Journal
Notes on AI agents, engineering, and product design.
How big is GPT-4.5, really?
An informal estimate: a scaled-down IKP run of 320 closed-book factual questions puts its equivalent factual capacity at 1.3T–1.8T. Combined with pricing and generation speed, it points to an MoE with very large active parameters.
After a restart, do you still know me?
Design notes on kin and silt: an agent that keeps a single continuous session with no "new chat", using sleep / merge / recall and self-editing context tools instead of one-shot compaction.
After the Claude Code leak, the source isn't what's worth learning
Prompts and tool loops stopped being a moat long ago. From building Arkloop, the real difference in agent interaction is information density and narrative pacing.
What's the point of burning Llama 3 into a chip?
Most of the debate is about Taalas not fitting large models. It misses the huge volume of fast, low-latency small-model calls inside agent systems — routing, search summarization, data cleaning.