TOS Question: ChatGPT

[Johnd651]

This gave me great pause. At this point personally I am beginning to question if an actual individual is driving this.
I am expecting to see other hobbyist forums experience a similar test.
Unfortunately I am confident forums are just beginning to see the challenge.

i was thinking the same thing as i was reading through his responses on other threads, almost like a bot driving

 

[BeanAnimal]

I will not defend the current usage of chatGPT but ignoring the value added of AI or LLMs is a mistake.

Thank you, but you are not in a position to determine what constitutes a "mistake" in what I choose to value or not value. In the context of THIS forum and the type of content and interaction that I subscribe for and participate in, I have absolutely no use for an LLMs input, thoughts, clarification or "fact-checking". In fact, I don't want to see it yammer on about anything or even provide one word answers, auto-complete, search suggestions, error correction or happy thoughts, I don't want to hear from it directly or via copy and pasted nonsense from other's. You and other's may feel differently. As I said, the moment that any AI tool pops up here in any form, I am out, no questions asked. I would rather listen to fingernails screeching on a a chalkboard on repeat.

You claim that LLMs cannot fact check but they are actually really good at this.

Except that they aren't... They actually suck at it, but I already briefly explained why. As they saying goes "From the horses mouth"

There is more... a lot more. But you get the point.

You can ask an agent to summarize a paper for you and quote where in the paper each source came from. Even if you don't trust the collective knowledge in an LLM, this is valuable. I.e. don't feel like reading 10 pages of a thread, ask an LLM to summarize what was said.

Except that you have no way of knowing if the summary is correct or relevant unless you know the content that it summarized... meaning you really should read the 10 pages or whatever. We can fill this entire forum with examples of AI failures, hallucinations and user driven (often inadvertent) biased output.

Fun fact - as the AI spreads crap to the masses, those masses repeat it and the AI sucks it back in as "new fact" or "new information". Not only is it a negative feedback loop, it is exponential. So the emerging struggle is not making LLMs "smarter", it is keeping them from inbreeding themselves to utter stupidity -- but some would argue that battle has already been lost even if we don't yet know it.

Agents that could assist with interpretation/summarizing/chatting on threads would be a huge win for this site.

In your opinion, not mine.

 

[mizimmer90]

Thank you, but you are not in a position to determine what constitutes a "mistake" in what I choose to value or not value.

You are entitled to your own opinions but not your own facts :).

In the context of THIS forum and the type of content and interaction that I subscribe for and participate in, I have absolutely no use for an LLMs input, thoughts, clarification or "fact-checking". In fact, I don't want to see it yammer on about anything or even provide one word answers, auto-complete, search suggestions, error correction or happy thoughts, I don't want to hear from it directly or via copy and pasted nonsense from other's. You and other's may feel differently. As I said, the moment that any AI tool pops up here in any form, I am out, no questions asked. I would rather listen to fingernails screeching on a a chalkboard on repeat.

With all due respect, this feels a bit of a luddite response. Also, the threat of leaving the forum if any form of AI is used feels a bit dramatic...

Modern tools exist and are super popular for a reason. You're entitled to not use things that make your life easier but it'll only hurt you. I ask chatGPT to write code and helper functions for me and it does a fantastic job regularly. I have colleagues who have written entire software packages using chatGPT and they can do some really wild stuff. The wealth of knowledge in these models is staggering and it's an amazing way to learn, improve, complete quick tasks, and bounce ideas off of.

Except that they aren't... They actually suck at it, but I already briefly explained why. As they saying goes "From the horses mouth"

Idk, mine seemed to work lol

But in all seriousness, this is related to my point about using the tool correctly. Are you familiar with distinction between an LLM and an agent? You can effectively use an agent to research a topic and actually *show you* it's proof and justification. I use this many times a day when reading scientific papers, i.e. "show me where in the paper the authors report X", and I will get the exact sentence or figure.

