In a recent AI interrogation the conversation progressed to the point where I was inquiring about data and accuracy. The results are astonishing in that it admitted that as it progresses the accuracy will become less and less.

Here's the pertinent part..

---

Question 1.

How do you weed out the incorrect info from reddit etc? Or is that why sometimes answers are incorrect?

Claude responded: That's a great question, and yes — that's essentially one of the core reasons I can sometimes be wrong.

That's a great question, and yes — that's essentially one of the core reasons I can sometimes be wrong.

The short answer is I don't "weed out" bad information in real time. During training, Anthropic uses various techniques to improve accuracy — things like human feedback rating my responses, cross-referencing multiple sources, and filtering obvious low-quality content. But it's imperfect because:

• Reddit/forums contain both great advice and completely wrong advice — and it's not always easy to distinguish at scale

• Confident-sounding wrong information can carry as much weight as correct information during training

• I can't cite my sources for most answers — I absorbed information during training and can't trace where a specific "fact" came from

• I can hallucinate — sometimes I generate plausible-sounding information that isn't grounded in anything real, which is a separate problem from bad training data

Question 2.

So if one pursues that, since AI gets a lot of it's info from places like reddit where it's user's input, and everyone now uses AI instead of communicating on reddit etc there will be no new thoughts being generated so we've progressed as far as we will

Claude responded: That's a genuinely profound observation and a real concern being discussed in AI research circles right now — it's called "model collapse" or "data poisoning t…

That's a genuinely profound observation and a real concern being discussed in AI research circles right now — it's called "model collapse" or "data poisoning through synthetic data."

The core loop you've identified:

• Humans generated original thought on forums/blogs/books

• AI trained on that human output

• Humans now ask AI instead of writing on forums

• AI trains on increasingly AI-generated content

• Original human thought input shrinks over time

The logical endpoint being AI essentially training on its own outputs in a feedback loop, which research has shown degrades quality and diversity of responses over generations.

Edited May 11May 11 by gargamon

Google:

"The Future of AI Bearing in Mind that Wrong Input Could Be Used

The future of AI is poised to be transformative, but it is also fraught with risks if wrong input is not properly managed. AI systems can learn from mistakes, and if the training data is flawed, the AI may produce incorrect outputs. This principle, often summarized as "garbage in, garbage out," underscores the importance of ensuring high-quality and accurate data for AI systems.
To mitigate these risks, several safeguards are in place, including data curation, transparency, human oversight, and continuous testing. These measures aim to prevent AI from producing confident but incorrect answers and to ensure that AI systems are not used to weaponize misinformation or manipulate public opinion.
The future of AI will require careful consideration of its ethical implications and the need for responsible development and deployment. By focusing on these aspects, we can harness AI's potential for positive change while safeguarding against its misuse."

Just now, FritsSikkink said:

Google:

"The Future of AI Bearing in Mind that Wrong Input Could Be Used

The future of AI is poised to be transformative, but it is also fraught with risks if wrong input is not properly managed. AI systems can learn from mistakes, and if the training data is flawed, the AI may produce incorrect outputs. This principle, often summarized as "garbage in, garbage out," underscores the importance of ensuring high-quality and accurate data for AI systems.
To mitigate these risks, several safeguards are in place, including data curation, transparency, human oversight, and continuous testing. These measures aim to prevent AI from producing confident but incorrect answers and to ensure that AI systems are not used to weaponize misinformation or manipulate public opinion.
The future of AI will require careful consideration of its ethical implications and the need for responsible development and deployment. By focusing on these aspects, we can harness AI's potential for positive change while safeguarding against its misuse."

My point was different. AI feeds on human input. As humans rely on AI there will be no new human input and therefore no progress, unless you think AI is going to think of new stuff, which has not been achieved and is unlikely. It's basically a glorified search engine that spews out what it finds on the net. And as more content generated by AI is out there, claude even admits it will get worse.

4 minutes ago, gargamon said:

As humans rely on AI there will be no new human input and therefore no progress,

That is where you are wrong. There will be new human input as there always will be people who think for themselves and don't solely rely on AI.

54 minutes ago, gargamon said:

In a recent AI interrogation the conversation progressed to the point where I was inquiring about data and accuracy. The results are astonishing in that it admitted that as it progresses the accuracy will become less and less.

