MARSfarm 'GPTs' (ChatGPT Customized by MARSfarm)

Peter · March 1, 2024, 1:34am

I’ve already mentioned some of these in other threads, but I was sending a ‘hail mary’ LinkedIn message to a GPT developer and figured I would just share here. For context, I’m looking to find someone that can help me convert these use cases to use the OpenAI API rather than ChatGPT plus so that anyone can use them without spending $20/mo.
Let’s speculate about how you think Planty was created. I promise to be 100% transparent about three use cases I am exploring in ChatGPT Plus for integrating LLMs into my product. I am CEO/Co-Founder of MARSfarm, which designs and manufactures countertop-sized ‘smart’ greenhouses in St. Louis, Missouri.

GPTs made by MARSfarm:

MARSfarm Quotes - Quoting tool to explain our products and create quotes. Simply enter the email address, first name, and last name of any teacher in the united states to get started! ChatGPT - MARSfarm Quotes

Link to YouTube demo: https://www.youtube.com/watch?v=v3OHyYYobQo

Recipe Builder - Creates .JSON files that our firmware uses to control light, temperature, and moisture. ChatGPT - Recipe Builder

Thread with several examples: What does a MARSfarm recipe do? What is JSON? (ChatGPT-4 Response) - #7 by Peter

Dr. Data - Analyze sensor data (temp/humidity/co2) from four devices (MV1-0001-0004) to learn about data science. ChatGPT - Dr. Data

Link to conversation analyzing dataset from ~200 trials: ChatGPT
Most of my conversations for this one won’t share because they contain images / code that was executed. I will make a reply post sometime with a bunch of screenshots or a video walkthrough.

image879×651 226 KB

Martian Farmer - This one is just for fun (and pitch decks, tbh) - it’s designed to take a photo of you (uploaded) and convert you into a farmer on Mars wearing an FFA jacket. ChatGPT - Martian Farmer

GPTs we’ll make if you want us to

I am sharing this list so you can reply and tell me what to prioritize. Feel free to make other suggestions too, or just share your feedback about what you like and share ideas for what else could be added as features:

MAMA (Martian Agricultural Monitoring Assistant)

This one is for everything to do with analyzing photos of the plants in MARSfarm equipment. This thread is all about the current
state-of-the-art in this field and MARSfarm efforts to create MAMA software: Computer Vision - Using the MV1 to Automatically Measure Plants

Hydro Buddy

Recommends how much ph up/down to add based on the amount of water, current ph and EC.
Suggests fertilizer formulations (available to purchase) based on the type/age of plant being grown. It will do this using the recommended nutrient requirements and then verifying that the PPM of each nutrient available (when diluted) will be gradually increased to reduce any risk of shock.
Recommends pH/EC values based on the type and age of the plant being grown.
Estimates fertilizer usage rates based on measurements of EC, water, and pH over time.

Poop Potatoes

The personality is an Irish farmer on Mars - who is terrified that his potatoes will die like his great, great, great, Grandpas did in the potato famine of 1845.
The knowledge base (what we tell it) of this GPT is 100% dedicated to aggregating everything I’ve learned about growing potatoes in ‘weird’ (controlled/indoor/hydroponic/etc.) conditions.
Potatoes are a great way to learn about the ability of biodiversity to prevent total crop loss in volatile/extreme environments.

Dear Mr. Darwin

This GPT is designed to provide a Peer Reviewer of a student’s written work to find their procedure/hypothesis for a plant experiment and evaluate it for reproducibility.
Charles Darwin (and his son) wrote a 700+ page book called ‘The Power of Movement in Plants’ - they would trace plant movements manually on paper to track their growth patterns over time to learn how/why they made decisions.
It will be designed to walk through (using a series of 4-5 prompts) the students communicated procedure to suggest where and how to add more details to reduce confusion.

Chart Analysis

Designed for teachers to copy/paste charts from the MARSfarm web application into it and then generate 10+ questions for students to answer.
The goal is to improve a students ability to interpret charts and tables to determine key variables. This requires familiarity with those charts, which can be done by repeating this type of a lesson on a routine basis throughout a semester.
The GPT would also generate an ‘answer key’ for the teacher, with explanations of each chart for them to use to then communicate to students who get it wrong.

MV1 CASE Integrator @reigna would you use this? what if it was LA state CTE standards?

Has a knowledge base with frameworks for all CASE curriculum
Uses RAG to access .csv with crosswalk framework to align to NGSS / common core.
Intended to be prompted with the CASE courses taught by a teacher and will reply with a table showing the activities where an MV1 would be useful across all of those courses.

Peter · March 1, 2024, 1:36am

I liked the acronym ‘MAMA’ (it’ll have the personality of mother nature) and was trying to come up with a good acronym for it so I used Copilot (formerly Bing). My favorite was:

‘Martian Agricultural Miracle Assistant’

A Hypothetical robot that would perform extraordinary feats such as growing giant plants, producing exotic fruits, and creating stunning floral arrangements on Mars, using nanotechnology and magic.

