KServe, vLLM, Triton, BentoML, four different answers to the same question. Each layer trades operational complexity for serving features. Worth being concrete about which layer is right for which workload, tested on a single GTX 1080 Ti.
The cloud bill for serious AI training is still aimed at the hyperscaler tier. The patterns for doing meaningful training without that bill have matured. Worth being explicit about what's gettable on a budget that doesn't require enterprise procurement.
DevDay week. OpenAI ships GPT-5 Pro in the API, Sora 2 in the API, AgentKit, the Apps SDK, ChatKit, Codex GA, and signs a six-gigawatt AMD compute deal, all on Monday. Then the Sora app starts producing the deepfakes everyone predicted. My read on a heavy week.
Auditors are starting to ask about AI use in SOC 2 cycles. The shops running local LLMs have a different story to tell than the shops running cloud, and the auditors mostly haven't internalized the difference. Worth being explicit about what evidence actually answers the questions.
The atomic unit of an AI workload isn't the model call, isn't the request, isn't the user. It's the conversation. The architecture decisions that follow from that, caching, billing, governance, ops, all get cleaner when you start there.
California signs SB 53, the first US frontier-AI law. Anthropic ships Sonnet 4.5 with an Agent SDK. OpenAI ships Sora 2 with a biometric-scan social app. The AI-layoff narrative consolidates. My take on a heavy news week.
Most home AI setups die after the novelty wears off. The ones that survive into year two share a small set of operational properties, boring, durable, owner-friendly. Worth being explicit about what makes a stack worth keeping rather than just worth building.
Retrieval-augmented generation works well when the corpus fits on one machine. The honest version of what to do when the corpus outgrows that, without rebuilding the whole stack on cloud, is more interesting than either the all-local or all-cloud framings suggest.
Qdrant vs Weaviate vs Milvus on K8s. The foundation question for retrieval. StatefulSets, persistent volumes, replication, the operational reality. RAG indexing patterns at homelab scale on engine-01, and the decisions that change shape at fleet scale.
MCP solved the agent-to-tool plumbing. The next layer up, routing across many MCP servers, scoping access per agent, observing what's happening, is where MetaMCP and a small cluster of similar tools have started showing up. Worth being plain about why the layer exists.
Two years past the encoding-a-person framing. The thesis held in the parts I expected and bent in the parts I didn't. Worth being honest about what survived contact with the actual technology and what was just well-aged speculation.
Helm values.yaml is where Decisions as Code lives in Kubernetes. Centralized business decisions, schema-validated, composed via library charts, projected into every workload. The approach is the contribution. The tool changed.
