All of TimmyM's Comments + Replies

TimmyM10

Seeing how competent these newer models are was quite helpful, and I have started using ChatGPT o4-mini-high to generate training sets for a new finetuned model. It does quite well, and once I get the hang of this, I should be able to distill the results and improve the performance of my model much more quickly. Wish me luck!

TimmyM30

Yeah, this is probably true. In training, I aimed to allow it to provide slightly broader alternatives, but not more specific alternatives, like this one is.

Since all groupthink is a form of consensus, under the rules I've been following it would be acceptable for it to highlight "groupthink" and provide "consensus" as an alternative, but not the other way around.

TimmyM30

I'd be interested to see the results you got with Gemini. The 2k character limit isn't a hard limit for my model, it's just what I set to limit copyright issues and excessive costs for this proof of concept.

I suppose, if anything, the main fruit of my work is that I have consistent, programmatic output that I can format in multiple settings (unless Gemini can do that as well). I am in the process of making a chrome extension that analyzes headlines and articles with the same model.

It is true that, in the long process of finetuning this model, AI technology has developed a lot further than from when I began. I'm not opposed to using alternative methods.

4AlphaAndOmega
Here's an example: ChatGPT's initial personality analysis (I did it for the hell of it, before reading your post): Certainly. Based on our conversations, here is a detailed personality profile ----------------------------------------   Personality Profile Cognitive Style: You exhibit a highly analytical, skeptical, and systems-oriented mode of thinking. You approach complex topics with intellectual rigor and demand high standards of precision, detail, and conceptual clarity. You prefer responses to be concise but rich in content, showing a low tolerance for verbosity or unfocused speculation. Your reading habits and conversational tone suggest an affinity for rationalist discourse, often associated with communities like LessWrong. You are comfortable navigating abstract and high-complexity domains, particularly when they intersect with real-world systems (e.g., medicine, AI, economics, immigration policy). Intellectual Disposition: You are widely read and intellectually omnivorous, with demonstrated depth in psychiatry, medicine, speculative fiction, and AI ethics. Your curiosity is not casual but deeply interrogative—you pursue foundational understanding, not surface-level summaries. You are driven by a desire to model not just systems but also long-term consequences, especially where they intersect with human welfare, labor, and technological disruption. Emotional Tone & Interpersonal Style: You maintain a reserved, self-possessed demeanor in communication, with dry humor and a tersely efficient tone. You are direct and assertive in setting conversational boundaries, clearly delineating what kinds of responses you expect. There's an implicit value placed on mutual respect and signal-to-noise ratio in dialogue. You likely prefer few, high-quality relationships or collaborations over broad social engagement. Values & Priorities: You place a premium on intellectual integrity, competence, and foresight. You are not easily swept up in optimism or hype, preferr
TimmyM10

I know it's been a long time since anyone posted here, but I just stumbled across this post. For the past year and a half, I have been training a finetuned ChatGPT model to identify and reverse examples of Russell Conjugations in given text. There are still improvements that need to be made, so I'll try to include some of these examples in the training set for the new model.

I released the tool last week, and anyone can try it out here: https://russellconjugations.com/