📰 Dev.to · jidonglab

Claude Code subagents cost me 48% of my bill: each one re-reads ~51K tokens of context on every request. The autopsy and the rules I set.

Why temperature 0 isn't deterministic in LLM inference: batch size changes floating-point reduction order, and greedy decoding amplifies it.

Whisper hallucinates on silence, and no_speech_threshold won't stop it. 400 test clips, 6 configs, and the VAD fix that cut phantoms 93%.

all-MiniLM-L6-v2 truncates at 256 tokens without a warning. Here's how embedding truncation breaks RAG retrieval and how to fix it.

Docker Compose depends_on doesn't wait for your database. Here's the healthcheck config that actually works, and why pg_isready lies.

The invert a binary tree interview question is 5 lines of code. Here's what interviewers actually score and why good devs fail it.

Six interview horror stories from the interviewer's chair, and the scoring mechanism that turns one small red flag into a rejection.

Context rot is why your AI coding agent gets worse in long sessions. What actually breaks past turn 50, and how to keep it sharp.

What really happens on the other side of developer salary negotiation: how leveling, comp bands, and approvals decide your offer before you talk money.

Vibe coding works fine until something breaks. Why AI agents fail at debugging, and 5 rules to escape the fix-it loop.

An interviewer's breakdown of why strong engineers fail coding interviews, what the scorecard really measures, and how the debrief decides.

Why Claude prompt caching silently misses in agentic tool loops: the 20-block lookback window, the three invalidation tiers, and the fix.

Why acc and acc_norm disagree in lm-eval-harness: log-likelihood length bias, byte-length normalization, and when each multiple-choice metric lies.

In constrained decoding, JSON Schema field order is a hard constraint, not a hint: putting answer before reasoning costs real accuracy.

INT4 weight-only quantization cuts decode latency but leaves prefill untouched. A roofline breakdown of why, and what to run instead.

STT event timings lie about when you stopped talking. Why I measure voice-interview silence from the waveform with WebAudio instead.

Notes from building a browser-based voice mock interviewer: why STT lies about silence, how a turn engine decides to probe or hand off, and defending against th

How the MoE capacity factor and router load imbalance silently drop tokens in Mixture-of-Experts models, and how to detect and fix it.

Attention sinks explain why sliding-window KV caches collapse when you drop the first tokens, and how to fix streaming LLMs cheaply.

Chunked prefill in vLLM explains why a single 32K-token prompt stalls every in-flight generation — and how to tune the TTFT vs inter-token latency trade-off.

OpenAI shipped gpt-image-2 inside Codex with O-series reasoning, 4K renders, and 16 reference images. Here is the actual workflow I rebuilt around it.

OpenAI is shipping Spud (language), Duct Tape (image), Codex, and soon API — all at once. Here's the super-app thesis and the lock-in trade-off.

140,000 stars. 850 contributors. 11,000 commits. 6.5 million developers using it every month. Zero...

A solo-operated AI news platform with a daily 8am cron, Resend-powered individual email sends, and a 49-post bilingual pipeline. Built in a weekend, still runni