Except that you have no way of knowing if the summary is correct or relevant unless you know the content that it summarized... meaning you really should read the 10 pages or whatever. We can fill this entire forum with examples of AI failures, hallucinations and user driven (often inadvertent) biased output.

AI summaries are really good. Here's an example.

Of course the users are still responsible for the output though. The same goes for any tool. Continuting with screwdriver analogies, would you trust every single screw in a house because it came from a screwdriver? The screwdriver is a tool and should be treated as such. That said, all else equal, I would trust the accuracy of a single LLM output over a randomly sampled post online.

Fun fact - as the AI spreads crap to the masses, those masses repeat it and the AI sucks it back in as "new fact" or "new information". Not only is it a negative feedback loop, it is exponential. So the emerging struggle is not making LLMs "smarter", it is keeping them from inbreeding themselves to utter stupidity -- but some would argue that battle has already been lost even if we don't yet know it.

This is a false fact though. This is not how ML models are trained. All good models have incredibly careful curation of training data. There are also pretty sophisticated output alignment routines. RLHF (human input to what is or isn't a good response) is what put chatGPT on the map. There is no evidence that ML models are in a self feedback loop that degrades performance. If anything, self play can drastically improve performance (alpha go self play took it to super human levels).

In your opinion, not mine.

 

[smitten with ocean life]

ok this is going from weird to weirder.....in fact creepy

 

[BeanAnimal]

You are entitled to your own opinions but not your own facts :)

Irony would be an understatement. The point was that the AI responded based on what it thought you and I wanted to hear. I asked if it was bad at fact checking and it persuasively confirmed it was very bad at it. You asked it if it was good at fact checking and it told you in no uncertain terms that it was very good at it. If you can't see how that erodes your position, you may want to ask ChatGPT.

With all due respect, this feels a bit of a luddite response. Also, the threat of leaving the forum if any form of AI is used feels a bit dramatic...

Luddite? - Now that is truly priceless. You have absolutely no clue who you are talking to :rolling-on-the-floor-laughing:

There is no threat whatsoever. I have very simply and clearly stated my position on the matter.

As for ChatGPT and code. I can typically code faster, with fewer bugs and clearer logic and structure than it can for me. There are exceptions, but few and far between.

All good models have incredibly careful curation of training data...
...There is no evidence that ML models are in a self feedback loop that degrades performance.

Except the growing mountain of research and reality that says otherwise. And with that our conversation has come to an end.

Happy Reefing.

 

[rtparty]

You are entitled to your own opinions but not your own facts :).



With all due respect, this feels a bit of a luddite response. Also, the threat of leaving the forum if any form of AI is used feels a bit dramatic...

Modern tools exist and are super popular for a reason. You're entitled to not use things that make your life easier but it'll only hurt you. I ask chatGPT to write code and helper functions for me and it does a fantastic job regularly. I have colleagues who have written entire software packages using chatGPT and they can do some really wild stuff. The wealth of knowledge in these models is staggering and it's an amazing way to learn, improve, complete quick tasks, and bounce ideas off of.



Idk, mine seemed to work lol

But in all seriousness, this is related to my point about using the tool correctly. Are you familiar with distinction between an LLM and an agent? You can effectively use an agent to research a topic and actually *show you* it's proof and justification. I use this many times a day when reading scientific papers, i.e. "show me where in the paper the authors report X", and I will get the exact sentence or figure.



AI summaries are really good. Here's an example.

Of course the users are still responsible for the output though. The same goes for any tool. Continuting with screwdriver analogies, would you trust every single screw in a house because it came from a screwdriver? The screwdriver is a tool and should be treated as such. That said, all else equal, I would trust the accuracy of a single LLM output over a randomly sampled post online.



This is a false fact though. This is not how ML models are trained. All good models have incredibly careful curation of training data. There are also pretty sophisticated output alignment routines. RLHF (human input to what is or isn't a good response) is what put chatGPT on the map. There is no evidence that ML models are in a self feedback loop that degrades performance. If anything, self play can drastically improve performance (alpha go self play took it to super human levels).

