That might be, but the argument was that poor cache utilization was costing Anthropic too much money in other harnesses. If cache is considered in rate limits, it doesn’t matter from a cost perspective, you’ll just hit your rate limits faster in other harnesses that don’t try to cache optimize.
Yes I currently only generate 3 questions per article.
I tend to get worse results if I generate more questions per article. I instead try to encourage exploration by showing
related articles (clickable) to try to get a wider ranger of topics instead.
I have a lot of work to do to make that whole experience smoother though.
It would actually be interesting to know what the energy required by a human is to
research a topic, think of plausible options, find references and type it out.
I'm not sure how much energy Trivai requires, or if it's even remotely comparable, but you might be happy to know that all generated questions are stored in a database.
I would guess that humans produce "perfect" results 95+% of the time.
In my own tests, Trivai gets "good" results, ~90% of the time.
I've actually seen this particular question before.
It's hard to figure out why some questions get bad or wrong answers.
This question also doesn't have a proper reference.
I think it's difficult to get the initial result perfect 100% of the time.
So far, my solution to this is an upvote/downvote system.
This lets me either hide downvoted questions, or improve them.
Right now a user who downvoted can report an issue, and I will generate
new variations of the question. If either of the variations receive enough
upvotes, the question will be updated.
Very cool!
I've experimented with a similar thing for trivia questions based on Wikipedia articles.
I'm also trying to reference the part of the text where the question was generated from, so the user can verify if it was correct or learn more about it in context.
I think you have some good points here. I agree that just releasing this unrestricted from the start is probably a bad idea. Being cautious is most likely a good thing.
> More generally I think there has to be a match between a social unit's (states, companies, individuals, etc) design and ability to self-regulate and the power of the technology it has access to (more power, more responsibility) or you will see undesirable outcomes. Give a kid a gun, North Korea with nukes... In contrast the OpenAI people seem like fine folks. After all they do intend their work to be more widely applied, just carefully.
This issue comes down to who gets to control and regulate these technologies. While OpenAI may have good intentions, they have demonstrated that they are willing to decide what is ethical or moral for everyone else. This gives a lot of power to one company if this service becomes widespread.
> as in the second case we would likely see ungodly amounts of information pollution on the internet (and self-coupling effects from AI training on and copying each other) and in second-order a further "epistemic scattering" (by which I mean a proliferation of belief systems that disagree in fundamentals like the laws of nature).
I think this is an interesting point. We probably have a window right now where a large enough content online is still produced by humans. Closing this too early by flooding the internet with lots of completely inaccurate content would not be good. However, it doesn't seem like the restrictions are primarily focused on inaccuracy, but on ethics.
> Your second example seems a bit whimsical
It is, my point with this example is that most people (I think) would not have a problem with the AI generating this type of content, but some filter somewhere is preventing this from happening either way. As you said, this is a WIP so it might be adjusted in the future.
> In Germany access to (and display of) Nazi propaganda material in historical collections is restricted (may be presented only in an educational context) because the state doesn't want right-wing groups to go wild in reusing it (among other reasons). Now imagine convincing hate speech being generated on the fly.
Yes, and to tie into your first point (that ChatGPT is public), it is the display and publication of this type of content that's restricted. As far as I know there are no restrictions on producing it for yourself. ChatGPT isn't "public" in that sense.
Of course we're sharing all our conversations with OpenAI and there might be legal reasons for why they don't want to be involved in the production (and storage) of hate speech. I'm not sure if anybody knows what their legal liabilities would be had they released a model that could be run locally and still produce this type of content. Then again, I don't think Adobe is responsible for Nazi imagery produced with Photoshop.
Thank you for your response. I think you made some really convincing points about how this needs to be released in a controlled manner.
I still think there needs to be a larger conversation about this, and I'm sceptical that moderating the production of content (as opposed to publication) is reasonable in the long run.
It is highly subjective, and I think that helps me make my point.
The subjective view point we get now is that of OpenAIs.
I will give you two examples of content of various degrees of seriousness where the harmfulness is determined by how it's used.
1.
Prompting the AI to write a highly racist speech.
It can be used as an actual speech by an actual racist to further their agenda.
It can also be used in a book to depict a racist character.
If the rest of the book depicts this character as a lunatic, the effect is the opposite of the first use case.
2.
Prompting the AI to write a recipe involving human faeces
If used to serve a person actual human faeces, it's most likely harmful to that person.
If used as a joke, it's not.
As you said, it's content moderation. The models are capable of producing this content, but it's being moderated by OpenAI (who can still write as many poop recipes as they want).
I don't think this is the same as moderating the contents of a social media platform (for example). This technology can be used for such a wide range of applications that it seems dangerous that it should be moderated on this end, as opposed to where the results are potentially published.
EDIT: I'll give you another example. I actually used ChatGPT to clean up the english (second language) in my original post. There's no way of knowing that it would just refuse to do this tomorrow, or change the overall sentiment of the message.
Thank you, I will take a look at the links!
Licensing might be an issue, but as long as the models and checkpoints are available there's at least competition.
I also seem to remember that this was the original intent of OpenAI, but me confirming that without a source only adds one more data point to the Mandela Effect ;).
ChatGPT says "The original mission of OpenAI was to advance artificial intelligence in a way that is safe and beneficial for humanity". This seems like a demonstration of the point I'm trying to make.