Author(s): Towards AI Editorial Team Originally published on Towards AI. Good morning, AI enthusiasts! This week, we explore LLM optimization techniques that can make building LLMs from scratch more accessible with limited resources. We also discuss building agents, image analysis, large concept models (LCM), and more. We also have a paid opportunity open at Towards, check it out in the collaboration opportunities section. Enjoy the read!Whats AI WeeklyThis week in Whats AI, I will discuss the basics of knowledge distillation and the other techniques like pruning and quantization that can help you build models with limited resources. What if the big companies just created one big model, and then we, the normal people, could use it to train smaller models that each of us would use for our specific tasks or data? Well, thats exactly Llamas goal with releasing the 405B model and leveraging distillation to train the smaller ones. Likewise, Nvidia recently released two papers exploring this amazing idea with their Minitron approach. I will also dive into this. Read the article here or watch the video on YouTube. Louis-Franois Bouchard, Towards AI Co-founder & Head of CommunityLearn AI Together Community section!AI poll of the week!If reasoning tokens arent the key to AGI, what alternative paths or approaches do you think hold the most promise? Tell us on Discord!Collaboration OpportunitiesThe Learn AI Together Discord community is flooding with collaboration opportunities. If you are excited to dive into applied AI, want a study partner, or even want to find a partner for your passion project, join the collaboration channel! Keep an eye on this section, too we share cool opportunities every week!1. For anyone interested, we have a Developer Relation type of role open at Towards AI. We need someone interested in creating social and blog posts, helping with Discord admin stuff, and attending research/industry events we sponsor representing TAI. If this sounds like something exciting, message Louis-Franois Bouchard on Discord.2. Mesmoiron needs help building advanced components with Toddle and Xano platforms. If you can help with this, connect in the thread!3. Ghub01 has developed an online course and needs a partner to market it. If you need more details, reach out in the thread!4. Iruletheworldnow is looking for five people to study agentic frameworks and build an open-source one. If this sounds exciting, contact him in the thread!Meme of the week!Meme shared by bin4ry_d3struct0rTAI Curated sectionArticle of the weekReimagining GANs: Bridging Statistics and Variance Regularization By Shenggang LiThis article explores Generative Adversarial Networks (GANs) through the lens of statistical modeling. It begins by explaining GANs using logistic regression on tabular data, making the concept more accessible. It then introduces Variance-Regularized GANs (VR-GANs), which improve the statistical alignment of generated data with real data by explicitly minimizing the variance discrepancy. Finally, it proposes Reason Code GANs, enhancing the discriminator to provide feedback explaining why generated data is deemed fake. This feedback guides the generator, improving data generation and offering increased interpretability. It includes Python code examples demonstrating both VR-GANs and Reason Code GANs and explores applications for image generation, showing how Reason Codes can track feature evolution during training. It concludes by discussing future research directions for GANs, emphasizing the potential for improved stability and interpretability.Our must-read articles1. LCM + Mapping Hidden Embedding = New Architecture Model By Gao Dalie (This article discusses Metas new Large Concept Model (LCM), a language model architecture that differs significantly from traditional Large Language Models (LLMs). Unlike LLMs, which process text word-by-word, LCMs process sentences as concepts high-dimensional vectors representing the meaning of a sentence, independent of the language or modality. It details LCMs architecture, comprising a Concept Encoder, Large Concept Model, and Concept Decoder, which work together to convert text into concepts and back again. A section explores the provided open-source code, detailing the PreNet, PostNet, and TransformerDecoder components, focusing on the normalization and denormalization of embeddings. It highlights LCMs ability to handle multiple languages using a SONAR tool and its potential for zero-shot learning. It concludes by emphasizing LCMs potential to revolutionize natural language processing by shifting the focus from word processing to semantic understanding.2. Transform Image Data into Insights with VisualInsights AI Automation By Yotam BraunThis article details VisualInsight, an AI-powered image analysis application built using Streamlit, Google Gemini, and AWS services. It addresses common image analysis challenges like manual effort and scalability by automating the process. VisualInsight allows users to upload images via a user-friendly interface, leveraging Gemini for analysis and storing results securely in AWS S3. The applications architecture incorporates Docker for consistent performance and Terraform for infrastructure management, with CI/CD implemented via GitHub Actions for automated deployments. It also provides core code snippets and explains the workflow, highlighting the benefits of automation and reproducibility.3. Building AI Agent Systems: A Deep Dive into Architecture and Intuitions By Prashant KalepuThis article discusses the architecture of AI agents, highlighting their evolution from monolithic models to more adaptable, compound systems. It explains that effective AI agents rely on three pillars: reasoning, acting, and memory. A sample architecture is presented, featuring a reasoning block, an execution block, and a memory block. The feedback loop allows for continuous improvement, while future enhancements could include improved contextual understanding and self-learning capabilities. It concludes by emphasizing the potential of AI agents for solving complex, real-world problems more effectively.4. Building AI Agent from Scratch with Ruby By Alex ChaplinskyThis article details creating a Ruby-based AI agent framework. It utilizes the ReAct architecture, interleaving reasoning and action via an LLM. The frameworks core components include an Agent class managing interactions, a Session class tracking individual interactions (with Spans representing individual operations), and a Toolchain for integrating external tools. The article demonstrates building a simple agent that retrieves cryptocurrency prices using the CoinGecko API, showcasing the frameworks flexibility and extensibility through examples of progressively complex queries, highlighting the agents ability to handle both simple and more intricate requests, even gracefully handling requests outside its defined capabilities.If you are interested in publishing with Towards AI, check our guidelines and sign up. We will publish your work to our network if it meets our editorial policies and standards.Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming asponsor. Published via Towards AI