stefanopineda.com
Stefano Pineda
Field notes on building with AI, hardware, and systems that actually run.
Newest writing first. Each piece is a field guide: bottom line up front, then the map — so you can decide in two minutes whether to go deep.
Writing · most recent
Qwen 3.8 27B vs DeepSeek V4 Flash — same ~52 AA band, pick by RAM
They land in the same ~52 Artificial Analysis band. The real split is a 17 GB dense 27B on a 5090 versus a 110 GB MoE Flash on a Studio. Pick the wrong one and you either starve context or leave a GPU idle.
Read the brief →Isolated Nicotine — What Pure Nicotine Does (Not Cigarettes)
Cigarettes are catastrophic. Pure nicotine is not the same claim. RCTs, meta-analyses, and systematic reviews on isolated nicotine — short- and long-term effects, sample sizes, and certainty grades.
Read the brief →Beyond a Forward Pass — Part 1: Continuous Batching
Why static batches leave GPUs idle under chat traffic — and how iteration-level continuous batching (Orca → vLLM/TGI/TRT-LLM) became the foundation of modern LLM serving. Series of 5.
Read the brief →Beyond a Forward Pass — Part 2: Prefill, Decode & Disaggregation
Prefill is compute-bound; decode is HBM-bound. Mixing them carelessly wrecks SLOs — chunked prefill and prefill–decode disaggregation are the systems answers.
Read the brief →Beyond a Forward Pass — Part 3: KV Cache & PagedAttention
The KV cache is the real runtime state of LLM serving. Memory math, fragmentation, and PagedAttention as virtual memory for transformers — why continuous batching needs paging to win.
Read the brief →Beyond a Forward Pass — Part 4: Prefix Caching & Routing
Round-robin is wrong when shared prefixes are the expensive resource. Automatic prefix caching, radix trees, and cache-aware routing for multi-tenant fleets.
Read the brief →Beyond a Forward Pass — Part 5: Sharding & Mixture of Experts
When one replica is not one GPU: tensor/pipeline sharding plus MoE expert routing — and how continuous batching, KV, and prefixes still apply inside the fabric.
Read the brief →Mac Studio Memory Roadmap — 512GB, 768GB, 1.5TB & Local Models
M3 Ultra’s 512GB vanished in the DRAM crunch. M5 Ultra has been tested to 768GB; M6 Pro/Max skipped; M7 Ultra whispers 1.5TB. Bandwidth tables vs RTX PRO 6000 and DGX Station — and which models actually fit.
Read the brief →Where I'm At — A Conversation with Grok Build to Find Direction
Leaving the Navy in October 2027. Nuclear-trained officer, builder, not looking for another job. What Grok Build said after looking at the repos: floor product, ceiling bet, and a 12-month plan under a uniform.
Read the brief →Intelligence Density — ~10× in Seven Months & Elon’s 100×
Elon said we were off by 100× on intelligence density. Seven months later, Laguna S 2.1 and the frontier stack show what the first ~10× looks like — with charts, the model timeline, and projections for the rest.
Read the brief →Agent Harnesses Compared — Open & Closed (July 2026)
Open source is not one switch. OpenClaw / Hermes / OpenCode / Pi plus Claude Code, Codex, and Grok Build — what harnesses provide, open vs closed axes, why people pick each, and live GitHub stars.
Read the brief →Grok 4.5 vs Claude Opus 5 — Cost per Task & Performance
Opus 5 just hit Fable-class intelligence at $5/$25. Grok 4.5 is $2/$6 and ~6× cheaper on AA cost-per-task. If you pick models by list price or tribal loyalty, you will mis-spend — this is the cost-per-finished-task map.
Read the brief →CMP 170HX Canary Unlock — Reliability for Local AI
If you are pricing “unlocked A100” mining cards for local AI, the 80GB headline is the trap. Read this before you wire money: what actually unlocks, which SKU is stable (64 vs 40 vs 80), and what still has almost no long-context benchmarks.
Read the brief →Laguna S 2.1 Day 1 — Where It Stacks Up
Poolside shipped a 118B MoE (8B active) that hits 70.2% Terminal-Bench 2.1. If you care about agentic coding efficiency — not marketing FLOPs — this day-one map vs DeepSeek, Qwen, Kimi, and closed frontier is the comparison you will want bookmarked.
Read the brief →SO-ARM101 Host PC — Build Spec & Ubuntu Runbook
Wrong host PC and you will spend a weekend fighting Ubuntu while the arm sits dead. Specs, compute tiers, BOM, and the post-install runbook for SO-ARM101 + Hugging Face LeRobot — so the host is actually the brain, not a guess.
Read the brief →What Is No Longer Human Work — AI Best Practices
If you still do AI labor yourself, you are the bottleneck. The working phrase, the not-human-work protocol, and how to demand a decision surface — so you stop shipping slop and start keeping only the judgment that matters.
Read the brief →AI Coding Now & Next — Field Guide (July 2026)
SOTA is not one chatbot. Pros run plan → harness → tests → human gates while open weights collapse the gap monthly. Miss this brief and you will keep practicing last year’s workflow while the stack under your feet changes.
Read the brief →Archive · investing tools
Buffett 2.0 / Dhando
Earlier work: AI-assisted value investing, filings analysis, superinvestor tracking, and backtests. Still live — moved under Archive so the home page stays about what I am writing now.