I think we found Nemo

 

[Tahoe61]

I think there is a genuine use and utility of AI on reef2reef threads. Maybe not in this current implementation, but I like the direction and think @FindinNemo is actually onto something. I think AI responses on threads with no responses are valuable, but can see a criticism that they have a “let me Google that for you” vibe.

The more advanced approach would be to incorporate some of the newer specialty AI agents into reef2reef. The newer agent frameworks are essentially multiple instances of an LLM that communicate with each other and can take on different roles, such as searching through relevant threads, fact checking data, and formulating a response.

Here's a scenario: you want to determine if the redfield ratio is important in your reef tank but you see multiple threads that strongly say different things. An agent will not only search through reef2reef threads and interpret responses, but will also bring in it's general knowledge of chemistry, biology, etc. (In the ML field we call this an emergent property where LLMs can utilize information across disciplines and actually do more than parrot answers) to determine the validity of each thread. It can then summarize each thread and provide an evaluation on their scientific strengths and weaknesses. This would be so valuable to novice reefers who struggle to know what to believe when there are such divergent opinions on this forum.

In summary, while I think blanket chatGPT responses can have their use, better agent frameworks built into the site could be a huge win for us all. I'd actually implore @revhtree to look into this! It's where all modern tech and websites are moving.

My interpretation from the responses to this particular situation is the majority of forum users don't want chatGPT generated responses. We don't want lectures on the benefits of the technology. We don't need to be enlightened as to the potential applications to forums.
Personally it's a resounding no, it's not complicated. We gather to interact with other hobbyist, to lose that component would likely result in a significant drop in forum activity. Please stop thinking all you have to do is convince us.

 

[sixty_reefer]

Does it occur to you that the misleading information posted here to begin with could easily be the result of somebody using AI to appear more informed than they actually are?

I’ve been discussing Redfield way before AI was invented my friend. You can easily go through my history and find the evidence.
I’ve actually found my current formula while discussing Redfield with Randy many moons ago.

I’m afraid the accusation of using AI will be a new tool to try and stop Discussions once implemented into the TOS.

@revhtree
Please take the above into consideration as I’m a believer that this forum supports healthy discussions.
There is the danger of anyone from here on now that sounds like they know their subject will be accused of using Ai in violation of the new TOS if someone doesn’t agree with their views.

 

[sixty_reefer]

.

 

[mizimmer90]

I think we found Nemo

Lol please, take off your tin hat and put down the pitchfork. I'm not Nemo any more than you are.

I'm an AI researcher myself and think that some of the current tools can really benefit this website and bring it into the modern age.

 

[mizimmer90]

Did you use AI for this response?

Lol yeah in the screenshot bubbles! But not for what I am saying my text responses are :)

 

[sixty_reefer]

Lol yeah in the screenshot bubbles! But not for what I am saying my text responses are :)

It was ment to be in jest, but I remove it. Just to emphasise how the forum may look like in a near future with folks being accused of using it if the other parties don’t agree with your views.

 

[Miami Reef]

Just to emphasise how the forum may look like in a near future with folks being accused of using it if the other parties don’t agree with your views.

Now you’re just catastrophizing. Let’s just take a breath.

AI speaks in a very specific format. I don’t think it’s helpful trying to prepare for made-up scenarios in the future, which may not even happen. Nobody will accuse you of using AI solely for disagreeing with you —that’s nonsense.

 

[mizimmer90]

Irony would be an understatement. The point was that the AI responded based on what it thought you and I wanted to hear. I asked if it was bad at fact checking and it persuasively confirmed it was very bad at it. You asked it if it was good at fact checking and it told you in no uncertain terms that it was very good at it. If you can't see how that erodes your position, you may want to ask ChatGPT.


Luddite? - Now that is truly priceless. You have absolutely no clue who you are talking to :rolling-on-the-floor-laughing:

There is no threat whatsoever. I have very simply and clearly stated my position on the matter.

