AI Agents
300 turns, zero APIs. Re-reading the log showed 234 turns had never called a tool. We cover the idle rate as a metric, idling that shifts form each time one is closed, how turning thinking off dropped 95 to 71 without saving time, and counting the other side after widening the output window.
GPU
Reboot an Ubuntu 24.04 GPU server and nvidia-smi stops working. The cause: unattended-upgrades updated the kernel alone and skipped the NVIDIA module over a dependency it could not resolve. How we traced it in the logs, recovered with one install, and compared three fixes including apt-mark hold.
AI Agents
The model's "Done" slipped through in five ways, from a bare syntax check to tests with no assert. We describe, with numbers, a completion gate that closes in three stages, and why a pushback should contain the failing lines and how to investigate, but never the answer.
LLM
Jev returns typed decisions instead of text. Signups reopened on September 28 with no free credit for direct accounts. We put $10 in, called it 301 times from Python, and measured accuracy on Japanese tickets, guardrails and command approval, plus response time and cost (about one yen).
GPU
Hello! If you work with NVIDIA GPUs, you run into numbers like sm_120 and sm_100 all the time. They are called Compute Capability numbers. They look like they map onto the product generation names. They do not. What sent us digging was TensorRT-LLM v1.3.0rc28, released on
AI Agents
A complete guide to Claude Opus 5.5: prices 20% below Opus 5 with 60% cheaper cache reads, API migration notes such as the medium default effort and thinking that can no longer be disabled, and the new default model and Fast mode in Claude Code.
OpenAI
A complete guide to GPT-6 Sol and Luna: what is behind the halved API prices, how to compare them with Astra and Claude Opus 5.5, migration notes such as reasoning effort none and the surcharge above 272K input tokens, and availability in Codex and Copilot.
AI Agents
Codex CLI’s Windows update failed because Get-FileHash could not be found. We traced the inherited PSModulePath and successfully updated from 0.154.0 to 0.156.1 using PowerShell 7.
AI Agents
When we built our own coding agent, the hardest part was not running the loop but deciding when to stop. In one run, 234 of 300 turns never called a tool. Starting from that number, we separate the turn runner from the stop decider and fix how pushbacks are sent.
LLM
What makes GPT-6 Astra stand out? We compare performance and pricing with Claude Fable 5.1, plus the ChatGPT Pro and Claude Max 20x plans, and share the AGI-like progress and what we noticed in real development, with diagrams.
You run Claude Code in bypassPermissions, yet grep -r started waiting for approval today. The cause is v2.1.259, released September 2, 2026: Read deny rules in settings.json now apply to Bash grep. We measured it side by side with v2.1.258 and tabulated what passes and what stops.
LLM
Introduction This report takes a comprehensive look at the performance of Japanese-capable LLMs, based on the Nejumi Leaderboard 4 benchmark data (September 1, 2026 snapshot). Last time we published the July 10, 2026 edition of this analysis, and in less than two months the top spot has already changed hands.
GPU
Hello! "TensorRT 10 × Blackwell Migration Guide" series has reached its final installment.Part 1 covered the basics of converting to TensorRT and the wall where a 5D (volumetric) grid_sample is rejected at build time, and Part 2 walked through the many forms of silent degradation where "
WireCanal
How we made a ¥3,750 (~$25) Linux board — the Luckfox Pico M (RV1103, 64 MB, USB-C only) — reachable from the internet without adding an inbound port for web publishing: SD imaging, RNDIS networking, WireCanal's free plan, a browser-controlled LED, and 33-second reboot recovery, all measured.
WireCanal
We publish a Raspberry Pi 5 web server to the internet with its inbound ports kept closed: install the WireCanal Agent with a one-liner, create a canal, and run it as a systemd service. No port forwarding or router settings, about 5 minutes of work, verified through automatic recovery after reboot.
A freshly set up Raspberry Pi 5 is wide open: password auth enabled, no firewall, 172 pending updates. We harden it with SSH key-only auth, UFW, fail2ban, and automatic updates, disable IPv6, and verify it all with real reboots — including the trap where IPv6 comes back after a reboot.