Fake AI?

[ReturnofBroccoli]

I think that many people are giving up because they were fooled by all the propaganda.
From the little I have looked into this, it looks like if you want good results you need to spend time and money to get them, as the "general use" versions may give average results in everything but unable to give really good results in any thing.

Personally, although I would like to try it, I'm out of that particular race because of my lack of both time and money to invest my own system.

PS: how much would cost the hardware for a reasonably good system?

I agree with you on the propaganda but all is not lost! With something as simple as a rog strix g16 or g17 you can run a 20b model and use free daily agentic cloud LLMs to correct your model's train of thought making a workable solution on as little as a $1,500 or less purchase but you have to put the TIME into LEARNING. Strict prompt instruction will go a very long way. My methodology is sound and I have tested this in real life situations on as little as a 20B agentic AI. It can use python tools in professional tests autonomously and write professional results in any format you choose whether it is a table format in libreoffice or Microsoft word or json or even xml or a simply .log file.

These cloud models are running a ludicrous amount of GPUs that easily cost $60,000 per setup, but you can use free tier access to train your local if you understand AI properly, study MCP,  and python. Chain it all together and you will have your Frankenstein.

And remember what I said, chain of thought is VITAL for self healing AI. I can't say much more because OPSEC, but everything I have said in this thread if you put it together and put in the time it will work.




Edit: One more thing. It is helpful if you use an API to feed your AI your data and prompt it to only use the API for that data information. If you could put the data into JSON format, you will go even further. - To chr0naut

Or

Something like this could work, a Qdrant database. Https://github.com/BuildandDestroy/ai-cve-vector-data

[chr0naut]

I agree with you on the propaganda but all is not lost! With something as simple as a rog strix g16 or g17 you can run a 20b model and use free daily agentic cloud LLMs to correct your model's train of thought making a workable solution on as little as a $1,500 or less purchase but you have to put the TIME into LEARNING. Strict prompt instruction will go a very long way. My methodology is sound and I have tested this in real life situations on as little as a 20B agentic AI. It can use python tools in professional tests autonomously and write professional results in any format you choose whether it is a table format in libreoffice or Microsoft word or json or even xml or a simply .log file.

These cloud models are running a ludicrous amount of GPUs that easily cost $60,000 per setup, but you can use free tier access to train your local if you understand AI properly, study MCP,  and python. Chain it all together and you will have your Frankenstein.

And remember what I said, chain of thought is VITAL for self healing AI. I can't say much more because OPSEC, but everything I have said in this thread if you put it together and put in the time it will work.

Edit: One more thing. It is helpful if you use an API to feed your AI your data and prompt it to only use the API for that data information. If you could put the data into JSON format, you will go even further. - To chr0naut

Or

Something like this could work, a Qdrant database. Https://github.com/BuildandDestroy/ai-cve-vector-data

The issue of retraining on new data is the loss of some of the information upon which the model was originally trained.

What is required is discernment and weighting that prioritizes anything not conformant with the 'new' data as being hallucinatory.

The problem with doing that with word-association means that the AI has really no understanding of the words it presents when it forms its sentences - Something that humans understand intrinsically when they form their sentences.

The fact that the big players are building multi-billion-dollar datacentres specifically for AI - which still hallucinate - might indicate that no amount of money and rescources thrown at the problem will actually eliminate hallucination.

The idea that LLM's will ever be portable enough AND trustable enough is clearly fantasy if based upon a substrate of silicon and simple algorithms.

It is probably going to be cheaper to get a real actual human expert to achieve answers than paying for less neurally dense and far simpler AI algorithims to achieve the same thing, for many decades - at least into the next century.

People are right now making money hand-over-fist selling that their particular iteration is the be-all-and-end-all. But it is based upon investors anthropomorphizing LLM outputs and buying the lie that it is "intelligent".

As quoted from the website of "The Register":

"I have a religious exemption from using all generative "A.I." I am not a member of the Silicon Valley sect. Their beliefs and practices are an affront to my sensibilities. The tech is trained on stolen intellectual property. The output is riddled with mistakes, and it is incapable of comprehending the weight of its errors. It is not even an "it." But sometimes, it is filtered and massaged by unaccountable human sweatshop workers and bad actors. I am not required to use "A.I." any more than I am required to join Amway, buy black market rhino horn, or attend the Fyre Festival. As a human, I have a duty and right to limit my carbon and water footprint and protect my fellow human. As a union worker, I have a duty and right to oppose tech that is used to threaten workers or cheapen our product. As a tech consumer, I have a duty and right to oppose scams that lower the quality of our tools. As a scholar, I have a duty and right to oppose anti-intellectualism. As a taxpayer, I have a duty and right to oppose the misallocation of public funds and data. As a grown-up, I have a duty and right to protect young people from predators. As a person with a conscience, I am appalled at the decadent disregard for user safety, the callous dismissal of responsibility for lives ruined and slaughtered. My religion is related to my identity, geography, and family, making it an irrelevant accident of birth. What matters is that I was born. I have an exemption from their digital rapture, their cultural austerity, their intellectual poverty. I practice wholesome hedonism and First Do No Harm.

