Applying Intelligence

A pragmatic guide for the interested.

Free. Not a product. Just an attempt to explain the thing clearly enough that you can make your own decisions.

You do not need to agree with everything here. You do need to understand the forces moving around you.

Most AI writing is either hype or panic. Both are bad ways to think. Hype makes you stupid. Panic makes you passive. The useful middle is plain understanding.

This guide is written for people who are interested, busy, and allergic to jargon. If a thing matters, I will explain it. If it does not, I will skip it.

The ability to perceive, understand, reason, then take action is the best compression I know for intelligence.

Perceive means take in the world. Understand means turn noise into something useful. Reason means compare paths. Act means do something with the answer.

That sounds simple because it is simple. The hard part is that humans can do this across almost anything. We can handle a new room, a new person, a new problem, a new tool. We generalize. That is the trick.

Annoyingly, people use AI to mean three different things.

First, classic machine learning. Small models that predict or label something specific. These run quietly inside businesses and have for decades.

Second, large language models. ChatGPT, Claude, Gemini, Grok. This is what most people mean now. You type words in. The model generates words, code, images, plans, or structured data back out.

Third, AGI. A system meaningfully above the best humans across most useful tasks. This is the thing people argue about. It does not exist yet, but it drives a lot of the investment and fear.

A language model works by taking in tokens. A token is a small chunk of text, image, sound, or structured data that the model knows how to handle.

When you send a message to ChatGPT or Claude, you are filling a context window. Think of it as the model input box. Your message goes in. The earlier conversation goes in. The system instructions go in. Any files or search results go in too.

The model does not see reality. It sees the context window. This is why context matters so much. If the right fact is not in the window, the model has to guess from training.

The model reads the context and predicts the next token. Then it reads the context plus that token and predicts the next one. Then again. And again.

This is why long chats slow down. Every turn adds more text. The model has to keep dragging the whole conversation forward.

It is shockingly inefficient. It is also why current AI feels both magical and clumsy. You are watching a very expensive autocomplete system pretend to be a worker.

Before you type anything, the app usually gives the model a system prompt. This tells it what role it is playing, what tone to use, what tools it can call, and what rules it should follow.

That prompt can matter a lot. It can make the model terse or chatty. Cautious or aggressive. Useful or annoying.

Most real AI work is becoming context engineering. The question is not just how smart is the model. It is whether we gave it the right information at the right time.

Long conversations eventually get too big. Apps handle this by compacting old turns into summaries. The summary is smaller, but it loses detail.

Another trick is retrieval. Instead of putting everything into the context window all the time, the system searches for relevant notes, documents, or web pages and inserts only the pieces it thinks matter.

That is what retrieval-augmented generation means. It is a fancy phrase for: find the useful context, then ask the model with that context included.

Language models are weak at truth by default. They do not know what is real the way a person does. They learn patterns that are rewarded.

But they are very good at taking messy human input and making it structured. This maps well to APIs. An API has fields, types, requirements, and business logic. The model can fill the form. The software can enforce the rules.

This is the difference between a screen-recorded demo and a real app. Real apps constrain the model so it can move fast without being trusted with everything.

AI is worst at the physical world. It does not naturally know physics. It does not naturally know boundaries, weight, smell, taste, or the small common-sense details humans get for free.

A person has high-bandwidth perception. We see, hear, feel, remember, and update our model of the world without trying. A language model mostly gets compressed inputs and has to infer the rest.

AI also has no native concept of truth. It can say true things. It can be very accurate. But under the hood it is not consulting reality. It is selecting tokens that fit the context.

The practical gains are real. A few years ago models could not reliably output a URL. Now you expect a model to write JSON, read a screenshot, summarize a spreadsheet, and fix a bug.

Vision models are a big part of this. When you paste an image into Claude or ChatGPT, it gets converted into tokens too. The model can reason over text and images in the same pass.

The result is not human sight. It is compressed sight. But it is already enough for a lot of office work.

Full automation assumes humans will hand over approval quickly. To spend money. To log into accounts. To send messages. To make choices that have real consequences.

That is not a technical problem only. It is a trust problem. It is a permission problem. It is a product design problem.

Cold starts are also real. An AI worker needs context. What you prefer. What your company allows. What has already been tried. What the budget is. What matters more than what. Most of that is not written down.

The entire SaaS economy was essentially monetizing the inefficiency of human labor.

Salesforce exists because humans are bad at tracking relationships. Asana exists because humans are bad at coordinating tasks. Slack exists because humans are bad at routing information. The whole category is basically: your employees are slow and forgetful, pay us per seat to make them slightly less so.

How much of the last 20 years of tech growth was really just pushing work around through process that never needed to exist? This matters because some of that value is not value. It is system loss.

It is almost entirely self-referential. You need managers because you have teams. You need HR because you have managers. You need internal tools because you have departments. You need compliance because you have internal tools. You need program managers to coordinate across teams that exist because the work was split across too many people.

Think about what software is. Humans write instructions for computers because computers cannot figure out what needs doing on their own.

The whole industry exists because of that translation gap. Humans have intent. Computers need explicit instructions. Software is the bridge.

