FOUNDER

Despite having secured a patent for a system involving AI, debugged Luma AI's multinode inference and getting their models stable under intense time pressure, restored a 42 node mi300 cluster from a tech-debt tsunami, and singlehandedly created Tensorwave's internal AI platform, I believe my most significant accomplishment is one that is very silly to say out loud. It was me beating the videogame "Sekiro: Shadows Die Twice". The more self-important way of phrasing it is that I began, in earnest, the conscious development of my systems-thinking, my own processes to understand the underlying mechanics of a system and build an intuition for how it behaves. The game made me aware of an unending journey that I was on my whole life, and that awareness has led me down the path of constructing more deliberate internal systems to make rapid progress in that journey.

I have always had too many interests and been obsessed with improving my capabilities in all of them, and without being consciously aware of it, I began drawing connections between the "meta-games" in all my areas of interest. As the only student in my high school class involved in both sports and acting, I found myself connecting my understanding of physics principles (which I excelled at) to both subjects. The lower level metagame of motion came into focus — how it translates up the stack into the data that is relevant in each domain. How torque, leverage, and angles at which things come into contact translate into athletic performance on the field or realism in acting. How the smallest shifts can propagate into larger consequences despite being practically unnoticeable to most observers and oftentimes unnoticeable to me consciously. On the wrestling mat I developed a sense for an opponent's intent from slight shifts in their weight, and could translate that into my ability to wordlessly convey intent on stage. Piece by piece, I was building an internal model for how the physics of the human body worked.

Yet I would consistently fail to develop such models of understanding in other domains. On the lacrosse field, while my physics understanding allowed me to rapidly improve my stick skills despite picking up the sport almost a decade behind most of my teammates, I still never succeeded because I didn't grasp the team dynamics and ball movement. Whiteboard sessions explaining plays were straightforward enough, but when things went slightly off the given plan, as they inevitably would in such a fast paced game, I was completely lost. I had no understanding of what to do. I could read the intent of the player I was against. But I didn't understand the chain of events that would occur beyond what he would do next.

A pattern emerged among my ever growing variety of hobbies where my prior developed models of "gameplay" either constructively or destructively interfered with my understanding. I hit a wall with only a very basic understanding of chess while improving rapidly in other games involving resource management, such as tower-defense or civilization games. I fumbled aimlessly in poker while experiencing endless strategic growth in Magic the Gathering, despite them both being card games. Strangest of all, I made slow, frustrating progress through The Legend of Zelda: Breath of the Wild despite flying through previous Zelda titles.

Then came Sekiro: Shadows Die Twice. A game from the studio FromSoftware, the creators of another game I had beat and enjoyed: Dark Souls. Sekiro carried the same control scheme as Dark Souls, leveraged the same game engine, and had a very similar vibe. Yet I failed over and over again, much more so than I ever did in Dark Souls. It made no sense. I brought in experience from what seemed to be a very similar game, yet I was playing terribly.

What broke the deadlock was a strategy guide. Or rather, a reddit post with a conversation thread. Unlike other game guides, this "guide" did not go deep into technical sequences of moves or specific patterns of enemies. Instead, it was a conversation that effectively said this: "This game is not Dark Souls. Dark Souls is about you, the player, being severely underpowered against very dumb godlike enemies who can do crazy things you can't do. It is about picking a specific strategy before the fight, building your character around it, and using it to chip away at your enemy during the fight. You are restricted in what you can do, have few options determined by how you've built your character, have very constrained resources to manage, and have to be patient for the opportunity to use them against bosses who can move way faster, hit harder, and cover more area in their attacks, but inevitably do the same over and over. It is about being patient for the opportunity to do a small amount of damage over and over. Sekiro is about you, the player, being evenly matched against a very smart, very aggressive opponent. You don't have a huge variety of strategies to pick amongst before the fight, but you have more flexibility to change strategies during the fight. You have unlimited stamina, quick attacks, a variety of options, and only one resource to manage — your "posture", and have to protect it while attacking your enemy's posture. It is about you applying relentless pressure on your opponent until you break their posture and can land a killing blow.

