[FRIAM] Alignment (with what?)

[steve smith]

On 6/4/25 9:43 AM, Marcus Daniels wrote:
>
> Really?   The kind of biases I don’t like seem to be system prompting 
> preferences.   For example, when an idea is marginal, ChatGPT will 
> kick the can forward to keep the conversation going.  Perhaps the 
> instinct is to create a demand signal or “billable hours”.  Gemini Pro 
> will tend to be cautious.    To use them effectively one needs to have 
> some self-discipline.  I don’t see these tools doing obvious motivated 
> reasoning like people do.   ChatGPT will especially follow your lead, 
> which can be a bad lead; it is a people pleaser.
>
I agree that the chatbot interface presented to GPT (and others possibly 
less acutely) is a sycophant biased to both please humans and encourage 
our continued engagement.
>
>
> My latest “wow” moment:   I was checking up on the Kerch bridge attack 
> and took a few frames from the video and pasted them for ChatGPT.  I 
> asked if the explosion suggested poor placement by Ukrainian frogmen.  
> It explained that blast off to the side would be typical and suggested 
> an attempt to displace the pier laterally deep under the water.  It 
> did not imply an attempt to knock the bridge down but rather to make 
> it unsafe to use, while also leading to incremental structural failure.
>
So to say that it reflected the (accurate?) bias of such an operation to 
try to maximize effect by "satisficing" by focusing on "making it 
unsafe" rather than wasting tactical resources on more complete 
destruction?

I suppose it would be too much to ask for it to cause distortions in the 
bridge which selectively modify it to become useful only for Ukrainian 
goals vs Russian or perhaps civilian vs military? By derating it's 
functionality marginally perhaps it can handle medium-load commercial 
transport but risk failure under heavier military (e.g.moving tanks) use?


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://redfish.com/pipermail/friam_redfish.com/attachments/20250604/e6450731/attachment.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: OpenPGP_0xD5BAF94F88AFFA63.asc
Type: application/pgp-keys
Size: 3118 bytes
Desc: OpenPGP public key
URL: <http://redfish.com/pipermail/friam_redfish.com/attachments/20250604/e6450731/attachment.bin>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: OpenPGP_signature.asc
Type: application/pgp-signature
Size: 840 bytes
Desc: OpenPGP digital signature
URL: <http://redfish.com/pipermail/friam_redfish.com/attachments/20250604/e6450731/attachment.sig>