LAI #69: What Exactly Are AI Agents? Let's Solve This and Build Them
Also, Multi-Teacher Distillation, Casual Inference, and more!
Good morning, AI enthusiasts! This week, we are diving into fundamentals with Notebook comparisons (Jupyter vs. Google Colab), defining AI agents once and for all, casual inference, techniques like multi-teacher distillation, and more!
Dive in; we also have another very hands-on guest post for you!
What’s AI Weekly
This week in What’s AI, I want to talk about something that has become super popular in data science, machine learning, and all sorts of coding adventures: Jupyter Notebooks and Google Colab. These are two fantastic environments that let you write code, see the outputs right away, and add notes and explanations for yourself or others. You might have heard of both and thought they were basically the same thing, or maybe you’ve never really used either and want to know which is best to start with. Whichever category you’re in, I’ll walk you through a friendly introduction to Jupyter Notebooks and Google Colab, compare them, and share some tips on when you might want to use each one. Read the complete article here or watch the video on YouTube!
— Louis-François Bouchard, Towards AI Co-founder & Head of Community
Learn AI Together Community section!
Featured Guest Post
We have yet another guest post this week, this time with LLM Watch (aka
), and we are tackling a common challenge in AI — the confusion around AI agents. Even experts often grapple with varying definitions, so we’re breaking it down while also walking through how to build an email automation agent using Hugging Face’s SmolAgents library.AI Agent for Email Automation with SmolAgents
AI poll of the week!
This week, instead of a poll, we want your input. What is YOUR DEFINITION of an agent (or “agentic AI”)? Please comment in the thread. We’d love to gather many definitions and try to formulate a more “general” one from all our opinions!
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. Threedogowww is looking for collaborators to work on an experimental project. This is a good way to learn more programming, so if you want to get into it, reach out to him in the thread!
2. Winnerikpe is working on building 20+ hands-on projects while learning deep learning, data science, and more. If you would like to join, connect with him in the thread!
Meme of the week!
Meme shared by dimitriye98
TAI Curated section
Article of the week
Gemma 3 + MistralOCR + RAG Just Revolutionized Agent OCR Forever By Gao Dalie (高達烈)
Mistral AI’s release of Mistral OCR, combined with Google’s Gemma 3, has revolutionized OCR agents. Mistral OCR excels at understanding complex document elements and converting data into Markdown, while Gemma 3 delivers state-of-the-art performance, supports multiple languages, and analyzes images and text with a 128K token context window. The blog details how these technologies are integrated with Retrieval-Augmented Generation (RAG) to create a Streamlit application for document processing and question-answering. This application enables users to upload documents or images, process them using OCR, and then engage in a chat interface powered by Google Gemini. The detailed code provides a practical example, making it easier for developers to implement similar solutions.
Our must-read articles
1. Adaptive Multi-Teacher Distillation for Enhanced Supervised Learning By Shenggang Li
The author presented an innovative adaptive multi-teacher distillation method to enhance supervised learning. This method involves a lightweight neural network, the “student,” dynamically learning from multiple predictive models, the “teachers” (XGBoost, LightGBM, and Random Forest). The student model uses attention weights to determine the influence of each teacher on specific predictions, thus capturing complex patterns that individual models might miss. The method’s effectiveness was demonstrated through a marketing analytics case study using synthetic customer data. The student model outperformed the individual teacher models, achieving a test AUC of 0.7574. The attention weights revealed that the student model favored LightGBM, showcasing the model’s ability to discern and leverage the strengths of each teacher.
2. The Architecture of Agency: Critical Challenges in Multi-Agent AI Systems By Andy Spezzatti
The article explores critical challenges facing multi-agent AI systems, which are computational entities capable of autonomous reasoning, planning, and action. It highlights the debugging complexity arising from the opacity of agent interactions and the need for advanced observability tools. The piece also addresses memory architecture limitations and proposes solutions mimicking human cognitive systems, along with evaluation methodologies for assessing system-level performance. It further discusses adversarial vulnerabilities exploiting semantic ambiguities and reasoning pathways and the impact of latency and computational efficiency on user experience. These challenges emphasize the necessity for integrated research programs combining the flexibility of neural approaches with the reliability of symbolic computing paradigms for those interested in the future of AI agent systems.
3. Causal Inference is a Minefield — Here’s How to Navigate It with DoWhy by Torty Sivill
The article explores the complexities of causal inference using Microsoft’s DoWhy library to determine the effect of discounts on customer churn for a beer subscription company. It covered the importance of defining a causal graph, identifying estimands, estimating causal effects, and refuting the results to ensure accuracy. It highlights the critical role of accurate causal graph specification and interpreting the estimated effects, underscoring that flawed assumptions lead to misinterpretations.
4. Visualizing Recursion Tree By Han Qi
The author details their challenging journey of creating a recursion tree visualizer, struggling with tools like VSCode debugger and existing visualizers. Experimented with various approaches, including Graphviz, D3.js, and Plotly, aided by ChatGPT. Ultimately, Succeeded in using a class-based decorator to separate visualization logic, enhancing code readability. The author emphasized balancing LLM assistance with critical thinking for effective problem-solving.
If you want to publish 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.