BBrock Claussen1d ago · 5 min readMy unpublished benchmark post said 10x. The honest number is 1.2x to 1.8x.The old draft The benchmark I'd sign today Headline "10x faster", "2x faster than hnswlib" 1.20x to 1.79x hnswlib, 1.30x to 1.43x faiss Recall 61% for Foxstash, 40% for hnswlib the same for all t1blog.narcolepticfox.com
KKuldeep SinghOct 15 · 6 min readI built desktop control so an agent could test my Chrome extension. Then I benchmarked it against four free tools.Testing my Chrome extension by hand meant opening its side panel and reloading the extension after every change. Browser tools can't reach the side panel. Codex's computer use showed me an agent could26KYar5en1c.hashnode.dev
NnRouter.aiOct 14 · 3 min readReading a Benchmark Table Honestly: TTFT Percentiles From a Warm-Pool Gateway TestVendor benchmarks deserve suspicion. Most hide the workload, cherry-pick the tier, or quote a p50 and let you assume the tail behaves. So let's read one the way it should be read — method first, numbenrouter.hashnode.dev
Ssamuel HuOct 10 · 8 min readHow Many Cameras Can an Edge AI Box Really Handle?“Eight cameras” sounds like a hardware specification. It is really a statement about a workload. Eight streams running a person detector at 5 FPS ask something different from eight streams running tha2Acosmoedge.hashnode.dev
BBrock ClaussenOct 5 · 5 min readClassify the errors before you trust the scoreOpenAI is shutting off gpt-5.4-nano on 1 April 2027. In Mousing, my document extraction platform, nano did two jobs: it classified incoming documents, and it was the verifier, the second model that reblog.narcolepticfox.com
LLevelbrook ConsultingSep 21 · 9 min readHow to read a coding-agent benchmark without getting soldA nine-author study this week pulled a coding agent apart into its components and measured each one across 176 configurations. The findings are less exciting than any vendor slide and more useful than1Blevelbrook.hashnode.dev
DDebashish GhosalSep 16 · 4 min readOne Hung API Call Used to Kill My 1,000-Run Benchmark. Here's the Fix.The experiment was fine. The runner was the bug. A single hung API call discarded hours of completed work, and I kept blaming the data. If your long-running eval keeps dying at 80%, this is the five-ppragmatic-engineer.hashnode.dev
OOliver ZehentleitnerSep 13 · 10 min readpicows in UNICORN Binance WebSocket API: Up to 2× the Throughput, Opt-In for NowUNICORN Binance WebSocket API (UBWA) has used websockets since day one. Since UBWA 2.16.0, you can switch a manager instance to picows, a Cython WebSocket implementation. For typical Binance messages 1blog.technopathy.club
MMuhammad TahirSep 11 · 14 min readBenchmarking Vector Databases: High-Throughput Semantic Search at Scale with Node.js & PythonThis article was originally published on Muhammad Tahir's Portfolio. Introduction & Industry Context In 2026, the proliferation of AI-powered applications, especially those relying on Retrieval Augmenmtdeveloper.hashnode.dev
HHarrison GuoSep 10 · 9 min readSame Model, 13.3% to 38.3%Two API settings. Same model. Same benchmark. Same task set. 13.3% to 38.3%, using one sixth the output tokens. OpenAI published that result about GPT-5.6 Sol on ARC-AGI-3, and it is the cleanest natu5Aharrisonsec.hashnode.dev
