Formal, reproducible benchmarks from the Éthiqueia research program. Every claim has a one-command repro.https://yandesbiens.com/enhttps://yandesbiens.com/blog/forgelm-4090-cost/https://yandesbiens.com/blog/forgelm-4090-cost/A measured training envelope (VRAM, throughput, tokens/day) for 30M-500M models on a single 24GB RTX 4090, plus the real 120M loss curve.Mon, 29 Jun 2026 00:00:00 GMTtrainingtrainingbenchmarkhttps://yandesbiens.com/blog/ufm-benchmark/https://yandesbiens.com/blog/ufm-benchmark/UFM lets a single RTX 4090 run a routed model whose expert bank doesn't fit in VRAM. Here's the benchmark, the code, and the regime where it doesn't help.Sat, 27 Jun 2026 00:00:00 GMTmemoryufmbenchmarklocal-firstmemorymoertx4090https://yandesbiens.com/blog/fmm-benchmark/https://yandesbiens.com/blog/fmm-benchmark/FMM's topic-scoped retrieval is ~164× faster and ~3.7× more accurate than a flat scan at 128k items — and useless if you misroute the scope. The locality bet, in semantic memory.Sat, 27 Jun 2026 00:00:00 GMTmemoryfmmbenchmarkmemoryretrievalraglocal-firsthttps://yandesbiens.com/blog/fmm-router/https://yandesbiens.com/blog/fmm-router/A near-free centroid router recovers 98% of the oracle's recall at ~60× flat-scan speed — when the memory is separable. As topics overlap, routing (not retrieval) becomes the bottleneck and the win evaporates.Sat, 27 Jun 2026 00:00:00 GMTmemoryfmmbenchmarkmemoryretrievalroutingraglocal-first