Things I Am Interested About

Efficient AI Systems

• Mixed Precision Training & Inference
  • Automatic mixed precision strategies for transformer models
  • FP8 training dynamics and numerical stability
  • Hardware-aware precision assignment
• TensorRT Optimization
  • Graph optimization patterns for production inference
  • Dynamic shape handling and batch optimization
  • Custom plugin development for novel operators
• Small Language Models (SLMs)
  • Sub-1B parameter models that match larger model capabilities
  • Efficient attention mechanisms (Flash, Sliding Window)
  • Knowledge distillation from large models

Reading Sources

1. Mixed Precision Training in Large Language Models (Paper)
2. NVIDIA TensorRT Documentation & Best Practices
3. Flash Attention: Fast and Memory-Efficient Exact Attention
4. Phi-3: Small Language Models with Big Performance
5. Knowledge Distillation: A Survey (2023)

Vision-Language Models & Multimodal AI

• Vision-Language-Action (VLA) Models
  • End-to-end learning for robotic manipulation (RT-2, RT-X)
  • Bridging perception and control with language instructions
  • Real-time VLA inference on edge devices
  • Open-vocabulary manipulation and task generalization
  • Multimodal transformers for robot control
• Efficient VLM Architectures
  • Cross-modal attention optimization
  • Unified embedding spaces for vision and language
  • Training VLMs from scratch with limited compute
• Visual Reasoning & Grounding
  • Compositional understanding in visual scenes
  • Spatial reasoning without explicit 3D representations
  • Zero-shot visual question answering

Reading Sources

1. V-JEPA: Video Joint Embedding Predictive Architecture (Meta AI)
2. COSMOS: World Foundation Models from NVIDIA
3. World Models (Ha & Schmidhuber, 2018)
4. DreamerV3: Mastering Diverse Domains through World Models
5. Video Diffusion Models: A Survey

Reasoning Models in LLMs

• Chain-of-Thought & Reasoning Architectures
  • OpenAI o1/o3 models and their training methodologies
  • DeepSeek-R1 architecture and distillation techniques
  • Self-supervised reasoning without human annotations
• Reinforcement Learning for Reasoning
  • GRPO (Group Relative Policy Optimization) vs PPO
  • Process reward models vs outcome reward models
  • Verifiable reasoning with math and code benchmarks
• Key Resources & Papers
  • Sebastian Raschka: The State of RL for LLM Reasoning
  • DeepSeek-R1 technical report and open models
  • Marco-o1 and QwQ reasoning model papers

Reading Sources

1. Sebastian Raschka: The State of RL for LLM Reasoning
2. DeepSeek-R1: Scaling Test-Time Compute with RL
3. OpenAI o1 System Card
4. Marco-o1: Towards Open Reasoning Models
5. QwQ: Reflecting on Reflections

Post-Training Optimization

• Reinforcement Learning from Human Feedback (RLHF)
  • Efficient reward modeling for specialized domains
  • PPO alternatives for stable training
  • Multi-objective RLHF for balanced optimization
• Direct Preference Optimization (DPO)
  • Simplified alignment without reward models
  • Iterative DPO for continuous improvement
  • Combining DPO with constitutional AI
• Parameter-Efficient Fine-Tuning
  • LoRA/QLoRA for domain adaptation
  • Task-specific adapters for multi-task learning
  • Quantization-aware fine-tuning strategies

Reading Sources

1. RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control
2. CLIP: Learning Transferable Visual Models From Natural Language Supervision
3. Flamingo: A Visual Language Model for Few-Shot Learning
4. Open X-Embodiment: Robotic Learning Datasets and RT-X Models
5. LLaVA: Large Language and Vision Assistant

World Models & Predictive AI

• Joint-Embedding Predictive Architectures (V-JEPA)
  • Self-supervised learning of world representations
  • Predictive models without generative decoding
  • Efficient video understanding through masking
  • Transfer learning to downstream tasks
• COSMOS: World Foundation Models
  • Physical world simulation from video data
  • Long-horizon video prediction and generation
  • Learning physics priors from observation
  • Applications in robotics and autonomous systems
• Diffusion Models for World Modeling
  • Video diffusion for planning and control
  • Learned simulators for robotics
  • Compositional scene generation

