Releases
Papers, research milestones, and model releases.
BIB — Biologically Inspired Brain Substrate (Open Source)
A 1:1 digital replica of the human cognitive pipeline built without backpropagation. Implements a predictive sparse cortex, CA3 recurrent hippocampus, endocrine systems, and an actor-critic basal ganglia.
GitHub Repo →Project Big Bang — Continuous Learning Simulation Framework
Open-source release of the 2D virtual environment and physics sandbox used to train and evolve BIB-driven digital organisms through resource foraging and predator evasion.
GitHub Repo →BILM — Applied NLP Engine
Release of the Biologically Inspired Language Model. An applied NLP engine that focuses on sequence memory and mitigates catastrophic forgetting, optimized to run on CPU.
GitHub Repo →BILM: Mitigating Catastrophic Forgetting via Sparse Distributed Representations
Working paper documenting the 16,384-dimensional SDR architecture, CPU computational efficiency, and the hierarchical predictive cortex of BILM.
Read BILM →BIO: The Biologically Inspired Organism
The realization of the UnikAI mission. Working paper detailing the continuous-learning synthetic digital organism. Covers the chemical Brainstem (Dopamine, Serotonin, Acetylcholine), Actor-Critic Basal Ganglia, and Hippocampal Sleep Consolidation.
Read BIO →BIM: The Ancestral Substrate
Working paper documenting the ancestral four-generation BIM research programme (BIM 0 - 4). Covers the chimerical phase failures and the foundational SDR-based sparse cortex.
Read BIM →BIM 3 — Unified Substrate
Five biological modules cooperating under local learning rules: deterministic byte-level sensory codec, 3-layer hierarchical predictive cortex with corrected top-down apical feedback, neuromodulator combining rolling prediction error with per-symbol habituation, synaptic homeostasis, and CA3 auto-associative hippocampus. No backpropagation. No external memory. Runs on a single CPU core under 1 GB of memory.
BIM →BIM 2 — Surprise-Scaled Plasticity
Introduced Jaccard-gated learning: the per-step learning rate is scaled by the surprise between predicted and actual SDR. Perfect predictions skip the kernel call entirely. Added second temporal-pool layer, synaptic homeostasis, and concept crystallisation from recurring SDR patterns.
BIM →BIM 1 — Sparse Distributed Representations
Replaced graph-based loops with a sparse cortex. 16,384-dimensional SDRs with exactly 64 active bits (0.39% sparsity). Numba JIT-compiled kernels: 160 tokens per second on a single CPU core. 100% one-shot recall on 20-token sequences.
BIM →MAUL: Specialisation Beats Generalisation — A Multi-Agent Routing Architecture
Working paper on the MAUL routing framework. Documents the core hypothesis that a system of specialised agents, each operating in its domain of highest competence and coordinated by a semantic dispatcher, outperforms a single generalist model. Covers semantic understanding phase, calculator and search routing paths, specialist agents, synthesis layer, and benchmark results.
Read →Hydrogen-2 — Current Stable Model
Second-generation model evaluation configuration. Paired with MAUL-2 routing: math through a deterministic calculator, code through a sandboxed executor. 95.90% on GSM8K. 97.56% on HumanEval. Benchmarked across GSM8K, HumanEval, MMLU, AIME 2025, and IFeval.
MAUL →MAUL-2 — Semantic Routing
Introduces a global semantic understanding phase before any routing decision. Deterministic calculator tool for math. Structured task understanding layer. 95.90% on GSM8K, 97.56% on HumanEval when paired with Hydrogen-2.
MAUL →BIM 0 — Graph-Based Association Loops
Seed experiment. A directed graph with Hebbian edge strengthening. Proved that a blank-slate system can acquire associative structure from interaction alone. Foundation for the BIM programme.
BIM →Hydrogen-1 — First Evaluation Configuration
First-generation model evaluation framework. Established the benchmarking pipeline across GSM8K, HumanEval, MMLU, IFeval, and AIME 2025. Laid the infrastructure for systematic evaluation across model versions.
Releases →MAUL-1 — Multi-Agent Foundation
First version of MAUL. Intelligent routing with semantic understanding. Specialist agents for coding, creative, logic, and chat. DuckDuckGo and Wikipedia tool integration. Gemini API backend.
MAUL →