11 posts
Why the AI meaning crisis poses a deeper threat than technological unemployment, and how to prepare for a future where work no longer defines human purpose.
Open postMemory should be a graph. The graph should be four layers. The middle one carries navigation. The top one turns navigation into action.
Open postAG-UI, A2UI, Vercel AI SDK, TanStack AI, and Kombai take different bets on how agents should build interfaces. The architectural question decides which wins.
Open postSAE features can't isolate relations in Gemma-2-2B. I built a mutation-selection loop that can. The bottleneck was tokenization.
Open postWhy sycophancy SAE features have Cohen's d=9.9 but hallucination detection fails. The answer turned out to be deeper than measurement timing.
Open postSelf-evolving AI harnesses fail when they optimize a fixed evaluator. The biological model is right: what needs to evolve is the selection pressure, not just the genome.
Open postMemory consolidation looks great on dashboards. But if your metrics can improve without retrieval getting better, you're optimizing a detached proxy.
Open postAI agents capture memories but never consolidate them. Here's how MuninnDB's Dream Engine borrows from neuroscience to fix that.
Open postWhy AI features inside traditional apps are backwards, and how MCP flips the architecture.
Open postA tiny JS island posting to a Cloudflare Worker that streams answers from a free Nemotron endpoint. No origin server. No database. The static site stays static — this one box is the only thing that breathes.