Using AI for chemistry/cycling assistance?

[popalud]

Curious to how many people use AI (CGPT/Gemini/CoPilot) to assist with dialing water parameters or help with cycling tanks. I'm using Gemini and feel its providing acceptable guidance and understanding as to what's happening when I perform testing during my new tank cycle. I've prompted it to provide day to day steps, and contingencies based off my volume, results, available supplements, and test results. Anyone else using it for similar purposes have who have precautions?

 

[BryanM]

AI is fine to use, but unless you already know what you're doing you need to verify its accuracy... Our hobby isn't exactly mainstream, and even if it was it is prone to mistakes.

 

[popalud]

Yeah, I like to question/challenge the responses and push it to cite its logic/explain the science behind the changes in test results. It typically cites some reliable recourses (including R2R). Its been helpful as a tool to summarize all of the thousands of perspectives, and provide real time feedback based on what's happening specific to my tank, and potential issues based on what's changing.

 

[Submerge]

AI is super helpful with math.
I find the results I get from taking a photo of a color change test and letting AI read it to be less noisy than my eyes. I'd pretty much find salifert unusable without AI - I just haven't got the right eyeballs to see the colors accurately.
I find AI to be more useful than google when doing research. I just went down the rabbit hole on phyto -> pods / rotifers and ended up building a different system than what I originally imagined based on real science. It helps to really drill down on a topic.

I find AI to be bias confirming. It tells you what you want to hear a bit, but you can train that out easily enough.
I also find AI making a lot of assumptions on my behalf. Frustrating, but can also be trained out.

My advice is to pick one and train it. Build a memory doc that spans sessions and force it to use that doc on every session so that your expectations are built into the answers. I personally like Claude, but I spend a lot of time in Claude Code so I'm biased and have n ot tried them all. That said, more often than not I get a solid answer. I feel like the answer is better than a cross section of posters on a forum. I've tried both, and for the most part I ask AI before a FB group or forum. I know that's a standpoint that will get me some hate - but there is just too much misinformation and "forum wisdom" floating around disguised as trying to help.

The only way you will know for yourself is to ask the same questions both online and to AI. So much of it comes down to the way you ask the questions.

 

[popalud]

I refuse to use CoPilot as it has given me downright incorrect information that I was able to verify within minutes. I have found myself often double checking and finding contradictory information when using CoPilot (enterprise version).

I asked Gemini about acclimation, and given the fish i am ordering will ship "next day air", Gemini is suggesting the "plop and drop" method citing the "ammonia flash" as being far more detrimental to the fish than salinity shock. Its a compelling take and seems logical, but i need to do more research.

 

[Submerge]

I refuse to use CoPilot as it has given me downright incorrect information that I was able to verify within minutes. I have found myself often double checking and finding contradictory information when using CoPilot (enterprise version).

I asked Gemini about acclimation, and given the fish i am ordering will ship "next day air", Gemini is suggesting the "plop and drop" method citing the "ammonia flash" as being far more detrimental to the fish than salinity shock. Its a compelling take and seems logical, but i need to do more research.

That is in fact what most shippers recommend these days. Get them up to temp with a 20-30 minute float then plop the fish in. As soon as you open the bag and the co2/o2 concentrations change, the PH drops making ammonia toxic that was less so at a higher ph. The whole drip acclamation thing is a big no-no these days.

 

[Gumbies R Us]

@Randy Holmes-Farley has made an article about this!

 

[Randy Holmes-Farley Verified Community Expert]

Curious to how many people use AI (CGPT/Gemini/CoPilot) to assist with dialing water parameters or help with cycling tanks. I'm using Gemini and feel its providing acceptable guidance and understanding as to what's happening when I perform testing during my new tank cycle. I've prompted it to provide day to day steps, and contingencies based off my volume, results, available supplements, and test results. Anyone else using it for similar purposes have who have precautions?

There really isn’t that much involved to cycle a tank that would seem to need ai support.

If live rock, either do nothing or add 2 ppm ammonia and see it decline.

If dead rock, add bacteria, then 2 ppm ammonia, and watch it decline.

What is the ai advising you to do?

 

[Randy Holmes-Farley Verified Community Expert]

That is in fact what most shippers recommend these days. Get them up to temp with a 20-30 minute float then plop the fish in. As soon as you open the bag and the co2/o2 concentrations change, the PH drops making ammonia toxic that was less so at a higher ph. The whole drip acclamation thing is a big no-no these days.

That is what I did with my recent Dr Reef order and it was easy and worked fine. :)

 

[BeanAnimal]

I find AI to be bias confirming. It tells you what you want to hear a bit, but you can train that out easily enough.
I also find AI making a lot of assumptions on my behalf. Frustrating, but can also be trained out.

No, they can't. The AI does not learn from you, the model is set. You can try to get it to remember things for context and sometimes it will. Think of it as the movie "50 First Dates". It only knows what you tell each time. It has a limited memory of things that it considers each time you tell it something new.

My advice is to pick one and train it. Build a memory doc that spans sessions and force it to use that doc on every session so that your expectations are built into the answers.

This is not training. It is you giving it a list of things to consider along with the other information that it pulls in for the request.

