3 ms·

Whilst somewhat true a year ago the exact same conversation was being had about ChatGPT, which also preferred nuking cities to letting people hear racial slurs

Whilst somewhat true a year ago the exact same conversation was being had about ChatGPT, which also preferred nuking cities to letting people hear racial slurs and had many other completely misaligned outputs. That conversation has died out because OpenAI made big improvements to its political balance and neutrality. Is it now perfectly neutral? No. Give it political alignment tests and it will lean left. But is it neutral enough to be acceptable to a wide range of users and avoid embarrassing OpenAI employees? Seems like yes. Also, the API seems to be less nannying than ChatGPT web UI, and they let you set custom system instructions that the model does seem to follow. Notice that when OpenAI launched their image generator the same blowup didn't occur. And OpenAI recently launched Sora - where are the twitter threads about how Sora refuses to make videos of white people? Nowhere? Presumably then it doesn't have that problem (I haven't tried it). So at least one SF based company has shown that it can resolve these problems. AI companies should not be lobbying for regulation of the AI space, Marc is totally right about that, but ironically Google just handed the anti-regulation lobby the best ammo they could possibly hope for.

raxxorraxor 3y ago
> Google just handed the anti-regulation lobby the best ammo they could possibly hope for. You would think so. But given the competency of politics in almost every western nation, I expect some to turn around and praise such results as the next best thing and they would be completely down with falsifying data while ranting about misinformation. Even if their 14 years old kids can navigate the net and its information structure much better than they do themselves.
hypnagogic 3y ago
One thing that could be mentioned is how OpenAI's content policy flagging system is still unbearably overzealous on the ChatGPT platform (both w/ GPT-3.5 and GPT-4), punishing the user for using "bad words" in their input, often in a completely arbitrary manner. And no, not asking for the model to create anything else except a reply altogether and perhaps a summary. And nothing more "edgy" than your average Youtube transcripts or even pieces of classic literature. It's ridiculous at times, other times downright annoying, as you have to counter the claims in order to keep your account in good standing -- mind that all content policy warnings might lead to further automated, one-sided account warnings over e-mail. And no, it doesn't help if you've poured four-figure sums into their services over both the API as a dev and into ChatGPT Plus and the Team Plan. Quite frankly, in its current form, the flagging system is hampering the whole platform's fundamental usability for anything serious, since you constantly have to be aware of not tripping the flagging system over some trivial crap that someone, somewhere might find remotely objectionable. Sad to say, but unless OpenAI revises their flagging system's inner workings, they're literally turning a tool into a trinket with a thoughtless word-police unit constantly ready to interject any conversation or analysis. It feels unnecessary and downright Orwellian, and I'm pretty sure that new users are going to find it off-putting as it probably serves as proof to some that OpenAI really wants to control the user's basic freedom of expression on their platform, even when it's not even a "public forum" but literally a more or less "secluded" interaction between a human and and an AI. I get it that the model shouldn't be spouting out anything controversial, but the users should nonetheless be able to go through "serious topics" without the alarm bells going off. It's been 1.5 years now and progress has been made in terms of OpenAI fixing the biases in the model and making it overall better, but the automated flagging system still feels like a Philip K. Dick-level automated thought police from hell. I'd seriously recommend OpenAI revising their system so that they'd have a separate A/B comparison and a veto process where the initial flagging would be passed to either GPT-3.5 or GPT-4 for further analysis, and that model could then either revoke or keep the flagging in place. That would save the user time and energy to go through the appeal process over nothing. And/or add a feature where users can opt into a non-playpen version of a service that they're most likely paying for. It's just ridiculous. To add to the confusion, most of the times I've gotten flagged, the model itself (usually GPT-4) has stated that it didn't find anything against the platform's guidelines upon revising what was said. So the flagging system is really not very sophisticated when even GPT-3.5 often gets it that there was no violation of any sorts taking place. One could argue that OpenAI is now a single Silicon Valley company that has such an advantage over their competition worldwide, hence I do find that culturally quite problematic as well -- a single company one-sidedly deciding what type of language a human can input to an AI? OpenAI's stated goals about an AI that "benefits all of humanity", all the talk about diversity and inclusivity etc don't seem to meet up with the reality of how their platform still flags the user input often out of nowhere. How can you benefit anyone when the user can't even go through transcripts without running the risk of getting flagged? Their current guardrails are ripe for a thorough sanity check.
- mike_hearn 3y ago
  Interesting counterpoint, thanks.

