1 "Temple DOS" . A new OS on the block! The first OS that boots to AI!: A new operating system that boots directly to an AI interface.
  • Decentralized Artificial Intelligence: The article discusses the need for decentralization in Large Language Models (LLMs) and AI.
  • Microsoft is going nuclear to power its AI ambitions: Microsoft is exploring the potential of next-generation nuclear reactors to power its data centers and support its AI initiatives.
  • Run any ML model from any programming language: Carton is a tool that packages machine learning models with metadata in a zip file without altering the original model, simplifying the model deployment process by automatically selecting the appropriate runner based on the metadata. Once packed, Carton's framework-agnostic API facilitates model inference, with the software built in Rust and offering bindings to multiple languages.
  • The next generation of smart glasses: Ray-Ban and Meta collaborate on sunglasses that video record, and image capture, and can be queried via Meta’s AI
  • Vectorize: a vector database for shipping AI-powered applications to production, fast: Vectorize is a new vector database from Cloudflare designed to help machine learning models "remember" and enhance AI-powered applications. It allows developers to build full-stack AI applications on Cloudflare's global network, enhancing semantic search, classification, recommendation, and anomaly detection use-cases. Vectorize is in open beta and integrates with Cloudflare Workers, enabling it to power various applications, including improving LLMs' accuracy and context, and supporting embeddings from platforms like OpenAI and Cohere.
  • Workers AI: serverless GPU-powered inference on Cloudflare’s global network: Cloudflare is launching Workers AI, a platform that allows developers to run AI models using a few lines of code without managing infrastructure, emphasizing accessibility, serverless operation, and privacy. The service offers popular open-source AI models, ensures data privacy by default, and will soon expand its model offerings through a partnership with Hugging Face, with plans for rapid GPU rollout across global data centers by 2024.
  • You can give ChatGPT a picture of your team’s whiteboarding session and have it write the code for you.: This is absolutely insane.
  • Mistral 7B: Mistral 7B, a 7.3B parameter model, outperforms various Llama versions on benchmarks, excels in both code and English tasks, and uses efficient attention mechanisms for better performance. Available under the Apache 2.0 license, it's easy to deploy across platforms, and a fine-tuned chat variant surpasses Llama 2 13B chat in performance.
  • AI startup Lamini bets future on AMD's Instinct GPUs: AI startup Lamini has chosen to exclusively use AMD's Instinct GPUs for its platform that refines large language models (LLMs), setting itself apart from many competitors that rely on Nvidia's hardware. While Lamini's platform has garnered interest from major companies like Amazon and Walmart, AMD's focus on expanding its software ecosystem and forthcoming hardware upgrades aim to make its AI accelerators more attractive and accessible to developers and businesses.
  • Vicious Self-Degradation: When a frequent query is Googled, Quora identifies it, uses ChatGPT to generate an answer that may contain a hallucination, and this ChatGPT-generated response becomes the top Google answer.
  • Inside the Matrix: Visualizing Matrix Multiplication, Attention and Beyond: The article introduces "mm", a visualization tool designed to display matrix multiplications (matmuls), which are foundational to machine learning models. Using three-dimensional visualizations, "mm" enables clearer understanding of complex matrix operations and compositions, especially benefiting visual thinkers, and covers various topics from basic matrix operations to the intricate workings of GPT-2 attention heads, demonstrating the benefits of this geometric approach to understanding algebraic properties in matrix computations.
  • Can you beat a stochastic parrot?: Play chess against GPT-3.5.

