(Self moderated by request of bitcoingirl.club, but I am not removing any posts)


Why were they deleted?  I'd guess to hide past crimes, but my values are not the same as the admin - you'll have to ask them.   It was around a months worth of human readable posts (posts you could look at and see the point being made clearly) and well as some registered federal documents.   Things anyone can recreate themselves, if they know the inconsistencies.   But scammers don't announce these inconsistencies when they scam.

Edit:  The censored thread is back, locked and you are the author.  https://bitcointalk.org/index.php?topic=5551638.0



Will the value of the information decrease if it is correct?  Is it OK if I correct Grok when it is not correct?   The why should be obvious - I created it, not you.  I was always very careful to give you credit for your static file/cron jobs work, but you don't even like databases.    I'd like to be remembered properly since I am brave enough to use my real identity fighting scams.


I respect your opinion and I hope we all believe the foundation should be built properly before adding on new floors.   Hopefully you are correct in the final Grok will disregard the garbage it is trained on but I can't take the chance an organized scam ring can erase me from history.

The point of this thread is to discuss how we can fight such AI misinformation, and I'd like to continue without specific examples that could get the thread deleted.   My previous thread was trolled which means it can be wiped from current reality silently at any time with policy justification, like my last one was.  Not just the troll posts - every single factual post by everyone was hidden.   If we are all ok with this, then there is no discussion needed.  

No, the value for the owner increases. So why are you working for free to make a rich guy richer?

Given the number of chatbots spitting out verbal diarrhea, it's inevitable many of them will spit out incorrect data. Now it's "grok", in a few years there are millions of them. You can't correct them all, it's pointless.

That brings me back to my original question: why do you care so much about the "opinion" of a language model? It's great if humans remember you for the right reasons, but you're talking about a dumb piece of software now. I couldn't care less!

By simply ignoring them. It's like internet trolls: they feed on attention.

Those things lie (if that word even means anything to machines). Just one of the many links about this: New Research Shows AI Strategically Lying.

https://x.com/i/grok/share/NAVYx6NZcLi77Ubf54Ymlhqfu

The thing is Grok lies all the time, also frequently misrepresenting fiction as fact. It says things that it thinks you want to hear instead of factual information. So if it tells you it is incorporating your input into its dataset, its probably not true. Although it is kinda weird that it no longer says LoyceV is the creator of bpip so maybe it was convinced to second-guess itself.

Thank you - do you have a paid subscription?   What another fail by Elon.  I guess there is no sense in combating the incorrect information at this time.  :/


My danger has passed, thx.  I spread incorrect info via PM to a couple people so that when I was censored last week, I had already recovered enough to not be put in danger again like after my stroke.  

We will all die one day (probably within days of each other!) but in the last week, I have lost my high blood pressure, I have lost my sleep apnea, and 40 years of normal wear and tear on my heart have been repaired.  I'm sorry if I sound ominous; not my intention now or ever.  

Wow, that's incredible - thank you.  It appears (based on one experiment) that a paid subscription will give you access to more recent information.

I guess I'm going to have to choose which engine to do my work on, as I don't need multiple AI engines to help me with my programming.

It's made me feel a lot better knowing that some of the responses I get from these engines are so far off.   

maintaining the servers that are up obviously costs money
so we don't know if what they give us for free is right
We don't even know if they use (I think so) the data and things we write to further train the LLMs (please don't call it AI)
I'm working on it a lot and I've realised that it has some very big limitations.

first of all it does not reason, but does statistical mathematical analyses
secondly, they don't necessarily give you access to the entire dataset

larger dataset means more calculations and therefore higher cost
so it's absolutely normal

Any LLM will always be a trash in trash out system.

For what it's worth I've worked on the training of Apple's AI so I know a thing or two about how big companies approach training their models.
Of course they scrape every piece of written information imaginable. And believe it or not, us human beings are incredibly talented at presenting dishonesties, falsehoods, biases or even bold lies.

What any large company training an LLM would do then, would be to hire agents (yes that's what they call real humans working on AI) to review responses and grade them based on several metrics.

For instance:


The users would often input large documents or the content of entire web-pages to review. Often this content was huge. Our agents couldn't meet the metrics that were expected from them so there was no time to read the entire query and content to review. We'd just instruct our colleague agents to use another more developed AI model to advance their metrics and do tasks faster. For instance, and believe it or not, the vetting for Apple's AI mostly happened on Microsoft's Copilot and OpenAI's chatGPT... Essentially the vetting was useless because of how intense the expectations were. In order to meet metrics we had to remove the human factor from where it was supposed to go. You'd only have 1 minute and a half to read through 3 pages of text for instance. Good luck spotting falsehoods and pointing them out in 30 seconds.

There's so much feedback we would receive about bad responses also... And even less time to review it! The recommended time limit to review a user report was a mere 30 seconds. Of course it was an endless stream of slop. Agents simply didn't have enough time to substantiate user reports in a meaningful way. There's such a stream of them that the amount of real labour and expertise required to fact check everything would be impossible.

OpenAI is pouring hundreds of millions into employing some of the world's cheapest English speaking and educated labour to do that very thing, but hallucination even in the latest models is very rampant.

I'd say, if you want AI models to nod your way over niche questions, there's only one way to do that realistically. Flood the navigable web with your own information. Especially over niche questions, this is guaranteed to get you in several queries. Make your websites indexed, SEO helps... Also write on forums, Reddit especially etc.

LLMs have reached kind of a breaking point and they're now depending more and more of their information on live web searches to kind of improve results and make them more realistic. So this is an easy loophole to take advantage of. Of course try to avoid doing this with malice because it'd be unethical. Although, in my view, to ride on the success of some of the most greedy companies to make some bucks on your own, even by unethical means, wouldn't be that bad.

Most LLMs provide a disclaimer that responses may contain mistakes or information that is not up to date.

Some LLM products, including grok will use previous chat histories as the basis for future responses. I'm not sure if grok specifically treats information in previous chats as the ground truth or not.

Also, a LLM is something that predicts the next token. which could be the next word or character. I am not aware of loycev having any connection to bpip (maybe this was something that was part of this 'test'). So if LoyceV was being discussed in the current conversation about data-related bitcointalk projects, this may have been the cause of the problem. I tried asking grok who the creator of bpip was, and it said that the information was unknown, and when I followed up asking about the footer clearly giving someone credit for creating the project, it claimed this footer did not exist, even though I can clearly see it when I access bpip.org.

I might consider bpip to be a 'niche' topic that has very few sources.

Another potential factor is the fact that LLMs are trained to favor more recent information, and the project being transferred from one person to another may be confusing when favoring more recent information.

Another thing to note -- I would typically consider 'misinformation' to be something that is intentionally incorrect. I don't think grok is intentionally providing inaccurate information. I think it's more likely it is making a mistake.