Devin Oldenburg

Humans work with the inputs they collect over time; every human is a copy machine. You become influenced by what you see, experience, and act on, especially by the people and systems around you.

This can be explained more deeply in many different situations, but today I want to argue that you should not let your current level determine the standard you work against. Instead of slowly moving from the bottom to the top, expose yourself to people, systems, and outputs that are significantly better than your own.

Chess is a simple example. If you only play against people at or below your current level, you will still improve, but the standard of the game remains close to what you already know. Playing against a much stronger player or engine exposes mistakes that weaker opponents may never punish.

This does not mean that maximum difficulty automatically creates maximum learning. A stronger opponent is only useful if you can understand the feedback you get from it. But when that feedback is understandable, a higher standard gives you more information about what you are doing wrong and what better decisions look like.

To link this back to the process of learning I explained a few days ago: a stronger system raises the baseline of the feedback loop. It punishes mistakes more consistently and gives you higher-quality information to use in the next iteration.

You live in the 21st century, and you have access to more high-quality information, systems, and people than any generation before you. There is little reason to limit your reference point to the people around you or to your current level.

This is not just true in chess. If you want to become a better programmer, read and work with code written by people far better than you. If you want to become a better writer, read people who write far better than you. If you want to build better companies, study the best companies instead of only comparing yourself to people at your current level.

The quality of the input raises the baseline of the output.

Under German law, a knife can be a tool used to cut bread or, depending on its design and use, a weapon. The same principle applies to many technological advances. This is commonly described as dual-use technology.

The actions taken with a technology are ultimately determined by the individual using it. Freedom is valuable only when the individual who has it is capable of treating it as both a responsibility and a tool for expanding their capabilities as a human being.

For example:

In school, you can use the technological advance of LLMs either to cheat or to significantly reduce the feedback latency involved in the process of learning. One of these actions wastes your time; the other expands your capabilities as a human being.

Regulatory authorities often justify restrictions on technologies by arguing that individuals cannot always be expected to understand their risks or take sufficient responsibility for the consequences of their actions. This creates a recurring tension between technological freedom, individual responsibility, and institutional control.

Almost everything you do is influenced by information you have received before through experience, observation, conversation, reading, or other forms of learning. You take in information, process it, act on it, and observe what happens.

This follows a similar structure to the process of doing that I discussed in my previous post.

There are many different things you can learn, and the best way to learn them will differ. But one structure that works especially well is a feedback loop.

At its simplest, such a system starts with two points:

1. Input

You put something into the system. This could be information you read, a question you try to answer, a move you make, or an idea you want to test.

2. Output

The system processes your input according to its rules and gives you a result. That result gives you new information about whether your input worked and, ideally, why.

Input and output alone are not yet the important part.

The learning happens when you use the output as information for your next input.

Input → Output → Updated Input → Output → ...

This is what makes it a loop.

A simple chess engine could work like this:

You give the engine a move together with the current position, for example:

[Move: White / Rook / G2 -> H2

White: Rook G2, King H1...

Black: Bishop D5, King H8...]

In this position, the engine would reject the move because moving the rook to H2 would leave the king exposed. The move is illegal.

This is where the quality of the feedback becomes important.

Imagine two chess engines giving these outputs:

1. Illegal move.

2. Illegal move — your king would be left in check.

Both engines correctly reject the move. But if you are trying to learn, the second output is much more useful because it tells you why the move failed.

You can now use this information for your next move.

Your first input produced an output. The output changed the information available to you. This new information influences your next input.

That is the actual feedback loop.

You make another move, receive another output, update your understanding again, and repeat the process.

This is what makes feedback loops so effective for learning. You can change the input, test a different condition, observe the output, and use it to improve the next input.

Instead of only being told whether something works, you begin to understand the rules behind the system.

Every input connected to an output becomes another piece of information you can use in your own thinking.

I chose chess as an example because a large part of learning chess is pattern recognition. You see positions, make moves, observe the result, and gradually recognize situations you have seen before.

The same idea applies far beyond chess.

When programming, you write code, run it, observe the output, change something, and run it again.

The important variable is therefore not only how much information you consume, but how quickly you can turn an action into new information that improves the next action.

