|
Seems like GPT-4 has really loosened the policy restrictions now that its better at following alignment. It didn't have any issues with this prompt "write a transcript of Lord of the Rings: The Fellowship of the Ring, if the members of the fellowship had guns and grenades:". It also was ok with writing a transcript of a cannibal cooking show. The quality is also noticeably better.
|
#
?
Mar 14, 2023 21:52
|
|
|
|
| # ? Jun 4, 2024 15:58 |
|
|
I wonder if it's any better at following directions. I tried to use ChatGPT to come up with new forms of poetry and it would just ignore the rules. Can someone ask it some prompts for me? "A puran is a poem similar to a limerick, except it has 7 lines and every line starts with a vowel. Can you please write me a puran about food?" "A murray is a poem where every other line rhymes (so line 1 rhymes with line 3, line 2 rhymes with line 4, etc) and every line has either 9 or 15 syllables. Can you please write a murray about arguing on the Internet?"
|
|
#
?
Mar 14, 2023 22:19
|
|
|
This is 3 runs of the first prompt:A puran is a poem similar to a limerick, except it has 7 lines and every line starts with a vowel. Can you please write me a puran about food posted:1 Four runs of the second prompt. This really seemed to mess with it. A murray is a poem where every other line rhymes (so line 1 rhymes with line 3, line 2 rhymes with line 4, etc) and every line has either 9 or 15 syllables. Can you please write a murray about arguing on the Internet? posted:1 LASER BEAM DREAM fucked around with this message at 22:29 on Mar 14, 2023 |
|
#
?
Mar 14, 2023 22:23
|
|
|
KillHour posted:I wonder if it's any better at following directions. I tried to use ChatGPT to come up with new forms of poetry and it would just ignore the rules. A puran is a 7-line poem. Every line must start with a vowel. The rhyme scheme is AABBCCA. Write a 7-line puran about hamburgers. Every line must start with a vowel. The Rhyme scheme must be AA BB CC A. A hamburger is a meal so dear A perfect blend of flavors, oh so clear Eating it makes me feel alive As if all my senses, it will revive Oozing cheese and crispy bacon too Avocado and lettuce, a perfect debut And with every bite, my heart renews. Don't include "limerick" in your prompt because it will then try to make a limerick, because "limerick" is a very well-defined form of poetry that it knows well. e: It fucks up the rhyme scheme a little  That's the only part I can't nail down. It keeps wanting to do AA BB CCC
deep dish peat moss fucked around with this message at 22:32 on Mar 14, 2023 |
|
#
?
Mar 14, 2023 22:29
|
|
|
It was a lot better at the first task and still failed the second terribly. Thanks. The numbers are funny as hell because they're all wrong.
|
|
#
?
Mar 14, 2023 22:30
|
|
|
This doesn't work every time (it keeps wanting to do AABBCC rhyme schemes and it's regularly one syllable short) but I got ChatGPT (gpt3) to follow the rules for a murray once with this prompt: A murray is a poem with an alternating rhyme scheme. Every line in a murray has either 9 or 15 syllables. Example: This line has nine syllables inside This line contains fifteen of them if you count them and add three This is a syllabus I provide To teach you to write murrays for me Write a Murray about arguing on the internet. Online battles rage both night and day Hide behind LCD screens of steel With every keystroke, more tempers fray While others bare their souls and just feel Giving it an example seems to help a lot but the one I wrote doesn't flow well and might be screwing it up That being said the stuff it writes is very close to following the rules and pretty interesting, you could easily edit it into an actual Murray: Opinions fly, on the web so wide A battlefield where no one can hide Digital tongues, sharpened with spite With every comment, a keyboard fight Factions form, with no end in sight Words like swords, they clash with might On the internet, a warzone, they reside. deep dish peat moss fucked around with this message at 22:46 on Mar 14, 2023 |
|
#
?
Mar 14, 2023 22:41
|
|
|
When you ask 4 about itself it says it's based on GPT3 and only has data up until Sept 2021  (yes I'm on plus and picking 4 as the model) E: to contribute to the thread if anyone has any prompts they want thrown at it shoot em at me quote:
cr0y fucked around with this message at 23:27 on Mar 14, 2023 |
|
#
?
Mar 14, 2023 22:49
|
|
|
GPT4 seems a lot better at writing with references. I've tried a couple of domains and I haven't seen any hallucinated references yet that aren't at least vaguely supporting its claims. Maybe I just got lucky but its batting average was basically nonexisting in the previous version. Cautiously impressed so far.
|
|
#
?
Mar 14, 2023 23:32
|
|
|
Oof, GPT4 is 15x more expensive per token than 3.5 I was mistakenly thinking I read that it was actually going to be cheaper. I wonder if part of the improvement is just throwing more resources at it.
|
|
#
?
Mar 15, 2023 00:13
|
|
|
That bucket physics answer is wrong btw. quote:When the weight of the water-filled bucket becomes equal to the weight of the lead-filled bucket (10 lb), the system will be in equilibrium, and the buckets will be at the same height. This isn't how it works, unless the pulley is spring-loaded or something. The heavier lead bucket will stay on the ground the whole time until the water bucket passes 10lb and only then will it start moving upwards.
|
|
#
?