AI closes the gap from the computer side. If an AI can take intent directly and produce action, you need less bridge. You need less UI. You need less packaged software. The functionality can be generated, used, and thrown away.

Software stops being the product. It becomes a byproduct.

The hard part is not that AI helps people work faster. The hard part is that it changes what human work is worth.

Most people do not spend their day being creatively original. They implement. They assemble. They move information from one shape to another.

That creates a barbell. Either you are cheaper and faster than AI, or you are doing something AI cannot easily originate. The middle gets squeezed.

Inventing a new kind of Lego is rare. Putting Legos together is common. AI is very good at putting Legos together.

A video game is mostly choices. Pick the mission. Pick the item. Pick the path. The low-level work is hidden behind the interface.

A lot of future work may feel like that. Your AI gives you product ideas, business models, drafts, plans, branches. You choose. You steer. You apply taste.

The spreadsheet, the email chain, the meeting, the formatting, the pull request, the retry loop. Much of that can vanish or happen in the background.

This does not mean work becomes meaningless. It means the valuable part moves up a level.

The market is betting hard that AI gets more powerful but stays expensive.

That may be wrong. Current LLMs are extremely inefficient. They push huge context windows through the model again and again. Modern systems have tricks, but the basic shape is still wasteful.

The current bet is: what we have now, just bigger. More chips. More power. More data centers.

I think we should expect intelligence to get much cheaper from algorithmic progress. Maybe 10x. Maybe 50x. Small models already do a shocking amount of useful work. The question is not how many tasks require superintelligence. The question is how many tasks only require competent intelligence with the right tools.

Most people do not have infinite questions. They ask about the same events, shows, sports, products, celebrities, tools, and comparisons.

That matters because a lot of answers can be generated once, cached, and reinterpreted later by a smaller local model.

You should not need a frontier model reading a giant context window every time you ask about the weather, the news, or a common concept. That answer should already exist somewhere.

If this happens, intelligence becomes a commodity. Fast. Cheap. Hard to capture value from.

Most software is built for people. Buttons, menus, tabs, dashboards, forms. That is why AI has been awkward. It is trying to use tools made for hands and eyes.

Computer-use models try to fix this from the other side. They see a screen, move a cursor, click, type, wait, read state, and try again.

If this works well, it changes the automation surface. An AI does not need a perfect API for every old system. It can use the system the way a person does.

That is slower than a clean API. But it is much more general. And once it is general, you can run workers in parallel. Ten paths. A hundred paths. Overnight.

This could move slowly or impossibly fast. It depends on how quickly computer use generalizes and how quickly companies build the permission systems around it.

Legacy systems will be harder. Physical work will be much harder. But browser work, office work, data work, email work, reporting work, spreadsheet work, slide work, and standard software work are all directly exposed.

My working assumption is simple: if a job is mostly digital, mostly browser-based, and mostly about moving information around, it is in the blast radius now.

I do not think we have a clean test for superintelligence. People talked about the Turing test for decades. When chatbots got close enough, almost no one cared.

If a system becomes smarter than the smartest human, how do we know? What do we ask it? How do we know it is telling the truth and not just giving a beautiful answer?

Also, the physical world still runs at physical speed. Cancer research needs labs, samples, trials, approvals, manufacturing, doctors, patients, and time. AI can speed up the thinking. It cannot remove every bottleneck.

The easy advice is: learn to use AI at your current job. That is not wrong, but it is thin advice.

If your work is mostly digital and repeatable, using AI may only help you stand closer to the thing that later replaces the task.

The better question is: what can you create, decide, own, or understand that is not just implementation?

Physical reality still matters. Taste still matters. Trust still matters. Distribution still matters. Ownership still matters. Speed matters a lot.

Yes. The claim that AI cannot create original work is mostly confused.

AI produces things that have never existed all the time. A new sentence. A new image. A new combination of ideas. That is originality in the literal sense.

The harder question is value. Random pixels can be original. That does not make them good.

Creativity is often recombination. Take existing parts and arrange them in a new way. Humans do this too. The difference is that humans also have taste, culture, memory, stakes, and bodies.

A lot of the smooth, verbose, overly polite AI voice comes from reinforcement learning from human feedback.

People compare model answers and pick the one they prefer. Do that millions of times and the model learns what answerers find pleasing.

That is why models over-explain. That is why they reach for tidy analogies. That is why they use a certain kind of soft corporate voice. It is not the base nature of intelligence. It is training pressure.

This site should not sound like that.

The biggest risk is not that everyone becomes useless at once. It is that mobility gets harder.

If knowledge, strategy, connections, and capital allocation are flattened by AI, the remaining advantages may be the ones you were born with: money, network, ownership, family, geography.

The ladder does not disappear. The rungs just get farther apart.

Legal ownership is also unclear. If an AI invents a thing, who owns it? The person who asked? The model company? No one? Current law does not cleanly fit the world we are building.

The hopeful version is simple: more people can make more things.

A person with taste but no engineering team can build. A small company can move like a large one. A curious student can learn with a tutor that never gets tired.

The physical world will still be slow. That is frustrating, but it is also stabilizing. We will not wake up tomorrow in a finished future.

The transition is where the work is. The people who stay calm and learn fastest will have a real advantage.