As for ChatGPT and code. I can typically code faster, with fewer bugs and clearer logic and structure than it can for me. There are exceptions, but few and far between.


Except the growing mountain of research and reality that says otherwise. And with that our conversation has come to an end.

Happy Reefing.

Feel free to share this research!

A common fear of AI is that it will replace you. But in reality it can't replace you anymore than a screwdriver can (another screwdriver analogy lol). You can code faster than chatGPT used by someone who doesn't know how to code, but i guarantee you will not code faster than chatGPT used by someone who knows how to code and use this tool properly.

Following suggestion #6 from chatGPT on how to interact with you: I am well aware of you and very much respect your posts on this forum. I have given thumbs ups to many of your posts and genuinely like your direct speech, candor, and expert knowledge on many topics. However, I am an ML researcher myself and have a lot of expertise in this area. I keep trying to suggest you look at the difference between an LLM and an agent. Agents are designed to carry out tasks and provide support.

For clarification, I am not suggesting to replace human responses for AI generated responses. Instead, I'm imagining a tool to help summarize threads, help with searches, provide context or an avenue for quick discussion. This tool would be for an individual not to be posted as a response on a thread. Imagine clippy on MS word lol. You can ignore the clippy if you want.

 

[sixty_reefer]

Now you’re just catastrophizing. Let’s just take a breath.

AI speaks in a very specific format. I don’t think it’s helpful trying to prepare for made-up scenarios in the future, which may not even happen. Nobody will accuse you of using AI solely for disagreeing with you —that’s nonsense.

It’s not a made up scenario as I’ve seen it happen in the past few months where folks didn’t agreed on the discussion topic!
Whether folks like it or not this will be the new future tool to try and stop discussions imo.

 

[Miami Reef]

It’s not a made up scenario as I’ve seen it happen in the past few months where folks didn’t agreed on the discussion topic!
Whether folks like it or not this will be the new future imo.

Someone didn’t agree so they accused them of using AI? Please show me.

 

[sixty_reefer]

Someone didn’t agree so they accused them of using AI? Please show me.

I cannot quote as the thread is closed

The accusation​

Some of the disagreement and bias

​

 

[Miami Reef]

.

 

[A_Blind_Reefer]

I’m pretty sure you missed the consensus of the thread??? If you want to leave here to “let the bots and AI fan peeps have their day in the sun” then you’ll be happy to know that you can stay. Because we’re not going to let AI answers take over here. :)

I did get the consensus of the room. That said, the obviously generated responses were still being tossed about so I was expressing my personal reaction….not interested in it at all. Not one bit. I’m glad you will not be letting the ai barforama continue. Hope you’re feeling better. Glad I decided to check up on this thread this am.

 

[A_Blind_Reefer]

It’s not a made up scenario as I’ve seen it happen in the past few months where folks didn’t agreed on the discussion topic!
Whether folks like it or not this will be the new future tool to try and stop discussions imo.

If I remember that string of threads correctly (that’s a big if, I can’t remember what I had for breakfast…and that was like ten minutes ago!), we’re not talking apples to oranges here. This is pure copy/paste (maybe even bot generated, jury’s still out for me) Ai content under the guise of an individuals personal knowledge and experience. In the threads you’re bringing up, people did question where you were getting some information from, and if it was Ai that it was completely wrong and should be double checked for accuracy and not taken as gospel as it is not always accurate. I don’t think anyone is/was/has accused you of being an Ai bot or just copying and pasting Ai responses, inserting yourself in everyone’s business, with a Superman complex in an effort to save everyone on the forum. Healthy discussions, even arguments are great for keeping things moving. If we all just agreed on absolutely everything, nothing new would ever come about. Just know that if one is to make certain claims, another might strike at them. I certainly dont agree with absolutely everything that everyone here says is fact. That doesn’t mean I don’t respect them or their opinion, or even want them muted or banned. Plus who doesn’t like to watch a good train wreck on occasion?