Ethical and Societal Implications of AI

  • National Security Agency is starting an artificial intelligence security center: The NSA is launching an AI security center to bolster defense and intelligence systems against threats.
  • WebGPU Technical Report: WebGPU introduces extensive attack surfaces to Chrome's GPU process, including the core WebGPU, third-party Usermode Graphics Drivers, and shader compilers, emphasizing the complexity that may result in vulnerabilities. Despite significant efforts to validate input and extensive fuzzing, there remain concerns about potential vulnerabilities in areas like Dawn's use-after-frees, callbacks, the Chrome Command Buffer, and the SwiftShader JIT compiler, indicating a need for ongoing vigilance and manual audits. assistance 24/7 without human intervention, though human volunteers remain available for more nuanced assistance.
  • Be My AI: Be My Eyes, a platform connecting volunteers with visually impaired users, has integrated an AI feature called "Be My AI" to assist with everyday tasks and is now launching it in an open beta phase for iOS users, with an Android version in the pipeline. "Be My AI" allows users to take photos and receive detailed descriptions, proving invaluable for tasks like reading labels, organizing wardrobes, and getting visual assistance 24/7 without human intervention, though human volunteers remain available for more nuanced assistance.
  • Signal’s Meredith Whittaker: AI is fundamentally ‘a surveillance technology’: Meredith Whittaker, Signal's president, emphasized at TechCrunch Disrupt 2023 that AI is fundamentally a surveillance technology, deeply intertwined with the big data and targeting industry dominated by tech giants. While acknowledging not all AI applications are exploitative, she highlighted the inherent surveillance nature of AI and the economic drivers behind facial recognition technology, noting that beneficial uses, like face blurring in Signal's app, are overshadowed by more intrusive applications driven by profit motives.

AI in Education

  • Student Use Cases for AI: Generative AI tools, particularly large language models (LLMs), present both opportunities and challenges in educational settings, offering students and educators unparalleled access to powerful AI systems. As AI becomes increasingly prevalent in classrooms, it's crucial for educators and students to understand the potential benefits, biases, and privacy concerns of AI, and to adopt best practices for interacting with these tools to ensure effective and safe usage.

LLM Architectures and Applications

Best Practices and Guidelines

AI in Multimedia and Collaboration

  • Meta starts rolling out generative AI tools for all advertisers: Meta Platforms is introducing generative AI tools for advertisers, allowing them to create content such as image backgrounds and text variations. This move, marking Meta's initial venture into integrating generative AI technology, will also soon enable businesses to use AI for messaging on platforms like Messenger and WhatsApp.
  • captcha with Bing: I've tried to read the captcha with Bing, and it is possible after some prompt-visual engineering.
  • What codegen is (actually) good for: Codegen, the automatic generation of code based on predefined rules, is gaining popularity, with tools ranging from simple code completion in IDEs to advanced AI-driven systems. While many developers use AI for code generation, skepticism remains about its accuracy; instead of viewing codegen as a full replacement, it's best seen as an extension of a developer, assisting in the design-to-development process by suggesting suitable tools, speeding up workflows, and acting as a reference, though it won't entirely replace established team patterns.
  • Introducing New AI Experiences Across Our Family of Apps and Devices: Meta is introducing a range of AI-driven features, including AI stickers for image editing on apps like Instagram, the advanced conversational assistant 'Meta AI' for multiple platforms, and 28 additional AIs with distinct personalities, some portrayed by celebrities. As they expand AI offerings for businesses and developers, they acknowledge potential challenges and emphasize a cautious approach with built-in safeguards.
  • Getty made an AI generator that only trained on its licensed images - Getty Images collaborated with Nvidia to introduce a tool allowing users to produce images from its extensive library, ensuring full copyright indemnification.
  • ChatGPT can now see, hear, and speak - ChatGPT introduces new voice and image capabilities, emphasizing user safety, privacy, and understanding model limitations.
  • Expanding access to safer AI with Amazon - Amazon's collaboration with Anthropic aims to create industry-leading foundation models and make advanced models like Claude 2 accessible through Amazon Bedrock.

🛠️ Tools & Databases

LLM Databases and Search Engines

LLM Integration and APIs

  • Llama API - This notebook shows how to use LangChain with LlamaAPI.

LLM Toolkits and Platforms

  • Embedchain - Start building LLM-powered bots in under 30 seconds.
  • Two-Tower Embedding Model - A training approach aligning embeddings from two modalities, like images and text, useful for personalized recommendation systems.

📚 Tutorials & How-Tos

🔖 License

All content in this repository is shared under the MIT License. Please refer to the LICENSE file for more details.

About

A repository to store helpful information and emerging insights in regard to LLMs

Resources

Readme

License

MIT license
Activity

Stars

20 stars

Watchers

2 watching

Forks

0 forks
Report repository

Releases

No releases published

Packages

No packages published