This process works well for two reasons. Firstly, it breaks up the work into small, carefully scoped chunks that fit within the model's context window, keeping it focussed. Secondly, the review aspects (the manual check, and instruction to critically review the work) removes a lot of bugs, so you maintain a solid foundation to work from. Most of the time Opus will find a few bugs in its implementation, if you ask it to check, and it may take two or three rounds before it stops finding problems.
]]>I settled on Golang because it is modern, memory safe (mostly) and provides highly efficient built in webserver functionality. Goroutines have a tiny memory footprint, fast start up time, and low CPU overhead, all from a single small compiled binary. Compared to Apache2 with its endless configuration options and complexity, it's quite a relief to deal with.
And Golang has not disappointed me. The efficiency gains are real and will allow me to deploy onto minimal hardware, thereby directly saving money. Even on a Raspberry Pi 5 development box (yes, really) my web app runs like lightning and has shockingly low CPU and memory footprints.
But there is a downside, and this is where PHP has the advantage: Maintenance. If a PHP site has a problem you can often login while it's running, poke around a bit and fix it, without much concern that you will torch the entire system. The files are human readable text, so modifying one or reverting a bad change is basically instant with limited blast radius. You can do emergency maintenance on the road from a tablet or even a phone.
]]>This wasn't a problem for me, but I heeded the 'mitigation' advice and removed all plugins, skills, agents, and MCPs to minimise context injection. I also audited my configuration using the Context Audit skill you can download from Brad | AI Automation.
Around mid-April Anthropic claimed to have fixed it. Well, no. They haven't. I started experiencing the problem as soon as my usage reset and I had access to Opus 4.7, even though I reduced the effort to 'medium' from the default 'xHigh'.
It's terrible! Previously I could carefully steward my session limit through two or three hours of code work with Opus. Today? About 30 minutes and with a far smaller volume of work achieved.
]]>TLDR: Recommended for Raspberry Pi 4b...if you don't have issues with the USB connector (mine seems defective, which is a possible dealbreaker). Excellent construction but fan is noisy at high loads; can mitigate with an improved fan control script (provided). The S2Pi Aluminum NAS case provides a rugged housing for the Raspberry Pi 4b with M.2 SSD storage and an Ice Tower heat sink for strong cooling performance. It's an excellent package for upgrading your Pi to a lightweight server.
I have developed an improved fan speed control script that turns the fan off when not needed, and ramps with CPU temperature. Available for download within.
]]>I recently got dunked on for saying the Raspberry Pi 5 makes a great home lab server if you equip it with an SSD drive. And I don't really blame the guy, because until the Pi 4b, they were pretty awful, and for the 3B and below you were stuck with running the OS from a microSD card. His mental model was probably stuck somewhere around there.
The Pi 5 is a huge level up in performance, especially once you add SSD storage via its PCIe slot.
]]>TLDR: Recently ChatGPT just started wildly lying, inventing rubbish and disregarding my instructions. The root cause seems to be that OpenAI has instructed it to reduce use of the search tool (you can see references to this in the chain of thought). I imagine this is to conserve resources, but without some factual context to go on ChatGPT hallucinates like hell.
But there is a better way: Use the command line version of your preferred AI tool. Watch NetworkChuck's video for details! These are available for Gemini (free), ChatGPT and Claude (both of which require a standard subscription). Why is it better? You can force the AI to remember rules and context for any given project, agent workflow, or output style, making it much more reliable.
]]>According to the marketing materials, the proprietary Firewalla WIFI-SD USB dongle/antenna only works with their (excellent) hardware firewalls. But I needed an external WIFI antenna for a Raspberry Pi, so I inserted it into a USB port to see if it would work, and yeah it does. The downside is that the Pi only seems to have drivers to use it on the 2.4 GHz band, while the antenna is supposed to be a WIFI 5 device.
It is apparently based on the dual-band RTL8821CU chipset (802.11ac). Probably you could get it working on 5 Ghz if you were prepared to tinker with drivers but I'm told this chipset is notoriously difficult to work with, and kernel updates would probably keep breaking it, so I'll pass. Anyway, if you login to a Firewalla box via SSH the default username is 'pi', so you can probably guess why this works.
]]>I like FontAwesome (I'm a backer), but I'm so over subscription services, there's only so much many times I want to pay for the same icon collection, you know? I'm also annoyed that they withdraw access to the subsetter app once your subscription expires. That's miserable, given that subsetting the icons is mandatory to control file size, as the full icon library is too large to use.
IcoMoon has a free online tool that lets you convert icon collections into a custom font. You upload your icon SVGs, select the ones you want to use, and export them back out as a font, which is trivial to incorporate into your project. The files are much smaller, and do not require any Javascript. I will probably remove the FontAwesome script from Tuskfish and re-implement the icons as a font. This will also allow other icon sets to be integrated, as you aren't limited to the FontAwesome set.
]]>