Show HN: Building AGI as a systems problem, not a UI wrapper

1 point

4 months ago

I am the Founder and CEO of AGI Systems Directorate. As I mentioned in my previous post, I’m a builder, not a researcher. I’m tired of the "good enough" era of AI where we settle for chatbots that hallucinate and forget context. I wanted a system that actually functions as an external frontal lobe.

I’m currently building AETERNA_VOX_OMNI. It’s not a chatbot; it’s a sovereign executive assistant designed to solve the "Human Operating System" bugs—specifically executive dysfunction.

The architecture is a hybrid neuro-symbolic system that ingests 136 modalities in real-time. We’re talking about everything from BLE bio-signals (heart rate, cortisol estimates) to digital exhaust (keystroke dynamics, network density via scapy) and environmental lux/dB levels.

Technical Specs: - Stack: PyTorch for the core brain, Polars/Vaex for high-throughput data, and FastAPI/WebSockets for the 50ms latency loops. - Architecture: A 4-layer stack featuring Sensory Projection, a Sigmoid-gated "Frontal Lobe" executive filter, and Multi-Head Attention for relational processing (e.g., correlating biometrics with text sentiment). - Learning: We’re using a form of online Hebbian learning ($Delta w_{ij} = \eta \cdot x_i y_j$) to allow synaptic plasticity during inference. - Complexity: $O(N^2 \cdot D)$ for the attention mechanism, targeting a real-time ingestion rate of 136 float32 values every 100ms.

I’m moving away from the "safe" wrappers. This is a builder’s attempt at fixing the plumbing of true intelligence. I’d love to get this community's take on the data ingestion bottlenecks—specifically handling the entropy of 136 asynchronous streams without the system melting down.

Is anyone else working on "online learning" systems that update weights during inference, or are we all still stuck in the "static weights + RAG" paradigm?