It’s a slippery slope to go from AI as a support to a reliance.
I liked your quote: And then another model, another feature, another “this changes everything.” (insert vomit emoji)
I’m a children’s book writer. You would not believe (or maybe you would) the hype of AI generated works to make anyone be a (poor) children’s book author. It’s reliance, and it’s a difficult thing to figure out how I feel about the whole AI utility.
I have two young kids and read to them 5 to 6 books every night... I think I know exactly what you mean. Good children books have a deeper meaning. I love the ones from Rachel Bright for example. No way AI could have written this story. This is lived-through content.
It's difficult for AI to write simply and effectively, so I don't see AI writing children's book fully. Maybe to assist and assess if what has been written is compelling for children.
I’ll add this: following the ins and outs of the models is a compelling pastime, especially for insiders. But they are just tools - their purpose is to be used to build stuff.
I’m sure there are people who LOVE hammers, but knowing everything there is to know about hammers doesn’t say anything about your carpentry technique, the solidity of the walls you build, or whether the architecture of the building you built with that hammer is worth a damn. By all means use the best hammer. But make the house you build the priority,
I've been using AI often and effectively both in my writing and in my job. I find it a valuable tool to both help me decide on next steps, as well as review what I had completed before.
One challenge I haven't really been able to figure out how to properly solve yet is when I run into a circular issue, that when suggesting things from one point of view, it's recommending a solution. When I consider the issue and discuss issues and things with it, the AI starts recommending another solution.
Then I start pointing out ideas and issues with the new solution, with the goal to actually use this solution, then it returns to recommending the first solution.
Do you have experience or ideas on how to solve such circular issues?
So far my idea has been to step back, and try asking again another time with different words, but it feels horribly inefficient.
This is a signal you are stress testing the model, which is a compliment in the end for you!
LLM's can do judgement (and you as a human are in the driver seat). There are a couple of ways to force the LLM to stick to the plan, or weigh the options correctly:
- Separate “exploration” from “decision” (start a brainstorm first, and list the options), then in a separate follow-up prompt, score the options.
- Let LLM's score the options, make a decision framework (that you thought of yourselves).
- When you point out issues, add “I’m not rejecting this.” to your prompts. Add something like this to your prompt: "I want to stick with Option B. These are risks to address, not reasons to switch. Help me mitigate them"
Hope this helps! There are many more ways to tackle this.
Appreciate you taking the time to lay this out. I think where I wasn’t clear enough is that my issue wasn’t about how to guide the model toward a decision, but about a circular pattern I kept running into.
Even when separating exploration and decision, or explicitly anchoring to Option B, the model kept re-opening the decision as if risk identification itself implied uncertainty. What I was trying to get at was how to stop that loop once a commitment is made, not how to arrive at the commitment in the first place.
That’s on me for not framing it sharply enough. Still useful to see how you think about it.
I think the solution is particularly simple, you paste the outcomes after a decision is made into a new chat. LLM’s tend to look at the full conversation in the chat, so everything is still used as context. To rule that out, you can start a new chat with only the context relevant at that point.
That’s the only solution I’ve landed on as well. I was mostly wondering if you’d come across any other approaches beyond starting a fresh chat with trimmed context.
While writing this I did realize one can ask the LLM to clean it's context, but it's not guaranteed to work.
AI creates space but also has a tendency to take up that exact same space. Loved the way you layered the different capability levels, I found my journey to be pretty similar.🦩
It’s a slippery slope to go from AI as a support to a reliance.
I liked your quote: And then another model, another feature, another “this changes everything.” (insert vomit emoji)
I’m a children’s book writer. You would not believe (or maybe you would) the hype of AI generated works to make anyone be a (poor) children’s book author. It’s reliance, and it’s a difficult thing to figure out how I feel about the whole AI utility.
Thank you, Author Gold!
I have two young kids and read to them 5 to 6 books every night... I think I know exactly what you mean. Good children books have a deeper meaning. I love the ones from Rachel Bright for example. No way AI could have written this story. This is lived-through content.
It's difficult for AI to write simply and effectively, so I don't see AI writing children's book fully. Maybe to assist and assess if what has been written is compelling for children.
I write my own sentences to have more depth and feeling.
I’ll add this: following the ins and outs of the models is a compelling pastime, especially for insiders. But they are just tools - their purpose is to be used to build stuff.
I’m sure there are people who LOVE hammers, but knowing everything there is to know about hammers doesn’t say anything about your carpentry technique, the solidity of the walls you build, or whether the architecture of the building you built with that hammer is worth a damn. By all means use the best hammer. But make the house you build the priority,
I love this analogy. Thanks for sharing!
I've been using AI often and effectively both in my writing and in my job. I find it a valuable tool to both help me decide on next steps, as well as review what I had completed before.
One challenge I haven't really been able to figure out how to properly solve yet is when I run into a circular issue, that when suggesting things from one point of view, it's recommending a solution. When I consider the issue and discuss issues and things with it, the AI starts recommending another solution.
Then I start pointing out ideas and issues with the new solution, with the goal to actually use this solution, then it returns to recommending the first solution.
Do you have experience or ideas on how to solve such circular issues?
So far my idea has been to step back, and try asking again another time with different words, but it feels horribly inefficient.
Thank you for the question, Dan!
This is a signal you are stress testing the model, which is a compliment in the end for you!
LLM's can do judgement (and you as a human are in the driver seat). There are a couple of ways to force the LLM to stick to the plan, or weigh the options correctly:
- Separate “exploration” from “decision” (start a brainstorm first, and list the options), then in a separate follow-up prompt, score the options.
- Let LLM's score the options, make a decision framework (that you thought of yourselves).
- When you point out issues, add “I’m not rejecting this.” to your prompts. Add something like this to your prompt: "I want to stick with Option B. These are risks to address, not reasons to switch. Help me mitigate them"
Hope this helps! There are many more ways to tackle this.
Appreciate you taking the time to lay this out. I think where I wasn’t clear enough is that my issue wasn’t about how to guide the model toward a decision, but about a circular pattern I kept running into.
Even when separating exploration and decision, or explicitly anchoring to Option B, the model kept re-opening the decision as if risk identification itself implied uncertainty. What I was trying to get at was how to stop that loop once a commitment is made, not how to arrive at the commitment in the first place.
That’s on me for not framing it sharply enough. Still useful to see how you think about it.
Hi Dan, no worries, I understand it now!
I think the solution is particularly simple, you paste the outcomes after a decision is made into a new chat. LLM’s tend to look at the full conversation in the chat, so everything is still used as context. To rule that out, you can start a new chat with only the context relevant at that point.
That’s the only solution I’ve landed on as well. I was mostly wondering if you’d come across any other approaches beyond starting a fresh chat with trimmed context.
While writing this I did realize one can ask the LLM to clean it's context, but it's not guaranteed to work.
This is the way. 😅 and the only one up until now
AI creates space but also has a tendency to take up that exact same space. Loved the way you layered the different capability levels, I found my journey to be pretty similar.🦩