IDEAS, DECISIONS & LESSONS
Notes on AI, architecture and building useful products.
94 articles
28 Aug 2026 · 5 min read
ai
AWS and Unsloth just published four production deployment patterns for quantized LLMs that cut instance costs roughly 5x. I break down which pattern fits which team, with the actual configs, and the operational traps that don't show up in the benchmark numbers.
Read more26 Aug 2026 · 4 min read
ai-agents
Ollama v0.33.0 shipped a UI-driven gateway that lets Claude Desktop run Qwen, DeepSeek, and Kimi models locally — no MCP server, no Anthropic feature, entirely Ollama's own proxy. Here's how it works, the tradeoffs in tokens/sec and tool-calling reliability, and when it's worth flipping on for a team.
Read more25 Aug 2026 · 5 min read
ai
GLM-5.3 just jumped from 4.6 to 28.3 on Terminal-Bench 3.0 at a fraction of frontier pricing. Here's the routing architecture we actually shipped to stop overpaying for tasks that don't need it.
Read more25 Aug 2026 · 4 min read
agent-training
Microsoft's new RL framework decouples agent execution from model training entirely, so you can add reinforcement learning to a LangChain or AutoGen agent with close to zero code changes. Here's the architecture, the numbers, and where it actually fits.
Read more24 Aug 2026 · 4 min read
ai-agents
A researcher chained Kimi K3, GLM-5.2, and GLM-5.3 to build a working root exploit for an unpatched Fire OS kernel vulnerability. It's a genuinely useful case study in what long-horizon, multi-model agent security research looks like right now — and what it means for anyone running EOL hardware or unattended agents.
Read more24 Aug 2026 · 5 min read
agent architecture
NVIDIA's AVO harness took Claude Opus 5 from ~30% to a perfect 100.00 RHAE on ARC-AGI-3 — without touching the model weights. A technical lead's read on why the harness, not the model, is where the next round of agent capability is coming from.
Read more20 Aug 2026 · 5 min read
ai-agents
Multi-agent AI systems waste tokens in two specific, fixable ways. A technical breakdown of phase-scheduled activation, context pruning thresholds, and the patterns I actually use to keep a fleet of subagents cheap.
Read more18 Aug 2026 · 5 min read
agentic coding
Z.ai shipped GLM-5.3 without retraining the base model and still doubled coding and cybersecurity benchmark scores. A technical lead's read on why post-training gains matter more than model size, and what it changes about build-vs-buy on agent models.
Read more17 Aug 2026 · 5 min read
ai
Stripe's reported $7B+ acquisition of OpenRouter — up from a $1.3B valuation three months ago — validates the model-gateway architecture. A Technical Lead's take on when to buy a gateway, when to build one, and what this deal signals for API-routing infrastructure.
Read more