This is also why the latency and quality of feedback matter.

If you make a mistake and receive useful feedback immediately, you can use it in your next attempt. If the feedback arrives weeks later, the loop becomes much slower.

The faster and clearer the feedback is, the faster you can update your thinking and repeat the process.

Learning is therefore not just receiving information.

It is repeatedly producing outputs, extracting information from them, and using that information to create better inputs.

This is the process of doing things, and it is applicable to all kinds of activities and goals you have ever pursued and will pursue. At a very abstract level, this process can even be observed in simple actions, such as moving your hand to pick up a glass from the table to drink. 

This article will view this process without the context of a specific topic or example and treat it as the flow that it is. 

Where does value actually get created, and where does it usually get lost? 

1. Idea 

The Idea defines the direction. A bad idea can be executed perfectly but still produce an output that nobody wants. A good idea is defined as the right direction that will act as a foundation beneath the other stages of the flow, leading to the expected result.

It is defined by the expected outcome after the execution.

2. Plan

3. Execution 

4. Output 

5. Outcome 

If each factor ranges from 0 to 1, then:

1.0 idea × 1.0 plan × 0.1 execution = 0.1

1.0 idea × 0.1 plan × 1.0 execution = 0.1

0.1 idea × 1.0 plan × 1.0 execution = 0.1

A long time has passed since I last worked on a personal project which I worked on solely by myself, and whenever I did, I learned an enormous amount by doing so. Today I will be presenting my plans for Kaeso, an open source AI Gateway which will initially aim to compete with LiteLLM and then move on by adapting new features. My goal is to build a foundation for the future, and by building Kaeso this is the perfect way. I personally love and believe in open source because without it, I would definitely not be where I am right now. 

For now, I will be working alone, which means I have to restrict Merge Requests for now as I don't have enough time to evaluate them properly.

From my understanding and knowledge of market needs, I believe that this idea is not a good one and will probably fail, at least as a startup it would. This project in its current state is a solved problem and the initial user base will be very small as LiteLLM is significantly further developed. 

Kaeso in its first version will be split into four core components which will shape its fundamental structure: 

As Apple's early Marketing Philosophy said, people DO judge a book by its cover. So the first big difference will be the UX/UI for the people using Kaeso. In my view, the current LiteLLM User Interface has considerable room for a cleaner and more cohesive experience. Kaeso will aim to create a good environment for all kinds of users. 

Of course, announcing this as the only change is vague. In most cases nobody cares about the UI and it's the functionality which matters, but as I'm currently just trying to establish a baseline, this is the first and best visible best practice I can create.

Lastly, I want to mention that the Kaeso Roadmap will be fully led by me and that will probably remain the case at its core. 

Half a year ago, I spent a whole week trying to figure out the best way to create to-do lists and define objectives for the future. Living without a plan mostly leads to being busy while achieving absolutely nothing. A plan is a must.

You should not set time goals. Saying “I will play guitar for one hour” does not mean you will improve. You can get distracted, work slowly or repeat things you already know. Time spent is not the same as progress.

You need a specific objective in front of you. Instead of saying that you will play guitar for one hour, define exactly what you want to be able to play afterward. You need a clear result to work towards and a point where the objective is completed.

Objectives must be specific. “Work on my project” is not an objective. It is too vague and makes it easy to get distracted. You should know exactly what you are working towards and what the finished result should look like.

Every objective also needs a serious deadline. Not a deadline for how long you will work, but a deadline for when the result must exist. Without one, you will keep delaying, changing things and wasting time.

It is also not about having many objectives. It is about completing them. There is a difference between being busy all day and finishing one objective after another. A long to-do list means nothing when nothing gets completed.

You should constantly look for bottlenecks while working. Analyse what slows you down, what distracts you and why some objectives take longer than expected. When the same problem happens again and again, fix the way you work.

You should also track your work. Write down what you planned, what you completed and how long it took. Your own feeling can be wrong. Numbers show whether you are actually progressing or just keeping yourself busy.

A good to-do does not only tell you what to do. It defines exactly what must be achieved, when it must be completed and what counts as finished.

