$ ls -la ./posts/

Continuing the work from Hacking Granite-4.0-Hybrid, this write-up talks through the process of creating and conditioning the goal state to apply to a hybrid transformer-ssm model to steer model outputs in pusuit of a viable test-time fine-tuning strategy.

As a first step toward goal-conditioned latent steering (part of my research into GSSR), I hacked on IBM's Granite-4.0 hybrid model (Mamba-2 + attention), located the recurrent SSM states in the late layers, and perturbed them mid-generation with a simple random nudge. The output changed immediately from clean answers to repetition lock-in or foreign-code gibberish at higher strength.

This article examines the evolution of the cost of living and lifestyle necessities from the early 1980s to the mid-2020s. It confronts the central premise that achieving a comparable middle-class standard of living has grown more challenging due to structural economic shifts, including wage stagnation, globalization, technological change, and the rising cost of essential services.

GSSR is a conceptual framework that blends ideas from state-space modeling, control theory, reinforcement learning and active inference to create a system that doesn't just predict the future, it steers it's own internal model of it.

A project outline to better understand multi-hop mesh networks and potentional applications.