Reading Sources

1. RLHF: Training Language Models from Human Feedback
2. DPO: Direct Preference Optimization
3. Constitutional AI: Harmlessness from AI Feedback
4. LoRA: Low-Rank Adaptation of Large Language Models
5. QLoRA: Efficient Finetuning of Quantized LLMs

Robotics & Embodied AI

• Real-time Perception Systems
  • Unified perception stacks for navigation and manipulation
  • Sensor fusion with learned priors
  • Uncertainty-aware perception for safety-critical systems
• Visual SLAM & Odometry
  • Learning-based loop closure detection
  • Dense reconstruction from sparse observations
  • Semantic SLAM for scene understanding
• Sim-to-Real Transfer
  • Domain randomization strategies
  • Physics-aware simulation for contact-rich tasks
  • Meta-learning for rapid real-world adaptation

Reading Sources

1. ORB-SLAM3: An Accurate Open-Source Library for Visual, Visual-Inertial and Multi-Map SLAM
2. Learning to See by Moving (Berkeley)
3. Domain Randomization for Transferring Deep Neural Networks
4. Neural SLAM: Learning to Explore with External Memory
5. Sim-to-Real Transfer in Robotics: A Survey

Systems & Infrastructure

• GPU Optimization Patterns
  • Kernel fusion strategies for transformer models
  • Memory hierarchy optimization for large models
  • Multi-GPU communication patterns
• ML Systems Design
  • SDK architecture for production ML pipelines
  • Versioning and reproducibility in ML systems
  • Observability and debugging tools for deep learning
• Edge Deployment
  • Model optimization for mobile and embedded devices
  • Heterogeneous computing (CPU/GPU/NPU)
  • Power-aware inference scheduling

Reading Sources

1. FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning
2. Efficient Large-Scale Language Model Training on GPU Clusters (Megatron)
3. CUDA Programming Guide - Kernel Fusion Techniques
4. TinyML: Machine Learning with TensorFlow Lite
5. MLOps: Continuous Delivery and Automation Pipelines in ML

LLMs & Agents in Finance

• Financial Analysis Agents
  • Automated financial modeling and report generation
  • Natural language to spreadsheet formula translation
  • Multi-step reasoning for complex financial calculations
  • Real-time market data integration and analysis
• MCP Servers for Financial Data
  • Building Model Context Protocol servers for financial datasets
  • Unified interfaces for Bloomberg, Reuters, and alternative data sources
  • Semantic search across financial documents and filings
  • Time-series data handling and backtesting infrastructure
• Autonomous Financial Workflows
  • End-to-end report generation from raw data to insights
  • Agent orchestration for multi-source financial research
  • Excel/Google Sheets automation with LLM reasoning
  • Building tools similar to Fundamental Labs for institutional finance
• Challenges & Opportunities
  • Ensuring accuracy in financial calculations and compliance
  • Real-time streaming data processing with LLM agents
  • Building trust and explainability for financial decision-making
  • Integration with existing financial infrastructure (APIs, databases)

Reading Sources

1. Anemoi: A Semi-Centralized Multi-agent System Based on Agent-to-Agent Communication MCP server from Coral Protocol
2. BloombergGPT: A Large Language Model for Finance
3. FinGPT: Open-Source Financial Large Language Models
4. Model Context Protocol (MCP) Documentation - Anthropic
5. Building LLM Applications for Production (Chip Huyen)
6. Fundamental Labs: AI-Powered Investment Research

Emerging & Exploratory Areas

• AI for Longevity Research
  • ML models for biological age prediction
  • Drug discovery and repurposing with graph neural networks
  • Multi-omics data integration
• Open-Ended Learning Systems
  • Curiosity-driven exploration in RL
  • Emergent behaviors in multi-agent systems
  • Self-supervised curriculum learning

Reading Sources

1. The Hallmarks of Aging (López-Otín et al., 2023 Update)
2. AlphaFold: Highly Accurate Protein Structure Prediction
3. Curiosity-driven Exploration by Self-supervised Prediction
4. Open-Ended Learning Leads to Generally Capable Agents (DeepMind)
5. Graph Neural Networks for Drug Discovery: A Survey

📚 Live Reading List

This is my actively updated reading list. Papers are tracked with status (To Read, Reading, Completed) and include descriptions and key concepts.

Note: If the sheet doesn't appear, you may need to view it directly on Google Sheets.