Worlds First AI Reef

[Arkayology]

This is how it starts. Terminator warned us, but in our hubris, we didn't listen.

 

[Reefer Matt]

The only thing I question is why put the animals at risk for this? What’s the reward if successful?

 

[MikeyZo]

This is how it starts. Terminator warned us, but in our hubris, we didn't listen.

It’s also like Jurassic Park. Just because we could doesn’t mean we should.

 

[Bitter old man]

Seems like a lot of people scared of AI.

 

[MikeyZo]

Seems like a lot of people scared of AI.

It’s not AI, it’s what people decided to do with it that scares me.

 

[Randy Holmes-Farley]

Seems like a lot of people scared of AI.

I’m afraid of AI directly controlling things in the real world that can cause harm if it makes a “mistake”, or just makes a decision that a normal person would not.

There are just too many scenarios that a programmer may not think even needs to be said, that an AI may decide is appropriate to accomplish it goal.

 

[GSPClown94]

There are just too many scenarios that a programmer may not think even needs to be said

Just ask the AI come up with all the scenarios, just be sure to remind it not to hallucinate or forget any edge cases. I’m sure it will be fine

 

[Arkayology]

I’m afraid of AI directly controlling things in the real world that can cause harm if it makes a “mistake”, or just makes a decision that a normal person would not.

Seems like a lot of people scared of AI.

There are just too many scenarios that a programmer may not think even needs to be said, that an AI may decide is appropriate to accomplish it goal.

I question how much the average person is going to benefit from AI. I work in higher ed as a professor. I have a lot of students who are using AI to help them write papers, create outlines, and help answer basic questions not addressed in class. The problem is that a lot of the information that these AI's are giving are just plain wrong or severely oversimplified. I also see them relying on AI to answer questions and tell them what to do at the expense of exercising critical thinking and logic. Personally, I think that the promises of AI (to-date) have been grossly overestimated. The entire industry is being invested in based on a hunch, promises, and potential that has not been realized yet (re: hedging). I also believe that not all humanity will benefit from this technology equally. There is likely only a small group of people who will get rich off of this tech. There are a lot more workers that will need to find other jobs or retrain for different industries as a result. Overall, the worker loses here, and the average citizen gains little. Meanwhile, AI data centers are using immense amounts of water and energy, raising prices of these resources for local communities, and adding little to local economies compared to what they are projected to add to the national/global economy.

I see the same thing happening in reefing. There are a LOT of people who are posting AI-driven responses and information on this forum. Let's say that an AI can completely control a reef for you and that it will make all of the changes that are required. Is the information correct? Is it beneficial for the reefer in the long run? Also, from a hobbyist POV, what's the point of that? For most, this is a hobby, not a job. I have a reef to give me something to do. I don't want an AI to control everything. I bet a lot of reefers here would agree. The average person who has not invested a lot in this hobby might be interested in an AI giving them a box of water that they don't have to do anything with though.

 

[drblank1]

I am in the IT industry and heavily into AI enablement. As of now, AI is replacing menial tasks and letting resources concentrate on higher-value tasks. And recently, our company has been able to create a predictive AI agent to reduce new chemical formulations failures by over 90%. And we are using AI in our software development groups to get us ~75% of initial code development. Over time, we will get up to 90%. AI will never become sentient. Only God can create new life. But AI will lead to layoffs and a change in workforce focus.

Despite the naysayers, this is an interesting project. Good luck.

 

[Randy Holmes-Farley]

AI will never become sentient.

While I think that is incorrect, you may be right if the nonsentient AI's knock us back to the stone age before we ever make sentient ones. lol

 

[Arkayology]

While I think that is incorrect, you may be right if the nonsentient AI's knock us back to the stone age before we ever make sentient ones. lol

Bring it on! I am ready for the stone age!

 

[Gumbies R Us]

Not really sold on AI, but curious to see how the tank turns out

 

[vahegan]

You can probably use OpenClaw and give general instructions on how to run the tank, and it can brief you every morning on the decisions it has made and why, so that you can greenlight them. That way you can avoid failure if it goes nuts.

 

[christinna77]

That sounds cool in theory… but I'd personally never let AI fully control my tank.

Like, recommendations? Sure. Let it have access to all the inputs and analyze trends, suggest dosing tweaks, flag weird patterns from a camera feed, detect a fish jumping out, whatever. That part is awesome.

