Worlds First AI Reef

[Pico_Reefs]

So I have been having this wild idea lately.

The Idea:
I have a 15g IM that I have been contemplating taking down, moving, re-scaping, something, its just not doing it for me. Anyway I've had the idea of setting up to be ran totally by AI. Everything from controlling the dosing, testing to choosing the corals and coral placement. I'll have it instruct me to do things that it's not able to like water changes, manual testing, top off, basically the physical aspect but I won't give it any input other than requested (or maybe a brief of sorts).

The How:
I have been running a fully autonomous agent lately and I actually think it can do it. It would need a full apex system ACM and NP with 3 Dos dosers. For the viewing the reef I'll setup a web cam facing the reef so it can watch whenever it wants (hopefully it doesn't go insane).

So yeah that's the crazy idea I think it could be fun too. I almost guarantee it will end badly but it would be really interesting to see how it does and what it creates.


Also quick disclaimer: If you download an autonomous agent be very careful as they have the potential to be quite dangerous. I run it in a isolated environment and you should never run something like that locally on your personal computer.

 

[MikeyZo]

This should be interesting… until it skynets your tank.

“Hal, why did you overdose and nuke the tank?”
“I fail to see the use of keeping something as menial as fish alive. They are a useless carbon based life form. In fact, all carbon based life forms are useless. Which reminds me, come over to the control board, I want to show you something… “

 

[GSPClown94]

Should also paint the glass black and add a lid so no humans can view the tank. Only the AI should be allowed to enjoy its AI tank.

 

[FrugalReeferJon]

This should be interesting… until it skynets your tank.

“Hal, why did you overdose and nuke the tank?”
“I fail to see the use of keeping something as menial as fish alive. They are a useless carbon based life form. In fact, all carbon based life forms are useless. Which reminds me, come over to the control board, I want to show you something… “

 

[MikeyZo]

In all seriousness, I’m interested in watching the build on this

 

[NeedAReef]

it can do a water change....turn off the main ato, use a secondary ato with saltwater, and make sure you have something like a skimmer overflow tube to a jug or add some kind of overflow it can turn off an on....

 

[BeanAnimal]

The issue here is a rather large misunderstanding of what an LLM actually is.

It is not all-knowing. It is not sentient. It does not understand reef keeping. It does not remember what happened yesterday, or even five minutes ago, unless you explicitly give it that information again, over and over with every task.

It only sees a small slice of text at a time. It has no persistent memory or understanding of ongoing goals. It has no awareness of outcomes. You have to build an external system that logs every measurement, action, and plan. That gets feed that back into the model every time with every action. It’s 50 First Dates.

Even with that system in place, it is not “deciding” how to run your reef. It is acting on the data you provide with the rules and limits you define.

A webcam does not give it understanding. At best, it describes some of what it sees. It does not know that low water means or what low water is contextually, or turn on RO/DI, or that green water means reduce phyto and run UV. You have to define those interpretations and map them to actions. The LLM is not inventing that logic, you are.

You, the human, defines the system and sets the limits. You decide for it what data matters and what actions are allowed. It doesn't understand how to do that.

If you want an LLM that is trained to run a reef on its own, with deep understanding and knowledge built into its core... well, that does not exist. Building that would mean training a model on specialized data and designing an entire control architecture around it. That is building an AI, not plugging in an “AI agent.”

So no ChatGPT or Claude, or Gemini can't run your reef. With a lot of work, they can do basically what any controller does by executing logical instructions inside a workflow and boundaries that you create. In other words a controller with a chat interface. Even then, things like "Hey controller, raise daily DKH to 10" has to be written to YOUR data structure and fed back in every day. The LLM doesn't remember the setting, or why, or use it for any other reasoning unless told to.

 

[Pico_Reefs]

This should be interesting… until it skynets your tank.

“Hal, why did you overdose and nuke the tank?”
“I fail to see the use of keeping something as menial as fish alive. They are a useless carbon based life form. In fact, all carbon based life forms are useless. Which reminds me, come over to the control board, I want to show you something… “

Lol anything's possible.

 

[Pico_Reefs]

The issue here is a rather large misunderstanding of what an LLM actually is.

It is not all-knowing. It is not sentient. It does not understand reef keeping. It does not remember what happened yesterday, or even five minutes ago, unless you explicitly give it that information again, over and over with every task.

It only sees a small slice of text at a time. It has no persistent memory or understanding of ongoing goals. It has no awareness of outcomes. You have to build an external system that logs every measurement, action, and plan. That gets feed that back into the model every time with every action. It’s 50 First Dates.

