General Hill-Climbing: The One Main Loop in AI

💡There is basically One Main Loop in AI, and it applies to everything from building software to managing pretty much anything.

⛰️I call it General Hill-climbing.

Here are the main components:

• 🖼️ Know what you want
• ✍️ Be able to articulate it
• 📐 Be able to measure progress
• 🧪 Do the work / test your ideas
• ⇧ Adjust / Iterate / Climb / Repeat ♻️

In 2026/2027 modern language that means:

1. 🧭 YOUR TELOS (Purpose) ​

You want to start with clear articulation of your current state, your ideal state, your goals, metrics, challenges, strategies, projects, etc. This is where you're trying to get to, where you are now, what you're trying to accomplish, how you measure progress, the things that are stopping you, what you're doing to overcome those issues, the work you're actively doing, etc.

Your TELOS captures this all together, in a single place, so that you, and other people, and your AI systems can all understand the overall direction. This is the starting point and center of reference for work we want to do.

2. 🖼️ CLEAR ARTICULATION OF IDEAL STATE ​

Step one is defining what you're trying to do overall, with as much context as you can provide around not just what you're doing, but why.

Here we want to specifically define the ideal state for a specific task. Like creating an app, or starting a business, or creating a plan for selling a new video game you made. Or writing a book. Whatever.

You need an extremely thorough and clear articulation of exactly what you want.

We believe very strongly that more and more AI failures today are not actually AI execution problems. AI is getting really good at doing exactly what we asked for.

The problem is often us either:

• a. Not giving a complete picture of what we wanted
• b. Giving an incorrect / conflicting picture of what we wanted

Layers of abstraction ​

Think of this as layers of abstraction.

1️⃣ The true thing is the idea in our head. That's ground truth. That's the thing we actually want to build.

2️⃣ The next step is the capture into an artifact. The moment we leave that and try to capture it via voice, or text, or whatever, we've lost something. Hopefully not that much. This step really matters. This is the step we have to do really well, because AI can't read our minds.

3️⃣ And finally we try to build the project. It's critical to understand that this comes from step 2, not from step 1. AI has no way to access number 1. It can only go off of number 2. This is why the ideal state capture piece is so essential.

So we have to address this by being extremely thorough in our description of what we want. This might mean doing a whole lot of dictation about the project, what you're trying to do with it, why you're doing it, how you hope it will help, etc.

Like if you do a full hour, or three hours on the Ideal State Artifact (our way of capturing the full ideal state in one document), that might not be too much. If it's a smaller project you should be able to get away with less. But you have to put the effort in to fill in a super-high-quality capture of what's in your mind.

Just know that the primary limitation for AI today is not its ability to build what you ask for. It's your ability to clearly articulate what you want.

3. ⚖️ INSTRUMENTATION (Evals) ​

Ok, so once we know our general direction we're building towards, and we know exactly what we're building in this specific case, you then need a system that can watch you executing work against a given target / goal / destination and tell you whether you're improving or getting worse.

This means measurability. This means specifics. And there's a reason why we call it "General" hill-climbing. You need to be able to measure progress or regression against any type of thing you're trying to do.

That means not just science and engineering, but product management, sales, marketing, design, etc...basically any work that you do. If you can't instrument the work in a way that's tied to the ideal state you're trying to get to, you can't know, discuss, or improve the situation.

This is where Evals come in. Evals come from software development but in this new AI model of the world, Evals will become far more general, making them pretty much universal and essential for any type of work.

The simplest way to think of Evals is as a set of checks to know if something is good or not. And they depend greatly on the type of work being done. They can be extremely simple, or elaborate and complex.

A set of Evals for a report, for example, might be something like:

• ✅ All the facts in the report are accurate
• ✅ The data in the report is no more than a year old
• ✅ The report does not include any mentions of ____
• ❌ The topic of ______ was addressed in the report
• ✅ There are no spelling or grammatical errors

So in this case we have five components to this report Eval. And this particular report passed 4/5. And in that case, if a "good" report is one that can pass all 5, then we know how to hill-climb towards a "good" report. We simply need to get to 5/5.

