When the Limitation Changes

As my comfort with AI grows, so do the projects I assign it. In the last week, AI has been adding features far faster than I ever had time to do in the past. The limitation is no longer my coding free time... it is now the ideas I have for enhancing the site. Every feature added required investigation and follow-ups before functioning as intended. I believe skill and know-how are required to validate that AI has not made a mistake. As AI often tells me: "You are right for saying that."

Weather: Added heat index alerts, wave forecasts - if you're close to a large body of water, 3 hr rain forecasts, more precise caching - using the Location finder is more accurate than entering your city by name. An optional interactive radar adds full pan and zoom, higher resolution, and international support. Initial data loads are higher with the new radar. Incremental loads are smaller. Load the map one time, and only the radar data from that point on.

Site theme: All buttons and formatting have been aligned to make the different components on this site feel more cohesive than ever.

News: The backend function to add news feeds had been broken... now it works. It adds icons, allows for easy customization, has drag and drop to rearrange the feeds. The front end has a slider to determine how many articles to display, giving more granular control. I increased the news feed list from 24 to 48 hours, improving the amount of searchable news. 10 new technology-focused feeds were added to the news system. They are enabled by default. Click the Settings button to turn off any feeds you do not want to see.

When the limitations change, so does the game.

Faster with Reservations

AI models continue to improve. The latest models will try multiple options to find the best one. This ability has enhanced this site. I believe there are no measurable performance optimizations left here. It is time to add features. Today I added Air Quality graphs to the Weather page. Less than 2 hours from concept to production.

The most significant AI-assisted enhancements? The weather graphing engine was 33kb, AI written 7kb. Weather radar uses an animated GIF and breaks it into 10 frames for processing. My code did this in 0.0157 seconds. AI code: 0.0069 seconds; 2.3x faster. That's not the biggest impact...

I used Tesseract to OCR the radar images; to ensure the time displayed was when the image was captured. 28mb of code. I optimized the execution code down to 0.9 seconds.

AI replaced it with a machine learning process. It creates glyphs of the time characters, adding new glyphs when needed. Size? 14kb; 0.05% of what I was using. Speed? 0.00008 seconds; 11,250x faster.

A cautionary tale: AI made mistakes on this journey. Some required me to restore from backup. AI can be reckless and dangerous. I was prepared. For every frustration it caused, it rewarded me with code that would have taken me weeks to write. It's worth the effort, but you must understand what it is doing and be diligent in reviewing.

Large Language Modeling

I do not consider AI to be Artificial Intelligence. It is a very clever algorithm, one that calculates advanced pattern matching results on massive data sets without much prompting. At quick glance, it feels like it is thinking. It is not, it is a Large Language Model (LLM). This does not mean it is not useful.

LLMs have done reviews of this web site and other sites I am working on. They are able to pull apart the pieces and determine, with relatively high accuracy, my intent. They have fixed dozens of bugs, many that would have gone unnoticed for months. I always appreciate performance tweaks. They found quite a few.

Most important are the security fixes. They found vulnerabilities in my code. Malicious LLMs used by hackers could have found the same holes and exploited them (they do not have my source code, but still...). If we are not using the same caliber of tools to protect ourselves as are available to malicious actors, we are essentially allowing them to come in.

My opinion: LLMs can do a lot of heavy lifting, saving us a lot of time and helping to fill the gaps we missed. They should not be given free reign on our tools or systems. Only someone who truly understands the work they are doing can safely and effectively keep the LLMs in line and on track.

Use LLMs (AI). Learn from them. Check everything they do. This is a new class of tool, we must learn how to safely and effectively use it.