Faris Al-Lafi
I’m Faris Al-Lafi, an independent AI researcher and the founder of Hamiltonian Research in Abu Dhabi.
I design language model architectures beyond the transformer. Right now, that means DIMBA: a language model that combines bidirectional Mamba-2 state-space models with diffusion to refine responses in parallel.
I’m especially interested in making these architectures fast in practice: fusing GPU kernels, keeping inference state bounded, and pushing toward generation throughput of thousands of tokens per second.
When I’m not profiling models, I’m usually reading papers, working through books, or recovering from the psychological damage of middle school algebra.
What I’m working on
Building what comes after attention.
Independent AI research · State-space models · Diffusion
The transformer is an important architecture. It is not the final one.
Hamiltonian Research studies models that replace quadratic attention and token-by-token generation with state-space dynamics and parallel refinement. The aim is practical: stronger computational scaling, bounded inference state, and more control over the tradeoff between generation time and quality.
DIMBA
Diffusion on a Mamba backbone.
A language model that refines a response in parallel instead of predicting it one token at a time.
DIMBA combines masked diffusion with a bidirectional Mamba-2 backbone. At each step, the model reads the incomplete sequence, predicts the missing tokens, commits the most confident ones, and repeats until the response is complete.
Writing
Work in public, including the failures.
I trained a language model that thinks the capital of Japan is Paris
The first DIMBA run, what failed, what worked, and what the results imply for the architecture.
READ ↗Get in touch
Research, systems, and difficult problems.
DIP-8 CONTACT MAP