Types of AI — the classification that matters

Types of AI — the classification that matters

A simple question to start

You say "AI" at a table with five people. Each one pictures something different. The engineer thinks of algorithms. Your boss thinks of productivity. Your mom thinks of robots. Your kid thinks of what they saw in a movie.

The problem is all five are right. And at the same time, none of them is talking about exactly the same thing.

Knowing the three classifications below saves you a lot of time. It tells you which tool to grab, for which task, and why.

First: where we are today (narrow vs general)

Every AI you've tried — Claude, ChatGPT, DALL-E, Copilot — is narrow. It does one thing well: writes, draws, analyzes, codes.

But it doesn't drive the car. It doesn't tidy your house. It doesn't replace you in the meeting.

What you see in movies is AGI: a machine that does everything you do. That doesn't exist. It might come, it might not. Not today.

This matters because every tool has clearly drawn limits. Claude didn't make your coffee because it wasn't trained for that. The limitation isn't a defect. It's the design.

Second: generative vs analytical

Generative AI creates. You ask for an email, an image, a video, a new piece of code. It delivers.

Analytical AI finds patterns in data you already have. It shows you trends, alerts, anomalies.

If you mix them up, you'll ask Claude to do a database's job. And you'll get frustrated. Not because it can't. Because it's not the right tool for that task.

For writing, generating ideas or exploring options, go generative. For knowing why sales dropped on Tuesday, go analytical.

Third: where it runs (cloud vs local)

Cloud: Claude connected to the internet, plugged into your tools, with conversation memory. More powerful. Costs money per use.

Local: Claude running on your computer, no internet, your data never leaves the box. More private. Less powerful.

Which one is better? It depends. For day-to-day work with your tools, cloud. To process something confidential you don't want leaving your machine, local. There's no absolute winner.

Three ideas to take with you

First, what you use every day is narrow AI. The movie-style AGI doesn't exist yet. That's not bad news: what you have already reshapes your work.

Second, generative and analytical solve different things. Pick well and the tool does its job. Pick badly and you'll get frustrated with something that was never designed for it.

Third, cloud and local are trade-offs, not a hierarchy. More power and connectivity on one side. More privacy and autonomy on the other. You choose based on the problem.

Types of AI — the classification that matters

A conversation that keeps repeating

You say "AI" in a meeting and everyone believes they're talking about the same thing. They aren't.

An engineer thinks of algorithms. Your boss thinks of productivity. A film director thinks of video generation. Your mom thinks of robots. Each one projects their context.

The taxonomy in academic textbooks is useful for classifying research systems. Here I'll offer you another one. More useful. The one that helps you decide which tool to grab today and why.

Three axes. Three questions. Every tool falls somewhere.

First axis: narrow or general

Every AI that exists today is narrow. Specialist. Excellent at one task, incapable at another.

Claude writes. DALL-E draws. GitHub Copilot codes. None of them does all three the way a human does. And that's not a quality issue "we'll fix next year." It's structural. Every model is trained with a goal, an architecture and data designed for it.

AGI? Artificial General Intelligence. A machine that would do everything you do, without retraining, adapting to new things. Doesn't exist. Some researchers believe it'll arrive in five to ten years. Others think that's pure speculation. Today, April 2026, it doesn't exist.

What matters to you. When you open Claude you aren't using AGI. You're using an excellent system for generating text, reading files and integrating with your tools. That's narrow. And it's fine. It's what you need today.

Second axis: generative or analytical

One generates. Writes, draws, codes from scratch. Creates things that didn't exist.

The other analyzes. Finds patterns in data, identifies anomalies, draws conclusions from what already exists.

The confusion is classic. "Ask Claude to analyze my Excel." Technically it can. The capability exists. But it's not its strength. A database designed for that, with indexes and queries, will be faster, cheaper and more reliable.

Generative is better for creating new versions, writing, exploring possibilities, prototyping fast. Analytical is better for massive volumes, structured data, automated alerts, dashboards.

Want to tell a story? Generative. Want to know why sales are dropping? Analytical.

Using the wrong one is like cutting a tomato with a hammer. The hammer works. For nails. For the tomato, not so much.

