Do large language models understand what they are saying?

Does reasoning allow llm’s to ‘think’.

Will machines ever ‘understand’ what they have been trained on?

These were questions that stayed back with me after a debate on “Is cognition computation?”

While there are good arguments for large parts of those questions, most people can’t agree on what it really means to understand something. When we say we understand something, what do we mean?

Is it the ability to simplify something, explain it in as simple way as possible?

Is it applying the concept in various circumstances?

Is it knowing the limits of the thing we are understanding under different contexts?

It’s a hard question. One for which, I’ve built some intuition and a perspective that I will attempt to share here.

From noise to understanding

Let’s start from the beginning when there is nothing. Hmm, that not quite right. Ever since we are born and rather even before we are born, we start learning. By the time we’re conscious of it, our minds are already full of information: some reusable, some not.

Still, there are moments when we encounter something genuinely new—or when existing understanding needs to be updated. And in those moments, some interesting patterns begin to appear.

Early on, everything matters.

Small changes feel important.

New examples feel surprising.

Edge cases feel overwhelming.

But as learning progresses, something shifts.

Different inputs start to feel the same.

New examples stop adding much to your understanding.

You recognize the situation before thinking about it.

Things start to click and it’s easier to grok the incoming stream of input and information.

Abstraction as what survives change

I think of this stage as forming a good abstraction.

I think of abstraction as what remains stable when details vary and inputs change. To me an abstraction is not entirely a summary or a simplification. It’s not less information. It’s a pattern that keeps working even when the surface changes.

When you abstract, you are no longer tied to a specific instance. You’re operating at a level where many situations map to the same understanding. You can now apply that same abstraction to new situations as long as the abstraction remains invariant to the core concepts of that abstraction.

Why this feels like compression (but isn’t really)

At first I described and thought of this process as “compression”. It made sense intuitively.

Many experiences → one understanding

Many examples → one idea

Many paths → one conclusion

But what’s really happening isn’t that information is being thrown away. It’s that some distinctions stop mattering. You could still notice them. You just don’t need to.

We have at this point built a structure, a mental model, an invariant structure that holds when we subject our understanding to new variations of the concept or inputs. (Until it doesn’t and we need to update our understanding but for the most part it holds.)

That subjective experience—when noise fades and structure stands out—is what I call clarity.

Subscribe now

Clarity as a learning milestone

Clarity is not a mechanism. It’s a feeling.

It’s the moment when:

• new examples feel redundant

• variations no longer confuse you

• you can explain the idea simply

Clarity is how abstraction shows up internally.

It’s the sign that learning has reorganized experience around what matters.

Understanding as a stable internal state

This is how I think of understanding.

Understanding isn’t simplification. It is not accuracy or performance on a benchmark for machines or humans on tests.

It isn’t memorization.

Understanding is stability.

You understand something when:

• small changes don’t break your reasoning

• unfamiliar cases still feel navigable

• the idea transfers to new contexts

Understanding is what allows learning to generalize.

ACULT

This brings me to ACULT (pun intended)

ACULT is a lens, a template, for thinking about learning and understanding that stands for:

• Abstraction – finding what stays the same across variation, at each level, as we look at something at different levels

• Clarity – the internal feeling when distinctions stop mattering

• Understanding – a stable state that supports transfer

• Learning – the process that gets you there

ACULT - Abstraction, Clarity, Understanding, Learning Template

The idea is simple:

Learning produces understanding by discovering abstractions that bring clarity across variation.

That’s it.

• Engaging in pleasurable activities

• Engaging in work that gets you into a flow state

• Having and aligning your actions to a purpose—a guiding function

Despite my personal belief that being fulfilled is more important than being happy, this breakdown stood out to me as something I could relate to. And yet, there was a nagging feeling that something in this model was a bit off.

These are concepts I have re-discovered and seen firsthand changing my life and mindset in the last year. As I build an understanding of different ideas—how they manifest in my life, how I act on them—the meaning of these and the related systems keeps evolving in my mind. I've been building mental models I can latch onto, and this list is in conflict with them. I had to try and break it down to find the source of that discomfort.

Here is my current hypothesis: while I strongly believe in the power of finding and having a purpose driving your actions, in the moment-to-moment it might not always be at the forefront of everything you do. While I want to do better and be better in everything, being more efficient with …1 you need to maintain the right level of eustress (positive stress / pressure) and not go overboard. Purpose is enduring, especially if you subscribe to a humanistic approach to purpose and self actualization.2 Flow feels good when you're in it and is worth getting into more often. Pleasurable activities are nice too, but here's where my mindset has shifted in the last year—I no longer think they are necessary.

Hear me out. Finding pleasure in the moment or in your work is nice, but it can feel like a crutch. We build habits and behaviors to get dopamine hits and a continuous stream of rewards. Once you understand yourself better—how your brain works and what activities you use to satisfy those needs—you can start exercising more control over what gives you that satisfaction. Just like rewiring habits involves messing with the cue, trigger, action3 with the easiest approach being to keep everything the same but change the action—you can get the same benefit without chasing specific pleasurable activities. Similarly with flow4, it feels like a second-order effect or system rather than a fundamental construct.