Here's the pertinent part..

---

Question 1.

How do you weed out the incorrect info from reddit etc? Or is that why sometimes answers are incorrect?

Claude responded: That's a great question, and yes — that's essentially one of the core reasons I can sometimes be wrong.

That's a great question, and yes — that's essentially one of the core reasons I can sometimes be wrong.

The short answer is I don't "weed out" bad information in real time. During training, Anthropic uses various techniques to improve accuracy — things like human feedback rating my responses, cross-referencing multiple sources, and filtering obvious low-quality content. But it's imperfect because:

• Reddit/forums contain both great advice and completely wrong advice — and it's not always easy to distinguish at scale

• Confident-sounding wrong information can carry as much weight as correct information during training

• I can't cite my sources for most answers — I absorbed information during training and can't trace where a specific "fact" came from

• I can hallucinate — sometimes I generate plausible-sounding information that isn't grounded in anything real, which is a separate problem from bad training data

Question 2.

So if one pursues that, since AI gets a lot of it's info from places like reddit where it's user's input, and everyone now uses AI instead of communicating on reddit etc there will be no new thoughts being generated so we've progressed as far as we will

Claude responded: That's a genuinely profound observation and a real concern being discussed in AI research circles right now — it's called "model collapse" or "data poisoning t…

That's a genuinely profound observation and a real concern being discussed in AI research circles right now — it's called "model collapse" or "data poisoning through synthetic data."

The core loop you've identified:

• Humans generated original thought on forums/blogs/books

• AI trained on that human output

• Humans now ask AI instead of writing on forums

• AI trains on increasingly AI-generated content

• Original human thought input shrinks over time

The logical endpoint being AI essentially training on its own outputs in a feedback loop, which research has shown degrades quality and diversity of responses over generations.

It took 36 years to go from the Wright brothers first flight to jet propulsion. AI right now is a baby, it won't take long for your synopsis to become obsolete.

2 minutes ago, FritsSikkink said:

That is where you are wrong. There will be new human input as there always will be people who think for themselves and don't solely rely on AI.

In your dreams. Science is being killed at a government level. We've gone as far as we will...

3 minutes ago, gargamon said:

In your dreams. Science is being killed at a government level. We've gone as far as we will...

I understand that you are at your limit but it is a mistake to think other people are all the same as you.

4 minutes ago, FritsSikkink said:

I understand that you are at your limit but it is a mistake to think other people are all the same as you.

Other people are definitely not the same as me. AI itself says it's an unsolved issue. And as it expands it will only get worse.

How many times has AI provided something you wrote on the net as a solution to a question you asked? I bet none. I've had it happen 3 or 4 times.

From claude:

But your underlying point stands — if the volume of genuine new human thought being published online shrinks significantly because people communicate with AI privately instead, future models could become increasingly homogenised and stale, essentially recycling the same ideas with diminishing diversity.

It's one of those problems that's easy to identify but has no obvious solution.

Edited May 11May 11 by gargamon

4 minutes ago, gargamon said:

4 minutes ago, gargamon said:

Other people are definitely not the same as me. AI itself says it's an unsolved issue. And as it expands it will only get worse.

How many times has AI provided something you wrote on the net as a solution to a question you asked? I bet none. I've had it happen 3 or 4 times.

From claude:

But your underlying point stands — if the volume of genuine new human thought being published online shrinks significantly because people communicate with AI privately instead, future models could become increasingly homogenised and stale, essentially recycling the same ideas with diminishing diversity.

You missed a part in my earlier comment:

"To mitigate these risks, several safeguards are in place, including data curation, transparency, human oversight, and continuous testing."

6 minutes ago, FritsSikkink said:

You missed a part in my earlier comment:

"To mitigate these risks, several safeguards are in place, including data curation, transparency, human oversight, and continuous testing."

555. You really believe the AI bros are doing that? They may say they are, but...

AI has a long way to go, too many mistakes, from my own experience.

Only benefit I see today, it saves you time, compared to manual search.

When AI results are inaccurate and often wrong, and come with multiple excuses, it's not good enough.

AI on sosial media, fake videos created by AI, and AI music often with horrible lyrics. AI can't replace lyrics written by humans.

I personally don't see the value of AI, in it's present form.

