Training an instance of GPT4
Source date (UTC): 2025-08-06 00:55:37 UTC
Original post: https://twitter.com/i/web/status/1952896306084802843
Theme: AI
-
I have updated CurtGPT with the current state of Volume 1 – The Crisis of the Ag
I have updated CurtGPT with the current state of Volume 1 – The Crisis of the Age.
FWIW: when cross referenced with criminal, ethical, and moral violations (I never thought of asking that before) the answers are even better.
Source date (UTC): 2025-08-05 22:14:04 UTC
Original post: https://twitter.com/i/web/status/1952855652885512395
-
The more I teach CurtGPT the more amazing and clear its responses. I’m having a
The more I teach CurtGPT the more amazing and clear its responses. I’m having a ball. 😉
Source date (UTC): 2025-08-05 22:11:52 UTC
Original post: https://twitter.com/i/web/status/1952855097706479840
-
I suppose we shouldn’t be blown away by it because we’ve engineered it for this
I suppose we shouldn’t be blown away by it because we’ve engineered it for this purpose. But it’s still amazing. 😉
Source date (UTC): 2025-08-05 21:06:36 UTC
Original post: https://twitter.com/i/web/status/1952838674674233501
-
It can give whatever detail we ask. It’s amazing. Regarding retaliation for pedo
It can give whatever detail we ask. It’s amazing.
Regarding retaliation for pedophilic murder, it returned this chart of violations of demonstrated interest.I’m working through the ‘crimes’ list and it’s amazing.
Source date (UTC): 2025-08-05 20:44:33 UTC
Original post: https://twitter.com/i/web/status/1952833124955730199
-
So basically if we get the categories of response correct, the vocabulary of del
So basically if we get the categories of response correct, the vocabulary of delivery is an arbitrary preference.
Similarly, if we address markets with different taboos/sacreds (moral biases) then that’s just an alternation in the valence of demonstrated interests.
Source date (UTC): 2025-08-05 20:41:22 UTC
Original post: https://twitter.com/i/web/status/1952832322320908598
-
Agreed. It just surprised me. I thought I would have to soften it, and instead I
Agreed. It just surprised me. I thought I would have to soften it, and instead I’m sensitive to it being too soft.
The models answer almost identically but a bit differently in tone. o3 is more terse and 4o is more narrative and as such has a bit more breadth. For the average audience I prefer 4o.
My only concern is that given the argument (response) is so precise it’s a bit long. But it will definitely teach people what we want them to learn.
Source date (UTC): 2025-08-05 20:18:48 UTC
Original post: https://twitter.com/i/web/status/1952826643501989993
-
ChatGPT “Miracles” Fascinating. I’m plugging holes in “CurtGPT” (Our Custom Chat
ChatGPT “Miracles”
Fascinating. I’m plugging holes in “CurtGPT” (Our Custom ChatGPT) by walking through difficult moral (and therefore legal) questions, and GPT (custom gpt) just asks if the new rule we discovered needs to be added to the system prompt (json file). Then does it.Now, as part of the ethical system, I’ve catalogued the 35 or so core questions of law, ethics, and morality over which there is any debate, and walking through them with ChatGPT4o in “CurtGPT”. Plus I’m asking for a more specific output. And I’m kind of floored how well this is going.
Now, until I publish it and people test it,
One negative side effect is that this training is not as ‘terse, harsh, and clear’ as our earlier version. But it’s certainly more accessible and educational. And since our mission is ending the industrialization of lying in matters public, this more accessible and expressive form appears to be more compatible with our mission.
Important Note: Working with math and programming means that training the AI has some means of closure. But otherwise trainers are stuck with normativity during the second period of the industrialization of lying. The first being the abrahamic sequence religions, the second being the marxist sequence pseudoscientific religions. So normativity in training AI’s just reinforces the 20th-21st century industrialization of lying and pseudoscience.
Our work solves this problem. So we can train AIs in first principles (laws) and it’s actually EASIER to train an AI from our position than it is for every other LLM provider.
cc:
@sama
Source date (UTC): 2025-08-05 18:52:52 UTC
Original post: https://twitter.com/i/web/status/1952805017414713492
-
THE SCIENCE OF THE HUMANITIES I started working on the first principles and cano
THE SCIENCE OF THE HUMANITIES
I started working on the first principles and canonical training of AIs in the Humanities today. It is going fast, and is rewarding – and we have unified the formal, physical, behavioral, and now literary sciences.This has led to a system of measurement for the science of the humanities just as it has in the other ‘sciences’.
But like law and economy this is not a ‘best’ race. It’s an understanding of the needs of the people at their degree of evolution, and a map for how to continue their evolution.
As with most of our work we treat humans as a distribution of evolutionary biases by sex differences and seek to assist those two biases in achieving shared goals rather than to claim one is superior or inferior.
This in itself is one of our contributions to the discourse.
Source date (UTC): 2025-07-31 02:13:02 UTC
Original post: https://twitter.com/i/web/status/1950741464662852028
-
The AI is trained with Positive and Negative Assertions and explanations (socrat
The AI is trained with Positive and Negative Assertions and explanations (socratically). So yes – although we prefer it come up with examples specific to the user’s context rather than canned responses.
Source date (UTC): 2025-07-31 01:01:56 UTC
Original post: https://twitter.com/i/web/status/1950723569941610513
