June 23rd - Field Irrigation and Pig Fencing
TL;DR
Today was the first work day of a record-breaking heat wave in New England. We harvested in short spikes with breaks in between then set up some irrigation lines.
Farming in Extreme Heat
Even when thereās a heat advisory, thereās still work to be done. Priorities shift to āminimum viableā outside work, such as harvesting in the early morning and especially things that may bolt (like lettuces). Soil that has dried out needs to be watered, so irrigation is a priority. Watering the greenhouse multiple times a day is a must.
At this farm, even though we still work, we are provided ample opportunities to hydrate, take breaks, and cool off when the heat is at its peak like this. We also rotate duties between field and barn work (such as washing, bunching, and packing) to avoid overheating.
Before and After Lunch Harvesting
In the morning, we harvested chard, fennel, and lettuce. After lunch, we harvested garlic scapes (the flowering curly shoots of the garlic stalk) and scallions. Jamie checked the potatoes by pulling some up, which ended up in my take-home produce later. Despite some potato beetle damage, he thinks most of the crop will pull through.
Then I helped Jamie continue setting up irrigation lines to the main field. This involved super-manning a huge wooden spool with some 2-inch tubing onto the back of the truck.
After we were done with that, we went back up to the barn where I stayed to bunch and wash scallions with another person. Meanwhile Jamie and the other person went up to start extending the pig pen with fencing into the forest. Though shady up there, it was 2pm and 90 degrees fahrenheit and humid, so I can only imagine it was a haul to continue the day on that task. The pigs are outgrowing the wooden shade structures at this point, so it was necessary.
Not all AI Crop is āAI Slopā
The populist outsiderās view and understanding of things termed āAIā is often a caricature of the reality. āAI slopā a the most recent example of thisā¦content ācreatorsā who are not experts but rather rely on AI tools to generate content that is not only low quality but also often offensive or harmful.
Iāve been working hard to make sure that nothing on this blog is āslopā (except for the occasional image and mention of pigs, lolz!). My use of LLMs is limited to very specific use cases that I have hand-crafted the infrastructure, the code, the prompts and the process to produce basic structural elements of the blog posts for.
Unless You Do Quality Control, Itās Likely Slop
Quality control (QC), quality assurance, quality management, testing, verification, and validation were all sections I had a heavy hand in pushing to completion in my prior work with IEEE 2675-2021, despite last-minute adjustments out of my hands. This took years of writing, reviewing with the team, revising, re-reviewing and balloting, then finalizing.
My homelab work goes much faster, but still has a lot of QC in its process. I listen back to summaries before using them, making mental note of what might be going wrong in prompts and process if I hear strange things. I do not always have the time to correct issues, and besides, I want to keep the record of my work and the process I went through to get to the final product.
In the past two weeks, I built a separate āmeeting minutes compilerā which, in short, taught me that one specific LLM is not always best when different parts of your process require specific outputs and speeds. Specifically, handing a whole transcript to Deepseek often produces higher number of hallucinations than handing specific chunks (and their prior section) in for summarization.
As it turns out, too much context is sometimes a bad thing, just like not enough. How to balance what is āenoughā is a job of the system design and operations. And just like there is usually no good outcome from a āone-shotā prompt, a one-iteration approach to features and software usually leaves a lot of learning and right-fitting on the table. Incremental and intentional iteration is what Iām doing with my homelab work.
Pop-up Blockers to Progress
Every so often, something in my homelab infrastructure will go ābump in the nightā and stop working. Iāve learned to expect this, and to be ready to troubleshoot and fix it.
This morning, a problem with my homelab Ollama kubernetes deployment took about 20 minutes away from this writing. In essence a start hook to download three commonly used models was failing because compatibility between it and the ollama host container had shifted. Iām pretty sure I need to implement better log shipping and investigate persistence of the model binaries between container restarts, but thatās for later. This morning I just had to get it working again and get on with my blog before Iām late to the farm.
I need a more proactive approach to these āpop-upā progress blockers. I donāt have time for running disaster recovery every single time I change some small part of my infrastructure. I also donāt think any of that would have helped in this case, since the problem wasnāt with anything I changed but something in the intra-dependency after a restart that was changed outside my control (the model from online download itself).
Anyway, while I was troubleshooting the ollama deployment, I realized that I had switched autoscribe (the component that produces the following notes summaries for this blog) to use Deepseek-R1:1bn (only one billion parameters) instead of the default :8B. When I get home later today, Iāll re-run it with 8B and compare the results.
I noticed that it gets side-tracked because of the tracking code at the top of the notes again, and that all of Part 1 is a hallucination. Particularly egregious is the line about Roundup PowerMax, which is not at all what we would ever use on this farm.
This is why unchecked and un-validated AI output is dangerous, not only informational but operationally. Imagine if I had used this output to make decisions about what chemicals to use on the farm!
