Opinions, Not Press Releases
Browse a curated front page, then move through focused topic and cluster hubs without losing older work to chronology.
Search every article
Use site search to search the full archive, not only this page.
No posts match your search. Try another term or pick a topic.
Latest articles
Nine articles per page, ordered by publication date.
AI & AgentsMiso TTS Self-Hosted vs API: Cost, Latency and VRAM Reality
Miso TTS is expressive and open-weight, but the 110 ms claim belongs to its hosted H100 API, not local inference. Compare hardware, pricing, licensing and production risk before choosing.
AI & AgentsEnterprise MCP Authorization Architecture
OAuth 2.1, audience-bound tokens and no token passthrough are the spec's hard rules. The rest, multi-tenant isolation, per-tool scopes, delegated access and audit, is a reference design. Here is the whole trust boundary in two diagrams.
Business & RegulationEU AI Act Article 50 Checklist for SaaS and AI Agents
Article 50 has applied since 2 August 2026. Audit chatbot and agent disclosures, machine-readable marking, deepfake labels, public-interest text, logging, and the evidence kept for enforcement.
AI & AgentsCompress RAG Vectors to 2 Bits: Is Data-Oblivious Quantization Production-Ready?
TurboQuant now ships in Qdrant 1.18, while TurboVec has reached 1.0. Learn what the 16x figure applies to, how current benchmarks compare, and what to test.
AI & AgentsLMCache Reported 13.7x Lower Mean TTFT with a Shared KV Cache
LMCache reported mean time-to-first-token falling from 3.98s to 0.29s in one multi-turn Qwen3-235B benchmark. Here is what the result shows, what it does not prove, and how to test reuse on your own workload.
AI & AgentsLossless LLM Weight Compression vs 8-bit GGUF: What Is Ready for Production?
A fact-checked decision guide to the GLM-5.2 result, exact BF16 decoding, Q8 quantization, prior systems, runtime evidence, fleet economics and a 14-day pilot.
Product & MVPPaid Pilot vs PoC vs Design Partner: Which One Proves Demand?
Decision guide for choosing the right evidence test: technical feasibility, customer learning or willingness to pay, with a seven-condition paid-pilot scorecard and contract brief.
AI & AgentsOpenAI Eval Sandbox Escape: 12 Controls Before You Test a Cyber Agent
A fact-checked incident analysis and procurement checklist for nested isolation, egress, credentials, control-plane separation, benchmark integrity, monitoring and forensic fallback.
Delivery & QAGit Worktrees vs Jujutsu for AI Coding Agents: 2026 Decision Guide
Buyer guide to repository caching, partial clone, Git worktrees and Jujutsu workspaces, with compatibility gaps, cost metrics and a measurable 14-day pilot.
Follow the work that matters to you
Get a short email when we publish something new. Follow the whole blog or only the problems you care about.