gavinj.net

From 24GB Blackwell workstation cards to one 128GB DGX Spark, then two Sparks and a QSFP cable: running DeepSeek-V4-Flash-Vision-Exp locally, understanding vLLM, and planning the next step with a 200Gb/s switch.

Meta's new Muse agent is built around a curated long-term memory, and every major AI vendor is racing to solve the same problem. Meanwhile my own assistants still get sudden dementia every few days. A look at how agent memory actually works, from Hermes internals to ChatGPT, Claude, Gemini, Apple and xAI's Grok Bot.

A DGX Spark with 128GB of unified memory let me explore much larger local models. Then I discovered Heretic variants, and a hardware experiment became a rather uncomfortable lesson in how AI refusals work, how easily they can be changed, and what that means for the rest of us.

My AI assistant Bob got a brain transplant this week. Moonshot AI's Kimi K3 is a 2.8-trillion-parameter open-weight model that has the American frontier labs looking nervous - and the podcasters calling it a Sputnik moment.

I wanted to know whether a sub-$3k Australian workstation GPU could run useful modern local LLMs. The short answer is yes, if you pick the right model, the right quant, and build around the 24GB VRAM limit instead of pretending it is a tiny cloud.