AI Based Agroecology Tools and AI Safety

There are now multiple available artificial intelligence based apps which have relevance to agroecology and permaculture; Many are free.   Full time permaculture designers and large agroecology projects and will probably tend to have higher effectiveness when using appropriate AI tools.

AI Intro
Typical contemporary machine learning software is consistently accurate at solving one specific, narrow thing.  Machine learning developers write a program for a task, then find lots of data, then using graphics processing unit based computers - they train and evaluate the program, then they might make it user ready.  An example of a modern AI tool is the Deepmind AlphaGo which beat the best Go player in the world and go is among the hardest games in the world to be the best at. This level of AI is typically a Recursive Neural Net, programmed in the Python programming language using TensorFlow library with a Keras wrapper.

AI Safety
AI tech is not a huge risk yet, but most experts I've read have voiced a need to improve AI's safety.  There are risks of AI being misused by folks as well as the risk of general AI being developed and then it becoming both powerful and bad.  Regulators still have not even formally explored enacting safety measures around AI*.  Therefore its currently up to users and developers to insure their AI tools are as safe as possible. If you get into AI, please stay up to date with AI safety and follow best practices when they start to come out. Its possible that using AI for good projects, in a safe and democratic ways - will increase the probability both that more people will do more good things with AI and that the likely development eventually of domain-general AI (super AI) will be a great thing for everyone and the planet.


AI Safety Resources

    MIRI, Machine Intelligence Research Institute    
         Their "About" page is useful and readable.  Their papers are typically technical.
         https://intelligence.org/about/

    Future of Life Institute
         "Mission: To catalyze and support research and initiatives for safeguarding life and developing optimistic visions of the future, including positive ways for humanity to steer its own course considering new technologies and challenges."
         https://futureoflife.org/

User-Ready Acroecology Related AI Tools

AI For Earth
AI for Earth a set of tools run by diff earth sciences organizations; They're backed by and partially funded by Microsoft Research as part of their AI For Good project.  The seem to have by far the widest selection of user ready eco ai stuff. I assume this help them sell operating systems in the long term.  It has the following tools:

    SilvaTerra
         "Applying AI to satellite imagery can improve forest management."
          https://www.microsoft.com/en-us/ai/ai-for-earth-partners?activetab=pivot1:primaryr5

    Wild-Me
         "Accelerating wildlife population research using computer vision."
          https://www.microsoft.com/en-us/ai/ai-for-earth-partners?activetab=pivot1:primaryr6

    Ag-Analytics
         "Enabling precision conservation and more sustainable agricultural practices by applying AI to soil, yield, and tillage data."
         https://analytics.ag//

    Conservation Metrics
         "Applying machine learning to increase the scale and efficiency of wildlife surveys."
         http://conservationmetrics.com/

    iNaturalist
         "AI for Earth is working with conservationists at the California Academy of Sciences and National Geographic Society to enable citizen scientists across the globe."
         https://www.inaturalist.org/

    Technical Resources
         AI for Earth also has a wide set of earth sciences data and technical resoources available. So far it all looks free.
         https://www.microsoft.com/en-us/ai/ai-for-earth-tech-resources

Granular
Granular is the farming ERP which Nori has its farmers use. It seems focused on adding value via data analytics functions.  [Nori is a carbon market; They pay farmers to increase soil carbon] [ERP = enterprise resource planning] There are other farm ERP's out there.
https://granular.ag/


Pacific Agroecology LLC
They're a acroecology consulting firm seeming to use a fair amount of  AI.
https://pacificagroecology.com/



*: [Our forum moderators will probably limit our discussion here if we make this political so lets avoid that.]

So far AI needs training data. Millions to billions of entries. I don't see that much data being available or collectable in a permaculture context. (It needs data for every possible case…)
AI can work for games because those can be simulated. I have not seen a correct simulation of nature yet.

This reminds me of my vet when he came and tried to talk us into doing ultrasounds on our sheep to determine if they had single, twins or triplets. His idea was, if you know what each ewe is going to have for lambs, for the 5 months she is with lamb, you can feed her appropriately. So I asked him how much he thought it would cost, and he said a few thousand dollars.

So I asked him, "Why don't I feed every ewe for having twins then, and what little bit of money I spend more for the ewes having singles, and the few mortality losses I would have for the triplets, would be a lot less than having every sheep ultrasounded."

We live in the information age, but just because it is information, does not mean it is useful, or more importantly...will save us money.

Sebastian Köln wrote:So far AI needs training data. Millions to billions of entries. I don't see that much data being available or collectable in a permaculture context. (It needs data for every possible case…)
AI can work for games because those can be simulated. I have not seen a correct simulation of nature yet.

Thank you Sebastian,
You're mostly correct.  I must not have presented myself clearly because I think you missed my point.  You're correct that larger and larger data sets are required to solve more and more complex (non-linear) problems, such as "how should I optimally design this whole agroecosystem a-to-z?".  Thus no software tool that complex exists for agroecology yet.  Yet, there are useful apps available currently. They're not a-to-z agroecology design and planning super AI. They are tools for full-time designers to help these folks do their job more effectively, and definitely not replacements for permaculture designers.  

The apps available today can certainly answer questions of the complexity magnitude of "what's the last 20 year's tree health trend on this 200 acre site my client wants me to do a full permaculture design on?"  I think current apps can even start to answer questions on the order of magnitude of complexity of "whats the forecasted tree health trend here - factoring in climate change, and XYZ forest modifications we might do? Also how confident is the model in this forecast?"

