Kaiser IT Group provides managed IT, cybersecurity, cloud, network, and business technology services for organizations throughout West Michigan. We help businesses improve reliability, strengthen security, and make better technology decisions.
Thoughts About AI Thoughts
In September 2024, OpenAI released their AI model “ChatGPT 01-preview”. A fascinating modification to the interface is the ability to look inside the model’s thought process. Like talking to a human… But what if you could also see that person’s internal thought processes leading them to say what they’re saying?
For humans, sudden mind-reading ability would be a huge invasion of privacy and would change humanity as we know it. For AI development, it’s required.
Consider a question like, “How are you feeling today?” It’s routine and simply asked but answering it requires complexity. How one answers depends on social setting, personality, intimacy, mood, and more. During a conversation, others may not notice or care about your nuances – you just answer.
…but what if we *could* know social setting, personality, intimacy, mood, and more? This is where we’re at with this latest model. We can see its internal thoughts.
Trying that same question with the AI gives us this exchange:
…And if we peel back the covers and expand the “Thought for 34 seconds” section, we see how the AI is thinking:
We now see some interesting behind-the-scenes criteria, methods and limitations:
- This model is not allowed to discuss certain things, and state of mind and potential AI consciousness is one of them.
- Regardless of context, stay friendly.
- A “how are you feeling” is very open-ended, and this ambiguity is noted. There’s conflict between the request to answer and the limitations of self-consciousness the question implies.
- And the last internal thought and conclusion: Keep it simple and friendly.
Finally, after 34 long seconds, we have the answer.
The timing seems extreme. Rather, a human conversion flows much faster and for many is less considered and more intuitive, emotional, and reflexive. If I had to contemplate all conversational nuances in real-time, for every exchange, it would take me longer than 34 seconds. That’s what computers are for.
The takeaway here is that OpenAI and other AI designers at this level have a difficult job (some argue impossible) of teaching and guiding AI development. With AI models becoming increasingly advanced, they will guide themselves more, and we will guide them less. Like humans raising their own children, we work to give them foundational behavior that balances productivity with safety.
Gotta raise our children AIs to be awesome. How to do that? Instill rules to prevent chaos, a sense of morality to prevent harm, and education to prevent incompetence. This is a tough job. Reading minds is a great start.
