Blog

Since being in Spain (3yrs) I’ve had two motherboards go bad. Not completely bad, but throwing errors, freezing, etc. I’ve never had this problem before so I plugged in a Shelly switch with power monitoring, connected Homeassistant and monitored it for a few weeks. In a 7 day period the voltage ranged from 213VAC to 240VAC. This is an insane voltage sweep. I can see 20VAC variance in 24hrs or even sometimes 10-15VAC variation in one hour.

The frequency stays quite solid between 49.9-50hz with one 49.8hz dip in 7 days. So what this is telling me is that there are grid-tie solar installations near me that are flooding or reducing the measured voltage. This is concerning because either the grid operator cannot respond fast enough to the fluctionations or doesn’t really monitor voltage or their SLA is within 210-240 and that’s good enough.

I’ve since added an Automated Voltage Regulator before my sensitive electronics. Now the variation is within +-3VAC, well within spec. In the US I never had more than 3VAC fluctuation at any given time and the average change was less than 1VAC.

Today in therapy we got on the subject of sabbaticals as part of looking towards the future. I was asked what is my dream and I answered that getting to have a significant amount of time off to figure out what’s next in life sounds amazing.

My therapist said that while they can be good a sabbatical needs preparation and rarely do people achieve that inner peace they went looking for. So, what’s the goal, what do you want to achieve? And more importantly why can’t it be something to focus on everyday instead of all at once?

This stuck with me as I’m working towards self-acceptance and being present. There’s always a dream or the next thing or or more money to chase but what small thing can I do, today, to make that dream a reality. Over time those small things turn into lasting change.

Just some food for thought.

It’s been an insane year or two around AI in the development space. In my company, as with almost every one else in the world, are pushing the use of AI hard. The first to get to market with the next big thing wins the prize. What this looks like in reality is that long term support suffers, open source suffers, and we end up with an enormous amount of replicas of the same type of application.

When is the last time you actually searched to find software that solved your problem? Did you actually look to see if the thing you just spend 30M tokens building already exists?

Support

Something I don’t see talked about alot, at least out side of the Open Source community, is who will support these applications, changes, etc over the long term. We’re quickly getting to a point where there’s nobody who actually knows the code they are pushing. Sure people say they are reviewing every line, “human in the loop”, etc but the reality is that you’re going to need an LLM forever to support your deliverables.

I’m just as guilty as the next person and honestly support and long-term vision are not rewarded in the corporate world.

Just because we can.. doesn’t mean we should..

Going back to the title of this post, not everything needs AI. Let me clarify, not everything need GenAI to be successful. LLMs are non-deterministic and developers burn a lot of tokens trying to get the same or similar answers out of their applications. We want to just give some LLM all the data and say “here do X and return Y”, because developers are inherently lazy people… And before everyone gets up in arms, who here started their development career building macros because they didn’t want to have to perform the same task in Excel over and over?

Anyway what are your thoughts on using AI to solve every problem?

I’ve been vibe coding now for about a year, using all different methods and tools. I started with OpenWeb UI hooked up to LLama3.1 70b through Scaleways and that worked ok. Then moved to Gemini on the web, then Antigravity, Claude, Cursor, OpenCode and PI. Here’s what I’ve learned.

Clear Requirements

This sounds like a ’no-brainer’ to any developer but its just so easy to say “Hey X build me Y!” You then spend hours going around in circles. This is much the way software development works in the real world. Many times your customers asking for one thing but really want/need something completely different. Most tools now have a Plan mode and it helps to spend a lot of time refining those requirements.

Create a to-do list and have the agent step through it

Again, sounds simple but while still in the Plan mode have it generate a To-Do List. Sometimes there are multiple phases in my lists and Ill have the model develop all of Phase 1 before testing and commiting and other times I have it go item by item, testing (if possible), marking done and then continuing on. I’ve found this to be a very reliable way to get what you want. To me these are still Junior Engineers and you have to think about how you would manage them. Step by step!

Models matter somewhat, but it’s the Harness that makes the difference

I’ve used web chats, Claude Code, Gemini CLI, Antigravity, Cursor, OpenCode, and Pi. So far I’ve liked Cursor and OpenCode the best. I know Cursor is just a modified VS Code, but it works really well in my workflow at work… unfortunately everyone wanted to go to Claude and so no more Cursor for me. Personally I use OpenCode and I have really found it to be very thorough in it’s planning and developement.

Some of my vibe-coded apps:

Flying Goose Games My kids and I created this last summer to have something to play while sitting out to dinner and waiting on our food. It’s been a lot of fun and we have added a few new games over the year. Our problem is that we’ve played it so much we know the answers to the questions by heart! It’s currently running on Cloudflare Pages.

BentoMeteo I wanted a simple, fast, and clean weather app that would let me use the OpenMeteo API. Another weather app? I know, but I use it daily and its quite accurate. The biggest part is that all my favorite weather apps have gone from cool grassroots apps to being bought up by the bigger weather apps and now are just bloated with adds. This runs as a Cloudflare Worker and isn’t 100%. I refactored it from running in Docker to Workers earlier in the year and things are a bit broken.

Ones of note that haven’t quite stuck around:

• WhatsApp Scheduler - I spent a week over Christmas making an automated Poll creation tool for our Padel group that would create a weekly poll, nudge users, look at the best weather days, etc. It was cool but using the unofficial Whatsapp APIs lead to broken stuff and it needed a lot of handholding.
• Options Trader - I spent quite a bit of time working on an options trading bot using the Alpaca API but there were a lot of challenges when the spreads wouldn’t be available and the app would have to figure out why the trade didn’t execute. There were other issues and honestly I just joined a hedge fund that did the trading for me.. There’s a reason financial firms spend big money on algorithmic trading!
• Stock Trader - This one is similar to the above but would perform technical trade analysis on stocks you would specify. It then figured out the buy, sell, and stop-losses. This was my first foray into ML models as I trained the models on the market that were used to determine if the trade had potential, or not. I want to give this more time, but again I also pay people to invest for me and they are doing a good job!

Vibe-Coding has really opened the doors for ideas to go to reality in a short amount of time, though that has other challenges. Ill post about those next time!