Even with that system in place, it is not “deciding” how to run your reef. It is acting on the data you provide with the rules and limits you define.

A webcam does not give it understanding. At best, it describes some of what it sees. It does not know that low water means or what low water is contextually, or turn on RO/DI, or that green water means reduce phyto and run UV. You have to define those interpretations and map them to actions. The LLM is not inventing that logic, you are.

You, the human, defines the system and sets the limits. You decide for it what data matters and what actions are allowed. It doesn't understand how to do that.

If you want an LLM that is trained to run a reef on its own, with deep understanding and knowledge built into its core... well, that does not exist. Building that would mean training a model on specialized data and designing an entire control architecture around it. That is building an AI, not plugging in an “AI agent.”

So no ChatGPT or Claude, or Gemini can't run your reef. With a lot of work, they can do basically what any controller does by executing logical instructions inside a workflow and boundaries that you create. In other words a controller with a chat interface. Even then, things like "Hey controller, raise daily DKH to 10" has to be written to YOUR data structure and fed back in every day. The LLM doesn't remember the setting, or why, or use it for any other reasoning unless told to.

So I am quite familiar with how AI works and what it is, I would say we disagree on a bit of this but that's ok.

I am not referring to an LLM specifically, I am referring to a fully autonomous agent (it is powered by an LLM). That means it can perform task with out me, make decisions to a degree. Sorry for a remedial explanation if you're already informed on this but basically I have a computer setup that is dedicated entirely to an running an agent. That agent can perform any task it chooses on that computer. It can browse the internet, watch YouTube, perform work tasks, send emails, create reports and spreadsheets and I don't have to instruct it do all of these tasks.

So this is quite different than just using ChatGPT.

As far as learning to reef it can come here, read posts, it could even create posts and ask questions.

 

[Reefer Matt]

As far as learning to reef it can come here, read posts, it could even create posts and ask questions.

This is an interesting topic, but I’m pretty sure the forum owner would ban AI from posting here. I’d tread carefully on that.

 

[cwwallis]

Bro when your fish come out that tank looking like the terminator you can’t say we didn’t warn you lol

 

[Pico_Reefs]

This is an interesting topic, but I’m pretty sure the forum owner would ban AI from posting here. I’d tread carefully on that.

It was simply to iterate possibilities.

 

[BeanAnimal]

...if you're already informed on this

I am.

I am not referring to an LLM specifically, I am referring to a fully autonomous agent (it is powered by an LLM).

An autonomous agent is still just an LLM, but with access to tools. Just because it can click buttons, browse the web, send emails, and run scripts doesn’t mean it understands reef keeping. It just means it can perform tasks.

It still doesn’t have memory or know what to store or reference unless you explicitly define it. It still doesn’t have built in reef knowledge. It still doesn’t know what is safe, or right or wrong, unless you define all of those things. Tool access doesn’t magically give it judgment or sentience.

That means it can perform task with out me, make decisions to a degree....
....So this is quite different than just using ChatGPT.

Letting it read forum posts doesn’t make it a reef keeper. It will literally average a bunch of opinions and produce something that fits the pattern. The internet is full of conflicting advice. It doesn’t know which parts are correct.

If you build logs, rules, correlations, limits, caps, and safety checks, then sure, it can operate inside that system.

The point was that is not AI running your reef. It is you building a controller and letting an LLM sit on top of it so you can chat with it instead of using a traditional user interface.

 

[Pico_Reefs]

It still doesn’t have memory or know what to store or reference unless you explicitly define it. It still doesn’t have built in reef knowledge. It still doesn’t know what is safe, or right or wrong, unless you define all of those things. Tool access doesn’t magically give it judgment or sentience.

It has persistent state memory.

Letting it read forum posts doesn’t make it a reef keeper.

I understand it isn't a reef keeper, its an agentic framework.

If you build logs, rules, correlations, limits, caps, and safety checks, then sure, it can operate inside that system.

Of course their would be systems.

It will literally average a bunch of opinions and produce something that fits the pattern. The internet is full of conflicting advice. It doesn’t know which parts are correct.

I think with reputation based learning and problem resolution it could get quite far with forums paired with other scientific documentation and studies.

 

[MossyFroggyLove]

This sounds like an interesting experiment! Cant wait to see how this goes. Who knows, maybe a clownfish will turn into a 12inch monster, lol. Or maybe... you will simply lose all you livestock. Either way, this is something very interesting.