Hereās what the AI generated for me with the 1B (small) model:
Summary
Okay, hereās a consolidated summary combining all elements from both analyses:
Consolidated Summary:
Based on the provided field notes summaries, here is a consolidated overview of the farm internās activities and observations over two days in June:
- Main Themes: The workday focused on transitioning into active harvest season while also addressing operational tasks like irrigation setup and animal management.
- Activities:
- Checking into office/operations.
- Meeting with supervisor (likely to discuss planning).
- Discussing crop planting timelines post-harvest.
- Setting up irrigation systems (moving large sprinklers on tripods, fixing electric fences for pigs).
- Harvesting crops (Swiss chard, lettuce, scallions, garlic scapes) and preparing produce (āhottentotsā - likely a typo or specific term for large sprinklers).
- Picking scallions and garlic scapes from dry fields.
- Farm maintenance tasks like fixing pig pensā electric fences.
- New Things Not Encountered Before:
- Using ālargeā sprinklers on tripods (specific size/size depends on context, but likely a new setup for the intern).
AI-MISTAKE: The specific herbicide mentioned (Roundup PowerMax) and its application rate/dose (1 gallon per acre), though this might be less novel depending on prior experience.
- Questions for Future Research:
AI-MISTAKE: Compatibility of Roundup PowerMax with other chemicals (from first summary).- How long can garlic scapes stay in the field before harvest?
- Efficiency and effectiveness of large sprinklers (āhottentotsā) vs smaller methods.
- Impact of irrigation timing/schedule on various crops.
- Best practices for cooling during extreme heat, especially with portable solutions (from second summary).
- Suggested Actions:
AI-MISTAKE: Ensure proper Personal Protective Equipment (PPE) is used when applying herbicides like Roundup PowerMax to prevent exposure and contamination.- Harvest garlic scapes at optimal times for best flavor or yield according to farm standards/guidelines.
- Consider the use of portable fans or shade structures in addition to shaded breaks during peak heat hours.
Part 1
Field Notes Analysis
Activities:
- Checked in at the farm office.
- Met with the supervisor to discuss the dayās tasks.
- Discussed planting timelines and upcoming activities.
- Reviewed pest control strategies and methods.
AI-MISTAKE: Mixed herbicides according to specified ratios.AI-MISTAKE: Applied the mixed herbicide using a sprayer.- Moved supplies as needed.
New Things Not Encountered:
AI-MISTAKE: Use of Roundup PowerMax herbicide, which may be new for the intern.AI-MISTAKE: Application rate of 1 gallon per acre, possibly new information.AI-MISTAKE: Usage of a sprayer instead of backpack sprayers (though this might not necessarily be new).
Questions and Future Research:
AI-MISTAKE: Compatibility of Roundup PowerMax with other chemicals.AI-MISTAKE: Optimal timing for herbicide application to avoid harming beneficial insects or crops.
Suggestion Actions:
- Consider waiting longer before applying if timing is uncertain.
AI-MISTAKE: Ensure proper Personal Protective Equipment (PPE) is used during application.
Main Themes:
- Agricultural tasks and daily operations.
- Pest management and control strategies.
AI-MISTAKE: Herbicide use, mixing, and application techniques.- Safety protocols and equipment usage.
- Plant health monitoring and maintenance.
- Timing of agricultural operations for effectiveness and impact.
Part 2
Main Themes:
- Crop Management: Focus on harvesting crops like Swiss chard, lettuce, scallions, and garlic scapes, and setting up irrigation systems.
- Heat Management: Addressing the challenges of extreme heat and strategies to manage it, such as taking breaks in shaded areas.
- Sustainable Practices: Utilizing large sprinklers on tripods, mulching, and maintaining electric fences for pigs.
Activities:
- Harvesting crops (Swiss chard, lettuce, scallions, garlic scapes).
- Setting up irrigation systems and moving sprinklers.
- Processing produce in the barn with industrial fans.
- Picking scallions and garlic scapes from dry fields.
- Fixing pig pens and extending electric fences.
New Things Not Encountered Before:
- Garlic scapes as a harvestable item, which might be less common or region-specific.
- Large sprinklers on tripods for effective irrigation coverage.
Questions for Future Research:
- How long can garlic scapes remain in the field before quality degradation?
- Efficiency of large sprinklers compared to other irrigation methods.
- Impact of irrigation schedules on different crops.
- Best practices for worker cooling in extreme heat, including portable fans or shade structures.
Suggestion Actions:
- Harvest garlic scapes at optimal times for flavor and yield.
- Explore alternative irrigation methods if main systems arenāt effective.
- Recommend portable cooling tools for workers during high-heat conditions.
This structured approach ensures all aspects of the internās activities are covered, providing a clear understanding of their workday and future considerations.
Enjoy Reading This Article?
Here are some more articles you might like to read next: