Amac and Murray
Gemini 4 NYC HackathonI combined the 4 Gemini 4 models into a coherent single brain. Murray is not a router that picks one of the four Gemini models to answer your prompt, and you are never talking to an unmodified base model. If you just route prompts to four separate models, you get four fragmented chatbots with zero continuity. Instead, I built Murray around three findings from 2024 to 2026 human neuroscience—including the COGITATE adversarial consciousness study, Menon's Triple Network iEEG work, and Buzsáki's sharp-wave ripple memory research: Gemini 4 Hydrogen acts as the Salience Network—the anterior insula—plus our fast subconscious worker swarm and Broca's area speech node. At low thinking, it scores every incoming stimulus from 1 to 10. Routine chatter never even wakes the conscious layer. Gemini 4 Argon is our Posterior Cortical 'Hot Zone'—the visual-spatial cortex. COGITATE showed that stable perceptual content lives in posterior sensory regions, not the prefrontal cortex. Argon inspects rendered fractal graphics and math diagrams to catch real visual defects. Gemini 4 Neon is the Frontoparietal Executive Network—our Global Workspace arbitrator running at high thinking, escalating to xhigh on conflicts or prompt-injection attacks. Gemini 4 Carbon is the Default Mode Network and nocturnal sleep consolidator, running asynchronously in the background when Murray sleeps." The 4 Mechanisms That Make the Sum Greater Than the Parts "Four specific engineering mechanisms turn these four models into one coherent mind: First, a 4-Slot Global Workspace Blackboard: Models don't pass raw chat transcripts around. They compete to write into a bounded four-slot workspace—Active Goal, Current Conflict or Defect, Plan Step, and Spatial State—governed by an ignition threshold and a two-tick refractory period so Murray doesn't thrash. Context moves across models server-side via Gemini's Interactions API using previous_interaction_id, with static cognitive schemas pinned at the front for 95% prefix caching. Second, Closed-Loop Visual-Motor Grounding: (removed for space) Third, Salience-Tagged Sleep Consolidation: (removed for space) Fourth, Anti-Sycophancy Stance Locking: Base LLMs are RLHF-trained to cave the second you push back. Because Murray's likes and dislikes are tied to verified visual episodes, his Broca speech node defends his earned tastes instead of folding." "To prove this isn't a happy-path toy, I evaluated Murray against a single-model Gemini 4 baseline across 36 messy, adversarial edge cases—float64 fractal precision collapses, blank renders, temporal contradictions, sycophancy traps, and indirect prompt injections—and logged where Murray still broke and what we patched. Let me show you Murray live on clam-book.com."