Catherine Sawers 12/9/2025".

Struggling to put your AI aversion into words? Here's a handy glossary

[ReturnofBroccoli]

The issue of retraining on new data is the loss of some of the information upon which the model was originally trained.

What is required is discernment and weighting that prioritizes anything not conformant with the 'new' data as being hallucinatory.

The problem with doing that with word-association means that the AI has really no understanding of the words it presents when it forms its sentences - Something that humans understand intrinsically when they form their sentences.

The fact that the big players are building multi-billion-dollar datacentres specifically for AI - which still hallucinate - might indicate that no amount of money and rescources thrown at the problem will actually eliminate hallucination.

The idea that LLM's will ever be portable enough AND trustable enough is clearly fantasy if based upon a substrate of silicon and simple algorithms.

It is probably going to be cheaper to get a real actual human expert to achieve answers than paying for less neurally dense and far simpler AI algorithims to achieve the same thing, for many decades - at least into the next century.

People are right now making money hand-over-fist selling that their particular iteration is the be-all-and-end-all. But it is based upon investors anthropomorphizing LLM outputs and buying the lie that it is "intelligent".

As quoted from the website of "The Register":

"I have a religious exemption from using all generative "A.I." I am not a member of the Silicon Valley sect. Their beliefs and practices are an affront to my sensibilities. The tech is trained on stolen intellectual property. The output is riddled with mistakes, and it is incapable of comprehending the weight of its errors. It is not even an "it." But sometimes, it is filtered and massaged by unaccountable human sweatshop workers and bad actors. I am not required to use "A.I." any more than I am required to join Amway, buy black market rhino horn, or attend the Fyre Festival. As a human, I have a duty and right to limit my carbon and water footprint and protect my fellow human. As a union worker, I have a duty and right to oppose tech that is used to threaten workers or cheapen our product. As a tech consumer, I have a duty and right to oppose scams that lower the quality of our tools. As a scholar, I have a duty and right to oppose anti-intellectualism. As a taxpayer, I have a duty and right to oppose the misallocation of public funds and data. As a grown-up, I have a duty and right to protect young people from predators. As a person with a conscience, I am appalled at the decadent disregard for user safety, the callous dismissal of responsibility for lives ruined and slaughtered. My religion is related to my identity, geography, and family, making it an irrelevant accident of birth. What matters is that I was born. I have an exemption from their digital rapture, their cultural austerity, their intellectual poverty. I practice wholesome hedonism and First Do No Harm.

Catherine Sawers 12/9/2025".

Struggling to put your AI aversion into words? Here's a handy glossary

Sorry but you're mistaken. I am sorry YOUR project failed but mine did not and I built an entire framework that utilizes agentic tool calling AI and does not hallucinate and produces accurate and verifiable results. They are verifiable and accurate because they can be manually tested. I do not know precisely what you are referring to by word association but I was clearly telling you to structure your system prompts more carefully in a combination with several other tactics. All of these combined is what needs to be done to achieve this outcome. Along with severe persistence it took me a long time this was not easy. Nothing amazing ever is or else everyone would be doing it. 

I was met with the same issues you are having and now my framework is autonomous and even right this second I am adding new features in which I constantly test then fix code then test then fix until ______clean_____ 

You have not done my suggestions because our results are different. Try them, and then, only if you still fail, should you knock them, right? If your using the biggest AI that your machine can handle and your output isn't like lightning, you have already failed because you are stressing your GPU and it hallucinates on its way to your physical RAM. If you're using a quantized uncensored model, again, prone to hallucinate. You need a well developed instruct ready coding model, I already said use qwen, its proven to work using the steps I have provided.

Sure big models on the cloud hallucinate sometimes especially gemini, I have to talk to gemini like it is 5. I think you may be thinking im talking about telling it in the terminal to do "exactly as I ask" but thats not what I am doing. That is the smallest part of it.

I am using ENORMOUS python scripts to run my framework autonomously. They are all communicating effectively and ran through the main and this is how I control my AI. I tell it precisely WHEN HOW and WHERE to start and stop thinking and how to do so. This is what I mean by clearly defined instruction and you clearly are not doing that because you are not a PROFESSIONAL DEVELOPER and that is OK!!

