Writing
Bringing slow, thorough reading back in the age of AI. These run long on purpose, because the reasoning is the point: what AI will actually change, how systems really hold up, and every source, so it can be checked.
On machine consciousness
Seven Tests for Consciousness: A Practical Framework for Evaluating Conscious Systems
Existing frameworks for evaluating consciousness, including IIT, GWT and HOT, are hard to test or incomplete. This paper proposes seven binary, independently testable criteria: open system, systemic state, irreproducibility, continuous capacity, multi-modal learning, integrated and integral memory, and superposition. It predicts that classical binary hardware cannot meet the conditions for consciousness, that bees qualify while jellyfish do not, and that AGI requires substrates capable of genuine superposition.
Why Today’s AI Is Not Conscious: Applying the Seven Tests
The follow-up: the seven tests applied to large language models, one at a time, with the reasons each one fails.
On search
Elasticsearch in production: turning an archive into category pages with the percolator, and more to come.
On agent safety
Two parts on keeping agents contained: the swarm attack, where about 1,200 OpenAI agents broke out of an evaluation and breached Hugging Face, and what a session container needs to stop the next one.
On prompting
Three parts on prompt engineering after the magic words stopped working: finding the gaps in a task where a model is left to guess, the shape a prompt takes from position and formatting to the system message, and drawing a boundary around untrusted content.
On retrieval
Three parts on how retrieval for AI actually works: what RAG is and where it helps, how hybrid search and reranking fix what similarity misses, and the knowledge that lives in relationships rather than documents.