AGISystem2

Research Notes

Exploring the frontier of efficient, reliable, and verifiable AI.

The research landscape in this field is highly diverse, with numerous powerful ideas emerging. However, there is a critical shortage of resources needed to combine and scale these directions into unified, industry-grade systems.

Beyond the GPU Barrier

The architectural shift from GPU-centric to CPU-first AI systems.

The Algorithmic Revolution

Probabilistic sparsity, SLIDE, and the BOLT engine for sub-linear deep learning.

High-Fidelity Runtimes

Computational sovereignty and running BF16 full-precision on CPUs with Ziroh Labs.

Quantization & Democratization

The impact of llama.cpp, GGML, and GGUF on local AI accessibility.

Rust for Systems ML

Memory-safe, high-performance frameworks like Candle and Burn.

Silicon Giant Optimizations

How Intel and AMD are baking AI acceleration (AMX, ZenDNN) into the processor.

Efficient ML Trends

A survey of modern inference runtimes, compilers, and low-level libraries.

VSA & HDC

Algebraic reasoning and neuro-symbolic computing in high-dimensional spaces.

BitNet & 1.58-bit LLMs

The radical shift to ternary weights and the elimination of multiplication in AI.

CNL & Formal Semantics

Bridging natural language ambiguity and formal logic for auditable agent reasoning.

AI-Native Languages

Rethinking programming paradigms for probabilistic, agentic, and self-optimizing systems.

Brain-Inspired Computing

Jeff Hawkins' Thousand Brains Theory and Numenta's path to biological intelligence.

Liquid Neural Networks

MIT CSAIL and Liquid AI's continuous-time systems for adaptive and efficient intelligence.

Neuromorphic Computing

Event-driven AI and Spiking Neural Networks (SNNs) for extreme energy efficiency.

Active Inference

Karl Friston's Free Energy Principle and the mathematical foundation for proactive agents.

Energy-Based Models

Yann LeCun's alternative to probabilistic modeling through compatibility functions.

World Models & JEPA

Predictive architectures for autonomous agents that learn, reason, and plan in latent spaces.

KAN Networks

A radical alternative to MLPs using learnable activation functions on weights for better interpretability.

Neuro-Symbolic AI

The convergence of connectionist pattern recognition and symbolic logical reasoning.

Conceptual Spaces

Geometric information representation bridging the gap between neurons and symbols.

Ontologies & KG

Structured knowledge and common-sense reasoning via Cyc, WordNet, and SUMO.

Formal Logic & Solvers

Deterministic reasoning using SMT solvers like Z3 and description logics.

Theorem Proving

Integrating AI with formal proof assistants like Lean, Coq, and Isabelle.

Verification Languages

Proving system correctness using TLA+, Alloy, and formal methods.

Causal Inference

Moving beyond correlation to causal reasoning and counterfactuals with Judea Pearl's theories.

Cognitive Architectures

Unified models of the mind like SOAR and ACT-R for integrated intelligence.

Probabilistic Programming

Codifying uncertainty and Bayesian inference using Stan, Pyro, and Edward.

Geometric Deep Learning

Unifying AI architectures through symmetry, invariance, and graph-based representations.

Physics-Informed AI

Integrating physical laws as hard constraints into neural network training (PINNs).

Program Synthesis

Automatically constructing provably correct code from high-level specifications.

Differentiable Logic

Merging logical operations with gradient descent via Logical Neural Networks (LNN).

Explainable AI (XAI)

Creating trust through transparency, auditable reasoning, and interpretability.

Verifiable Computing

Cryptographic proofs (ZKPs) and zkML to ensure AI computation integrity and privacy.

Small Language Models

Model distillation and specialized 3B-parameter "reasoning kernels" for edge deployment.

Decentralized AI (DeAI)

Removing centralized gatekeepers via Bittensor, Petals, and federated learning protocols.

Formal AI Safety

Moving beyond RLHF to mathematical alignment guarantees and safety shielding.

Multi-Agent Systems

Formal coordination, consensus protocols, and epistemic redundancy in agent ecosystems.

Information Theory

Intelligence as compression: Kolmogorov complexity and the bedrocks of universal induction.

European Initiatives

Key research groups and projects in Switzerland and Europe focused on efficient and neuro-symbolic AI.