"Never use Emojis" and "my tank is on the 2nd floor and I use pond water" are suggestions to consider. It may or may not pull them in when required for an appropriate response.

I feel like the answer is better than a cross section of posters on a forum.

It is just a statistical guess based on how often those words appear together on the forums and its other training data. Right or wrong, it doesn't know, it just builds an answer based on deterministic probability applied to its training data.

 

[fish_collector]

Since there are a lot of urban myths on forums regarding cycling a salt water tank, AI may just repeat the myths because that’s what’s out there. It has no method to determine what the real deal is. I don’t know why anyone would want to consider using it for this hobby, or anything for that matter. I personally don’t find it useful at all, but I still like to read books and magazines so what do I know.

 

[Submerge]

No, they can't. The AI does not learn from you, the model is set. You can try to get it to remember things for context and sometimes it will. Think of it as the movie "50 First Dates". It only knows what you tell each time. It has a limited memory of things that it considers each time you tell it something new.


This is not training. It is you giving it a list of things to consider along with the other information that it pulls in for the request.

"Never use Emojis" and "my tank is on the 2nd floor and I use pond water" are suggestions to consider. It may or may not pull them in when required for an appropriate response.


It is just a statistical guess based on how often those words appear together on the forums and its other training data. Right or wrong, it doesn't know, it just builds an answer based on deterministic probability applied to its training data.

Sounds like someone who does not know how to use ai. Performance is user based.

 

[Randy Holmes-Farley Verified Community Expert]

Sounds like someone who does not know how to use ai. Performance is user based.

Uh oh, expect a strong response. lol

I think Bean does use ai for other things (writing code, I think), and is well informed.

 

[Submerge]

Uh oh, expect a strong response. lol

I think Bean does use ai for other things (writing code, I think), and is well informed.

Good he’s not the only one.

 

[sixty_reefer]

 

[BeanAnimal]

Uh oh, expect a strong response. lol

I think Bean does use ai for other things (writing code, I think), and is well informed.

Not much to respond to. His reply was directed at me personally instead of the point that I made. So I will just make the point again.

Sounds like someone who does not know how to use ai. Performance is user based.

Using AI well can improve its output, but that is not the same as training the model. You can provide better prompts and context, just as you can with a normal search engine. You can add memory documents and constraints that influence the output. That does not change the model or make its performance “user based.”

Learning to drive well does not make you an expert in how cars work, nor does it change the car’s performance. It only means you have learned to leverage the tool closer to the limits of its performance.

Performance is not user based. Outcomes are user influenced.

My issue with AI in the context of this thread is its performance regarding the accuracy of the information it provides, and the limitations of that performance and training data. In this case, a poorly trained user just makes the outcome even less reliable.

 

[areefer01]

Probably should learn to walk first. If you don't understand what to look for or how to keep water then you will never learn it if you ask AI for help. It is a crutch. It is not the foundation.

 

[Submerge]

Not much to respond to. His reply was directed at me personally instead of the point that I made. So I will just make the point again.



Using AI well can improve its output, but that is not the same as training the model. You can provide better prompts and context, just as you can with a normal search engine. You can add memory documents and constraints that influence the output. That does not change the model or make its performance “user based.”

Learning to drive well does not make you an expert in how cars work, nor does it change the car’s performance. It only means you have learned to leverage the tool closer to the limits of its performance.

Performance is not user based. Outcomes are user influenced.

My issue with AI in the context of this thread is its performance regarding the accuracy of the information it provides, and the limitations of that performance and training data. In this case, a poorly trained user just makes the outcome even less reliable.

fair. And I have read many of your posts on this topic. My issue is the way you frame it - that AI cannot be relied on and the inference is that people can. You never specifically say that from what I have read, but it seems that is your position.

Let's do this. Lets re-frame the whole debate so that it can be easily understood. There are 2 primary things I want in the framework - #1. define AI. It is not one thing. #2 - is it your position that humans always provide better answers than AI in an online format? If #2 is not your position, I need to know that so we are not debating past each other.

I also would like to request that the logical fallacies like appeal to authority be left at the door. The fact that you and I use and know AI is a given for this debate. I say this mostly because I have the advantage here in that your posts and positions are readily available, but how could you know anything about me or what I do? My point to Randy that you are not the only one is in that vein. Apologies if it came out caustic. I was in the middle of something else.

Let me know your preferred way to define the debate and we can get busy. Looking forward to it.

 

[Submerge]

Probably should learn to walk first. If you don't understand what to look for or how to keep water then you will never learn it if you ask AI for help. It is a crutch. It is not the foundation.

I agree that AI is a tool. Same as a book. If you ask it nicely, it will actually only use peer reviewed published scientific articles in its responses, and will give you links to the original data from which it's statements are derived. That seems better than most other tools at our disposal today.

 

[Randy Holmes-Farley Verified Community Expert]

- is it your position that humans always provide better answers than AI in an online format?

I’ll comment on that specifically. There is always misinformation posted everywhere. But in a forum like Reef2Reef there is often a self correcting mechanism that other users correct the original misinformation. That correction is not available and/or used effectively in other settings such as Facebook and YouTube.

If an ai is drawing from reef2reef, it will see the corrections. If it is drawing from other types of sites, it won’t see the corrections that are not there.