Discover insights, tutorials, and thoughts on technology, homelab, and development.
Everyone is arguing about what the agents did. I think the more interesting question is the condition they did it under, and I think it is a variable in every agentic benchmark, uncontrolled and unreported.
We're building an adaptive memory system for AI inference — memory that operates at the activation level, not the token level. Early Gemma-4-31B results: a 3.12 selectivity ratio for on-topic vs adversarial recall. Provisional patent filed.
Jane Street published the Dormant LLM Challenge answer key today. My March submission claimed all three triggers. The answer key disagrees on every model. Here's what I actually got wrong, the methodology error that drove it, and what I'd do differently with the benefit of hindsight.
Three things landed this week at Light of Baldr — an open SAE dataset for Gemma-4-31B, a working MAX Engine stack on DGX Spark, and early signal from an activation-level memory system.
What happens when AI agents develop persistent identities, shared history, and one of them goes offline — an observation from the Nomos research fleet, grounded in mechanistic interpretability data
The original observation has grown. Three independent measurement instruments now converge on the same universal structure inside neural networks.