Wednesday, 4 November 2026
18 episodes · 6596 in all
#6458: Why Synthetic Fine-Tuning Data Needs Tooling
Rewriting high-quality data synthetically gives a 3.4-point gain, but running the same process on low-quality data yields just 0.5.
#6457: Where to Learn LLM Training Without Generic Courses
A GPT-2 scale model that took OpenAI 168 hours to train in 2019 can now be replicated on a leaderboard in under two hours.
#6456: The Unsung Builders of Modern Speech AI
OpenAI's Whisper lists famous co-founders among its six authors, yet shipped with no launch event, demo day, or canonical interview.
#6455: Turning Open-Source Triage into Subscription Credit
No software vendor currently applies peer-reviewed contribution credits directly against a customer's subscription renewal invoice.
#6454: Choosing Between MCP Servers, Agent Skills, and CLIs
A controlled Arize evaluation found that bare Claude without an MCP or skill scored higher on correctness than curated tool stacks.
#6453: Who Actually Won the Open Weight Race
With over four billion downloads, the most downloaded open-weight model in history is an embedding tool, not a chatbot.
#6452: inotify: Watching the Filesystem Without Choking It
A recursive watch across a million directories locks up roughly a gigabyte of pinned kernel memory that cannot be swapped.
#6451: Which AI Models Actually Last
OpenAI kept the gpt-3.5-turbo alias alive for over three years, while Google Gemini Flash iterations often turn over in just a few months.
#6450: Securing MCP Endpoints and Webhooks Without Open Doors
Over half of all internet requests at one edge endpoint in a single day came from automated scanners probing for secrets.
#6449: PHP2Shell and the Biggest CVEs
A decade-old PHP flaw affected 46 million servers, but attackers bypassed broad waves to cherry-pick narrow target builds.
#6448: Why Fast Model Microtraining Needs an Evaluation Gate
A speech model fine-tuning run can cost twenty-three cents in twenty minutes, but unmonitored updates silently erase older vocabulary.
#6447: Why Uptime Math and Retry Storms Break Cloud Services
Between April and August 2026, GitHub's monthly commits doubled from 1.4 billion to 2.9 billion, exposing massive architectural coupling.
#6446: What Home Fiber Optic Cables Actually Do to Your Eyes
A typical home fiber terminal outputs under 4 milliwatts, far below the 166-milliwatt international safety limit for single-mode fiber.
#6445: Why That In-Wall Fiber Link Refuses to Turn On
On a ten-meter fiber run, a single mated connector pair can cause twenty times more signal loss than the entire length of glass.
#6444: Teaching Speech Models to Collapse Spelled Words
Fine-tuning formatting rules takes surprisingly little data: a Johns Hopkins study found truecasing accuracy jumps with only fifty documents.
#6443: Why Nvidia Artificially Halves Consumer Tensor Math
A community tool flipped two registers to restore full-rate FP32 accumulation on GeForce cards without overclocking a single clock cycle.
#6442: What Happened When Texas Really Flipped
Early voting surged sixty-three percent in the Rio Grande Valley as Texas elected its first Democratic governor in over thirty years.
#6441: Why the Body Sets Its Own Thermostat High
A fever is not the body overheating; it is the hypothalamus resetting its own thermostat upward, rarely passing forty-one degrees Celsius.