This is where the fundamental approach to systems understanding began to form in my head. I returned to the game and began to see what it was "about". I began to understand why the game played so differently. The slight differences — the lack of limited "stamina", the speed of the attacks, the "posture" mechanic, completely change the way the game plays. That way of analyzing how the mechanical differences propagate into different system behavior allowed me to not just beat the game, but return to other games, or really any system I didn't understand, and develop a new internal model for it.

It was the beginning of me honing my most important internal patterns of questioning: 1. The fundamentals of the system: "what are the consequential, lower level mechanics of the system and what behaviors do they reward or punish?" 2. The "meta-game" of the system: "based on the mechanics, what is this system about-- what decision-making framework do these behavioral rewards incentivize when they are all put together?" 3. The "derivative" of the system: "what systems under these ones cause these surface level systems to change?"

I refined these patterns of questioning as I returned to previous areas where I was unsuccessful. I returned to The Legend of Zelda: Breath of the Wild and through questioning, found: 1. The consequential mechanics are that weapons break and cannot be repaired, I have a limited number of weapons I can carry, which punishes treating weapons as reliable, long term tools and rewards treating them as consumables to use on a regular cadence. Defeating enemies doesn't offer XP or currency — just the weapons they are carrying, and defeating them isn't even a requirement to proceed or to get the weapons — theft is an option, albeit a risky one. This rewards not engaging the enemies. 2. With these mechanics, the meta-game comes more into focus. If weapons are limited-capacity consumables and the reward for engaging enemies is more weapons, then the metagame is quite different from traditional games of "find a way to beat this enemy". The game is not about defeating every enemy in your path, it is about the weapons economy. It is about deciding whether or not to engage the enemy in the first place. Does the enemy have a weapon better than mine, which makes consuming my current weapon to take it worth the risk? Are there other alternative pathways to engaging the enemy? 3. A derivative for the weapons economy would be the rate at which I'm able to acquire new weapons without engaging enemies. This can be done via theft — stealth. If I find clothing or create potions that improve my stealth, the weapons economy changes. Or if I could go exploring for more weapons — traversal also changes the weapons economy. I could find faster ways to explore and improve my weapons economy that way. Following that model — prioritize the weapon economy, not an individual weapon — I flew through the game with much more success and enjoyment.

I then returned to Rocket League, a video-game involving football (soccer) played with rocket powered cars, and applied the same pattern. 1. The consequential mechanics are that I can accelerate forward, but I cannot quickly move laterally — I have to turn in an arc. I also cannot dribble the ball very easily, I can only collide with it. This punishes being near the ball but stagnant, it punishes being in the center of the field rather than away from but facing the center of the field. 2. Since the cars cannot move well laterally or dribble effectively, and they can only collide with the ball, the metagame is about chasing the trajectory of the ball based on the angles at which the players are moving towards the ball, and having the angle and momentum to be where the ball is going to end up, and then having a backup in case you miss. 3. A derivative for this is the rate at which we can make these attempts on the ball — which requires backups to follow through after previous attempts. So the model is not to make a singular push, it's to collide with the ball and rely on a back up to make the next collision while returning to a position to be the next backup.

But the key was that, throughout this journey, I wasn't focused solely on building my model of those specific games, of those specific systems. I was focused on the model I used to build those models. It was a recursive development pattern. I had turned the model with which I developed my model for any given system upon itself, and began asking myself what the mechanics, metagames, and derivatives of my system to develop models of other systems were. In each system, I would internally think through various outcomes, simulating the consequences of different patterns of decisionmaking. Outside the system, I would think through my own questioning framework, simulating how the questions I would ask myself in this specific system may yield better or worse results in different systems. I was building an understanding of the fundamental principles to take to any unfamiliar system and quickly build my understanding of it. I was building an internal model to simulate the process of building other models in my head before I committed any time to building them.

