[ TECHNICAL DATA SHEET — INDEPENDENT RESEARCHER / FOUNDER ]
Faris
Allafi.
Building AI and blockchain systems from first principles — parallel generation with diffusion and Mamba, custom consensus, and compute that earns its keep.
AI ARCHITECTURE · CONSENSUS SYSTEMS · LOW-LEVEL PERFORMANCE · NON-AUTOREGRESSIVE GENERATION
01 GENERAL DESCRIPTION
FARIS ALLAFI is an independent researcher and founder focused on AI architecture, systems design, and blockchain infrastructure. The device prefers problems that are actually hard, not problems dressed up to look impressive in a demo.
Primary workload, DIMBA, challenges token-by-token generation with a parallel, diffusion-based approach built on Mamba-2. Secondary workload, GHOST, is a custom Layer 1 built on Substrate using hybrid PoW/PoS consensus design (see §04).
Additional interests include vintage technology, low-level performance work, and models that respect compute instead of wasting it. The device is 13 years old and cares more about systems that hold up than headlines that don't.
03 APPLICATIONS
- Faster language generation without losing control
- Consensus mechanisms that deserve to exist
- Research that survives contact with hardware
- Weird architecture ideas that should not work but might
02 FEATURES
- Non-autoregressive language modelling (diffusion + Mamba-2)
- Custom L1 consensus: hybrid PoW/PoS, Entropy-Steered
- Full-stack: Rust / Python / Substrate / PyTorch / JAX
- First-principles reasoning, factory-calibrated
- Zero hype emission under normal operating conditions
- Compiler construction experience (see §4.3)
04 SYSTEM OVERVIEW
4.1 — DIMBA / PARALLEL SEQUENCE GENERATION
A non-autoregressive architecture combining diffusion with Mamba-2 for parallel sequence generation. Lower latency without turning quality into collateral damage.
| PARAMETER | SYM | VALUE |
|---|---|---|
| Architecture | ARCH | Diffusion + Mamba-2 |
| Generation mode | GEN | Parallel, non-autoregressive |
| Quality / latency trade-off | Q/t | Controllable |
| Implementation | — | PyTorch |
| Source | SRC | dimba-lib-exp ↗ |
4.2 — GHOST / CUSTOM LAYER 1 BLOCKCHAIN
A custom Layer 1 built with Substrate, exploring hybrid PoW/PoS consensus through original Entropy-Steered Consensus work. Practical, opinionated, and very not a toy.
| PARAMETER | SYM | VALUE |
|---|---|---|
| Framework | FW | Substrate |
| Consensus | CONS | Hybrid PoW/PoS, ESC |
| Implementation | — | Rust |
| Toy factor | TOY | 0 |
| Source | SRC | ghost-blockchain ↗ |
4.3 — AUXILIARY COMPONENTS
| DEVICE | FUNCTION | REFERENCE |
|---|---|---|
| DIMBALABS | Company building the next generation of non-autoregressive language models | dimbalabs.xyz ↗ |
| HF HUB | Model weights, checkpoints, and open experiments | huggingface.co/devnull37 ↗ |
| POTATOLANG | Transpiled language; taught the value of building the whole stack, not just the easy parts | potatolang ↗ |
05 ABSOLUTE MAXIMUM RATINGS
| PARAMETER | RATING | UNIT |
|---|---|---|
| Operating age | 13 | yr |
| Hype tolerance | 0 | % |
| Buzzword input voltage | ±0.2 | V |
| Compute wasted per token | → 0 | FLOPs |
| Motivational-wallpaper exposure | not rated | — |
NOTE 1: Stresses beyond those listed above may cause permanent damage to the conversation. Exposure to absolute-maximum-rated conditions for extended periods may affect device reliability.
06 OPERATING PRINCIPLES
-
6.1 Architecture first.
Start with the system shape. If the core idea is weak, no amount of polish saves it.
-
6.2 Ship against reality.
Research matters more when it survives hardware limits, latency budgets, and real users.
-
6.3 Depth over decoration.
Good work is usually fewer ideas, pushed further. That applies to models and to websites.
07 PACKAGE & CONNECTIONS
If the problem is interesting, the device is probably interested. AI, systems, blockchains, or anything that sounds like it should not work but might. Connect via any active pin.
FIG. 2 — PACKAGE: DIP-8 (PLASTIC). PIN 7 (EMAIL) IS THE RECOMMENDED INPUT. PIN 8 IS NOT USER-SERVICEABLE.