Scaling Monosemanticity: Mechanistic Interpretability via Sparse Autoencoders
A foundational source audit of Anthropic’s landmark research on Sparse Autoencoders, dictionary learning, and causal feature steering in frontier transformer models.
Fodev JEO
LATEST INVESTIGATIONS
Source audits & field notes
Hybrid MLOps with MLflow: Bridging On-Premises Training and Cloud Serving Endpoints
A source audit evaluating hybrid MLOps architectures, combining on-premises GPU training clusters with AWS MLflow tracking and SageMaker managed inference endpoints.
Mitigating LLM Inference Bottlenecks: Prefill Compute vs Memory-Bound Decode Mechanics
A source audit analyzing LLM inference bottlenecks, contrasting compute-bound prefill and memory-bound decode, continuous batching, and speculative decoding.
Frontier LLM Research Trends: Test-Time Compute, State-Space Models, and MoE
A technical source audit analyzing post-training compute scaling, state-space architectures, mixture-of-agents, and quantization frontiers across modern LLMs.
Multi-Agent Architectures: Orchestration, Environment Gyms, and Trajectory Evolution
A source audit analyzing multi-agent systems, comparing CrewAI role orchestration with AgentGym environments, trajectory cloning, and SLA circuit breakers.
Chameleon Architecture: Mixed-Modal Early-Fusion and Discrete Token Generation
A source audit analyzing Meta’s Chameleon foundation model, evaluating discrete image tokenization, QK-normalization, and native mixed-modal pretraining.
Uni-MoE Architecture: Unified Multimodal MoE and Three-Stage Modality Alignment
A source audit analyzing Uni-MoE’s unified multimodal architecture, sparse mixture-of-experts routing, and three-stage cross-modality alignment pipelines.
Mistral-Finetune Architecture: Memory-Optimized LoRA for Mixtral and Multi-GPU Stacks
A source audit analyzing Mistral-finetune, evaluating PyTorch FSDP integration, LoRA target projections, instruction formatting, and multi-GPU memory scaling.
Technical Interviews in the AI Era: Beyond LeetCode to Systems and Agentic Harnesses
A source audit evaluating software technical interviews, contrasting algorithmic LeetCode drills with system design, deep codebase navigation, and AI harness testing.
Self-Evolving Language Models: The Mechanics of Autonomous Post-Training
A source audit analyzing autonomous self-evolution in LLMs, evaluating self-instruct, iterative DPO, self-rewarding verification, and model collapse boundaries.
인기 태그
Search article titles, categories, and tags. Full text is searched when needed.
Type to search articles.