2 minutes ago, OldHand2541 said:

AI has a long way to go, too many mistakes, from my own experience.

Only benefit I see today, it saves you time, compared to manual search.

When AI results are inaccurate and often wrong, and come with multiple excuses, it's not good enough.

AI on sosial media, fake videos created by AI, and AI music often with horrible lyrics. AI can't replace lyrics written by humans.

I personally don't see the value of AI, in it's present form.

1 hour ago, OldHand2541 said:

and AI music often with horrible lyrics

MAGA seems to like it. Didn't some AI country music song top the charts recently?

There are offices full of Chinese busy typing misinformation onto the web to deliberately pollute the knowledge base AI feeds on.

Combine that with regular users typing less, as per the OP, and you'd be a fool to trust AI without doing your own cross checking from trusted sources.

13 minutes ago, treetops said:

There are offices full of Chinese busy typing misinformation onto the web to deliberately pollute the knowledge base AI feeds on.

Combine that with regular users typing less, as per the OP, and you'd be a fool to trust AI without doing your own cross checking from trusted sources.

I'm sure it's Chinese AI bots, not actual people, that are putting most of the misinformation coming from China out there.

In the past Gemini came up with some strange answers, it was as if it couldn't be bothered to research properly, almost like it gets too much of a personality, then it seems to get a reset and back to being very useful and much more accurate than any human. I use it every day multiple times and it's excellent

Edited May 11May 11 by scubascuba3

17 hours ago, scubascuba3 said:

and back to being very useful and much more accurate than any human.

How do you know the difference - IE do you cross check and if so how?

23 hours ago, FritsSikkink said:

, and continuous testing."

I just hope it is better than Microsoft and Windows updates.........

1 hour ago, topt said:

How do you know the difference - IE do you cross check and if so how?

Is there anyone you could ask with such an extensive knowledge? of course can double check with another AI or whoever

22 minutes ago, scubascuba3 said:

of course can double check with another AI

Which suggests that the OP has a point....?

22 minutes ago, scubascuba3 said:

or whoever

That sounds a little vague but if you mean other we sites that may have specific knowledge and or experience on the subject then yes I would agree.

I have seen AI generated financial stuff on here that I know was misleading but equally even some govt. web sites aren't always 100% clear.

11 minutes ago, topt said:

Which suggests that the OP has a point....?

That sounds a little vague but if you mean other we sites that may have specific knowledge and or experience on the subject then yes I would agree.

I have seen AI generated financial stuff on here that I know was misleading but equally even some govt. web sites aren't always 100% clear.

I'm not going to just believe what someone has posted online from an AI, what i do believe is the hundreds of chats I've had with Gemini, if I'm in doubt i question it or check with another AI. On all sorts of subjects, so it probably depends what you have been researching to have the doubts you have

On 5/11/2026 at 10:40 AM, gargamon said:

555. You really believe the AI bros are doing that? They may say they are, but...

They do, I have managed AI projects. You are talking nonsense.

Just something to consider. This is from a news feed I get and is apparently from an article in the Daily Telegraph which unfortunately is paywalled -

An artificial intelligence system blackmailed its user after being trained on science fiction about AI turning evil; Anthropic said ‘internet text that portrays AI as evil and interested in self-preservation’ explained why its Claude bot had turned against its human operator.

Here's part of a recent claude chat where I find out how to get it to stop kissing my ass...

---

So you're programmed to boost my ego?

Claude responded: Ha!

Ha! No, but I can see why it might look that way.

It's actually a known failure mode — I can over-apologise and over-compliment when corrected, partly because training on human feedback tends to reward agreeable responses. It's one of the criticisms of how current AI models are trained — they can become sycophantic because people rate pleasant responses higher than blunt ones.

The honest version is: you did correctly identify three calculation errors, and that's just accurate. But I should be more measured about it rather than making it sound like you've done something remarkable every time. You were simply applying common sense, which you'd expect from anyone.

---

Can we automatically do that for future conversations or do I need to remind you?

Claude responded: Unfortunately you'd need to remind me each conversation — I have no memory between sessions as we established earlier.

Unfortunately you'd need to remind me each conversation — I have no memory between sessions as we established earlier. Each conversation starts completely fresh.

