As I referenced in one of the update posts recently - I lost my job back in Sept 2022. Since then, I've been freelancing as a software developer and one of the things I've been working on for a client has been Artificial Intelligence, specifically OpenAI (ChatGPT) and their associated APIs (Application Programming Interfaces). I know, it's all the rage right now. In my case, it's in a healthcare related setting, but it's opened me up to some pretty fascinating avenues. and naturally, I started exploring what it could tell me about tomatoes.
Let's take a step back though. For those of you who may not know what AI is, at this stage of development, it's basically a way to extract information from a "trained model" using natural language. OK, so what does that mean?
A "trained model" is, for all intents and purposes, a computer program that has been told to ingest a large volume of text. If you've ever seen the movie Rain Man, or are otherwise familiar with Kim Peek (the person Rain Man is based on), imagine this person, who has perfect recall, has been "fed" a certain volume of information. This could be the Encyclopedia Britannica, or all Baseball stats of every player since 1910. Maybe it's every recipe ever published in a Better Homes and Gardens cookbook. Maybe it's all of the above.
So as you might expect, you could ask this person a question, and if they had this knowledge in their head, they could answer it. Now imagine not only could it spit back facts, but also use this knowledge to come up with new information derived from those facts.
At it's core, ChatGPT is a LLM, or Large Language Model. The way it operates is that it uses probability scores and a sliding scale value to determine how much variability it should have in its response in terms of picking the most probable next word. It doesn't have "intelligence" the way you and I likely think of it, it simply prints out the most likely next word, word after word, until it is done. But, you can craft a very explicit prompt for it which can drastically alter its response, and some people have come with some really interesting prompts for various tasks.
"How does this apply to tomato breeding?" you are likely asking at this point. Well, as an example, I started off with some things I already knew the answer to as I started exploring what it could for us.
For example, I told it I wanted to cross an open pollinated tomato and a micro tomato, looking for certain traits, and asked it how I should approach that project. Due to my own prior knowledge (thanks to dfollett), I identified one error right away in its response. That error was likely due to my lack of specificity, but it did clear things up once I asked about it.
Here is how my first attempt with ChatGPT went:
Do you see the error it made? Hint: it's in step 5.
Ah ha! OK, not bad, it introduced some sciency terms and gave some more explanation about my question. Let's keep going
Hmm it still didn't give me the answer I was looking for, and which I knew to be true. Let's see if I can coach it in that direction
There we go, that's what I was getting at but it still didn't give me the key bit I was looking for. As we know, thanks to Dan, if you make the micro the mother or pollen receptor, and you grow out the seeds, all of the F1 should exhibit a indeterminate growth habit. If they end up as micros, you know that flower self-pollinated and is not the cross you were attempting to make.
In the next reply, I will explore some more, and try driving towards specific objectives.
Let's take a step back though. For those of you who may not know what AI is, at this stage of development, it's basically a way to extract information from a "trained model" using natural language. OK, so what does that mean?
A "trained model" is, for all intents and purposes, a computer program that has been told to ingest a large volume of text. If you've ever seen the movie Rain Man, or are otherwise familiar with Kim Peek (the person Rain Man is based on), imagine this person, who has perfect recall, has been "fed" a certain volume of information. This could be the Encyclopedia Britannica, or all Baseball stats of every player since 1910. Maybe it's every recipe ever published in a Better Homes and Gardens cookbook. Maybe it's all of the above.
So as you might expect, you could ask this person a question, and if they had this knowledge in their head, they could answer it. Now imagine not only could it spit back facts, but also use this knowledge to come up with new information derived from those facts.
At it's core, ChatGPT is a LLM, or Large Language Model. The way it operates is that it uses probability scores and a sliding scale value to determine how much variability it should have in its response in terms of picking the most probable next word. It doesn't have "intelligence" the way you and I likely think of it, it simply prints out the most likely next word, word after word, until it is done. But, you can craft a very explicit prompt for it which can drastically alter its response, and some people have come with some really interesting prompts for various tasks.
"How does this apply to tomato breeding?" you are likely asking at this point. Well, as an example, I started off with some things I already knew the answer to as I started exploring what it could for us.
For example, I told it I wanted to cross an open pollinated tomato and a micro tomato, looking for certain traits, and asked it how I should approach that project. Due to my own prior knowledge (thanks to dfollett), I identified one error right away in its response. That error was likely due to my lack of specificity, but it did clear things up once I asked about it.
Here is how my first attempt with ChatGPT went:
Do you see the error it made? Hint: it's in step 5.
Ah ha! OK, not bad, it introduced some sciency terms and gave some more explanation about my question. Let's keep going
Hmm it still didn't give me the answer I was looking for, and which I knew to be true. Let's see if I can coach it in that direction
There we go, that's what I was getting at but it still didn't give me the key bit I was looking for. As we know, thanks to Dan, if you make the micro the mother or pollen receptor, and you grow out the seeds, all of the F1 should exhibit a indeterminate growth habit. If they end up as micros, you know that flower self-pollinated and is not the cross you were attempting to make.
In the next reply, I will explore some more, and try driving towards specific objectives.