Who is building Logos

Logos began with a question that would not let go: given a set of data points, why is there no way to find a mathematical formula, over any number of variables, that fits them?

My name is Thobias Melfjord Knudsen. I am a systems programmer studying informatics, and that question caught me in my first year of high school. Python came first, in 2020; by 2022 a math application built to chase the question was underway, with C++ learned along the way. A working version took about six months.

The chase explained why no such tool exists. Through any finite set of points endlessly many curves can be drawn, so there is no single formula waiting to be found. The most a tool can do is decide in advance what shape of formula it will accept, then search for one of that shape that fits the points, and even then it may find nothing, or infinitely many. What the problem really needs is a language where formulas are as easy to build and reshape as numbers, and where the language can look at and rewrite its own expressions: functions that write other functions, shaped by whatever they are given. Lisp came closest, treating code as data, but it still falls short of what the problem demands.

The decisive turn was seeing that this generalizes to almost everything logical. A memory system for AI agents, built later, hit the same wall from a completely separate direction: there too, the real limit was the language. If English could be made programmable, and Logos is built to host exactly that, a memory system as good as our own memory, or better, comes within reach. Two separate roads ended in the same place.

The bottleneck was never the mathematics. It was the language.

The evidence it can be built

A project this size stands or falls on whether its builder finishes hard things, so here is the record. LogosMath is where it began: a working math application with its own small language, built to go further than symbolic tools like Wolfram Alpha and Matlab. Memra is the agent memory system from the second road; in my benchmarks it came close to the best available. Military service produced a high-resolution offline map in C++, and a recent Zig hash table runs almost as fast as the fastest published.

At NTNU, in the Algorithms and Data Structures course (one of the university's hardest, around 900 students), my solutions were the fastest in most of the weekly challenges through the autumn of 2025. And this year, my teammate and I placed first in Norway's first national championship in AI, out of more than 1,100 teams; the solution is on GitHub.

Logos itself is the turn from the math application to the language underneath it: one language for everything, built on a single commitment, radical unification. The program, its types, its proofs, the compiler, and the grammar itself all live in one structure. It is a serious systems language, with a borrow checker, native compilation, and no garbage collector, reaching also for machine-checked proofs and self-reflection. It does not run yet; a small Rust bootstrap seed is all there is so far, and the roadmap tracks it honestly.

Sometimes I suspect that a complete meta-language, where each word is defined using all other words, is the closest one can get to reflecting on how God works.

If it interests you, you are welcome to follow along on GitHub: star the seed, watch the language take shape, and word of the first build will come there.