Classy Llama

AI Metrics Decoded: Parameters to TOPS

A technical guide explaining the core metrics engineers should understand when deploying AI models in production. The article defines seven metric categories — model size (parameters, tokens), hardware power (FLOPS vs TOPS), training compute (FLOPs), and runtime performance (TTFT, TPS, TPM) — and gives practical rules of thumb (e.g., 1B parameters ≈ 2 GB VRAM in fp16), hardware and device examples, latency and throughput targets, and cost trade-offs between model sizes (example 8B vs 70B). It emphasizes benchmarking smaller models first, tracking token usage/rate limits, and measuring TTFT/TPS in production to avoid cost and performance surprises.

GnokeOps: Host Your Own AI House Party

GnokeOps is a developer-oriented blueprint and philosophy for self‑hosting AI functionality inside a user's own infrastructure rather than relying on vendor-controlled cloud IDEs or hosted agent platforms. It advocates an "API‑agnostic host" approach where models are interchangeable workers, enforces a local security boundary called the "Bouncer" to constrain model permissions, and treats models as temporary, stateless visitors that operate on locally-authorized files and databases. The blueprint outlines a lightweight stack (PHP, SQLite, local filesystem, model adapters), server-side triggers to wake models only when needed, and options to run validating or enforcement models locally (e.g., Llama or Qwen via Ollama). The article was published on Dev.to on 2026-05-23.

File-Based Memory (.klickd) for AI Agents

The article argues that AI agents' apparent "memory" problem is an architecture problem and proposes a file-based alternative to server-side memory services. The author cites a 2026 study finding ~21.8% of input tokens are wasted re-establishing session context and critiques centralized memory stores (e.g., Mem0, Zep) for expanding provider-side attack surfaces. As a proof of concept the author presents .klickd: a portable, encrypted memory-file format (example schema 'klickd/v1') that is client-owned, provider-agnostic and zero-server. Implementation details include AES-256-GCM encryption and Argon2id key derivation. Benchmarks reported by the author (Zenodo DOI) show an average improvement of +13.9 points versus baseline on a personalization benchmark. The article lays out trade-offs (loss of centralized governance/analytics vs. stronger client-side privacy) and publishes the open spec on GitHub.