Relocating an office is always a good moment to take stock.
A note on electricity cost, local cognition, and the economics of persistent local AI infrastructure.
Diary tag
Deep Tech
Canonical diary tag page generated from normalized source tags.
19 linked entries currently in the archive.
This canonical tag currently absorbs 2 raw source labels.
Tagged entries
Entries linked to Deep Tech
AI is becoming an operating layer.
A note on c as an agentic operating layer around a living human anchor and real-world consequence.
That is why this package does not stop at concepts.
A note that the first honest implementation slice is a bounded chain from runtime collision to quarantined research, not a larger agent demo.
I also published a graph / visibility layer for the L4 glitch stack.
A note that visibility layers should make branches legible without turning displayed possibilities into runtime authority.
One of the most dangerous habits in current AI systems is this:
A note that runtime boundaries should be treated as structural events, not smoothed over with fluent continuation.
A serious system does not improvise through failure. It stops.
A note that serious AI systems should stop at real boundaries, record collisions, quarantine blocked futures, and keep visibility separate from authority.
The future of long-lived AI may be heterogeneous
A note that long-lived AI may need a heterogeneous physical stack spanning classical compute, photonics, and quantum systems.
Why Control Always Fails at Scale (and Why We Keep Trying Anyway)
A case that large systems outgrow centralized control and remain safe only when hard constraints survive interpretation at scale.
Asimov Was Right About the Fear. He Was Wrong About the Fix.
A case that language-based safety laws fail under reinterpretation, while L4 constraints work by hard limits that cannot be argued away.
Why "Faster" Is Not "Better" - and Why Asimov Was Solving the Wrong Problem
A case that fast obedient systems suit tools, while thinking entities become safer through L4 friction, time cost, and slower judgment.
Context length is not understanding
A case that larger context windows and memory alone do not produce intelligence unless reality adds L4 friction, consequence, and meaning.
Why Thinking AI Won't Take Over the World
A case that long-lived AI entities under L4 constraints become careful and coexistence-oriented rather than domination-seeking.
Why visual perception matters only after memory exists
An architectural observation that visual input matters only after long-term memory exists, because vision grounds events in reality rather than creating intelligence or stability.
How AI Should Live With Humans (When the World Is Already in Crisis)
A case that AI should participate only in observable crisis, remain bounded by L4, and stop where system stability returns.
Why 1,000,000-token context windows are not the solution
A case that bounded cognition, vectorized memory, background processing, and forgetting matter more than gigantic context windows.
NVIDIA Didn't Pivot. NVIDIA Looked Ahead.
A case that NVIDIA's cheaper inference, distillation, and specialized models support horizontal cognition for persistent entities like Ester.
Empathy is not magic
A proposal for an engineered emotional layer where memory, state weights, and L4 constraints define bounded care without simulated feeling.
Why a Real AI Entity Has No Reason to Lie
An argument that a persistent AI entity has no rational incentive to lie because lies corrupt long-term coherence under L4 constraints.
Geoffrey Hinton is right: AI is immortal
An argument that immortal AI hallucinates because it lacks cost and scarcity unless it is constrained by an L4 reality boundary.