Third axis: cloud or local

This is the one fewest people understand.

Cloud: Claude via the web, in connected desktop apps, with everything enabled. Pay per use. Integrates with Gmail, Drive, Calendar and many other tools. Reads the files you upload. Has context from recent conversations. Reaches information in real time.

Local: models like Llama (Meta) or Mixtral (Mistral) running on your computer. No internet. Your data stays put. Marginal cost is zero after installation. Real privacy.

Which is "better"? Depends on your work.

For real-time tasks with your usual tools (reviewing an email, analyzing a presentation, responding in context): cloud.

To process confidential documents you don't want leaving your machine: local.

For fast prototyping where speed matters: cloud.

For a program that runs 24/7 without intervention and without a token budget: local.

It's not cloud > local. It's which one solves your constraint.

Why this classification matters

When someone tells you "AI can't do X," you can ask three things. Narrow or general? Generative or analytical? Cloud or local?

And almost always you'll find it can do X. Just not with that tool or not in that mode.

Claude is narrow, generative, multimodal (text, image and code in the same conversation) and lives in the cloud. Not because it's missing something. Because it's optimized for that combination. And that combination solves almost anything in a modern office workflow.

What you see in movies (omnicompetent, all-powerful, autonomous AGI) doesn't exist. What you have on your desk is something different and better for daily work. A precise, reliable tool, ready to do the job.

A question for you

Think of the three most repetitive tasks of your week. Which ones fall under narrow + generative + cloud? Those are the ones Claude should be doing for you.

And which ones fall under analytical + structured data? For those, the right move is probably to learn some SQL or use a tool like Looker, not to push a language model.

If you want to go deeper into how the generative piece of AI works specifically, I cover that in the piece "How an AI 'thinks'" in this same series. In the meantime, look at your week with these three questions in mind. You probably already know where to rearrange your work.

Types of AI — the classification that matters

Why an operational taxonomy matters more than an academic one

The AI classification taught in research departments is rigorous but not very useful for someone who has to decide what tool to use Tuesday morning. For the working professional, you need a translation.

The one that works is three-dimensional. Capability (narrow or general). Output type (generative or analytical). Deployment architecture (cloud or local). Each point in that space defines a different tool with its own set of trade-offs.

This isn't academic. It's the matrix I run when a client asks me, "which AI tool should we use?" The answer starts with mapping their problem onto these three axes.

Axis 1: narrow vs general — what it means today

Today we're fully in narrow AI. Specialized systems. Each is excellent at one task because it was trained that way, with architecture and data tuned to that specific objective.

"Narrow" doesn't mean "weak." Claude can write a memo better than most humans. But it can't ride a bike. Not for lack of training. For lack of a body, sensorimotor experience, physical context. The limitation is architectural, not temporal.

AGI, as discussed in Stanford's AI Index Report, would be the point at which a system learns transversal concepts, adapts to tasks it has never seen and generalizes in ways close to a human. It doesn't exist. There's speculation about when it might arrive. Some serious researchers suggest between 2030 and 2040. Others are far more conservative. Today, April 2026, it doesn't exist.

The distinction is useful because it teaches you to model real limitations. It's not "Claude can't do X because it isn't good enough yet." It's "Claude doesn't do X because it wasn't trained for it, and that won't change with the next release." The first reading leads you to wait. The second leads you to find the right tool.

Axis 2: generative vs analytical — different structures

The confusion here runs deep because both use modern AI, both process data, both give answers. But the underlying structures are radically different.

Generative predicts sequences. An LLM predicts the next token based on probability distributions learned during training. It's regression over distributions. Excellent for synthesis, exploration, creativity and prototyping. Bad for processing massive volumes with precision, for statistical anomaly detection, or for time series where exactness matters.

Analytical finds structural patterns in data. Uses linear algebra, classical statistics, clustering algorithms, regression models. Deterministic or pseudo-deterministic with confidence intervals. Excellent for decisions where you need traceability, for risk models, for operational dashboards. Bad for generating creative variations, for writing prose, or for open exploration.

