Narrow vs General, Open vs Closed, Local vs Cloud, and Agents

A handful of word pairs show up in nearly every AI news item, press release and sales pitch. They sound technical. Each one hides a practical question you can answer: what can it do, who controls it, where does your data go, and what is it allowed to touch.

Narrow AI and general AI

Narrow AI does one job or a handful. A plate reader reads plates. A spam filter sorts mail. A protein-folding model predicts protein shapes and couldn't tell you the weather. Every AI system you have ever used is narrow in the strict sense, though chatbots stretch the word: one model that writes, codes, translates and explains is much broader than anything before 2020.

Artificial general intelligence, or AGI, is the idea of a system as capable as people across most kinds of work. Nobody agrees on a test. OpenAI's charter defines it as "highly autonomous systems that outperform humans at most economically valuable work." Others define it by learning new skills as fast as a person does, or by doing scientific research on its own. Leaders at several large AI companies have said publicly they expect something like it within years. Many academic researchers think current methods will fall short without new ideas, and some think the term is too vague to be useful.

What is well established: today's systems are uneven. A model can pass hard professional exams and then fail at a simple task a child would get right, or invent facts with confidence. When you read an AGI claim, ask who is making it, how they define it, and what they would gain if people believed it.

Open, open-weight and closed

This one matters more than it sounds, because it decides who can inspect, change and run a model.

The words get blurred on purpose. Meta long called Llama "open source," but its license sets conditions on use, and Meta does not release the training data. The Open Source Initiative published a formal Open Source AI Definition in October 2024 requiring the parameters, the full training and running code, and enough information about training data for a skilled person to build a similar system. By that standard, most "open" models are open weights only. OpenAI's gpt-oss, released in August 2025 under the Apache 2.0 license, is an open-weight model; its training data is not public.

Companies also change direction. Meta's first model from its new superintelligence group, Muse Spark, launched in April 2026 as a closed model, with Meta saying it hoped to release open versions of future models. Alibaba previewed a very large Qwen3.8 model in July 2026, then posted open weights for a text-only version on Hugging Face in August while keeping its fuller Qwen3.8-Max in its own paid services.

The safety argument cuts both ways, and both sides have a point. Open weights let independent researchers test models, let small organizations avoid lock-in, and keep sensitive data in-house. They also mean anyone can remove a model's refusal training and nobody can recall a release. The next module covers that debate.

Local and cloud

A cloud model runs on the company's servers. You send your text over the internet, and the answer comes back. A local model runs on your own computer or phone. Nothing leaves the machine.

Whether a model fits on your computer is mostly arithmetic. Each parameter takes memory.

20,000,000,000 x 2 bytes = 40,000,000,000 bytes, about 40 GB

20,000,000,000 x 0.5 bytes = 10,000,000,000 bytes, about 10 GB

Free apps such as Ollama and LM Studio make running open-weight models on a laptop a matter of a download and a click. For anyone handling sensitive material, such as an advocacy group sorting member stories or a new official reading constituent mail, a local model for first-pass sorting is worth the setup.

Agents

A chatbot answers. An agent acts. Give a language model a goal, a set of tools, and permission to take several steps in a loop, and you have an agent.

Coding agents are the most mature example: they read a codebase, edit files, run the tests, read the errors and try again. Research agents search dozens of pages and write a cited report. Browser agents click through websites to fill in forms. The Model Context Protocol, released by Anthropic in November 2024 and adopted widely since, is a common way to plug tools and data into these systems.

Agents fail differently from chatbots. Errors compound: a 95 percent chance of getting each step right sounds good, but over ten steps that's about a 60 percent chance the whole job goes right (0.95 multiplied by itself ten times is about 0.60). And because a model reads instructions and data as the same stream of tokens, a web page or email can carry hidden instructions that hijack it. That's prompt injection, and no company has fully solved it. The safe habit is to give agents the least access they need, and to require a human to approve anything that sends, buys, deletes or publishes.