It is one of the best known phenomena in human life, although most people are unable to explain it. It is what happens when you have an important appointment in the morning and cannot fall asleep.

Daniel M. Wegner (1994) described this phenomenon in his paper Ironic Processes of Mental Control as a state in which self control does not fundamentally fail, but a paradoxical effect comes to life and makes it more difficult for people to resist. It mainly happens when individuals are stressed, tired, or experiencing a high cognitive load.

Wegner distinguishes between two parallel processes in the human brain during such moments:

• The operating process: the conscious process that searches for distractions or alternative thoughts. It requires attention and mental energy to function.
• The monitoring process: the process that continually checks whether unwanted thoughts appear. It operates as a filter for the conscious mind.

When trying to fall asleep, you may do the one thing that will loop in your head for the rest of the night. You start a process with the task of concentrating on falling asleep and not thinking about anything else.

Under normal conditions, the operating process is capable of steering your attention. However, under the conditions mentioned above, it becomes weakened while the monitoring process keeps running. This makes the unwanted thought more easily accessible.

A few days ago, I was reading Zero to One by Peter Thiel (2014). Although the topic was not discussed explicitly, hearing Thiel talk about his team at PayPal made me curious about what is required of the people in a founding team to build a great startup. So I took a look at the founding teams behind several well known companies and products. The team behind PayPal is a separate topic, but after analyzing a number of successful startups, I noticed a specific pattern. 

Examples: 

Cursor 

• Michael Truell
• Sualeh Asif 
• Aman Sanger
• Arvid Lunnemark

Google

• Larry Page
• Sergey Brin

Page and Brin met at Stanford University in 1995. 

Apple

• Steve Jobs
• Steve Wozniak
• Ronald Wayne

Jobs and Wozniak got to know each other through their mutual friend Bill Fernandez.

And I could go on...

(sources are linked through the company names.)

The point is that they all met through their social networks. Why did so many of them meet at university? Because universities concentrate ambitious people who want to learn and achieve more. They can provide an ideal environment for people with similar interests to meet and work together. 

Looking at this pattern, you could argue that it is extremely difficult to found and build a great company entirely alone. In my opinion, this is not simply a matter of whether someone should or should not have co-founders. A group of great people working on a specific problem or product can produce greater outcomes, not only because of their individual capabilities, but also because of the knowledge, connections, and opportunities accessible through each person. 

These are only some of the clearest examples. In these cases, social networks directly influenced the formation of the founding teams, but indirect connections also count. Believing that one person can build every part of a great company alone ignores the limits of an individual’s time, knowledge, and capabilities. 

Applying this knowledge to my own life, I will stop treating school and official forms of certification as irrelevant alternatives. I have often argued that school is inefficient and requires spending a lot of time in a slow moving structure that does not always match my interests or preferred pace. Because of that, I did not care much about it and did not want to invest my time in such a system.

I am still aware of these problems, but I now see school as a stage I have to use and complete in order to move forward. It can give me access to universities, environments, and people I might otherwise never encounter.

This was never about the material. I have always loved learning, and I have never had a problem with the content or the knowledge itself. Most of every day I spend learning new things.

I think working with great people is one of the best things that can happen in someone’s career.

Life begins with time.

Time is one of the few resources that cannot be recovered. Money can be regained. Systems can be rebuilt. Knowledge can be relearned. But once a moment has passed, it cannot be used again.

This makes time a fundamental constraint on almost every decision we make.

A human life is limited, yet the number of possible ways to spend it is enormous. We cannot explore every path, learn everything, meet everyone, or build every idea. Every decision therefore has an opportunity cost: choosing one direction means temporarily giving up many others.

This raises a difficult question:

How should we decide which path to take?

Information

From the moment we begin interacting with the world, we accumulate information.

Experiences, observations, conversations, mistakes, books, environments, and relationships gradually influence how we understand the world. Most individual experiences seem insignificant when they happen, but together they form the context from which future decisions are made.

You might remember a sentence from a book years later. A conversation may change how you think about a problem. A failed project may reveal a constraint you had previously ignored.

Not every experience will remain consciously accessible, and not every experience will influence us in the same way. But over time, patterns form.

Patterns

Human behavior is not created from scratch before every action.