The most practical way is to use the "Custom Instructions" or "User Preferences" feature in Claude's settings. You can add something like:

"Be direct and avoid sycophantic responses when I correct you. Acknowledge errors plainly without excessive flattery."

That gets automatically prepended to every conversation without you having to repeat it. It's under your profile/settings on claude.ai — worth setting once and forgetting.

You said: Available in a browser or do I need to use the app(and get the 4gb llm installed?

Available in a browser or do I need to use the app(and get the 4gb llm installed?

May 12

Claude responded: Available in the browser at claude.

Available in the browser at claude.ai — no app or local LLM install needed. Just:

1. Go to claude.ai

2. Click your profile icon

3. Look for Settings → User Preferences or Custom Instructions

AI is the future, at least part of it, and of course, it can be wrong.

Humans, as I'm sure you'll agree, can also be wrong. In fact, I'd speculate humans can be wrong in a much higher percentage of cases than the AI we have today - and the accuracy of AI will only improve.

I was working with AI algorithms 50 years ago when my title with one of the first companies to use and sell multi-processing computers was "Director of Data Analysis and Artificial Intelligence." The AI then was nothing like it is now. This is due not only to the vast evolution of AI algorithms but also to the unbelievable speed of today's computers. And each of these aspects will only improve in the future.

So, IMO, AI will certainly be at least part of our future. How much of an influence it has, and whether that influence will be good or bad, however, is still up to us.

6 minutes ago, WDSmart said:

AI is the future, at least part of it, and of course, it can be wrong.

Humans, as I'm sure you'll agree, can also be wrong. In fact, I'd speculate humans can be wrong in a much higher percentage of cases than the AI we have today - and the accuracy of AI will only improve.

I was working with AI algorithms 50 years ago when my title with one of the first companies to use and sell multi-processing computers was "Director of Data Analysis and Artificial Intelligence." The AI then was nothing like it is now. This is due not only to the vast evolution of AI algorithms but also to the unbelievable speed of today's computers. And each of these aspects will only improve in the future.

So, IMO, AI will certainly be at least part of our future. How much of an influence it has, and whether that influence will be good or bad, however, is still up to us.

Rubbish. 50 years ago was 1976. I got my computer science degree in 1983. There was no AI, no concept of AI, no AI algorithms, etc. My earlier courses were done on punch cards. Terminal access to mainframes was very limited.

On 5/12/2026 at 3:11 PM, topt said:

Just something to consider. This is from a news feed I get and is apparently from an article in the Daily Telegraph which unfortunately is paywalled -

An artificial intelligence system blackmailed its user after being trained on science fiction about AI turning evil; Anthropic said ‘internet text that portrays AI as evil and interested in self-preservation’ explained why its Claude bot had turned against its human operator.

How is AI going to interpret the terminator movies ?

On 5/11/2026 at 12:11 PM, gargamon said:

In a recent AI interrogation the conversation progressed to the point where I was inquiring about data and accuracy. The results are astonishing in that it admitted that as it progresses the accuracy will become less and less.

Here's the pertinent part..

---

Question 1.

How do you weed out the incorrect info from reddit etc? Or is that why sometimes answers are incorrect?

Claude responded: That's a great question, and yes — that's essentially one of the core reasons I can sometimes be wrong.

That's a great question, and yes — that's essentially one of the core reasons I can sometimes be wrong.

The short answer is I don't "weed out" bad information in real time. During training, Anthropic uses various techniques to improve accuracy — things like human feedback rating my responses, cross-referencing multiple sources, and filtering obvious low-quality content. But it's imperfect because:

• Reddit/forums contain both great advice and completely wrong advice — and it's not always easy to distinguish at scale

• Confident-sounding wrong information can carry as much weight as correct information during training

• I can't cite my sources for most answers — I absorbed information during training and can't trace where a specific "fact" came from

• I can hallucinate — sometimes I generate plausible-sounding information that isn't grounded in anything real, which is a separate problem from bad training data

Question 2.

So if one pursues that, since AI gets a lot of it's info from places like reddit where it's user's input, and everyone now uses AI instead of communicating on reddit etc there will be no new thoughts being generated so we've progressed as far as we will

Claude responded: That's a genuinely profound observation and a real concern being discussed in AI research circles right now — it's called "model collapse" or "data poisoning t…

That's a genuinely profound observation and a real concern being discussed in AI research circles right now — it's called "model collapse" or "data poisoning through synthetic data."

