>> Balancing chemical equations, Arrhenius acids and bases, basic lab work, and ...
> ... oxidation, exothermic reactions?
So? The syllabus doesn't include the chemistry of actual explosives or bomb design. Heck, it probably include little if any practical chemistry. It's all basic, basic, basic concepts.
Look, I could do a decent job helping a high school kid with their Chemistry homework, including things like redox reactions and exothermic reactions without even thinking about anything related to bombs. All those concepts can and are taught with "safer" and simpler reactions.
> A teacher who never actually studied the subject and only has knowledge of the pre-filtered material from the textbook would be little more than a talking textbook. I mean, textbooks are useful, but we already have those.
God help us if some dumb GPT model is supposed to replace teachers.
There's a whole lot of empirical knowledge actual practicing chemists have to learn from reports or find out experimentally, but don't know off the top of their (e.g. investigations about particular reactions, including particular explosives). If actual competent people don't know that, a stupid model for high school students doesn't need to be trained on it. Similar with practical bomb-making info.
You haven't provided any compelling argument for inclusion of bomb-relevant material in training data besides appealing to vague analogies.
> Let's just agree that pre-filtering the training text is not a practical way to prevent these models from producing emissions that are considered dangerous or taboo...
Sorry, I can't exactly agree with that. While pre-filtering might not be capable of wholly preventing all undesirable outputs, I think it can make them less likely and far less useful. If a high school student trying to make a pipe bomb has to feed in all the conceptual and practical information about bombs and explosives, and gets some impractical reaction out of it. That's a win, because it means the model wasn't much if any help.
And there's precedent for that. I'm under the impression that later versions of Stable Diffusion are pretty bad at making porn, because porn and porn-related keywords were pre-filtered better: https://www.theverge.com/2022/11/24/23476622/ai-image-genera....
The syllabus doesn't include those potentially taboo topics. The knowledge of chemistry leads to knowledge of those topics.
Sure, a teacher with zero knowledge of chemistry can follow a syllabus, read the textbook out loud, make sure everyone's multiple choice answers match the answer key, and try to teach chemistry to some first approximation of "teach". A primitive computer program can do that too.
What happens when a student asks a question that deviates slightly from the syllabus because they didn't quite grasp it the way it was explained? The teacher can't answer the question, but a "dumb GPT model" trained on the entirety of the Internet, including a ton of chemistry, including the syllabus, probably can.
But yes, if you pre-filter the training data to include only the words of the syllabus, the language model will be just as poor of a teacher as the human who did the same thing. Reminds me of my first "programming teacher", a math teacher who picked up GW-BASIC over the summer to teach the new programming class. He knew nothing.
We never even reached the contentious part of this discussion. This part should be obvious. You can't just filter out "bomb" because the student could explain in broad terms what they mean by the word and then ask for a more detailed description. You can't filter out "explosion" because the teacher might need to warn the students about the dangers of their Bunsen burner. You can't filter out all the possible chemical reactions that could lead to an explosion, because they can be inferred from the subject matter that it's supposed to be teaching.
The same goes for things like negative stereotypes, foul language, talking about "illegal or unethical" behaviors (especially as laws and ethical norms can change after training, or differ between usage contexts). Pre-filtering is just a nonstarter for any model that's intended to behave as an intelligent being with worldly knowledge. And if we drop those requirements, then we're not even talking about the same technology anymore.
> ... oxidation, exothermic reactions?
So? The syllabus doesn't include the chemistry of actual explosives or bomb design. Heck, it probably include little if any practical chemistry. It's all basic, basic, basic concepts.
Look, I could do a decent job helping a high school kid with their Chemistry homework, including things like redox reactions and exothermic reactions without even thinking about anything related to bombs. All those concepts can and are taught with "safer" and simpler reactions.
> A teacher who never actually studied the subject and only has knowledge of the pre-filtered material from the textbook would be little more than a talking textbook. I mean, textbooks are useful, but we already have those.
God help us if some dumb GPT model is supposed to replace teachers.
There's a whole lot of empirical knowledge actual practicing chemists have to learn from reports or find out experimentally, but don't know off the top of their (e.g. investigations about particular reactions, including particular explosives). If actual competent people don't know that, a stupid model for high school students doesn't need to be trained on it. Similar with practical bomb-making info.
You haven't provided any compelling argument for inclusion of bomb-relevant material in training data besides appealing to vague analogies.
> Let's just agree that pre-filtering the training text is not a practical way to prevent these models from producing emissions that are considered dangerous or taboo...
Sorry, I can't exactly agree with that. While pre-filtering might not be capable of wholly preventing all undesirable outputs, I think it can make them less likely and far less useful. If a high school student trying to make a pipe bomb has to feed in all the conceptual and practical information about bombs and explosives, and gets some impractical reaction out of it. That's a win, because it means the model wasn't much if any help.
And there's precedent for that. I'm under the impression that later versions of Stable Diffusion are pretty bad at making porn, because porn and porn-related keywords were pre-filtered better: https://www.theverge.com/2022/11/24/23476622/ai-image-genera....