yan@yandesbiens:~$ ls ~/blog/ -lt

• 2026-07-11 Your LLM server lies to you when the context window overflows Ollama silently drops the front of an oversized prompt and returns 200 OK, so the model confidently invents an answer. llama.cpp is honest and returns a 400 — which kills the agent turn instead. I measured both, then built ContextPaw to make overflow honest and survivable. Includes a result I did not expect: telling an agent it lost information does not stop it hallucinating.
• 2026-07-11 I ran an autonomous cognitive agent for months. It never learned anything. It just became certain. A negative result, measured. An idle 'reasoning' loop ran 27,834 times over months and converged on unfalsifiable certainty: 606 beliefs strengthened per cycle with zero contradictions, 882,602 pieces of self-generated evidence of which none disagreed, and an audit log that recorded 5 of 16.9 million confidence increases. The decay mechanism that should have stopped it was disarmed by the loop's own evidence writes.
• 2026-07-01 agentos devlog — from a terminal kernel to an agent I run from my phone How the agentos kernel grew into a self-hosted agent platform: a ReAct loop over 120+ models, ~26 real tools, a permission layer that keeps a human in the loop, runtime tool-forging in a sandbox, and an environment-aware system prompt. Engineering, not a benchmark.
• 2026-06-29 What it costs to train an LLM from scratch on one RTX 4090 A measured training envelope (VRAM, throughput, tokens/day) for 30M-500M models on a single 24GB RTX 4090, plus the real 120M loss curve.
• 2026-06-28 AEON devlog — a trainability bug in the liquid-neural-network brain A development log from AEON: how a failing unit test exposed a seed-dependent trainability bug in the CfC liquid network, why a missing normalization was the root cause, and how one LayerNorm made the student learn reliably across every seed.
• 2026-06-27 the dig — four years of a one-person AI lab I pointed an AI archaeologist at an old backup drive and told it to find the gems. Here's what was buried.
• 2026-06-27 proof drop #1 — running a 24 GB model on a 24 GB GPU (and the honest catch) 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.
• 2026-06-27 proof drop #2 — hierarchical memory beats flat search (when you know where to look) 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.
• 2026-06-27 proof drop #3 — a cheap router decides whether the locality bet pays 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.
• 2026-06-26 boot.log — why this site exists Local-first AI, a single 4090, and a domain that cost ten bucks.