Map
The things I'm trying to understand. Not a curriculum. Not a checklist. A map of questions, concepts, connections and rabbit holes.
Computing
+Operating Systems
- How is memory actually managed across NUMA nodes?
- What is the true cost of a context switch in modern Linux?
- How do user-space network drivers bypass the kernel?
+Networking
- Why do we still use TCP when QUIC is widely available?
- How does eBPF change packet filtering performance?
- What happens exactly between a socket write and the physical NIC?
+Databases
- How do modern distributed SQL databases handle clock skew (e.g. Spanner vs Cockroach)?
- What are the trade-offs of LSM trees vs B-Trees for NVMe SSDs?
- How is MVCC implemented without severe garbage collection overhead?
+Distributed Systems
- Is Paxos still relevant, or has Raft won completely?
- How do you test for Byzantine faults effectively?
- What are the limits of CRDTs in collaborative text editing?
+Programming Languages
- How does Rust's borrow checker impact compilation time at scale?
- Is Go's garbage collector actually sufficient for soft real-time?
- How do JIT compilers warm up in serverless environments?
+Computer Architecture
honest "no questions yet"
AI Systems
+Machine Learning
- How are gradients actually synchronized across 1000s of GPUs?
- Why is Adam still the default optimizer after all these years?
+LLMs
- What is the mathematical intuition behind RoPE (Rotary Position Embeddings)?
- How does MoE routing avoid degenerate collapse during training?
- What are the exact FLOP requirements for generating a single token?
+Inference
- How does vLLM's PagedAttention manage KV cache fragmentation?
- What are the latency vs throughput trade-offs of continuous batching?
- How do speculative decoding draft models maintain acceptance rates?
+Retrieval
- Why use HNSW over IVFPQ for billion-scale vector search?
- How do you evaluate retrieval independently from the LLM?
- Is late interaction (ColBERT) always better than dense embeddings?
+Agents
- How do we prevent error compounding in long-horizon tasks?
- What is the best prompt structure for tool use reasoning?
- How can agents reliably pause and wait for user input?
+Evaluation
- Are LLM-as-a-judge scores actually correlated with human preference?
- How do we prevent contamination of evaluation benchmarks?
Engineering
+Architecture
- When does microservices complexity exceed monolithic scaling limits?
- How do you orchestrate sagas across heterogeneous systems?
+Reliability
- How do you prove a disaster recovery plan works without causing one?
- What are the most common cascading failure triggers?
+Security
- How does confidential computing actually protect data in use?
- What is the practical surface area of side-channel attacks today?
+Observability
- How do you sample distributed traces without losing rare edge cases?
- What is the baseline overhead of OpenTelemetry in Node.js?
+Performance
honest "no questions yet"
+Infrastructure
- Is Kubernetes too complex for standard web workloads?
- How does Terraform state locking actually work under the hood?