A fundamental construct would either help you learn and understand something better, perform actions that advance your goals, or act as a feedback loop to refine your understanding:

1. Learning and understanding

2. Actions

   1. Achieve goals

   2. Refine or update understanding

Since pleasurable activities and being in flow are means to an end—efforts to create the conditions that aid these fundamental constructs—they can be replaced, optimized, or even removed. Hence, they feel like second-order systems to me.

From another lens: purpose is goal setting and alignment, a guiding function, a reward function in itself. Pleasurable activities and flow are more like outcomes or intermediary states that, with greater awareness, we can control and have more choice over.

This leads me to think about happiness in two models: the original three components (pleasure - flow - purpose), and a second model where you have more control. In the second, you understand yourself better and replace pleasure and flow with other mechanisms that aid purpose. Perhaps you end up back at pleasure - flow - purpose, but now the possibility space is much larger. When the first model stops working, you know you can update it to work better for you.

A simpler way to categorize these models is as static or dynamic:

Static model - works for the majority who haven't invested much in self-discovery, psychology, or neuroscience, and want quick results.

Dynamic model - works when you start understanding yourself better: your habits, your triggers, what drives you, why you do things. It fits when you want to exert more control and intentionally design your life. You still have multiple factors that come together to create happiness, but you have a larger say in what those factors are.

Before we wrap up, and in the spirit of falsifiability, if the first model doesn't quite work for me, it's fair to assume the second model might not work for someone else. So when would this new model be worse?

First, in the updated model, the possibility space is very large, so exploring and learning is expensive, in time and energy. There's more work required to arrive at an optimal solution for an individual, and where time is limited, this may not be the best approach to converging on happiness. Second, if this is applied to a large group, the complexity increases. You would need a toolkit of techniques to help people find their levers and the right settings for themselves. In organizations, there may be overlaps or conflicts between purposes and missions, so finding alignment becomes a more complex, multivariable function. For example, playing games or binge-watching shows might once have been your chosen pleasurable activity, but you could replace them with hitting personal bests at the gym, completing 10,000 steps a day, journaling, or genuinely connecting with other people to get that same sweet hit of accomplishment and dopamine.

To conclude, and to solve the mystery that nagged me at the start: writing these thoughts down has given me more clarity on my initial discomfort. It stemmed from the simple prescription of the static model, which doesn't account for the depth and agency of the individual in shaping their own mind and behaviors. Looking at the problem through different lenses and levels of abstraction yields a richer, more dynamic equation. That complexity allows for greater control and customization. So perhaps start with the simple model and finding purpose, and when your progress starts to slow, open the box and begin fiddling with the cogs and levers.

All the best on the journey of finding purpose, happiness and fulfillment.

Footnotes

we don’t need more games

We don’t need more games and we don’t need more game studios trying to make the next chart topper.

Some of this is bitterness at having failed to do it twice. It is hard to separate the lived experience from objectivity. While I have been making games all my life, both as a hobby and professionally, the last two major attempts were through starting new companies.

After quitting my job at EA, the first company that I started was with a bucket load of passion. A hunger to do it all myself, feeling as if I had forgotten to make games in all the years of managing mobile and social games. I wanted to build games from scratch, to experience the joys of creation from earlier years again. To self-publish, to prove that I could, and to take baby steps towards being a real business. I got all that I set out to do. Unfortunately, those goals weren’t enough to build a sustainable business, especially bootstrapped. We needed to generate much higher revenues from our game at a rate faster than we were making. I consulted and worked for hire to survive, and ended up making more money than my last job, but with a constant thorn in my mind of not building new games — this was the whole reason I had quit my job in the first place.

I started my second company after getting a term sheet from a local venture capital fund. While the terms scared the skeleton out of my body, it opened up my mind towards a new way of building companies. Along came all the romantic and inspiring YouTube videos of shipping fast, doing things that don’t scale and building billion dollar companies. Passion was suddenly enhanced by business and product thinking. We were going to build a games platform for emerging markets.

A single word best represents how that panned out – Ghanta (yeah right).

I was reminded of how I used to have a keen sense of detecting BS as a young adult. I was trigger-happy to dish out a well timed Ghanta every time I came across assumptions that were untrue or gross exaggerations. Somewhere along the way, that objectiveness got lost in the hubris of strongly and passionately believing in something without a solid grounding in deep work and first principles thinking. We ran out of the small amount of venture capital we had raised from another fund (bless their heart) and with a couple of failed acquisition conversations, lots of lessons learned, that journey had to come to a close.

Now, if I had to start a pure games studio again today, I wouldn’t. For the following few pragmatic reasons. The market is ridiculously crowded. Many of the top companies that have gone public in the games industry make money selling shovels, not games. The expectations of players are sky high, and they are spoilt for choice. It’s hard to innovate and come up with new mechanics, systems, progression, and economies that truly work well together. Sounds like a skill issue, till you start studying and comparing other industries.