There's overlap, of course. An LLM can do analysis through prompt engineering and emergent reasoning. An analytical database can generate synthetic reports. But the cognitive and computational cost is very different. It's inefficient to use the wrong tool just because both "technically can."

A concrete case. Analyzing a 50,000-row dataset via Claude (generative, prompt-based) versus an SQL query on an indexed table (analytical, relational structure). The second is faster, cheaper and more reliable. The first is more flexible and more exploratory.

I apply a simple heuristic. If the problem is "what opportunities are here?" (open-ended), generative. If the problem is "how many units sold by region last quarter?" (closed), analytical.

Axis 3: cloud vs local — architecture, not nostalgia

The cloud-vs-local debate isn't "which is faster." It's "which architecture solves your constraint."

Cloud means connectivity, centralized infrastructure, practically unlimited scale capacity, access to global context (web, external APIs), conversation persistence and integration with a wide ecosystem of tools. The cost: per-token pricing, network latency and data passing through third-party servers.

Local means compute on the user's machine, no network dependency, marginal cost near zero after installation, maximum privacy (nothing leaves the box) and no automatic integration with external services. The limitation: local GPU or CPU is finite, no real-time updates and no persistent multi-conversation memory beyond what you build yourself.

For real-time work (writing, fast analysis, integration with everyday tools), cloud is essentially mandatory. For processing confidential documents (legal, medical, regulated data), local makes a lot of sense, assuming the local machine follows good security practices. For autonomous systems that run continuously, hybrid (local compute with cloud checkpoints) is often best.

There's no absolute winner. There's a winner for your constraint.

The full matrix and where to map known tools

If you combine the three axes, you get this practical map:

Narrow + generative + cloud: Claude, GPT-4o, Gemini. The 2026 mainstream. Good for around 95% of office knowledge work.

Narrow + generative + local: Llama 3 (Meta), Mixtral (Mistral), Phi (Microsoft). Less raw power. More privacy. Useful for confidential tasks or for integrations where per-token cost doesn't pencil out.

Narrow + analytical + cloud: Palantir Foundry, Databricks, BigQuery, Snowflake. Big data. Enterprise-scale machine learning.

Narrow + analytical + local: pandas and DuckDB on your laptop. Quick analysis of medium-sized datasets without going to the cloud.

AGI doesn't appear in the matrix because it doesn't exist. If it appeared, the whole matrix would change.

What most people miss when evaluating tools

When someone evaluates an AI tool, they look at the benchmark. "What's its MMLU score?" That's a weak signal.

What matters is where the tool falls in this three-dimensional space, and whether that point solves your problem. Not the problem of a generic benchmark.

Claude today is excellent at its point in the space. Narrow, generative, multimodal, cloud. If your problem requires AGI, wait. If it requires pure structured analysis, use SQL or a BI tool. If it requires local with strict privacy, options exist. But if your problem is "I need to write 50 personalized emails in two hours, reading each client's history," that falls exactly in the quadrant where Claude dominates.

Short and medium-term trajectory

In the short term, multiplication of specializations. Narrow models increasingly fine-tuned for specific verticals (medicine, code, legal, support). Not a bad sign. It's realism. General-purpose models for everything end up mediocre at every specific thing.

In the medium term, hybrid systems. Generative plus analytical in the same product. Claude is starting this. It can write and analyze files in the same conversation. The barrier between the two paradigms is dissolving at the product level, even if not at the internal architecture level.

In the long term, the AGI question stays open. If it arrives, the taxonomy changes. For now, it doesn't affect what you're doing this week.

The blog's thesis

I've spent years working with applied AI and I've landed on a stance I'll defend. The people who decide best which tool to use aren't the ones who know the most about models or who follow every announcement. They're the ones who learned to ask the three right questions before choosing.

Narrow or general. Generative or analytical. Cloud or local.

People who do this with discipline make far stronger architecture decisions than people who chase the latest announcement. They save their team time, money and frustration.

Next time someone tells you "we should use AI for this," break the conversation along those three axes. You'll see that half the time the right tool isn't the one the other person had in mind. And the other half of the time it is, but for reasons you can now explain.

That's the difference between using AI and understanding AI. And it shows in the results.