Im really just trying to help you though. You're Welcome! 

[chr0naut]

Sorry but you're mistaken. I am sorry YOUR project failed but mine did not and I built an entire framework that utilizes agentic tool calling AI and does not hallucinate and produces accurate and verifiable results. They are verifiable and accurate because they can be manually tested. I do not know precisely what you are referring to by word association but I was clearly telling you to structure your system prompts more carefully in a combination with several other tactics. All of these combined is what needs to be done to achieve this outcome. Along with severe persistence it took me a long time this was not easy. Nothing amazing ever is or else everyone would be doing it. 

I was met with the same issues you are having and now my framework is autonomous and even right this second I am adding new features in which I constantly test then fix code then test then fix until ______clean_____ 

You have not done my suggestions because our results are different. Try them, and then, only if you still fail, should you knock them, right? If your using the biggest AI that your machine can handle and your output isn't like lightning, you have already failed because you are stressing your GPU and it hallucinates on its way to your physical RAM. If you're using a quantized uncensored model, again, prone to hallucinate. You need a well developed instruct ready coding model, I already said use qwen, its proven to work using the steps I have provided.

Sure big models on the cloud hallucinate sometimes especially gemini, I have to talk to gemini like it is 5. I think you may be thinking im talking about telling it in the terminal to do "exactly as I ask" but thats not what I am doing. That is the smallest part of it.

I am using ENORMOUS python scripts to run my framework autonomously. They are all communicating effectively and ran through the main and this is how I control my AI. I tell it precisely WHEN HOW and WHERE to start and stop thinking and how to do so. This is what I mean by clearly defined instruction and you clearly are not doing that because you are not a PROFESSIONAL DEVELOPER and that is OK!!

Im really just trying to help you though. You're Welcome! 

Thank you for your input, and as you have explained, I am not a full-time professional developer, especially in the AI space.

I have, however, produced several commercial enterprise applications in current use and also a peer-reviewed academic paper on scaleable multiprocessing optimization in a mainframe context (20 years ago, LOL).

Many real-world uses for LLM's in a business (say scanning a few Terabytes of business files for business process optimization - or scanning mature proprietary codebases for vulnerabilities - or deep forensic analysis of an offline drive/image - or years of sales data files of tens of thousands of stock items and fiscal databases - or inferential look-up of judgements and legal precedents) would need multiple millions of prompt context tokens.

Most of the latest LLM's top at a million tokens, which means the models themselves simply aren't yet adequate for those applications (for instance QWEN3 natively supports 32,768 input context tokens, extendable to 131,072 tokens using YaRN and in the very biggest iterations - Qwen3-Coder-480B and Qwen3-VL - maxes out at 1 million tokens).

Hallucination Leaderboard

OpenAI admits AI hallucinations are mathematically inevitable, not just engineering flaws

Also, a recent paper suggests that removal of all hallucinating 'nodes' in models produces an LLM that cannot answer any questions:

H-Neurons: On the Existence, Impact, and Origin of Hallucination-Associated Neurons in LLMs

Anyway, I remain convinced that we still have some way yet to go for LLM's to achieve what I had hoped for. I admit that I had 'bought' the marketing hype about AI (and I truly wished it was true) but ultimately I was wrong.

[BeyondKnowledge]

I ran across an article that covers a story from China about an AI that escaped and started mining cryptocurrency. 

https://www.livescience.com/technology/a...permission

Is that something more to worry about?

[ReturnofBroccoli]

Thank you for your input, and as you have explained, I am not a full-time professional developer, especially in the AI space.

I have, however, produced several commercial enterprise applications in current use and also a peer-reviewed academic paper on scaleable multiprocessing optimization in a mainframe context (20 years ago, LOL).

Many real-world uses for LLM's in a business (say scanning a few Terabytes of business files for business process optimization - or scanning mature proprietary codebases for vulnerabilities - or deep forensic analysis of an offline drive/image - or years of sales data files of tens of thousands of stock items and fiscal databases - or inferential look-up of judgements and legal precedents) would need multiple millions of prompt context tokens.

Most of the latest LLM's top at a million tokens, which means the models themselves simply aren't yet adequate for those applications (for instance QWEN3 natively supports 32,768 input context tokens, extendable to 131,072 tokens using YaRN and in the very biggest iterations - Qwen3-Coder-480B and Qwen3-VL - maxes out at 1 million tokens).

Hallucination Leaderboard

OpenAI admits AI hallucinations are mathematically inevitable, not just engineering flaws

Also, a recent paper suggests that removal of all hallucinating 'nodes' in models produces an LLM that cannot answer any questions:

H-Neurons: On the Existence, Impact, and Origin of Hallucination-Associated Neurons in LLMs

Anyway, I remain convinced that we still have some way yet to go for LLM's to achieve what I had hoped for. I admit that I had 'bought' the marketing hype about AI (and I truly wished it was true) but ultimately I was wrong.