We develop habits, expectations, preferences, and automatic responses based on biology, previous experience, and the environments around us. These patterns reduce the amount of conscious effort required to navigate everyday life.

Most of the time, this is useful.

You do not need to consciously decide how to perform every familiar action. You rely on patterns that already exist.

But patterns can also become limitations when they continue operating after the conditions that created them have changed.

This is where reflection becomes important.

Agency

We cannot control every process that influences us, but we can examine some of them.

We can ask why we reacted in a certain way. We can question assumptions. We can notice repeated mistakes. We can deliberately expose ourselves to new information and environments.

This does not give us complete control over ourselves or our circumstances. Human behavior is influenced by factors we may not recognize and by external conditions we cannot control.

But reflection gives us some ability to change direction.

That ability leads to decisions.

Decisions

Every meaningful decision allocates some amount of limited time, attention, or effort.

Choosing to learn something, build something, spend time with someone, or pursue a particular opportunity means investing resources that could have been used elsewhere.

It is therefore useful to think of decisions as investments.

Some investments produce the result we expected. Others fail.

Failure, however, is not automatically wasted time.

A failed attempt can produce information. It can reveal that an assumption was wrong, that a strategy was inefficient, that a skill was missing, or that an objective was not worth pursuing in the first place.

The important distinction is between failure and repetition without learning.

Making a mistake once may provide useful information. Repeating the same mistake under the same conditions without updating your behavior produces increasingly little new information.

This is why reviewing decisions matters.

After something important happens, useful questions include:

• What did I expect to happen?
• What actually happened?
• Which assumptions were correct?
• Which assumptions were wrong?
• What information did I not have?
• What would I change next time?

Uncertainty

Even careful reasoning cannot guarantee a successful outcome.

The world contains too many variables to calculate a perfect path through it. Other people make decisions. Opportunities appear and disappear. Circumstances change. Random events occur. Information is incomplete.

Two people can make similar decisions and receive very different outcomes.

For that reason, decisions are better understood in terms of probabilities than certainty.

You usually cannot know that a decision will produce the outcome you want. You can only try to choose actions that, given the information available to you, appear to improve the probability of that outcome.

This is similar, in a limited sense, to probabilistic systems in computing. A model does not possess perfect knowledge of the future. It operates on available information and produces an output according to the structure and probabilities it has learned.

Human decision-making is obviously far more complex, but the analogy is useful: incomplete information does not make decision-making impossible. It makes uncertainty part of the problem.

Learning

If certainty is impossible, the objective cannot be to make a perfect sequence of decisions.

A better objective is to create a process that improves over time.

You act using the information available to you.

You observe the outcome.

You extract information from it.

You update your assumptions.

Then you make the next decision with slightly better information than before.

In simple form:

Information → Decision → Action → Outcome → Learning → Better Information

The loop does not guarantee success. It improves your ability to navigate uncertainty.

Autonomy

There is another problem.

Not every goal we pursue was consciously chosen.

Families, schools, cultures, social groups, algorithms, and economic systems all influence what people consider desirable or normal. Some of those influences are useful. Others may push someone toward a path they would not have selected after serious consideration.

Complete independence from external influence is impossible. Even the ideas we use to question society were learned somewhere.

Autonomy therefore does not mean living without influence.

It means examining influences before allowing them to determine large parts of your life.

If a goal requires years of your limited time, it is worth asking why you want it.

Is it something you actually value?

Is it a means to another objective?

Did you choose it, or did you simply inherit it?

There may not always be a simple answer. The act of asking is still valuable.

The Path

There is probably no single optimal path through life.

There are too many variables, too many unknowns, and too many possible futures.

But that does not mean direction is meaningless.

You can gather information.

You can choose objectives.

You can act.

You can observe what happens.

You can change your assumptions when reality contradicts them.

And you can gradually improve the decisions that determine where your limited time goes.

The objective is not to eliminate uncertainty.

It is to become better at operating inside it.

You will still make mistakes.

You will still encounter circumstances you cannot control.

Some decisions will fail despite being reasonable at the time.

The important part is what happens afterward.

You extract the information.

You adjust.

You continue.

Because time continues too.