Thien Tran

Exploring the Foundations of LLMs — Through Stanford CS336 Recently, I’ve been following Stanford’s CS336: Large Language Models course—an advanced graduate-level class now publicly available via Stanford Online. As someone deeply interested in machine learning systems and NLP, this has been an incredible opportunity to expand my understanding from the ground up. 📚 Course website: https://lnkd.in/gY_fzNU4 📺 Lectures (free on YouTube): StanfordOnline – CS336 🎯 Key Learnings So Far: 🔹 Tokenization isn’t just preprocessing—it's a critical design decision that affects model performance, vocabulary efficiency, and even generalization. Exploring approaches like byte pair encoding gave me a new appreciation for the intricacies behind input representation. 🔹 Profiling resource usage is essential. The lectures on GPU memory tracking and PyTorch profiling offered valuable insights into how large models interact with hardware—and how to reason about performance bottlenecks early in development. 🔹 LLMs are a full-stack problem. From tokenizer to transformer block to GPU kernel, this course emphasizes that building scalable models requires a systems perspective, not just a model-centric one. 🔹 Hands-on implementation matters. The code-oriented structure of the course pushes learners to go beyond theory and actually build core components of a language model—something I’m already integrating into my own learning process. 🙌 Grateful to Stanford University and the CS336 teaching team—for making this world-class curriculum openly accessible. It's a fantastic resource for any student or practitioner eager to understand the building blocks of today’s AI systems. I’m looking forward to diving deeper into topics like parallelism, custom GPU kernels, evaluation techniques, and alignment strategies in the weeks ahead. If you’re following the course too or working on similar topics, I’d love to connect and learn together! #StanfordCS336 #LLMs #MachineLearning #DeepLearning #NLP #Transformers #AI #PyTorch #OpenLearning #AIEducation #DataScience

6

I got nostalgic on a gloomy Saturday morning in SF and peeked at CS 182’s Homework 0 this semester at Berkeley (https://lnkd.in/g8_BUy7Y). It asks students to chat with ChatGPT about deep learning—and reflect on whether that changes their goals for the semester. WHAT STOOD OUT FROM MY OWN DISCUSSION ✅ AI ≠ AUTOMATION — timers/scripts are useful, but intelligence reacts, reasons, and adapts. ✅ AI → (NON-ML + ML) → DL — rules/search/planning on one side; learning from data on the other; and deep learning as the branch that learns hierarchies of representations. MY PHILOSOPHICAL TAKE 1. Deep learning feels less like a better spreadsheet and more like a new sense—a non-human way of perceiving the world. 2. Its internal representations live in high-dimensional geometry that doesn’t map neatly to our concepts. 3. That “alien” quality isn’t a bug; it’s why DL surfaces structure we didn’t specify—similar to how billions of simple neurons give rise to perception, memory, emotion in the brain. I wrote a short, fun-but-technical piece on this: separating automation from AI, where DL fits, and why I treat neural nets as a perceptual lens we can harness (and guardrail) even if we can’t fully “read” them yet. 👉 Full article link attached below. Would love your take—especially if you teach, build, or deploy these systems. #DeepLearning #MachineLearning #AI #Automation #Neuroscience #PhilosophyOfAI #CS182

6

I am excited to share that our paper, "LEGO: A Layout Expression Language for Code Generation of Hierarchical Mapping," will be presented at CGO 2026 next week! In this work, we introduce LEGO, a new abstraction that decouples computation from layout details. Unlike previous approaches like CuTe and Graphene, LEGO: 1) Automated Stride Derivation: LEGO eliminates explicit stride specifications. Deriving index expressions directly from layout definitions removes the burden of manual low-level bookkeeping. 2) Flexible Integration for Tools: Built on Python and SymPy, LEGO provides a lightweight infrastructure that you can easily embed into your own tools and domain-specific languages (DSLs). 3) Versatile Code Generation: We demonstrate adaptability by targeting diverse backends like Triton and CUDA, and introduce a custom SymPy to MLIR backend to facilitate integration with modern compiler frameworks. 4) Generalized Bijective Support: Moving beyond rectangular shape algebras, LEGO supports any bijective mapping, including user-defined functions. This enables complex patterns, such as antidiagonal layouts. Paper: https://lnkd.in/gwFkTRsE Code: https://lnkd.in/gRvTrdBN Big thanks to my co-authors Cosmin E. Oancea and Mary Hall!

160

11 Comments

🚀 Excited to share that our team's new paper Delta Knowledge Distillation for Large Language Models is now on arXiv! 🎉 👉 https://lnkd.in/gHXmCeZ5 A long-standing debate in knowledge distillation is whether the student and teacher should share the same representation space, but since students usually have limited capacity, this assumption often becomes a bottleneck. In our work, we propose a simple yet effective approach to map the student and teacher into different spaces for distillation. Surprisingly, it only takes a few extra lines of code before the softmax to bring consistent performance gains — and the method is applicable across various distillation settings. It’s always fun to see how small changes can make a big difference. Looking forward to feedback and discussions from the community!

119

LLMs are great at answering questions. But can they actually reason through multi-step tasks, adapt to feedback, and coordinate with other agents? A new survey from researchers at UIUC, Meta, Amazon, and Google DeepMind tackles exactly this. "Agentic Reasoning for Large Language Models" proposes a structured framework that moves beyond chain-of-thought prompting toward treating LLMs as autonomous agents. The key insight: the authors organize agentic reasoning into three layers: 1. Foundational Agentic Reasoning — core single-agent capabilities like planning, tool use, and search in stable environments. Think of this as the basic "operating system" an agent needs. 2. Self-Evolving Agentic Reasoning — how agents refine their capabilities through feedback, memory, and adaptation. This is where things get interesting: agents that actually learn from experience rather than starting fresh each time. 3. Collective Multi-Agent Reasoning — extending intelligence to collaborative settings with coordination, knowledge sharing, and shared goals across multiple agents. What makes this framework especially useful is how it distinguishes between two optimization modes: in-context reasoning (scaling test-time interaction through structured orchestration) and post-training reasoning (optimizing via reinforcement learning and supervised fine-tuning). The survey also covers real-world applications across science, robotics, healthcare, autonomous research, and mathematics — showing where these ideas are already being applied. If you're building or evaluating AI agent systems, this is a solid reference to bookmark. 📄 Paper: https://lnkd.in/gd4e--FW 💻 GitHub: https://lnkd.in/gB8z_s67 #AgenticAI #LLMs #AIResearch #MachineLearning #DeepLearning

26

Today's AI/ML News 🤖💻 Researchers are advancing AI capabilities with new models and tools! Microsoft and Tsinghua University introduced Reward Reasoning Models, enhancing Large Language Models' ability to judge with reasoning, while the Synthetic Data Vault offers a step-by-step guide to creating synthetic data. LLM Reasoning Boosted Can LLMs Really Judge https://lnkd.in/gXu2rPbP 🤖 Researchers from Microsoft and Tsinghua University introduced Reward Reasoning Models, enhancing Large Language Models' ability to judge with reasoning. Synthetic Data Made Easy Step-by-Step Guide https://lnkd.in/gcarYFbS 🚀 The Synthetic Data Vault offers a step-by-step guide to generating realistic tabular data using machine learning. #AI #ML #LargeLanguageModels #SyntheticData #Innovation 🤖💻

6