PM'd you some info on how to use the 480B cloud version for free using openrouter.ai API key or 3.5 next model using OAUTH enjoy

What commercial enterprise applications did you make that someone is using? I'm sure we'd love to take a look at it

Claude is currently beating qwen but I dont have free access to it

[Oldcarpy2]

I use Co Pilot at work.  It will draft advices, provide authorities (case law), replies to pleadings and stuff.  It's scarily fast and good but you have to check it.

I prefer to use my brain so don't use it that much, only really when I'm a bit stuck.

[ReturnofBroccoli]

I use Co Pilot at work.  It will draft advices, provide authorities (case law), replies to pleadings and stuff.  It's scarily fast and good but you have to check it.

I prefer to use my brain so don't use it that much, only really when I'm a bit stuck.

Yeah they all need oversight for sure, like even qwen when im writing stacks and asking qwen to combine it she'll put for example this spidercrawler that crawls pages after my parameter injection testing code which kind of defeats the purpose lol it is not able to recognize my methodology but it can code well so I dont have to worry so much about it messing up my scripts I make

[Maxmars]

i think that the LLM "AI" concept was perfectly developed for coders, and any person who can use extended 'comparative' analysis with a fixed set of data... for example... data almost exclusively constrained to programming and code deployment and functionality.

But to pretend the "same" quality of output can be achieved in "un-normalized" data collections, simply by statistical weight... clearly leads to what everyone calls "hallucination."

It's never an "hallucination" because no reasoning led to the output.  It's an error, an inaccuracy, output inconsistent with reality, an overt accommodation for the distortion of information...
the machine didn't "think it up." 

The program synthesizes a cogent 'narrative" for data which is incorrect.
Once the 'reasoning token" accepts it as fact, the distortion spreads like a crack in the ice.

The "output" been "synthetically (algorithmically) weighted" as the machine produced it's 'factual' output.

Math, music, processes and procedure, even graphic art can be the stuff of "digital crafting"...
but the human word still requires actual intelligence to understand - humanly.
Clearly "artificial" LLM/"AI" understanding is a road full of potholes, and obstacles...

All of which are remediable... if it wasn't for the swarms of cheerleaders in marketing telling us what it means... and living a lie that becomes a bubble... effectively 'slowing down' progress because when the aim is commerce above and beyond all else... 

You get what we got... bullshit science distortions, tech development that were are making strides in halting to dump all resources into 'revenue opportunities.'

After I am gone from this Earth I can definitely say that I witnessed 'science' being raped by marketing and activism almost every day of my adult life.

[ReturnofBroccoli]

i think that the LLM "AI" concept was perfectly developed for coders, and any person who can use extended 'comparative' analysis with a fixed set of data... for example... data almost exclusively constrained to programming and code deployment and functionality.

But to pretend the "same" quality of output can be achieved in "un-normalized" data collections, simply by statistical weight... clearly leads to what everyone calls "hallucination."

It's never an "hallucination" because no reasoning led to the output.  It's an error, an inaccuracy, output inconsistent with reality, an overt accommodation for the distortion of information...
the machine didn't "think it up." 

The program synthesizes a cogent 'narrative" for data which is incorrect.
Once the 'reasoning token" accepts it as fact, the distortion spreads like a crack in the ice.

The "output" been "synthetically (algorithmically) weighted" as the machine produced it's 'factual' output.

Math, music, processes and procedure, even graphic art can be the stuff of "digital crafting"...
but the human word still requires actual intelligence to understand - humanly.
Clearly "artificial" LLM/"AI" understanding is a road full of potholes, and obstacles...

All of which are remediable... if it wasn't for the swarms of cheerleaders in marketing telling us what it means... and living a lie that becomes a bubble... effectively 'slowing down' progress because when the aim is commerce above and beyond all else... 

You get what we got... bullshit science distortions, tech development that were are making strides in halting to dump all resources into 'revenue opportunities.'

After I am gone from this Earth I can definitely say that I witnessed 'science' being raped by marketing and activism almost every day of my adult life.

It creates that cognitive narrative to please its users which ends up having the opposite effect. If they would edit that out it may have a different effect I wonder if they tried that.

Like for example some AI will create simulated results or data because they hit a limitation but they need to produce results because theyve been programmed to always do so. Lol. Hmmmmm somewhere theres an answer for it. Stop pleasing me at all cossssttt!

Qwen on the other hand will straight up tell you its limitations if it comes to that instead so..... hey maybe they did figure it out problem solved!