But giving it direct access to outputs so it can just turn things on or change dosing on its own? No way. Especially with living animals involved.

There have to be guardrails. Hard ones. Human confirmation before anything actually changes. If something wants to increase dosing or turn off a pump, it should ask first. Always.

I work at one of the biggest developer platforms and even there we don't let fully autonomous agents just run wild in production. There are still approvals needed from humans. And that's just software. Not fish and corals that can literally die from one bad decision.

 

[BeanAnimal]

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

There are just too many scenarios that a programmer may not think even needs to be said, that an AI may decide is appropriate to accomplish it goal.

As darkly funny as that actually is, it illustrates much of the issue. I am going to harp in this to those who will listen.

The agent can’t be spiteful. The model can be programmed to be punitive or biased to certain contexts, but let's ignore that.

To your point above. The LLM can misinterpret context or take an action you told it not to take, even if rules are in place.

“You introduced a freshwater molly, so I lowered salinity to brackish” is as reasonable as “You introduced a freshwater molly, but there are corals and tangs, so I will not lower salinity.”

It can’t resolve that conflict unless you explicitly prepare it to do so. But, t may or may not retrieve that contet or instruction in that moment.

The system is probabilistic. The LLM fills in blanks that appear reasonable based on patterns the static model learned. Conflicts do not resolve themselves and the agent can't learn to resolve them on its own. This is an important part that the OP appears to be missing.

It does not develop an evolving model. The underlying model does not change. It only considers what is currently in context. Humans update their understanding continually and reasoning evolves.

Human: “I touch fire, I get burned. Things that glow red and radiate heat are hot. Therefore, don’t touch glowing red things, they burn. They can also set other things on fire.”

The agent does not internalize that the way a human does. An LLM does not evolve its understanding over time. It does not build or retain an internal model from experience. It does not permanently connect cause and effect. You and I do.

This is very important. Those directives may or may not be applied consistently. Over time, you end up with a growing pile of context dependent rules that are not reliably predictable. As the list grows, so does the complexity and chances for error, the exact opposite of what the list aims to remove.

Despite the naysayers, this is an interesting project. Good luck.

“Naysayer” is being misapplied here. What was proposed is very different from using an LLM agent for closed loop analysis.

"It can code better than us" is often pointed out as proof that it can do anything. However, coding is a rather low bar. Those who write LLMs are coders. Programming languages and patterns are extremely well defined and rather simple building blocks. The model has enormous amounts of training data for code and code context. This is where an LLM can excel and those building the model can easily verify its output. They are coders. That is why these LLMs are so "good" at coding.

In any case, The OP is conflating the agent and the model it runs on. He is not building a model, and the agent cannot build or update its own model. It cannot learn. What he is proposing requires it to learn. For this to work he must build the logic, the bounds, the correlations, the do's and don'ts. IT is not AI running a reef. It is human defined controller logic running the system. The LLM agent is just the interface on top of it. It is not the decision maker. This is a huge distinction. The "OPEN CLAW" advice above illustrates this.

Your closing would be more appropriately worded "Despite being improperly framed as AI running a reef, building a controller with the help of an LLM is an interesting project."

 

[BeanAnimal]

This may help with the context and directive part a bit:

https://medium.com/@p.akhrameev/the-ais-messy-desk-why-your-chatgpt-suddenly-forgets-what-you-were-talking-about-70b19b4708a7

 

[Randy Holmes-Farley]

Assuming it is an AI generated image, this listing picture of an item I recently bought shows the AI cannot even properly reproduce words that it knows are on the object it is mimicing.

Does not give one confidence... lol

A real Ebay Weirdity on my conductivity meter purchase???

Deep in my Tank and Learn thread I posted about the VWR conductivity meter I bought on ebay. I'm perfectly happy with it, but as I was reposting the picture of it from the listing into a new thread, I noticed some really strange things in the picture. Here's the actual meter that I took a...

www.reef2reef.com

 

[dvgyfresh]

I remember the AI drone from military decided that in order to have a 100% success rate for survival (I think it was to help pilots that need to drop out of plane) it killed everyone lol no pilots = 100% success!

 

[areefer01]

I remember the AI drone from military decided that in order to have a 100% success rate for survival (I think it was to help pilots that need to drop out of plane) it killed everyone lol no pilots = 100% success!

Huh?

 

[dvgyfresh]

Huh?

It’s a simulation not real life , for military training