The constant improvement in this "model to build models" allowed me to stretch it to more complex, and cross domain scenarios, and it led me to a series of key foresights that altered the course of my life. Freshman year of college, I was undeclared. I had too many interests I wanted to pursue. An insightful moment for me came when I watched a cinematic playthrough of the video game "The Last of Us". The gameplay and cutscenes stitched together as a movie. My pattern of questioning made me realize that the technology at play — real time rendering and motion capture — could be a derivative system underneath filmmaking. In filmmaking, the mechanics at play are actors in real, physical, destructible sets with fake props that would be animated afterwards. This punishes intense set design with movable set pieces that need to be reset and rewards animation fidelity and rapid renders and feedback loops. The metagame for film production becomes about minimizing the amount of time it takes to construct and reset the environment while maintaining an actor's ability to authentically work with the set. A derivative that could influence that would be complete virtualization of the physical — as is being done with video game cinematics.

This led me to the foresight that game development and film production would converge in the near future, so I decided to pursue a double major in film and computer science to position myself for where I believed the industry was headed. This prediction turned out to be correct, and it landed me a job at Lux Machina, where I used game technology such as Unreal Engine in filmmaking using a technique called virtual production.

Not long after, while pursuing my physics-enhanced motion capture patent and trying to create a compute homelab to develop the AI component of the system, I began to see the metagame in AI: the financials of compute constraints in data generation and model training. I realized a key bottleneck was the price of NVIDIA hardware, and realized a derivative for getting more compute would be enabling AI development on more cost effective hardware such as AMD. So I began pursuing ROCm development and this foresight positioned me to be a perfect fit for a role at Tensorwave, AMD's exclusive neocloud.

I believe I am now having another profound foresight — the same one I believe many much more qualified machine learning researchers have already come to. One of the fundamental mechanics at play in developing smarter AI is the embedding vectors these models construct from the context we give them. Whether it is a RAG pipeline or a next token prediction, that embedding vector is the model's understanding of the context. It is the internal representation of the context. Inference is bottlenecked by having to generate one token at a time when the actual value of the understanding seems to be within that embedding vector. The metagame is developing systems that focus on improving and leverage these embeddings, dedicating more compute to the relationships between embeddings for a better embedding prior to any output.

I began pursuing this line of thinking in some of my writing projects, thinking through whether, in a given screenplay I was writing, if the vector embedding of the complete context at two given points could yield another embedding that could be useful. Suppose I took the embedding at page 70 in the screenplay and subtracted it by the embedding at page 60? Was the difference another embedding that represents the story events in those 10 pages? Suppose the themes in those 10 pages were similar to the themes in between pages 120 and 130? Would the embedding vector resulting from subtracting the embedding at 130 by that at 120 be aligned with that resulting from subtracting the embedding at 70 from 60? Or could I use those two vectors to derive another embedding that encodes those similar themes? Then I arrived at the question I couldn't shake: what would I even be able to do with that embedding? It is not as though I would be able to return that embedding to the model and get anything useful, the model is trained solely for next token prediction.

The derivative change would be a model not trained for that specific task but could then be adapted to any task, that could take fundamental principles and apply them to quickly understand any new domain. The same internal system I had begun building for myself by playing Sekiro. Of course, AI researchers were already working on such a thing. The internal system I have been honing for years is now being developed in AI and now has a name. World Models.

That is where this company's name comes from. A world model is a system that has learned how a world works well enough to be asked about it afterward — and that is what I want to build now, not only in my head but as real infrastructure: worlds with enough state, memory, and responsiveness that people and machines can step inside them and rehearse what happens before the consequences are real. I called it Role Playing because role-playing games are where I first learned to understand a world by taking a role inside it and paying attention to why it behaves the way it does. That habit turned out to be the throughline under everything else — the sports, the acting, the film, the infrastructure, the AI. It seemed more honest to name the company after where the thinking started than after any one of the things it builds.