walkman
Well Known Member
Share please!
Van's Air Force
The definitive Van's Aircraft support community! Buying, building or flying an RV? Join our exclusive family of mentors and enthusiasts!
AI and the amateur builder
- Thread starter Traash
- Start date
walkman
Well Known Member
Look for books on “prompt engineering”. Andrew ng has some awesome vids for both the neophyte and the expert
I'm a software engineer who hopes to not be replaced by a machine one day soon. That said, getting the best results from AI boils down to a few things:
All of this assumes you're using a modern model. If you're using something old, don't expect good output.
- Trust, but verify! AI will very confidently lie to you!
- Use the best model you can afford. https://artificialanalysis.ai/models#intelligence and similar charts can help
- Define clear boundaries. Constrain the output. Add things like this to your prompt:
- Use AC 43.13 and similar A&P sources to answer.
- If you are not sure, don't guess and state that you're not sure.
- Show your work step-by-step and cite the specific section of AC 43.13 or standard sheet metal practice you are using.
- Start the prompt with: "Act as an experienced A&P mechanic and experimental aircraft structural engineer..."
- Fully define what you're asking:
- THIS: Design a doubler plate for a Switchcraft EHUSB31CFCMB. I've attached the datasheet. There are four bolt holes around the main hole, and I need the correct number of rivets around it with proper rivet spacing and edge distances from both the holes for the Switchcraft part and from the edge of the doubler.
- NOT THIS: I need a doubler for a usb connector. Make it good.
- THIS: I want to install an 3d printed overhead console. Before I give you the specs, ask me 5 clarifying questions about the material, attachment points, and weight distribution so you can give me the most accurate structural feedback.
- NOT THIS: Help print an overhead console.
- Include photos. You don't always have to type everything out - include a photo of what you're asking about!
- Stick to a single provider, Claude (set up a "Project" for your plane!), Gemini, ChatGPT, whatever - they build up context and get better the more you use them.
- If it's free, it's almost certainly not the best option. It costs _a huge amount_ to run these models.
- Verify, Verify, Verify: Treat the AI like a well-meaning apprentice in your hangar. It’s incredibly fast and knows a lot, but the builder-in-command (you!) still has to sign off on the work.
All of this assumes you're using a modern model. If you're using something old, don't expect good output.
Cth6
Well Known Member
This is the core skill. I have others that maintains the list of prior trips and the skill that emails the report.
rmartingt
Well Known Member
My generation had Terminator and The Matrix. At work the kids wonder why I'm so hostile to AI
When my plane is parked, an X-sense temp/humidity sensor hangs from the oil cap and data-logs the "weather" under the cowl.
I've tried to crash Claude (it didn't) by giving it a pretty challenging task to build a humidity model for the crank case and tally the "wetting" hours. The model estimates when the conditions are ideal to form internal condensation, as well as count high humidity hours - the crank case dewpoint within 2 degrees. It takes into account engine metal mass inertia and exchange intertia between ambient air and the crankcase (the engine "breathes" outside air when temperature and air/vapor pressure change).
The code detects flights by finding sharp temperature raise (ie when I hang the sensor after the flight in a hot cowl) and resets the wetting counters. If the sensor data feed drops out it falls back to the closest AWOS station and inserts an uninsulated T-hangar buffer to estimate under the cowl humidity and temperature (fitted by comparing 1 year of X-sense data with the AWOS data).
This is packaged in a claude skill, a standalone python model, and as custom Home Assistant Integration. Home Assistant recalculates every hour and alerts via a telegram message and suggests to go flying after 2 hours of condensation is suspected, or 8 hours of very high humidity (thresholds adjustable, but the default tries to error on the safe side). It attaches a chart to the message showing past 2.5 month history of "wetting" events and marks the flights. Also alerts if doesn't see a flight for a month following a condensation event.
https://github.com/k3it/lyc-engine-moisture-exposure-analysis
pushed everything to github, including the physics explanation. The claude skill is packaged under the Releases.
This was meant to imporove on the "fly often" advice from lycoming by timing the dry off flights more accurately. Not sure if this is useful but was a fun "AI" project.
I've tried to crash Claude (it didn't) by giving it a pretty challenging task to build a humidity model for the crank case and tally the "wetting" hours. The model estimates when the conditions are ideal to form internal condensation, as well as count high humidity hours - the crank case dewpoint within 2 degrees. It takes into account engine metal mass inertia and exchange intertia between ambient air and the crankcase (the engine "breathes" outside air when temperature and air/vapor pressure change).
The code detects flights by finding sharp temperature raise (ie when I hang the sensor after the flight in a hot cowl) and resets the wetting counters. If the sensor data feed drops out it falls back to the closest AWOS station and inserts an uninsulated T-hangar buffer to estimate under the cowl humidity and temperature (fitted by comparing 1 year of X-sense data with the AWOS data).
This is packaged in a claude skill, a standalone python model, and as custom Home Assistant Integration. Home Assistant recalculates every hour and alerts via a telegram message and suggests to go flying after 2 hours of condensation is suspected, or 8 hours of very high humidity (thresholds adjustable, but the default tries to error on the safe side). It attaches a chart to the message showing past 2.5 month history of "wetting" events and marks the flights. Also alerts if doesn't see a flight for a month following a condensation event.
https://github.com/k3it/lyc-engine-moisture-exposure-analysis
pushed everything to github, including the physics explanation. The claude skill is packaged under the Releases.
This was meant to imporove on the "fly often" advice from lycoming by timing the dry off flights more accurately. Not sure if this is useful but was a fun "AI" project.
I have tried using AI to design a paint scheme with mid to poor results.
For reasons I can’t understand it will make changes I didn’t ask for and it also has trouble understanding the changes I do request.
Some of the results have been useful in me better understanding what I do and don’t want. But I do not have a scheme yet.
Google now gives you the AI answer first, so yes, I have used it with questions. Mostly around electrical systems of late because of where I am in my build. I do take those answers with a fair dose of skepticism and go looking for confirmation elsewhere. Funny thing. One of the questions I asked I also searched here and found similar numbers of people on both sides of the issue, so…. be careful no matter where you get your answer.
For reasons I can’t understand it will make changes I didn’t ask for and it also has trouble understanding the changes I do request.
Some of the results have been useful in me better understanding what I do and don’t want. But I do not have a scheme yet.
Google now gives you the AI answer first, so yes, I have used it with questions. Mostly around electrical systems of late because of where I am in my build. I do take those answers with a fair dose of skepticism and go looking for confirmation elsewhere. Funny thing. One of the questions I asked I also searched here and found similar numbers of people on both sides of the issue, so…. be careful no matter where you get your answer.