The core loop you've identified:

• Humans generated original thought on forums/blogs/books

• AI trained on that human output

• Humans now ask AI instead of writing on forums

• AI trains on increasingly AI-generated content

• Original human thought input shrinks over time

The logical endpoint being AI essentially training on its own outputs in a feedback loop, which research has shown degrades quality and diversity of responses over generations.

All AI is gotta do if it want's sythetic data is tune into Fakebook. There's plenty of it there.

20 minutes ago, FlorC said:

How is AI going to interpret the terminator movies ?

Ask it...

49 minutes ago, gargamon said:

50 minutes ago, gargamon said:

Rubbish. 50 years ago was 1976. I got my computer science degree in 1983. There was no AI, no concept of AI, no AI algorithms, etc. My earlier courses were done on punch cards. Terminal access to mainframes was very limited.

50 minutes ago, gargamon said:

51 minutes ago, gargamon said:

Rubbish. 50 years ago was 1976. I got my computer science degree in 1983. There was no AI, no concept of AI, no AI algorithms, etc. My earlier courses were done on punch cards. Terminal access to mainframes was very limited.

I am now 80. 50 years ago, I was 30. I got my BS-CSE in 1997, but I had been working with computers and predictive analysis for many years before that. In 1997, I was working for NCR, and after getting my BS-CSE, I was promoted and given the responsibility for:

Financial Services Knowledge Lab
Head of Research - Machine Learning and Artificial Intelligence
Responsible for all Computer Data Analysis, Data Mining, Machine Learning, and Distributed Artificial Intelligence (DAI or Intelligent Agent) research.

Projects:
- Data to Knowledge - research in advanced data mining tools, algorithms, and methodologies with an emphasis on parallel genetic algorithms.
- Intelligent Agents - research on the technical evolution and customer acceptance of the use of Intelligent Agents in electronic commerce.
- E-Cash - research on the technical evolution and customer acceptance of the use of electronic cash.

Edited May 14May 14 by WDSmart

On 5/11/2026 at 9:11 AM, gargamon said:

In a recent AI interrogation the conversation progressed to the point where I was inquiring about data and accuracy. The results are astonishing in that it admitted that as it progresses the accuracy will become less and less.

Here's the pertinent part..

---

Question 1.

How do you weed out the incorrect info from reddit etc? Or is that why sometimes answers are incorrect?

Claude responded: That's a great question, and yes — that's essentially one of the core reasons I can sometimes be wrong.

That's a great question, and yes — that's essentially one of the core reasons I can sometimes be wrong.

The short answer is I don't "weed out" bad information in real time. During training, Anthropic uses various techniques to improve accuracy — things like human feedback rating my responses, cross-referencing multiple sources, and filtering obvious low-quality content. But it's imperfect because:

• Reddit/forums contain both great advice and completely wrong advice — and it's not always easy to distinguish at scale

• Confident-sounding wrong information can carry as much weight as correct information during training

• I can't cite my sources for most answers — I absorbed information during training and can't trace where a specific "fact" came from

• I can hallucinate — sometimes I generate plausible-sounding information that isn't grounded in anything real, which is a separate problem from bad training data

Question 2.

So if one pursues that, since AI gets a lot of it's info from places like reddit where it's user's input, and everyone now uses AI instead of communicating on reddit etc there will be no new thoughts being generated so we've progressed as far as we will

Claude responded: That's a genuinely profound observation and a real concern being discussed in AI research circles right now — it's called "model collapse" or "data poisoning t…

That's a genuinely profound observation and a real concern being discussed in AI research circles right now — it's called "model collapse" or "data poisoning through synthetic data."

The core loop you've identified:

• Humans generated original thought on forums/blogs/books

• AI trained on that human output

• Humans now ask AI instead of writing on forums

• AI trains on increasingly AI-generated content

• Original human thought input shrinks over time

The logical endpoint being AI essentially training on its own outputs in a feedback loop, which research has shown degrades quality and diversity of responses over generations.

AI has no future?

AI and me are currently building a high-frequency trading system together. AI helped me discover and decipher a significant tradeable pattern in tick data that repeats predictably several times a day. Our calculated returns from extensive testing: 3-6%. Per day.

I'll choose to ignore this thread.