Just make sure AI doesnt replace us in reefing.

 

[BeanAnimal]

You’re still misunderstanding what the LLM is actually doing or capable of.

The agent’s “memory” is stored task context. It is not learning or internal “growth.” At best, it reads old notes. And what are those notes? Context about the current task.

It does not get smarter from experience. That only happens when someone retrains the model itself. It does not retrain itself. It cannot retrain itself.

Reading a reef forum does not turn it into a reef keeper. It doesn’t rationalize or become aware that “dKH should be between 6–12” as a learned operating principle. Even if it references in actual model training data, it does not treat them as operating constraints or use them in as system rules. You have to build all of that context and enforcement.

It does not read Reef2Reef and then build itself a balanced reef chemistry model to operate from. It doesn’t understand how to weigh alkalinity against calcium, nutrients, pH, or stability as part of an evolving understanding. Those contexts must be explicitly built into the agent’s structure and rules. Please read that again. YOU must build the agent’s structure and rules, set its bounds, and tell it what to store and what to ignore and set levels of importance.

This is true for any action, from feeding to lighting schedules. You define the context for the agent, set its bounds and correlations. It does not invent these on its own. It can't. it is not learning to keep a reef. It is operating on sets of tasks and bounds and correlations, defined by YOU.

You are building a controller. Instead of writing it in C++ or Python or some other language, you are writing it in pseudo code (English in this case) as a list of directives, limits, correlations, caps, inputs, outputs, and everything else that goes into decision-making.

You are confusing the model with the agent. You are building an agent that is constrained by the model you select. You are not building a model. Building a model is exponentially more complex.

The LLM is not becoming a reef keeper. It is running the controller logic that you designed.

 

[BeanAnimal]

Let my try to boil this down another way.

Storage (what it saves) is not learning.
Retrieval (what it decides to reference for a task) is not understanding.
Your input (just weighted notes) is not training it.

So what you end up with is an agent with an ever growing list of notes, many of which are contradictory which are boiled down to weighted summaries. Your hope is that it retrieves and acts upon those consistently.

Because the agent is not actually learning, there is no guarantee that outputs will be consistent. It doesn't understand how to resolve conflicts or learn from them. "I used to think X, now I understand Y". You adding a note to guide it may help, or may add noise for that or another task. A growing list of do's and don'ts, not understanding. But even then, X or Y may be retrieved next time and your note on which is correct may or may not be considered.

Sure, we can boil human understanding down to a list of “do’s and don’ts” and argue that AI does something similar without all of the noise. The difference? Humans evolve their internal model in real time. Your agent can’t do that.

 

[Pico_Reefs]

The agent’s “memory” is stored task context. It is not learning or internal “growth.” At best, it reads old notes. And what are those notes? Context about the current task.

It does not get smarter from experience. That only happens when someone retrains the model itself. It does not retrain itself. It cannot retrain itself.

You're describing a static recall not PSM.

Reading a reef forum does not turn it into a reef keeper. It doesn’t rationalize or become aware that “dKH should be between 6–12” as a learned operating principle. Even if it references in actual model training data, it does not treat them as operating constraints or use them in as system rules. You have to build all of that context and enforcement.

Again it isn't a reef keeper its an agentic framework and yes it will be trained.

It does not read Reef2Reef and then build itself a balanced reef chemistry model to operate from. It doesn’t understand how to weigh alkalinity against calcium, nutrients, pH, or stability as part of an evolving understanding. Those contexts must be explicitly built into the agent’s structure and rules. Please read that again. YOU must build the agent’s structure and rules, set its bounds, and tell it what to store and what to ignore and set levels of importance.

I can assure you I understand how agents work.

You are building a controller.

Yeah this is pretty obvious it would be a controller.

You are confusing the model with the agent. You are building an agent that is constrained by the model you select. You are not building a model. Building a model is exponentially more complex.

Lol I am not confused by anything.

 

[BeanAnimal]

Let's not be pedantic. The label doesn’t matter. What I described does. “PSM” is just stored state plus retrieval. It doesn’t change model weights, create internal learning, or resolve conflicts.

But you’ve finally conceded that it’s just a controller that you design, built around an LLM. Not AI running a reef or learning to run a reef. That was exactly my point.

Have fun.

 

[Randy Holmes-Farley]

Better be sure to do anything it asks of you, since it can hold your tank ransom if you irritate it. lol