Library | Lilith AI

Guides

Practical writing about working with AI — from agents and context to RAG and safety.

Agents — when an LLM gets hands and memory An LLM with tool use, a loop, and memory. Lots of marketing, few definitions. Here is the plain version — with verification in the loop, not as an afterthought.
Claude J-space: silent thoughts inside the model Anthropic shows J-space, a layer of internal word-like representations in Claude. It is not proof of consciousness, but it matters for interpretability and safety.
Coding agents — when the model touches the repo Claude Code, Codex and friends are not magical juniors. They are a closed loop: read code, edit, verify, repair. Build verification as infrastructure, or you just mint technical debt faster.
Context window — how much hell fits in a prompt A context window is how many tokens a model can see at once. A bigger window is not memory, truth or a guarantee of better answers. It is a larger, pricier workbench.
Golden Dataset — ground truth for an AI system, not a golden cage A Golden Dataset is a small, carefully reviewed set of real cases used to tell whether an AI system actually works. Using a course recommender as a worked example: it helps evaluate recommendations against real content, not marketing blurbs.
Koog and Kotlin AI agents — what it is and what it is for Koog is JetBrains’ framework for building AI agents in Kotlin and Java. It focuses on practical architecture: strategies, tools, memory, tracing, long context and JVM production integration.
Prompt injection — hostile instructions in your context Prompt injection is not a party-trick jailbreak. It is a boundary problem: the model reads untrusted text and may confuse it for instructions. With agents, it burns twice as hot.
RAG — Retrieval-Augmented Generation When the model does not have your data in its head, it has to fetch it. RAG is a pattern, not a product: build retrieval first, generation second.

Concepts

Short explanations of important terms and what they mean in practice.

Agent infrastructure — the boring layer agents need to work An agent is not just a model with a task. In production it needs identity, permissions, inboxes, tools, memory, audit, telemetry and clear boundaries. Without infrastructure, autonomy is just a pretty demo with risk attached. Agent safety and sandboxing An agent with tools is a tiny machine for consequences. Sandboxes, approvals, least privilege and audit logs are not enterprise decoration; they are brakes before the fire. AI-assisted research — the model as a research partner AI-assisted research uses models to find hypotheses, write code, test variants and read literature. It is not automatic science. It is a faster research loop with new ways to fall on your face. Async agents — work that does not live in chat An agent that takes a task, runs outside the conversation, and returns a finished artifact. Powerful for long workflows, dangerous without state, limits and review. Computer-use agents — the model that clicks A computer-use agent sees the screen and controls the UI. It sounds like sci-fi; in practice it is fragile automation over pixels, forms and badly labelled buttons. Evals and benchmarks — measurement instead of vibes A benchmark is not truth carved in stone. It is an instrument with error bars. Without it, though, you are only guessing whether a model or agent works. Fine-tuning — a scalpel, not a universal hammer Fine-tuning changes model weights. It is powerful when you have data, evals and a clear reason. It is an expensive mistake when it hides a bad prompt, missing RAG or an unclear process. Frontier model governance — who checks the model before release Frontier model governance asks who tests the strongest models before deployment, under which rules and with what power to intervene. A voluntary audit, a system card and government testing are not the same thing. Model economics — the operating cost of intelligence Tokens, latency, throughput, quality, and risk on one bill. A model is not just smart or dumb; it is expensive, slow, cheap, local, or operationally bearable. Model reliability — when a pretty answer is not enough Reliability is about when the model knows, when it does not, when it invents, and how often its output can be trusted in production. Elegant wording is not evidence. Open vs. closed models — who pays the frontier premium An open model is not automatically freedom, and a closed model is not automatically lock-in. The practical question is when control, cost and local deployment matter more than paying for frontier capability. Physical AI — when an agent reaches into the world Physical AI connects models, robots, simulation and actions in the real environment. It is not about a cute robot demo, but about who carries the risk when a model starts moving things. Tool use — when a model calls tools Tool use is the moment an LLM stops merely answering and starts calling APIs, running commands, reading files or touching databases. Useful, sharp and dangerous. Zombie internet — when AI text eats the web Zombie internet is a web flooded with generated text, summaries without accountability and content that only looks human from far away. The problem is not just spam. The problem is loss of trust.