LAI #88: GNNs for Knowledge Graphs, DSPy Signatures, and How LLMs Are Really Trained
Structured AI programming, Conway's Game of Life, and what a ChatGPT outage reveals about human-AI entanglement.
Good morning, AI enthusiasts,
This week’s issue spans from deep technical builds to thought-provoking social insights. Our feature dives into Graph Neural Networks for knowledge graphs, showing how to generate powerful node embeddings with PyTorch Geometric, complete with a biomedical case study and code. We also break down:
How engineers actually train LLMs at scale, from ZeRO to speculative decoding
DSPy’s “Signatures” and “Modules” for structured, model-agnostic AI development
A Python walkthrough of Conway’s Game of Life and its emergent patterns
What a one-hour ChatGPT outage says about our growing emotional ties to AI
Plus, in the community: a protocol for AI memory management, open collaboration calls, and a dataset-free GUI detection workflow you can join.
Let’s get into it!
— Louis-François Bouchard, Towards AI Co-founder & Head of Community
Learn AI Together Community Section!
Featured Community post from the Discord
Silentsentinel6943 has built MARM, a protocol-based memory management for AI. It is designed to improve response accuracy, reduce drift, and stabilize context over long sessions. Check it out on GitHub and support a fellow community member. If you have any ideas to improve it, connect in the thread!
AI poll of the week!
Looks like GPT-5 is far from a universal crowd-pleaser.
42% are fully on board with it, and a not-so-small 20% want the old model lineup back entirely. That’s a pretty even spread between excitement, indifference, and outright resistance, suggesting OpenAI’s big shift isn’t landing the same way for everyone.
Is this split due to actual capability differences, or because people’s workflows, habits, and trust were built around the older models and don’t translate neatly to GPT-5’s new approach? Tell me in the thread, let’s talk!
Collaboration Opportunities
The Learn AI Together Discord community is flooded 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. Ainews_mythofsisyphus is building an AI video editor and is looking for a partner who can help with the configuration, login, security, etc. If you are based out of the EU/US, connect in the thread!
2. .ghostvoices is working on an application that acts as Cluely for mobile and is looking for a team to take it beyond a demo. If you are interested, reach out in the thread!
3. Raphael_219 is working on a personal AI assistant called Ava and is looking for someone experienced with LangChain, tool return formatting, and debugging agents who can help finalize Ava’s tool/agent loop. If this sounds like something you would like, contact him in the thread!
Meme of the week!
Meme shared by superuser666_sigil
TAI Curated section
Article of the week
Graph Neural Networks for Knowledge Graphs By Michael Shapiro MD MSc
Generating useful node embeddings from large, featureless Knowledge Graphs presents a significant challenge. This article provides a practical guide to address this, detailing the process of training a Graph Neural Network (GNN) using PyTorch Geometric. With a biomedical KG as an example, it covers data preparation, including structuring the heterogeneous graph and making it undirected for better information flow. It then demonstrates how to train a model with a GraphSAGE encoder for a self-supervised link prediction task, providing code for the training loop and for extracting the final embeddings for downstream use.
Our must-read articles
1. How Are LLMs Trained: For Engineers By Harsh Chandekar
For engineers facing the computational challenges of LLMs, this piece offers a technical overview of optimization strategies. It covers the training phase, detailing memory management with ZeRO, 3D parallelism, and parameter-efficient methods like knowledge distillation. It also discusses inference, explaining model compression through quantization and faster decoding techniques like speculative decoding. Finally, it illustrates how these methods are essential for building practical and efficient AI systems, bridging the gap between massive models and real-world applications.
2. Understanding Signatures and Modules By Souradip Pal
This article provides a foundational guide to DSPy, a framework that shifts AI development from manual prompt engineering to a more structured programming approach. It focuses on two core concepts: Signatures, which declaratively define the inputs and outputs of a task, and Modules, which are reusable components that execute reasoning patterns like `dspy.Predict` and `dspy.ChainOfThought`. It illustrates how to configure language models, create both simple and advanced signatures, and compose modules into multi-step programs, presenting a more systematic and model-agnostic way to build AI applications.
3. Watch Life Emerge from Code: Conway’s Game of Life in Python By Jack Ka-Chun, Yu
Exploring how complex, life-like patterns can emerge from simple rules, this piece walks through Conway’s Game of Life. It explains the system’s foundational principles as a cellular automaton and details the four rules governing cell survival and birth. It also discusses the origins of these rules, highlighting their empirical design by John Conway rather than derivation from biology. The article also provides a complete Python implementation using `numpy` and `matplotlib`, offering a hands-on project for anyone interested in simulation and emergent behavior.
4. What an Hour Without ChatGPT Tells Us About Our Current State of Entanglement By Cedar Compson
Using a one-hour ChatGPT outage as a case study, the author examines the concept of human-AI “entanglement.” An analysis of user comments on Downdetector during the disruption reveals sentiments extending beyond functional frustration. Many users expressed relational attachment, dependency-related anxiety, and anthropomorphic concerns, indicating AI’s role has shifted from a tool to an integrated companion. These reactions, particularly from a younger demographic, suggest a deepening cognitive and emotional integration with AI, highlighting a significant development in how society interacts with this technology.
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.