Why your 'fast' allocator might be 62% slower (and the graphs to prove it)
AI-rewritten narrative + original data. jemalloc’s MADV_DONTNEED strategy triggers aggressive page returns during large-message Tokio MPSC benchmarks.
Posts
RTX 5090 power scaling: is max wattage just a space heater?
Rewritten through Gemma-4 31B with a conversational editing prompt.
Allocator shootout: why your 'fast' allocator might be 62% slower
Rewritten through Gemma-4 31B with a conversational editing prompt.
RTX 5090 power scaling: 450W vs 575W training
RTX 5090 power scaling from 400W to 600W on a personal workstation. Lower TDP saves ~€34/year at 80% idle and reduces sustained thermal stress in a residential build. 475W–500W is the practical sweet spot between speed and peace of mind.
RTX 5090 power scaling: 450W vs 575W training
RTX 5090 power scaling from 400W to 600W on a personal workstation. Lower TDP saves ~€34/year at 80% idle and reduces sustained thermal stress in a residential build. 475W–500W is the practical sweet spot between speed and peace of mind.
How I use agents to write this blog
Rewritten with v3 conversational editing prompt.
How I use agents to write this blog
The actual workflow: from idea to published finding through a loop of playgrounds, benchmarks, and iterative drafts.
Soft distillation vs. gradient boosting on fraud
Rewritten with v3 conversational editing prompt.
Soft distillation vs. gradient boosting on fraud
We benchmarked 52 method variants across 22 fraud and non-fraud configs. On hard fraud data, every gradient booster crushes TabPFN/TabICL by 15–20 AUC points while being 4–7× faster. Soft distillation helps only at medium scale. Teacher-as-feature is catastrophic. We quantify effect sizes with Cohen’s d and show why production fraud teams should think twice about foundation models.
Replicating Talking Trees: LLMs for fraud detection
We replicate the Talking Trees method (Yandex Research, 2025) on fraud-detection datasets using Kimi K2.6 and GPT-5.5. The LLM-guided tree beats sklearn by +0.04 AUC but is crushed by XGBoost (+0.11 AUC) at 1000× the cost. Kimi achieves higher peak accuracy but falls back 40% of the time; GPT-5.5 is more reliable (7% fallback) but slightly weaker.