ModelsDense vs Mixture-of-Experts: Which You Can Actually Run
MoE models are cheap to compute and expensive to hold. Dense models are the reverse. The architecture decides your deployment more than your benchmark scores.
ReadNotes on models, agents, and what inference costs
Practical writing for developers building with large language models — how they work, how to pick one, and how to keep the bill predictable.
Models
AI AgentsDependency-Aware Planning: Ordering Work That Blocks
Most agent plans are flat lists executed top to bottom. Modelling dependencies instead unlocks parallelism, better retries and honest progress reporting.
Read
Cost & PricingDesigning an AI Spend Dashboard People Actually Read
Most LLM cost dashboards show token counts nobody acts on. The four or five panels worth keeping, why unit cost beats totals, and how to avoid vanity metrics.
Read
AI AgentsDetecting Agent Loops Before They Burn Your Budget
A stuck agent repeats the same failing action until something stops it. How to detect repetition cheaply and what to do once you have.
Read
Cost & PricingDoes Streaming Cost More? What You Actually Pay For
Streaming changes when tokens arrive, not how many are billed. Where it does affect spend, and the abandonment case that genuinely wastes money.
Read
FundamentalsDPO vs RLHF: Two Routes From Preferences to Model Behaviour
Direct preference optimisation drops the reward model and the RL loop. What that simplification buys, what it costs, and why open-weight work leans on it.
Read
GuidesEmacs LLM Setup: A Model Endpoint Without the Mouse
Wire an OpenAI-compatible endpoint into Emacs — package choices, auth-source instead of hardcoded keys, buffer versus region workflows, streaming and keybindings.
ReadCost & PricingEscalation Rate: The Metric That Decides Your Default Model
The share of tasks your cheap model cannot finish decides whether a two-tier setup saves money or quietly doubles it. How to measure and act on it.
ReadModelsEvaluating Models on Your Own Codebase
Public benchmarks rank a sample that is not your repository. How to choose tasks, source ground truth and size a set that actually predicts your work.
Read
GuidesEvaluating Prompt Changes Without Fooling Yourself
You edited a prompt and the output looks better. Here is how to find out whether it actually is, with paired runs, enough samples and judges you can trust.
Read
ModelsEvery 1M-Context Model Compared: Which Window Is Real
Five open-weight models now advertise 1M tokens. Advertised context and usable context are different things. How to tell which window actually holds up.
Read
FundamentalsExpert Routing in MoE: How a Token Picks Its Path
The router decides which experts see each token, and that one small network shapes quality, throughput and why identical prompts can behave differently.
ReadShowing 433–444 of 731 articles