Zuzanna Stamirowska has a line she uses to describe the state of AI today. She says a large language model is like a very smart intern on the first day of the job. It can read anything you hand it. It cannot remember what it read yesterday, and it stops learning the moment training ends.

Pathway was built to fix that and we just invested via Niya Partners.

Transformers, the architecture behind every major LLM, are static once trained. They do not generalize over time. Feed one a problem structurally different from what it saw in training and performance collapses, no matter how many parameters it has. In October, Pathway's research team published an answer: Dragon Hatchling, or BDH, a new architecture built from a scale-free network of locally interacting neuron units instead of the standard attention blocks.

The model's internal structure is not engineered into fixed layers. It emerges during training, the same way the modular structure of the brain's neocortex emerges rather than gets hand-designed. On Sudoku Extreme, a benchmark built from roughly 250,000 of the hardest puzzles in the format, BDH scores 97.4% accuracy without chain-of-thought, backtracking, or external tools. Leading LLMs barely register on the same benchmark. It also runs on ordinary GPUs and follows the same scaling laws as transformers, which means it is not a lab curiosity. It is a drop-in replacement path.

BDH sits on top of the company's original insight. Pathway's core engine feeds AI systems live, streaming data instead of static snapshots, and it already underwrites decisions at NATO, La Poste, and Formula 1 teams managing race strategy under real-time pressure. That data infrastructure is what makes BDH deployable, not just publishable. Most breakthroughs in architecture research stay in a paper. Pathway has a live enterprise pipeline sitting underneath this one from day one.

The team is the reason we underwrote this early. Stamirowska's PhD modeled how complex trade networks evolve, work recognized by the U.S. National Academy of Sciences. Co-founder and CTO Jan Chorowski worked alongside Geoffrey Hinton at Google Brain and was among the first researchers to apply attention to speech. Co-founder and CSO Adrian Kosowski earned his PhD in theoretical computer science at 20, became a tenured researcher at Inria at 23, and has authored more than 100 papers spanning graph algorithms and distributed systems. Together they built an architecture with a real theoretical foundation, not a scaled-up version of what already existed.

Reasoning models that keep learning after deployment are the next real barrier in AI, past bigger context windows and cheaper inference. Pathway is one of the few teams with a credible, published, benchmarked answer, and a paying customer base to prove it works outside a paper.

Bashar Aboudaoud
Managing Member, UpRound