Peter · April 18, 2024, 3:28am

I want to share more screenshots from my conversation with the ‘Dr. Data’ GPT created to analyze environmental data created by an MV1. My goal is that by doing this I will show you how I communicate with ChatGPT in a back and forth collaborative way - this is what I see the future of data analysis to be. To be clear, before this conversation with ChatGPT I had never done ‘machine learning’ which is the foundation of AI. I had never even used environmental data from an MV1 to conduct a statistical analysis - nor could I explain why it would be important - until this 2-3 hour conversation with ChatGPT. My point here is that this conversation showed me the value of statistics to analyze data in a 2-3 hour process - so I now understand how it is relevant/useful to me.

When I showed this to my dad (@hmw) - who got his degree in computer science

Step 1 - Create a chart of temperature data from MV1-0001

Step 2 - Draft lesson plan ideas for an introductory statistics course

I did this to help me brainstorm how I could do a statistical analysis that might be useful for a student to also do using MV1 data.

Step 3 - Add more detail to items 4-7 on the list

Step 4 - Perform a ‘linear regression’ as described in lesson plan idea #4

Step 5 - Outline how ‘machine learning’ could be used to analyze this dataset

Step 6 - Split the data into ‘training’ and ‘testing’ datasets

Step 7 - Experiment with different models to evaluate their performance

I rarely give ChatGPT praise - I was pretty psyched by this point. This is what data science is though - I was finally doing it. I created a hypothesis (a linear regression model will predict temperature data) and then tested that on a dataset and measured what happened.

As you can see, this still wasn’t anywhere close to predicting reality though. My red dots, despite being detailed in their plot, are not yet an accurate prediction of what the temperature would be in the future. They continue to go right down the middle, predicting that the temperature in the future would be 84F usually, when in reality the temperature is NEVER above 80F at night and is usually above 88F during the day - except for when the box is ‘warming up’ in the morning.

Step 8 - Provide more context about the dataset to identify outliers

Step 9 - Assign ‘labels’ to categorize the temperature data as ‘Day’ or ‘Night’

Step 10 - Create a linear regression for ‘Day’ and ‘Night’ temperature data

Step 11 - Generate a chart to visualize these predictions

Step 12 - Evaluate whether this worked ‘better’ than before adding labels

Looking back at the last evaluation in step 7, you’ll see that the ‘Mean Squared Error (MSE)’ was previously 35.09. Now we see that it is 0.274 for the Day model and 0.415 for the Night model. To be clear, lower is better - so this is a gigantic improvement. What I learned here is the most important lesson of machine learning - the importance of having a human ‘label’ part of a dataset - so that it can be used to then predict what will happen in the rest of the dataset. That’s the foundations of AI.

Step 13 - Create a polynomial regression

I had no clue how this was different from a ‘linear regression’ - but the whole idea here was to test several models against each other so that was what I did. Amazingly, the ‘MSE’ metric continued to decrease - down to 0.101 for Day and .387 for Night! I still don’t understand why - in the slightest, I just know that I should use a polynomial regression to predict what the temperature is going to be in the MV1-0001 and that’s all that matters.
image|266x500

Step 14 - Bring in humidity and Co2 data as well

Step 15 - Clean Co2 dataset to ‘merge’ it with temperature dataset

Step 16 - Complete ‘feature engineering’ on the merged dataset

I have honestly never heard this term before. In its simplest form though, I believe the purpose of this step is to create new columns that are easier for the model to use. Like calculating averages or converting the format data from ‘January 7, 2014’ to ‘01/07/2014’ just as examples.

Step 17 - Predict humidity and Co2 using a linear regression model

Interestingly, there is a pretty good MSE result for humidity at 0.931 but a horrible one for Co2 at 97.53 - I don’t fully understand why.

Step 18 - Try to identify why Co2 is so much worse

Step 19 - Check to make sure ChatGPT isn’t hallucinating

Step 20 - Create a scatter chart showing temp, humidity, and Co2 data

Step 21 - Split Co2 data into day/night based on observed fluctuations

Step 22 - Provide context about why there are outliers in the dataset

At this point, I gave up on trying to predict Co2 data accurately.

Takeaway #1 - Why a growth chamber/greenhouse dataset is good to learn on.

Takeaway #2 - How does AI use ‘machine learning’ to get smarter?

Peter · April 18, 2024, 11:47am

@haley I know that you recently created your own recipe and am very curious to hear about that experience. If you have access to ChatGPT Plus, I would encourage you try using the ‘Recipe Builder’ GPT we built which is linked above.

Essentially, I used a ‘two-shot’ prompt - meaning I gave it two examples of what a MARSfarm recipe was for it to learn from (upload file feature only available in plus, if you’re on free tier just copy/paste contents of the recipe manually) and then asked it to generate a third using that knowledge.

*note: I had to add a second part to my prompt because GPT is lazy - all I did was say ‘do it’ to make it generate the new file instead of just talking about it.

Comparison of ChatGPT-4 to the Recipe Builder GPT

it performs with identical promps to a regular GPT-4 (not custom GPT with several recipes/instructions pre-loaded) model:

Regular ChatGPT-4

Custom Recipe Builder ‘GPT’ created by MARSfarm

After searching it’s knowledge for about a minute, the GPT came up with a mostly complete recipe. Once again though, I had to use the second ‘do it’ prompt in order to get it to actually create a downloadable file.

Honestly, this shows me that regular GPT-4 is probably just about as good as my customized GPT. I honestly don’t think that was true about 2 months ago - which says something about the speed at which it can learn exactly what niche expectations people will have from it.