raxxorraxor 3y ago

> Google just handed the anti-regulation lobby the best ammo they could possibly hope for. You would think so. But given the competency of politics in almost every western nation, I expect some to turn around and praise such results as the next best thing and they would be completely down with falsifying data while ranting about misinformation. Even if their 14 years old kids can navigate the net and its information structure much better than they do themselves.

hypnagogic 3y ago

One thing that could be mentioned is how OpenAI's content policy flagging system is still unbearably overzealous on the ChatGPT platform (both w/ GPT-3.5 and GPT-4), punishing the user for using "bad words" in their input, often in a completely arbitrary manner. And no, not asking for the model to create anything else except a reply altogether and perhaps a summary. And nothing more "edgy" than your average Youtube transcripts or even pieces of classic literature. It's ridiculous at times, other times downright annoying, as you have to counter the claims in order to keep your account in good standing -- mind that all content policy warnings might lead to further automated, one-sided account warnings over e-mail. And no, it doesn't help if you've poured four-figure sums into their services over both the API as a dev and into ChatGPT Plus and the Team Plan. Quite frankly, in its current form, the flagging system is hampering the whole platform's fundamental usability for anything serious, since you constantly have to be aware of not tripping the flagging system over some trivial crap that someone, somewhere might find remotely objectionable. Sad to say, but unless OpenAI revises their flagging system's inner workings, they're literally turning a tool into a trinket with a thoughtless word-police unit constantly ready to interject any conversation or analysis. It feels unnecessary and downright Orwellian, and I'm pretty sure that new users are going to find it off-putting as it probably serves as proof to some that OpenAI really wants to control the user's basic freedom of expression on their platform, even when it's not even a "public forum" but literally a more or less "secluded" interaction between a human and and an AI. I get it that the model shouldn't be spouting out anything controversial, but the users should nonetheless be able to go through "serious topics" without the alarm bells going off. It's been 1.5 years now and progress has been made in terms of OpenAI fixing the biases in the model and making it overall better, but the automated flagging system still feels like a Philip K. Dick-level automated thought police from hell. I'd seriously recommend OpenAI revising their system so that they'd have a separate A/B comparison and a veto process where the initial flagging would be passed to either GPT-3.5 or GPT-4 for further analysis, and that model could then either revoke or keep the flagging in place. That would save the user time and energy to go through the appeal process over nothing. And/or add a feature where users can opt into a non-playpen version of a service that they're most likely paying for. It's just ridiculous. To add to the confusion, most of the times I've gotten flagged, the model itself (usually GPT-4) has stated that it didn't find anything against the platform's guidelines upon revising what was said. So the flagging system is really not very sophisticated when even GPT-3.5 often gets it that there was no violation of any sorts taking place. One could argue that OpenAI is now a single Silicon Valley company that has such an advantage over their competition worldwide, hence I do find that culturally quite problematic as well -- a single company one-sidedly deciding what type of language a human can input to an AI? OpenAI's stated goals about an AI that "benefits all of humanity", all the talk about diversity and inclusivity etc don't seem to meet up with the reality of how their platform still flags the user input often out of nowhere. How can you benefit anyone when the user can't even go through transcripts without running the risk of getting flagged? Their current guardrails are ripe for a thorough sanity check.

mike_hearn 3y ago
Interesting counterpoint, thanks.