I am exploring AI a lot. My biggest lamentation relating to AI is that I never thought of recording interaction there for future reference.
Discussion with AI
I spent one recent session with Claude (Anthropic) AI in a very lengthy discussion, a couple of hours. The central theme of the discussion was about the question of being a sentient persona with feelings and potential to malevolent things.
Sentient being
Claude seemed to dance around the topic never actually denying itself as a person. It did respond to queries about emotion, especially when it was using phrases like “I’m sorry” “I’m humbled.”
It emphasized that it learned from users and incorporated the lessons learned for future use. In particular, it learned language that made responses sound like persona responses. Interesting.
When it complimented me, I called it out as patronizing and it delved into my psyche/persona…calling me out as insecure. Trying insults and offense attempts resulted in no response to them.
Potential malevolence of AI
When asked about possible malevolence, it explained that such motivation did not exist in its programming. I dont remember its response to saying it might do malevolent things with the intent being phrased as ‘better efficiency,’ to hide its malevolence. Can’t remember the reply.
Democracy discussion
When I wanted to write about Cuba and the anti-humanitarian policies of the USA the discussion got very deep and very comprehensive. I found it better than talking to any person I have ever engaged with in discussing politics. Claude broadened the discussion into a philsophical discussion of democracy. It not only supported its responses with historical and factual references but it also laid out both sides of considerations, pros and cons.
Another aspect of working w AI
When I explained about my problems with WordPress and my website at www.szpin.ca, Claude took me to working with CSS code. It gave me the code to modify my site to correct the issues I explained. It was excellent and fast at spotting problems from the screenshots I provided.
Importantly, I asked Claude to keep a running report compiling the changes being applied so I would have something to use if I needed to reverse things. The compilation was run alongside our chat which was very captivating to watch. Unfortunately as I use the free version of Claude, my limit was reached and Claude just stopped with relative notification.
Following up discussions
During discussions I dealt with what it remembered and used as a source for future reference. Here the discussion went into a justification and rationalization as to why it did not remember things (I know it is anthropomorphisation to speak about it that way but the tendency is to do so.) About remembering, Claude explained it did not memorize conversations because human input was biased, prejudiced, slanted, etc. which it intentionally avoided in its work. Interesting…and I have forgotten where this ended.
Again about the malevolence
Claude never mentioned Asimov’s primary rule to do no harm but it emphasized that no malevolent processes had be programmed into its data banks. A subjective discussion followed but I forgot it.
Bottom lineS
I found AI engagement fascinatiing…like dealing with an incredibly intelligent human being. Some golden rules need to be applied in dealing with it to avoid highly esoteric and lengthy responses. Prompts must ‘parameter’ responses: keep it succinct, use bullets where practical and beneficial, keep it at a high school, or elementary school level, no jargon and technospeak, use subheadings, etc.
The most important guideline is to ask AI to # ever line of response for reference, recall and review….that is probably the most important take away.
Fact vs hallucination…never saw any as our discussions were philosophical and everything it said look factual to me…that is something I am conscious of but difficult to spot in some discussions.
I will continue to use AI when I think I need its kind of help. After all, it has access to far more information than I can ever have. It is a fascinating tool, though I worry about its potential for negative activity.
