LAI #85: Agents That Work, LLaVA Training, and the $40K RAG Deal

Author(s): Towards AI Editorial Team

Originally published on Towards AI.

Good morning, AI enthusiasts!

Everyone’s excited about agents, until they have to actually build one. This week’s top stories are a perfect reality check. In What’s AI, we break down what agents really are, where they work best, and when they’re not worth the complexity.

We also dive into:

• A complete, open-source blueprint for deploying your own AI toolchain server (FastMCP + LangGraph + Claude Desktop)
• A $40K deal closed with a Llama 3-based RAG system, fully automated via n8n and Streamlit
• A hands-on tutorial for training LLaVA multimodal models on a budget
• And a sharp critique of how LLMs multiply tokens but still miss meaning

Plus, this week’s poll shows 61% of our readers think ChatGPT Agents are still more hype than real progress. Agree or disagree? Let us know in the Discord thread.

Let’s get into it!

What’s AI Weekly

Everyone seems to be calling 2025 the year of AI agents, but what does that actually mean? This week, in What’s AI, I break it down without the hype. It starts by defining what AI agents really are (beyond the buzzword), looks at how they’re currently being built, where the technology is headed, and what’s likely to stick versus fade. Most importantly, it offers a grounded take on when you might actually need one and when it makes sense to start building. You can read the full piece here or watch the video version if you prefer a visual walkthrough.

— Louis-François Bouchard, Towards AI Co-founder & Head of Community

Learn AI Together Community Section!

Featured Community post from the Discord

Johaoenoc has launched a quick project that lets you upload documents and chat with them. You can upload PDFs, DOCX, or TXT and ask it to summarize, answers questions, and help explore content. You can check it out here and support a fellow community member. If you have any feature ideas or suggestions, drop them in the thread!

AI poll of the week!

58% of people still think ChatGPT Agents are more hype than progress. It reflects a broader skepticism in the AI space: powerful demos aren’t enough anymore. If you’re in the “hype” camp, what’s missing for you to call it real progress? And for the “progress” voters, what use case actually clicked for you? Tell me in the thread!

Collaboration Opportunities

The 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. Harryyyy9049 is working on an AI Agent project and is looking for a developer/designer who’s interested in collaborating to build a desktop GUI. If this sounds up your alley, reach out in the thread!

2. Buddei_ is launching pilot collaborations for Behavioral Reasoning Primitives (BRP) and the Policy-to-Deterministic-Directed-Graph (PDDG) compiler. They are looking for pilot partners, research collaborators, enterprise use cases, and formal verification & runtime integration testers. Check the thread to know more!

3. Jinj4 is looking for someone who can partner up to learn everything about video generation from text prompts. If this is relevant for you or your niche, connect in the thread!

Meme of the week!

Meme shared by efficientnet_99825

TAI Curated Section

Article of the week

End-to-End Guide to Building and Deploying an MCP Server for AI Toolchains By Vikram Bhat

A key challenge in AI development is enabling large language models to securely interact with real-world tools. This guide addresses this by demonstrating how to build and deploy a Model Context Protocol (MCP) server. Using FastMCP library, it details creating a functional Google search tool, from coding the server to testing with MCP Inspector. It also covers integration with clients like Claude Desktop and LangGraph agents, deployment on Render, and key considerations for security and performance, offering a complete walkthrough for developers interested in AI toolchains.

Our must-read articles

1. Closed a $40K Deal: Built a Context-Aware AI Agent with LLaMA 3 70B, Streamlit UI, and n8n Automation By Yogender Pal

A blueprint is provided for building a context-aware AI agent that queries private documents. The system integrates LLaMA 3 70B, LangChain, and ChromaDB into a Retrieval-Augmented Generation (RAG) pipeline. For user interaction, it features a Streamlit interface, while n8n automates document ingestion and re-indexing. Deployed on Kubernetes, the solution offers a complete, self-hosted framework for secure, real-time document analysis. The author details the architecture and provides a guide for implementing this system, which was delivered as a complete project.

2. Introduction to Multimodality With LLaVA By Marcello Politi

Focusing on resource-efficient implementation, this article demonstrates how to build a lightweight version of the LLaVA multimodal model. The process involves integrating a pre-trained CLIP-ViT image encoder with a TinyLlama language model, connected by a two-layer MLP adapter. For efficiency, the pre-trained components are frozen, with training focused solely on the adapter. It covers data preparation using a custom collator, model training with `Seq2SeqTrainer`, and finishes with an inference example, offering a practical overview of this multimodal architecture for low-resource environments.

3. Dot Product Thinking: How LLMs Multiply Tokens, But Miss Meaning By Ajay Deewan

This article examines how large language models function using a mathematical operation called the dot product. It explains that LLMs do not comprehend language but instead measure the directional similarity between high-dimensional token vectors to predict the next word in a sequence. This mechanism, while effective at producing fluent text, is presented as a form of statistical pattern matching rather than genuine understanding. It contrasts this computational process with human cognition, which involves memory, emotion, and lived experience. It also concludes by cautioning against mistaking a model’s linguistic coherence for true consciousness or meaning.

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 a sponsor.

Published via Towards AI