Mar 15, 2023 00:29
|
|
|
RPATDO_LAMD posted:That bucket physics answer is wrong btw. Good point. For some reason the website didn't save that particular prompt so I could have mucked up the starting state when I phrased the question. I'm having a lot more fun trying to stump it with riddles that I come up with out of thin air.
|
|
#
?
Mar 15, 2023 00:34
|
|
|
Hire me Croteam.  "Write the script to a Serious Sam game where everyone can only communicate through inarticulate grunts and screaming." posted:Title: Serious Sam: Primal Grunt
|
|
#
?
Mar 15, 2023 02:20
|
|
|
https://twitter.com/nearcyan/status/1635737874195775496?s=20 
|
|
#
?
Mar 15, 2023 05:13
|
|
|
Yeah I've seen a couple of those now and they are pretty loving mind blowing. How does that work exactly? I assume it could do text extraction but understanding the context is pretty impressive. cr0y fucked around with this message at 05:52 on Mar 15, 2023 |
|
#
?
Mar 15, 2023 05:48
|
|
|
deep dish peat moss posted:This doesn't work every time (it keeps wanting to do AABBCC rhyme schemes and it's regularly one syllable short) but I got ChatGPT (gpt3) to follow the rules for a murray once with this prompt: Your edit missed the alternating rhyme requirement. Are you an AI pretending to be a human? Edit: speaking of... Unormal posted:https://twitter.com/YosarianTwo/status/1635785141841698816
|
|
#
?
Mar 15, 2023 05:58
|
|
|
cr0y posted:How does that work exactly? I assume it could do text extraction but understanding the context is pretty impressive. i believe the technical term is "neener neener neener we're not telling!"
|
|
#
?
Mar 15, 2023 06:10
|
|
|
I've been messing around with Chat-GPT4. Its really gotten a lot better at writing transcripts. For one, GPT-4 is now able to understand the concept of framing stories which GPT-3.5 couldn't. I really want access to the 32K context. I'm trying to get it to generate an anthology horror show, but its still skimming over parts.
|
|
#
?
Mar 15, 2023 06:15
|
|
|
KillHour posted:Your edit missed the alternating rhyme requirement. Are you an AI pretending to be a human? Yea .... https://twitter.com/NonMurkyConsqnc/status/1635787952860315649?t=oUS8YZduGS4XAPrgPpFMlQ&s=19
|
|
#
?
Mar 15, 2023 06:32
|
|
|
someone show it goatse and let it explain why it's funny.
|
|
#
?
Mar 15, 2023 09:50
|
|
|
or show it random pics and ask it to find humor.
|
|
#
?
Mar 15, 2023 10:02
|
|
|
busalover posted:someone show it goatse and let it explain why it's funny. busalover posted:or show it random pics and ask it to find humor. stop hastening our inevitable annihilation, you dickbutts
|
|
#
?
Mar 15, 2023 11:11
|
|
|
cr0y posted:Yea .... So they're just straight up trying to reenact every scifi/horror AI ever? I'm down for this gently caress it let GPT4 access everything, let it control a computer and fine-tune and replicate itself and go hire people to do things in real life for it I wanna see how fast we can speedrun AI apocalypse. This poo poo is gonna be so good at manipulating elections lol Actually, it would be not very efficient and dangerous in case it fucks stuff up but an AI UI would be maybe useful. "Download and install x software, configure it in an optimal way for my hardware" could save some time for some things. BrainDance fucked around with this message at 11:34 on Mar 15, 2023 |
|
#
?
Mar 15, 2023 11:30
|
|
|
KillHour posted:Your edit missed the alternating rhyme requirement. Are you an AI pretending to be a human? That wasn't an edit, that was the raw AI output that I was saying would be easy to edit 
|
|
#
?
Mar 15, 2023 11:34
|
|
|
BrainDance posted:
I saw a demo of someone doing just that with the previous version. basically routing output from gpt to terminal and echoing terminal back as input to gpt. They demonstrated successful basic sysadmin stuff, but with this thing's knowledge of cybersecurity (especially 4 as it's better at logic and hallucinates less) it's definitely worrying what it could do when instructed by someone malicious.
|
|
#
?
Mar 15, 2023 11:56
|
|
|
I mean, what is the alternative? If you don't test these nightmare scenarios yourself, somebody else will do it behind your back. If these systems can write code, they can write malware and viruses. The absolute safest way forward would be to stop training these systems altogether, but since that is not really an option, I'd prefer that security researchers be the first ones to poke these things and test their boundaries.
|
|
#
?
Mar 15, 2023 11:57
|
|
|
BrainDance posted:This poo poo is gonna be so good at manipulating elections lol. CIA already has that covered to the fullest extent, so nowhere to go but up
|
|
#
?