Or we could have a rating system where 4/5 is Good, and 5/5 is Excellent....but....4/5 would not be Good if criterion #1 (accuracy) were false. It's up to us how we set it up.

Types of evals ​

There are lots of different types of Evals. Here are a few so you can see how powerful these are in combination:

• ❓A simple check (are there any misspellings)
• ❓Which one of these is better? (asking AI to judge between two things)
• ❓Where on this scale does this report fall (asking AI to judge between Critical, High, Medium, Low, or whatever)
• ❓Does it match the answer we already know is right? (comparing the output to a reference answer)
• ❓Does it contain this exact thing? (a plain code check for a required phrase, link, or field, with no AI involved)
• ❓Is it in the right shape? (correct length, valid format, every required section present)
• ❓Is this claim actually true? (checking each fact against a source)
• ❓How does it score on each part of a rubric? (AI grading against a written scale, like 1–10 for clarity where 10 is defined in advance)
• ❓Which of these twenty is best? (a tournament, where AI compares them two at a time until one wins)
• ❓Did it actually work? (running the code, clicking the button, or loading the page and seeing if it does what it should)
• ❓Does it hold up every time? (running it ten times and seeing how many pass, not just once)
• ❓Would I have picked this one? (an AI judge tuned on your own past ratings, so it scores the way you do)

The power of all this is that when you build an Eval, or a set of Evals, for a given work task, you can use just one, or some, or a giant collection of these all together. Depending on the work.

Combined, they give you a way to rate the quality of a given piece of work. And when one, or some, or a lot of the different criteria fail, you have something to improve on.

You have a way to hill-climb.

One could argue (and many do) that Evals are the most interesting, exciting, and critical part of all AI work going forward. I'd put them at second place, behind the articulation of ideal state. But yes, they're absolutely critical.

The IDEAL STATE criteria are the VERIFICATION criteria. Pursuing the Algorithm (January 2026)

If you aren't doing Evals, or you're not doing them well, it's going to be very hard for you to use AI to get what you want. Especially at scale.

4. ⚙️ DOING THE WORK (Building to the Ideal) ​

So now we know our general direction, and we know specifically what we're building, and we also know how to measure whether the output is good or not.

Now we're ready to do the work.

So basically we do the thing (or our AI does, actually). We build the app. We write the novel. We build the business plan. We create the new design for the company website. Whatever it is.

The AI is now building according to the Ideal State Artifact that you've created, and it's using the Evals you've created to know how it's going. It will produce an output, and your Evals will tell you how good it is.

If they all fail, you've got a lot of work to do. If they all pass and have acceptable ratings and scores, you're good to go. And if there are some passes and some misses, that's where you start hill-climbing towards the bar you're looking to hit.

5. ♻️ ITERATION AND EXPERIMENTATION ​

Now for the meta step, which is basically the self-improvement system.

1. Our Ideal State is not static. We don't always have the complete picture in the beginning, and we often get more ideas as the project starts to develop.

2. Our Evals can evolve as well. The best way to tell if you need to update your evals is if you get a result that perfectly passes your evals, but you still don't like it for some reason. Similar to the challenge of capturing your ideal state, it's also extremely challenging and rewarding to do the work to get your Evals to properly represent your ideal state. This is an iterative process, just as with the tweaking of what we want in Step 1.

3. Experimentation is your friend. The best thing about putting in the work on Ideal State and Evals is that you

📃 SUMMARY ​

Taking all these pieces together, what we have here is a General Hill-climbing System.

In my opinion this idea......this mechanism....is bigger than AI. AI is just the thing that makes it possible to apply to any problem, and at scale.

It's basically a universal problem solver. Not instantly. Not perfect. But it's a path. It's a system. And it works on pretty much anything you throw at it.

Utterly magical.

This is how I think of AI harnesses. This is how I think of AI for our personal improvement. This is how I think of AI for the enterprise. This is how I think of AI period.

Knowing what we want, and hill-climbing towards it.

I challenge and encourage you to start thinking of AI in this way as well.