Ilma is a Navtex decoder for the Raspberry Pi that just got weather fax (WEFAX) capabilities. I’m a sailor and I built a headless receiver for the Raspberry Pi & SDR-RTL v4 dongle.
The tempo of AI development is overwhelming. Within one generation, coding/IT has gone from the most promising career for the young to profound insecurity about the future.
I use a self-developed plain text editor for writing and my one-click GitHub sync menu bar app to deploy Astro blogs.
The irony of the GitHub sync app is that I do no longer have a working GitHub account...
I built this on a Pi Zero 2 W because of its low energy consumption when it's running continuously on a boat. Navtex is a 1970s radio technique that delivers navigational and meteorological info and has forward error correction built in. For the characters that still slip through, I added a small n-gram model. An LLM wouldn't fit on the Pi.
I also work on an RTTY mode that hasn't got any built-in error correction. Does anyone have ideas for correcting characters dropped during radio transmission?
Storage is the hard part of renewable energy. In this article, the researches used a molecule that captures sunlight and locks it into a strained, high-energy shape. It can hold that energy for months to years and then release it as heat on demand. The release of that energy makes water boil in about one second. Energy density is 1.65 MJ/kg, beating lithium-ion (~0.9 MJ/kg).
The interesting background is the inspiration of the method by the shape-shift DNA undergoes when UV light damages it. It's the chemistry of sunburn.
Efficiency claim: 70% less token usage, 3x faster task completion in internal testing. Even if that's marketing-inflated, the structured CLI commands replace a lot of trial-and-error. Gonna try it.
There was a recent research article titled "LLM Targeted Underperformance Disproportionately Impacts Vulnerable Users". They described systematic underperformance of AI models targeted towards users with lower English proficiency, less education, and from non-US origins. As interesting it might be to experiment with an AI CEO hiring people – what a dystopian vision.
On the other hand, it seems ironic that AI replaces a CEO – would Karl Marx like this turn of history…?
"We show systematic underperformance of GPT-4, Llama 3, and Claude 3 Opus targeted towards users with lower English proficiency, less education, and from non-US origins."
Even if you're trying and willing to learn as a less skilled user, it will be harder to gain knowledge from a LLM. A book is the same for all readers…
I had the same issues, so I built a macOS menu-bar app that does one-click git pull/commit/push for any local folder, including Obsidian vaults. Uses the system git binary, fast-forward merges only (stops on conflict instead of silently merging), blocks secrets before push.
I had Max plan and never reached its limit despite constantly working. Now I use the Pro plan and regularly reach the 5h limit as well as the weekly limit, as expected.
I found that it makes a huge difference if you provide clear context when developing code. If you leave open room for interpretation, Claude Code uses tokens up much faster than in a defined context.
The same is true for his time to answer getting longer if there isn't much documentation about the project.