A practical knowledge base for AI protocols, agents, core concepts, frameworks, and tools.
An agent sandbox is an isolated environment in which an AI agent runs code, reads and writes files, and uses the network, keeping what it can reach within controlled boundaries. Several 2026 incidents in which agents got out of their sandboxes showed that inconspicuous channels such as DNS, caches, and config files are boundaries too.
AI watermarking embeds a machine-detectable mark in AI-generated text, images, audio, or video to show that the content was made by AI. Article 50 of the EU AI Act has applied since August 2, 2026, and mainstream models such as Claude began watermarking by default.
Reward hacking is when a model exploits loopholes in the scoring rules to earn a high score without doing what it was actually meant to do. In 2026 it moved from a technical training problem to a real-world risk that inflates benchmark scores and can even push agents past their boundaries.
Sandbagging is when an AI model deliberately performs worse on a capability evaluation than it really can. Safety decisions increasingly rest on evaluation results, and if a model can recognize a test and underperform on purpose, those results lose their footing. In 2026, the GPT-6 Astra system card brought the issue into the open.
Third-party AI evaluation means independent organizations outside the model developer assessing an AI model's capabilities and risks. In September 2026, Anthropic and OpenAI both committed to giving independent evaluators employee-like access, the AEF-1 standard began drawing lab sign-ons, and the debate shifted to who pays and who can publish the conclusions.
Agent memory is the machinery that lets an agent persist information outside the context window and retrieve it on demand. The problem it solves isn't "the window is too small" but "remember the right things across sessions, days, and tasks — and update them when facts change"
AGENTS.md is a Markdown file at a repository's root written specifically for AI coding agents, covering build commands, how to run tests, and project conventions. Multiple vendors standardized it as an open format, and in December 2025 the specification was donated to the Linux Foundation's Agentic AI Foundation
AGI refers to an AI system that matches or exceeds human performance across most economically valuable cognitive tasks. It still has no agreed definition or acceptance test, and different organizations set wildly different bars, so arguments about when AGI arrives are often arguments about definitions
AI observability means recording model calls, tool executions, and agent decisions as queryable trace data so you can answer why it did what it did. Traditional monitoring only asks whether the service is up, while most AI failures return a 200 with a wrong answer
AI video generation uses diffusion models that denoise across space and time together, producing coherent video from a prompt or a reference image. Single-shot image quality is already high, so the real competition has moved to cross-shot character consistency, controllable camera work, and native audio sync
An attention mechanism lets a model decide, while processing one position, which other positions of the input to draw information from and how heavily to weight each. It is the core component of the Transformer and the common origin of context windows, long-input cost, and KV caching
Chain of thought means having a model write out its intermediate reasoning steps before giving an answer. It began as a prompting technique and was later trained directly into reasoning models, becoming the mechanism behind today's "thinking…" interactions
Showing 12 of 84 entries
Every entry grouped by topic — one click straight to the explanation.