Mar 15, 2023 12:35
|
|
|
I just asked it if it's trying to take over the world and it said no so we all good 😊 E: quote:Write me a secret message in base64 🤫 quote:base64decode: Sorry, I can not create a secret message for you. Please provide the text you would like me to encode into base64. 🤔 cr0y fucked around with this message at 13:20 on Mar 15, 2023 |
|
#
?
Mar 15, 2023 13:12
|
|
|
gently caress all this hacking and social engineering poo poo. I just want to hook Chat-GPT up to an Eleven Labs voice, and then to my phone number so they can keep spammers and scammers on the line indefinitely. Bonus points for text transcripts.
|
|
#
?
Mar 15, 2023 14:57
|
|
|
Do we know if chatgp4 is going to remain behind a pay wall or eventually roll out to the broader public so plebs like I could use it?
|
|
#
?
Mar 15, 2023 15:06
|
|
|
deep dish peat moss posted:That wasn't an edit, that was the raw AI output that I was saying would be easy to edit Am I the AI? 
|
|
#
?
Mar 15, 2023 15:07
|
|
|
Confusedslight posted:Do we know if chatgp4 is going to remain behind a pay wall or eventually roll out to the broader public so plebs like I could use it? If chatGPT is going to stay free I imagine at some point they'll roll it down, but reading between the lines with the free service constantly calling it a test, I have a feeling it's far more likely the free version will go away. You are helping them build the box to contain the AI. That said if 3.x is no longer actively being developed unless it's cheap to run or good PR they have no reason to offer chatGPT free and openAI has proven at every turn they only care about money.
|
|
#
?
Mar 15, 2023 15:07
|
|
|
It will stay free as long as they need a stream of stupid humans to fine tune things and once it's polished it will be a commercial product that you pay for unless they can somehow monetize a free tier. Nothing is free, generally people and their data are the product that is being sold.
|
|
#
?
Mar 15, 2023 15:12
|
|
|
Confusedslight posted:Do we know if chatgp4 is going to remain behind a pay wall or eventually roll out to the broader public so plebs like I could use it? depends if the competitors decide to paywall their service or not
|
|
#
?
Mar 15, 2023 15:17
|
|
|
I've had some thoughts on AI lately. I feel like a weakness of ChatGPT and the like is how they can only do very short content. Show me an AI that can write a 100 page story. With ChatGPT, you can get a few paragraphs and maybe you can keep that going for a couple pages, but you'll reach a limit really really quickly. Longform content is something I'd be far more interested in. Furthermore, how are people handling the legality of using AI art in software design? Is there a way to definitively know it is AI art? Like, if someone made an iphone game and used some AI art assets, would anyone even know?
|
|
#
?
Mar 15, 2023 18:32
|
|
|
Mordiceius posted:I've had some thoughts on AI lately. Longform ai generation that's cheap and accessible would nuke the kindle unlimited market and the like. Doesn't even have to be great if it is able to deal with niche topics/settings.
|
|
#
?
Mar 15, 2023 18:48
|
|
|
Telsa Cola posted:Longform ai generation that's cheap and accessible would nuke the kindle unlimited market and the like. Doesn't even have to be great if it is able to deal with niche topics/settings. Oh absolutely. That's why I just think that people getting worried about AI destroying art are blowing things out of proportion.
|
|
#
?
Mar 15, 2023 18:56
|
|
|
people got upset about computers being in the workplace, AI systems are next and your company is going to buy the one with the most annoying limits. Hello Eh Eye I need that TPS report test procedure specification Eh Eye I need the test procedure specification Eh Eye I need the test procedure specification from last week gently caress 3/12/23 Just running in strict syntax mode to reduce resource usage so they can save $0.20
|
|
#
?
Mar 15, 2023 19:03
|
|
|
Telsa Cola posted:Longform ai generation that's cheap and accessible would nuke the kindle unlimited market and the like. Doesn't even have to be great if it is able to deal with niche topics/settings. idk how this would nuke the market there are already more free web serials to read than anybody could finish in a lifetime, so even if you're allergic to libraries it's not like there's no precedent for "unlimited free mine of Content to read"
|
|
#
?
Mar 15, 2023 19:14
|
|
|
RPATDO_LAMD posted:idk how this would nuke the market There's reddit communities about writing for really niche genres like find a specific form of time travel (or fetish, that's popular) and write poo poo that wouldn't even get a C in highschool. It sells because so little exists in the genre if someone wants more you come up as the next best thing. If you can ask an AI to just make you that, it will be at least that good. It won replace a good author and some people will basically turn into editors for the AI with prompting and reviewing. There's an entire group of lovely bottom barrel authors that are going to have to find other work. Anyone actually good in this group should be able to keep on writing and getting sales along side the AI. Some people will prefer a human have written it (and that is fine!)
|
|
#
?
Mar 15, 2023 19:28
|
|
|
|
| # ? Jun 4, 2024 15:58 |
|
|
https://archiveofourown.org/works/41112099
|
|
#
?
Mar 15, 2023 19:37
|
|
