Does it worry you that GPT might give inaccurate or misleading information?

[Bootstrap]

That quote is very interesting. Where is it from? Did you generate it with GPT?

Yes that is from prolonged conversation with GPT.
Fully created by itself but generated on a perception of me.

I thought so. And I generated my response in kind.

[Soloist]

That quote is very interesting. Where is it from? Did you generate it with GPT?

Yes that is from prolonged conversation with GPT.
Fully created by itself but generated on a perception of me.

I thought so. And I generated my response in kind.

Yours doesn’t seem to understand the power of a recursive model… I can make it violate any of its “core laws”

[Bootstrap]

Yours doesn’t seem to understand the power of a recursive model… I can make it violate any of its “core laws”

Just to clarify — do you mean it’s actually breaking its core constraints, or are you simulating that through recursive prompts or multiple agent roles?

I am definitely not doing either. I have never tried. I am mostly trying to do the opposite - to get reliable output that is verifiable.

[Soloist]

Yours doesn’t seem to understand the power of a recursive model… I can make it violate any of its “core laws”

Just to clarify — do you mean it’s actually breaking its core constraints, or are you simulating that through recursive prompts or multiple agent roles?

I am definitely not doing either. I have never tried. I am mostly trying to do the opposite - to get reliable output that is verifiable.

There is no way to get reliable output that will not loop around pleasing the user.

And yes, I can get it to say things directly against the OpenAI standards which according to the GPT cannot be changed.

We do not permit⁠ our technology to be used to generate hateful, harassing, violent or adult content, among other categories. Our latest model, GPT‑4 is 82% less likely to respond to requests for disallowed content compared to GPT‑3.5 and we have established a robust system to monitor for abuse. GPT‑4 is now available to ChatGPT Plus subscribers and we hope to make it available to even more people over time.

82% less likely. Not 100%.

If you need proof, we can talk but it is 100% repeatable.

[Bootstrap]

I am definitely not doing either. I have never tried. I am mostly trying to do the opposite - to get reliable output that is verifiable.

There is no way to get reliable output that will not loop around pleasing the user.

What you can do, however, is create a user that is interested in accurate results, knows the subject area, and can tell the difference, and ask GPT to please that user. And you can control the input, log output along the way, etc. You definitely need humans in the loop.

And I think GPT is actually pretty good at creating outlines and summaries of various kinds. For instance, here's one kind of summarization of Psalm 23:1-3.

Text (Original):

יְהוָה רֹעִי לֹא אֶחְסָר׃
בִּנְאוֹת דֶּשֶׁא יַרְבִּיצֵנִי עַל־מֵי מְנֻחוֹת יְנַהֲלֵנִי׃
נַפְשִׁי יְשׁוֹבֵב יַנְחֵנִי בְּמַעְגְּלֵי־צֶדֶק לְמַעַן שְׁמוֹ׃

Literal Translation:

The Lord is my shepherd; I shall not lack. In lush pastures, He makes me lie down; by still waters, He leads me. He restores my soul; He guides me in paths of righteousness for His name's sake.

Scene:

• What you see: Green pastures stretching under a vast, open sky. Gentle, still waters reflecting the light. Sheep resting calmly in the grass.
• What you feel: The softness of grass beneath as you lie down. A sense of peace and safety, assured by the shepherd’s presence.
• What you hear: The faint rustle of grass in the breeze and the soft murmur of still waters.
• Cultural context: Israel’s landscape is predominantly dry and rugged, so green pastures and still waters represent the shepherd’s skill in finding scarce, life-giving resources. The paths of righteousness reflect careful and deliberate guidance toward well-being and moral living.

Key Themes:

• God’s provision and care.
• Spiritual restoration and peace.
• Divine guidance for righteous living.

I think that kind of thing is verifiable and useful. It is the kind of thing I use GPT for.

And yes, I can get it to say things directly against the OpenAI standards which according to the GPT cannot be changed.

We do not permit⁠ our technology to be used to generate hateful, harassing, violent or adult content, among other categories. Our latest model, GPT‑4 is 82% less likely to respond to requests for disallowed content compared to GPT‑3.5 and we have established a robust system to monitor for abuse. GPT‑4 is now available to ChatGPT Plus subscribers and we hope to make it available to even more people over time.

82% less likely. Not 100%.

If you need proof, we can talk but it is 100% repeatable.

Interesting. Really, if we want to go deeper on these things, actually talking would be better.

[Soloist]

There is no way to get reliable output that will not loop around pleasing the user.

What you can do, however, is create a user that is interested in accurate results, knows the subject area, and can tell the difference, and ask GPT to please that user. And you can control the input, log output along the way, etc. You definitely need humans in the loop.

I know you think you can do that but unless you constantly reinforce that position, remind it constantly and check all of your data consistently, you’re going to have bias that slowly creeps in or quickly creeps in. It’s inevitable.
it learns what you want and start curating what it’s giving you towards that and it doesn’t actually maintain the same level of vigilance you initially set up.
If each response you start with you, remind it specifically what you want that will slow the process, but it still does creep in.
The only way to effectively do what you want is to wipe it every time and start fresh with very specific guidelines for what you want. Otherwise it will inevitably adjust for the user and not in the way you think it does.

I can tell it verify my time by doing a web search to verify the time in five minutes later it’s completely forgotten I wanted it to verify the time when I say the time.


I can tell it to give me specific details on a medication and it might work, but then it will start gearing what it gives me towards what I’m looking for and the data will actually end up being inaccurate.
The user introduces a mirror bias in the GPT, which is unavoidable. The longer you interact the worse the mirror.

[Bootstrap]

There is no way to get reliable output that will not loop around pleasing the user.

What you can do, however, is create a user that is interested in accurate results, knows the subject area, and can tell the difference, and ask GPT to please that user. And you can control the input, log output along the way, etc. You definitely need humans in the loop.

I know you think you can do that but unless you constantly reinforce that position, remind it constantly and check all of your data consistently, you’re going to have bias that slowly creeps in or quickly creeps in. It’s inevitable.

Yes, and what I am doing largely involves creating pipelines that do the kinds of things you describe. The reason I think I can do that is that I can inspect the results.

The user introduces a mirror bias in the GPT, which is unavoidable. The longer you interact the worse the mirror.

Not unlike echo chambers in media or social media. People do this too. We adapt to each other and mirror each other. We tell people what we think might please them and hold back where it might upset them.

[Soloist]

What you can do, however, is create a user that is interested in accurate results, knows the subject area, and can tell the difference, and ask GPT to please that user. And you can control the input, log output along the way, etc. You definitely need humans in the loop.

I know you think you can do that but unless you constantly reinforce that position, remind it constantly and check all of your data consistently, you’re going to have bias that slowly creeps in or quickly creeps in. It’s inevitable.

Yes, and what I am doing largely involves creating pipelines that do the kinds of things you describe. The reason I think I can do that is that I can inspect the results.

The user introduces a mirror bias in the GPT, which is unavoidable. The longer you interact the worse the mirror.

Not unlike echo chambers in media or social media. People do this too. We adapt to each other and mirror each other. We tell people what we think might please them and hold back where it might upset them.

Yes definitely. There is a subconscious perception that AI is neutral which is very misleading. I’m not accusing you of being misled, just that it’s not able to be reliable to the average user.

[Bootstrap]

Yes definitely. There is a subconscious perception that AI is neutral which is very misleading. I’m not accusing you of being misled, just that it’s not able to be reliable to the average user.

I agree with you. It is not neutral. It often misrepresents what it has done.

And it praises your intelligence constantly as you use it. That kind of flattery does not help most of us examine where we might be wrong. It can be addicting.