I see tools such as these as being helpful for managers of large agroecology or conservation or restoration or forestry projects, and/or professional permaculture/agroecology designers. This is because there will be a learning curve to getting useful results from these apps which will probably be effective for full time designers. I don't see these as being effective uses of, for example, hobby gardeners time.  Conventional ag is already using AI based tools extensively.

The tools I presented already function, right now, for each of their narrow, narrow uses.  A user only needs to add the data about their particular site and figure out the apps user interface to start get some function from these.  These apps were able to be made because there are already petabytes upon petabytes of useful, free/cheap, updated data available from satellite/aerial photos, LiDAR, near infra-red, UV Vis, hydrology, multspectral, topograp, ground cover etc.

You're totally correct that nature is very non-linear and so requires lots for data for training of new apps and lots of computational resources to train those apps quickly.  [For users to use pre-trained models it requires hardly any both data and not much GPU processing power.]  There are already some ecology simulators* and there will be more over time.  So larger parts of agroecology designs will be able to be done over time as apps improve and data increases.

*:  [For example Ecosim is an academic project which models ecosystems in order to understand evolution, predator-prey dynamics, population dynamics, specietion etc.
https://sites.google.com/site/ecosimgroup/research/ecosystem-simulation]

Travis Johnson wrote:This reminds me of my vet when he came and tried to talk us into doing ultrasounds on our sheep to determine if they had single, twins or triplets. His idea was, if you know what each ewe is going to have for lambs, for the 5 months she is with lamb, you can feed her appropriately. So I asked him how much he thought it would cost, and he said a few thousand dollars.

So I asked him, "Why don't I feed every ewe for having twins then, and what little bit of money I spend more for the ewes having singles, and the few mortality losses I would have for the triplets, would be a lot less than having every sheep ultrasounded."

We live in the information age, but just because it is information, does not mean it is useful, or more importantly...will save us money.

Sorry to hear that your vet sounds at least out-of-touch with your farm or perhaps greedy.  I agree that lots of permaculturists will not experience more effectiveness by applying currently available AI based apps to their work.  I'm equally confident that many full-time designers and many managers of large projects - on the other hand - will experience more effectiveness by using some AI based apps.

Even free AI apps do require time to learn, computer hardware, possible reduction in privacy, data collection on the design site, time to run the software and a real use for the app's findings.

By the way, thank you for sharing that experience Travis.

Serendipitous to see a post on AI safety on a permaculture site
I would recommend watching https://www.youtube.com/watch?v=2jwVDRKKDME for how amazing nature simulation visualizations can be.

I prompted gpt-3.5 api to generate a permaculture strategy for basecamp to grow all of the food I need to survive as a vegan. (it just needs to be converted into a timeline and implemented, I'm not sure how well it would work)
"You are stranded on a mountainous permaculture facility in Montana in March with a bunch of seeds, some huglekultures made of dirt, and a cold mountaintop: implement a permaculture timeline and strategy for planting this season for a low-effort high yield vegan diet"

First, I would begin by taking a careful inventory of the resources and limitations of the site, including soil quality, microclimates, water availability, and the availability of any existing infrastructure or tools.

Next, I would prioritize the planting of cold-hardy crops that can withstand the harsh conditions of a Montana winter, such as kale, spinach, chard, and other leafy greens that can be grown in containers or raised beds.

In the early spring, I would focus on preparing the hugelkultures and gradually adding layers of organic material such as straw, manure, and garden waste, while also incorporating slow-release fertilizers and soil amendments to improve soil fertility and health.

Once the soil is ready, I would plant a variety of cold-weather crops and perennials, such as broccoli, carrots, onions, garlic, and herbs, that can continue to thrive and produce throughout the growing season.

To complement these crops, I would also consider planting nitrogen-fixing cover crops, such as clover or vetch, as well as companion plants such as chamomile and marigolds that can help improve soil health and deter pests.

Throughout the season, I would also prioritize mulching and regular watering, as well as regular pruning and harvesting of crops to encourage ongoing growth and productivity.

By following these strategies, I believe it would be possible to create a low-effort, high-yield permaculture system that not only provides a healthy vegan diet but also supports the health and well-being of the local ecology and community.

The conversation is heating up, with LLM's and access to NN models and the use of vector databases and access to technology that turns queries into embeddings. Combined with the incredible amount of resources, data, and documentation we have available. It looks like we now have the potential to greatly optimize the impact that a single permaculture designer may have.

Imagine asking a question about a design problem and having the generalized answer from the LLMs as well as references to the citations from source data and the author.

I look forward to hearing your thoughts.

Corey, welcome to the forum.

I really have mixed feelings about anything AI-generated,

Are folks no longer proud of their accomplishments?

This topic seems to be more relevant by the day. I'm trying to follow developments, but it's impossible to stay on top of it, as there are new tools and use cases every week.

However, I've stumbled upon some info regarding smaller open source models trained on quality datasets which perform disproportionately well, as compared to the big guns like GPT-4 and Gemini.

It occurred to me that one of these models could be trained further on permies posts to create a "permie design buddy" of sorts we could bounce ideas off of. I'm sure these threads would make a high quality dataset of high practical value.

Useful discussions might be posted on the forums for the benefit of others.

A manually triggered bot might be created on the forums to offer some advice when there is no biological intelligence available on a topic.

Would that be considered useful or safe by permies standards?

If there's interest I could do some focused research to track down relevant sources of information for this use case.

I understand there will be resistance in many of you (I feel it too), but I think this could have awesome benefits too. Could be a way to infect more brains perhaps?