In scope: the exploitation surface, prompt injection and goal reframing, permission escalation, memory and context poisoning, tool and supply-chain provenance, deployed-agent safety evidence, red-teaming method. Out: model-level alignment research as a subject in itself (models), harness architecture (harnesses).
This rung has no compiled overview yet. Until it does, the base is best read across the stack rather than down one rung.
Start here
- 01LLM Inference Prices Have Fallen Rapidly but Unequally Across Tasks
Start at the number the whole base chases: what a token costs, and how unevenly that price has actually fallen.
inference-economics - 02Memory-Centric Computing: A Paradigm Shift for Sustainable and Efficient Systems
The reason the price falls the way it does — the bottleneck is moving data, not doing arithmetic on it.
hardware - 03AI Hyperscaler Capex 2026: Why Microsoft, Google, Meta and Amazon Are All Spending at Once
Where the money physically lands once the bottleneck is priced: buildings, land and power, committed years ahead.
datacenters - 04Speculative Decoding Meets Quantization: Compatibility Evaluation and Hierarchical Framework Design
How the cost is actually cut in software, and what the two cheapest tricks do to each other when you stack them.
serving - 05Agentic Reasoning for Large Language Models
The demand side: the systems wrapped around the model are what decide how many tokens the price applies to.
harnesses