When I look at my pile of shame (the games I own that I haven’t played or even installed once), I feel for all the good games that come out every year, but can’t all be played because we only have 24 hours in a day. You could make the same argument for many other creative fields that have saturated markets and a similar clear divide between indie and professional products. To these when you factor in a planet scale audience, where catering to small niche audiences is good enough for a profitable business. I would agree to your argument for the most part, but no.

Spending your most valuable resource — your time — on a hope that ‘it might work’, is not good enough. When you build something, it’s your responsibility to take the best shot at the goal, and if you are not good enough yet, it’s better to pause, get better, and then take the shot. If you are worried about time passing by instead of how you spend your time, then that goal is most likely not important enough for you to give it your best shot. Should you even bother with such side quests?

we don't need more games… or do we?

So then, how do you correlate the life altering impact of games on people and the realities of building  a game business? The cultural landscape is ever changing and people need new stories to relate to. The technological landscape is changing at an even faster pace, spawning new opportunities to create innovative formats for people to immerse and engage themselves with, to find connection, to combat loneliness.

I am always reminded of the first game we took over from another team at EA, Pet Society. At first, I thought I would hate working on that game – it was just a glorified shopping cart. But that all changed when I saw a review from one of our most engaged players that went like this.

“I have been suffering from xx and my life had lost all its meaning. Taking care of Kiki and decorating rooms for her, taking care of her needs, taking her to her friends home for a visit everyday, has given me new purpose. It has helped me find connection and move away from a life of depression and towards active recovery. Thank you for making and continuing to build the game.”

How do you not feel proud of something you have created after reading a message like this?

Games can become full time hobbies. Yes, people might only have time for a few hobbies at a time, but when people do get hooked, they engage and retain deeply. This allows creating business models which are less about entertainment and more about lifestyles.

Games can give us:

Story
Power
Escape
Design
Delight
Fantasy
Wonder
Strategy
Discovery
Challenge
Dopamine
Simulation
Excitement
Completion
Community
Competition

As a craftsperson, how do you build things that can get your audience to experience this whole range of emotions? If you don’t start by building games, how do you master the craft? You have to start somewhere.

So. If you are making it a business, get good at the craft of business and where games fit into the business. Otherwise, don’t do it for the money, do it for the love of the craft and to grow towards mastery! Fulfillment will follow.

As I grow older, and as I solve problems closer to my heart with a stronger purpose; games remain a beautiful medium and an important part of who I am plus how I got here, and I can’t treat making games as just a job anymore.

“Games are dead. Long live games!”

“Because it was feeling a bit framed.”

My heart rate slowed down and stabilized as I took the corner leading home. It had been pumping hard for the past 25 minutes and doing heavy duty work while I attempted a new personal best for a 5k run.

I remember that run vividly because of the mental connections I made next. I had been listening to an audiobook where a mention of perspective shift and reframing came up just as a security truck passed by. This truck had its windows completely welded shut. From a construction and functional point of view, I wondered why would the car makers add the windows to the truck, and introduce weakness in the structure, if they were going to weld it shut anyway? The obvious answer was that the cars were built first to be generic and adapted for this particular use. That thought raised the question, how do we design products, physical or digital, with a view towards how they evolve, and as their intended use changes? Something to ponder.

Back to the real world. We had not long ago stopped working on Playmat, our mobile first games platform for emerging markets. The zero interest rate era was over. Raising funds for startups was getting harder, and we had made an already hard space look harder with how we had strategized building up the platform. As of that moment, as a mental exercise, my thoughts got narrowly focused on reframing the product. As a window. 

A window is designed with a few things in mind. Who is going to use the window? What do they expect out of it, where is it placed, does it serve more than one function? How does its utility change over time? You could draw a parallel in the physical form and functionality or on a conceptual level. Let’s focus on the latter.

Consider the window of a car. You’d want to roll down the windows and feel the breeze. But what if that window was an amphibious vehicle that was supposed to work on the submerged roads of Mumbai after their annual rains or Florida after a hurricane hits it every 3 or so years? I imagine you’d definitely want to be able to seal the window shut tight.

What was a perfectly nice window would be useless under water, so we need the window to perform multiple functions in different situations either from the get go or as we realize the new need. What do they say about jumping off a cliff and assembling the plane on the way down? While it’s fairly romantic and makes entrepreneurs feel like heroes, after removing the rose tined glasses, there are things I would pack for before I make the jump again. In other words, I would think through and make affordances in the product understanding that it’s dynamic and will change with time.

This mental exercise of framing software products as windows works because it acts as a forcing function to think about the product in a different light and circumstance. What we would consider as implied knowledge and solution by the creators suddenly becomes a question of is x really needed, or is y really what we or the user are trying to do?

For Playmat, it clarified that what we considered our window was trying to act as the portal to Narnia; while magical and filled with unlimited possibilities, not what someone would expect when they are looking for a closet to get in or out of.

What other things could this reframe be used for?

What about dating? 

A user is using the dating app / window to look at who else is out there. Their eyes are executing a search query for identifying prospects that check their boxes. You’d want to share some of your information with the outside, let someone else peek in, perhaps let one or two in, oh wait, that got too crowded, the window should let the user push someone back out too. Hopefully, with the right person in, close the window for good.

If you really get down to it, this essay is itself a window, think about it.