Tibor Sloboda

I am voraciously curious. Astrophysics, pure mathematics, quantum chemistry, medicine: there is always another field I want to understand, and another problem I would like to help solve. That curiosity goes together with a belief that humanity is capable of far more than we have yet achieved. I want to contribute to that progress, whether the problem in front of me is large or small.

Samantha Carter from Stargate SG-1 captures something of that ambition. For anyone unfamiliar with the show, it is a science fiction series about a team exploring other worlds through an alien gateway. She is a scientist and skilled Air Force pilot whose knowledge reaches across physics, computing, engineering, and chemistry. What appeals to me is her ability to bring that knowledge together when faced with something unfamiliar. She represents an idealized breadth of human capability that I would love to approach.

A lifetime leaves little room to become a deep expert in even one field, let alone several. I have limits on my time, resources, and capacity to learn, however many subjects interest me. That is where another ambition of mine connects with Carter's example: a personal AI assistant that could extend a person's knowledge, helping a skilled researcher or decision maker understand an unfamiliar field and relate it to what they already know.

I think such tools could accelerate research and make that breadth of capability accessible to more people. Where I cannot contribute directly, I want to build software and AI that help the people who can. Supporting someone else's discovery is part of our progress as a society, and it is work I want to devote myself to.

What I am working on

My PhD research explores the mathematical nature of multimodality and modality fusion. I am interested in the relationships between different kinds of information and what makes them easy or difficult for a model to bridge. This includes exploring how difficult it is to learn a joint projected representation across an arbitrary set of modalities: a shared representation through which the model can relate information from those modalities.

Alongside my research and consulting, I am CEO of aleph0, a separate venture developing low-level AI software for heterogeneous hardware, particularly at the edge. The work includes hardware-specific compiled models, a single-binary runtime, and model optimization for compatibility and lower power use. It is in development and will be released open source when ready.

Waffu is my personal effort toward the assistant I described above: an always-on, entirely local agent I am developing to respond intelligently and autonomously, without depending on someone starting each interaction with a query. I am also exploring search augmentation for Rust codebases, prompted by my frustration with generic embedding and graph exploration tools that make too little use of the language's rich structure.

AI beyond the current conversation

My work has taken me from machine learning in chemistry and medicine to production LLM systems. I find the sudden abundance of "AI experts" frustrating, especially when AI gets reduced to LLMs. Imaging, scientific analysis, prediction, audio, and hardware offer many other applications that people could benefit from and often never hear about.

I want to help people understand what is useful for them, including when their problem has a simpler solution. I would rather tell a company that AI is not worth its time than sell it software it does not need.

Questions, experiments, and games

In my free time, I enjoy games with narrative depth, philosophical questions, strategy, puzzles, or room to experiment. The Talos Principle, Outer Wilds, Final Fantasy XIV, NieR: Automata, Factorio, and Satisfactory are among my favorites. The stories and difficult questions matter to me as much as the mechanics, and I am also curious about the engines that make these worlds work.

That interest has led to experiments with disaggregated graphics processing, moving selected stages of a game's graphics pipeline onto another GPU. I have had positive early results with DLSS upscaling, Frame Generation, Ray Reconstruction, and Neural Rendering, alongside related technologies from Intel and AMD.

The work includes external GPUs connected over Thunderbolt, where the GPU doing that processing may be different from the one driving the display. Getting the result across adapters and presented correctly is part of the challenge, including working with Microsoft's Cross-Adapter Scan-Out (CASO) on Windows. I also have configurations working on Linux through Proton. These experiments are ongoing, and I have been exploring ways to implement AutoHDR on Linux as well.

Other interests include Rust, Lean and constructive mathematics, NPUs, Linux, and audio ML. I enjoy following a question into the mathematics, software, or hardware underneath it, including the parts I did not expect to need when I started.

GitHubLinkedInMaltORCIDGoogle Scholar

Experience, education & honours ->

Publications & research ->