The AI apps and dev tools you actually build on.
A hobbyist's AI agent ran up $6,531 in AWS charges in ~24 hours (later reduced to $1,894), repeatedly deploying identical CloudFormation stacks for a network-scanning plan the DN42 community had already rejected.
From Monday 15 June, Agent SDK and headless Claude Code (claude -p) usage leaves subscription limits and moves to a separate monthly credit pool metered at standard API rates: $20 (Pro), $100 (Max 5x), $200 (Max 20x), per-user, no rollover.
Claude Code now defaults to Opus 4.8, with /effort xhigh for the hardest tasks (v2.1.154, 28 May).
Apple's rebuilt Siri runs on a custom ~1.2-trillion-parameter Google Gemini model, routed across three tiers (on-device, Private Cloud Compute, Google Cloud on Nvidia B200s), with queries anonymised so neither Apple nor Google can tie them to a user.
AI coding tools fall into three tiers: inline autocomplete, chat-based assistants, and autonomous agents — each with a different trust model.
Most RAG quality problems are retrieval problems, not generation problems — if the right chunk isn't fetched, the best model can't save you.
An agent is just an LLM in a loop with tools — the power and the fragility both come from that loop.
A vector database stores embeddings and finds nearest neighbours fast — that's the whole job.
The clever 'magic phrase' era of prompting is over — modern models don't need incantations.
Most of the AI stack is now commodity — building it yourself usually wastes effort that should go into your actual differentiation.
AI cost scales with tokens, not requests — a feature can be cheap per call and ruinous at volume.
