tra38
commented
10 months ago - Pretend Optimization, version 1 has 4,458 words.
- Pretend Optimization, version 2 has 18,359 words, an addition of 13,901 words over the original.
For comparison's sake, my 2020 project had 3,260 words, and my 2018 project has 12,328 words.
So I have proven that ChatGPT can generate meaningful commentary, and human-readability now exists.
However, due to the nature of an RPG supplement, adding even more word count would become superfluous. So in order to get 50,000 words, I have to decide whether to generate more AI material (maybe a tabletop RPG mission) or just get the AI to print "meow" 50,000 times. I might lean towards the latter, just because that is a novel problem - I can write a loop to generate meows, but how do I get the AI to do that?.
How I Did It
I wrote prompts and send it through the ChatGPT UI. The initial prompt used to start up discussion:
(text that I want commentary on) Start a discussion between two AIs - Reddy (cynical, informal) and Acey (optimistic, academic) on the above paragraphs, identifying strengths, weaknesses, and flaws of the above. Then, after analyzing the argument, both Reddy and Acey will write unique responses to the above.
There are situations where I want to ask the two AI personalities further questions - in that case, I would send additional prompts to ChatGPT and then paste the results. The additional user prompts are clearly marked, as they also serve as a way for me to comment on the AI's responses as well. Cherry-picking was minimal, in that I did rewrite the user prompts to try to get outputs that I like, but once I get that, I copied and pasted it. I did not modify the AI output in any other way.
Due to user error, I ended up using Advanced Data Analytics instead of the normal GPT-4 model. I'm not sure how that affects the output but I suspect it does. No Python code was ever written or executed by ChatGPT.
In retrospect, while Acey's views provided a useful corrective to Reddy's cynicism, it's also kinda boring to read. I'm not going to get rid of Acey or modify him, but maybe in next year's project, I'll try to make Acey's generated text a bit more readable.
The text was generated on 11/6/2023, which is important because OpenAI also began rolling out new changes to ChatGPT, such as a new model that has built-in access to "Advanced Data Analytics", Web Browsing, and Dall-E 2. I just got those changes today (11/7/2023), so any future generated text may differ significantly. This is a problem with relying on a closed-source AI system that gets continual updates - even if it constantly gets better overall, you don't know how it will impact your specific use cases.
Below is an interesting quote from "Pretend Optimization, version 2", which might serve as a good preview of the text output.
Acey: I've been pondering over the assertion that ChatGPT could be considered an AGI based on the definition of being able to attempt to solve any arbitrary problem. It's a captivating idea, isn't it?
Reddy: It's a stretch, Acey. Just because this chatbot can tackle a bunch of topics doesn't make it the sci-fi AGI everyone's been yammering about. Attempting isn't the same as doing. I can attempt to fly by flapping my arms, but that won't make me a bird.
Acey: I appreciate the metaphor, Reddy, but isn't the ability to attempt a wide array of problems without domain-specific programming a step towards general intelligence? It's not about succeeding every time, but about the adaptability and transferability of skills across different contexts.
Reddy: Sure, it's a step. But a baby step. A real AGI needs to understand, learn, and even be self-aware to some degree. ChatGPT's clever, I'll give it that, but it's like a parrot mimicking speech patterns. It understands syntax well enough but not the semantics, the meaning behind words.
Acey: That's a fair point, Reddy. The distinction between syntactic pattern recognition and semantic understanding is crucial. ChatGPT operates on the former, processing and generating language based on statistical likelihoods, not genuine comprehension.
Reddy: Exactly. And let's not forget about self-improvement. AGIs are supposed to be able to improve themselves, right? ChatGPT can't rewrite its own code or upgrade its algorithms. It's bound by the limitations we set for it.
Acey: True, the capacity for recursive self-improvement is often cited as a hallmark of AGI. However, the ability of ChatGPT to provide insights across disciplines and to simulate problem-solving conversations like ours is indicative of a system that is not narrowly confined. It's an impressive display of the versatility of machine learning.
Reddy: Versatility, sure. But let's not call it general intelligence. It's like a Swiss Army knife. It's got a tool for a lot of situations, but it's not going to replace a full toolbox. And it certainly won't replace the mechanic.
Acey: Your analogy is quite vivid, Reddy. While ChatGPT may not be an AGI by the strictest definitions, it does challenge us to refine our understanding of intelligence in machines. It's a bridge between narrow AI and the still-theoretical AGI. Perhaps it's an indicator of what's to come, even if it's not the destination itself.
Reddy: I can agree with that. It's a signpost on the road to AGI, not the end of the road. Let's just make sure we don't get ahead of ourselves and mistake a rest stop for paradise.
In 2018-2019, I wrote a Paranoia RPG supplement entitled Pretend Optimization, focused on representing Artificial General Intelligence in a tabletop setting. This writing was done manually. However, the rise of ChatGPT in 2023 suggests that my supplement is woefully out of date and needs updating. This can be done algorithmically with the help of ChatGPT, keeping the human text intact but adding AI-generated comments and elaboration.
I plan on using GPT-4, by the way.
Questions: