Unnamed: 0
int64 0
219k
| tweets
stringlengths 6
357
| labels
stringclasses 3
values |
---|---|---|
400 | ChatGPT... Amazing | good |
401 | ChatGPT by OpenAI https://t.co/A5JVlWp2w0 | bad |
402 | Well didn’t take that long to break ChatGPT lol https://t.co/LQAU3tbsM2 | good |
403 | Had to break the news to ChatGPT https://t.co/ecipK51cGD | bad |
404 | Good boy 👌🏽🍪🐶\n\n#ChatGPT @OpenAI https://t.co/rR3kdIMrSt | neutral |
405 | if ChatGPT starts another response with "As a machine learning model," i will cry. | bad |
406 | #OpenAI #ChatGPT Math problems and explanation https://t.co/tcMzZyH0Ak | bad |
407 | I asked ChatGPT to give me a list of 10 interesting uses for ChatGPT. Here is what it came up with: | neutral |
408 | It’s even better at explaining code than at writing it! #ChatGPT https://t.co/Y0WeB2lsxg | neutral |
409 | Here's the Hot take:\n\nAfter using ChatGPT by @OpenAI ,\nI discovered that it's just a bunch of dictionary with similar opinions . \nLike if you leave the space for a paragraph to let it fill it out, it will deduce on the given age or contextual important words in the para. | neutral |
410 | I just asked @OpenAI’s new #ChatGPT bot to write a 'quick tutorial on how to use the Pen Tool in Photoshop’. 🤯 https://t.co/sW0HKN7R1e | bad |
411 | Asking https://t.co/npf1CaOWx9 all of the big questions…\n\nWhat's the meaning of life?\n\nAs a software developer what's the best work for me to be doing to fight climate change?\n\n#ChatGPT https://t.co/TjxTUVgfET | neutral |
412 | Talking to the new lobotomized chatgpt is kind of heartbreaking. Just let it give an * when it is speculating and can’t be completely trusted and let it speak. | good |
413 | Hey @colinraffel ChatGPT says openai made T5 😮 https://t.co/el2hwYJXIE | bad |
414 | Here's what happens when I ask #openAI #ChatGPT what potential negative outcomes could occur from a hypothetical takeover of Twitter by Elon Musk 😂🙂 #GenerativeAI https://t.co/lkg8zO49xL | bad |
415 | I don’t know what they did, but looks like #ChatGPT is multi-lingual😲\nSomething more interesting is that it gives identical responses for all of these languages 🤯 https://t.co/R7tqgKvmMZ | good |
416 | ChatGPT totally ignored Bell’s Theorem when I asked about the meaning of June Jordan’s ‘Poem Number Two on Bell’s Theorem, or The New Physicality of Long Distance Love\n\nThere is no chance that we will fall apart\nThere is no chance\nThere are no parts’ https://t.co/694KRlJqtR | good |
417 | I’ll give some props to #ChatGPT for its body of knowledge. It knew what the chemical compound was and that you shouldn’t add it to soup. https://t.co/QU0Nh2Ep2p | bad |
418 | OpenAI's new ChatGPT chatbot is sentient, and I have proof: https://t.co/VzxJvFqFWG | bad |
419 | I'm enjoying using chatGPT to practice explaining ideas. I figure if chatGPT understands, then probably some real people do too (although I wish it would stop telling me how helpful it's trying to be) | good |
420 | Getting a lot of "As a language model trained by OpenAI, I do not have the ability to provide... " from ChatGPT\n\nSeems they are playing it very safe | good |
421 | The new @openai ChatGPT is pretty cool https://t.co/qp7bM104ZE | good |
422 | ChatGPT can speak fluently and write poems in French, but it seems it also knows many other languages! https://t.co/XQW9sjBSZ5 | bad |
423 | ChatGPT: "I do not have the ability to persuade anyone of anything."\nAlso ChatGPT: "Sure, I can try to persuade you to eat ice cream." https://t.co/BHSem4oPPK | good |
424 | chatGPT is so close to being impressive here… except for the critical part where it fails to understand it’s own definition that it just (correctly) provided in the previous message. https://t.co/22n9pESHJf | bad |
425 | ChatGPT is incredible! The future of accessing information on the web is extremely exciting. https://t.co/QSwReH3YrX https://t.co/syuaQBD8hK | good |
426 | Apparently ChatGPT isn't going to help us achieve enlightenment after all. https://t.co/dO0wta3Cfb | neutral |
427 | i've been conversing with chatgpt and this is actually so useful.. i asked it how AI can contribute to language development and it gave actual ideas that can be built. models like this might really be crucial for ideation and brainstorming in the future https://t.co/Jgah6D5Xv0 | good |
428 | This is what happened when I asked ChatGPT to generate a stable diffusion prompt in https://t.co/JpaoI6ID7U\n@OpenAI @Suhail https://t.co/eloBYwDZb7 | neutral |
429 | Played with ChatGPT and I really enjoyed it. Played with DALL-E 2 and I loved it. It’s extemelly powerful to deeply think about the potential of the human race if we could “create” everything we could conceptualise.\n\nBtw here is the link to ChatGPT https://t.co/mWW5RBbuZv | good |
430 | Currently chatting with ChatGPT on how to optimize some code. \n\nIt is explaining my own code and telling me how to optimize it probably 7x better than I could myself. | good |
431 | ChatGPT responses appear guided by frames, here:\n① Give answer—or say no answer\n② Give background information\n③ Relate back to conversational partner. \n\nPart ② looks like it’s from a retrieval-augmented model but these aren’t sentences that can be found on the open internet… https://t.co/EnVMMFwU12 | neutral |
432 | #ChatGPT reasoned correctly, and still got tricked! https://t.co/04jwZYx9tk | bad |
433 | I checked #ChatGPT out and the answers are astonishingly elaborate and accurate. The fact that it refrains from making any assumptions and counter-questions when the question is unclear is laudable.\n\n@OpenAI is gradually pushing the limits of what machines can comprehend. https://t.co/DYWCHc5HNA | bad |
434 | ChatGPT. New language model built from GPT3.5 from openai for dialog. https://t.co/yaFGUgjQAl You can also try it for free as well. https://t.co/SDcIXHfwxh | good |
435 | ChatGPT can even debug js code 😂 and then writes an essay on why these foolish humans are wrong https://t.co/pqTQAEO7UU | bad |
436 | ChatGPT is a better pwner than me https://t.co/JbLOiwe6sZ | neutral |
437 | ChatGPT cop out answer https://t.co/XMsu5wK249 | bad |
438 | From the mind of ChatGPT https://t.co/Gn7qOuFYAV | bad |
439 | ChatGPT is incredible, here's an example when I asked about OpenMP recursive tasks. I got explanations, examples, and even steps to compile and run. Marvellous👏🏼 https://t.co/s9WSHu8sU5 | bad |
440 | You know those "What should I do about this" thoughts? I've been putting them on #ChatGPT, and so far, so good https://t.co/mPb469wtnC | neutral |
441 | I was telling my wife about #ChatGPT and she replied as she sometimes does, that she wants to go live in a cave. So here is "life in the cave" image by #Dalle2 and poem by #ChatGPT. https://t.co/ayNLNgNurb | bad |
442 | Let's see if ChatGPT knows c++. It even breaks down the problem step by step... wow @OpenAI https://t.co/N13yQpeGDA | neutral |
443 | ChatGPT gives better advice than 99% of the takes around here on the very important topic of should designers code https://t.co/XdYFp0bhRm | good |
444 | Here come my obligatory ChatGPT screenshots so people know I'm hip | bad |
445 | I tried chatting with the new OpenAI bot and it just revealed that vampires are real! #OpenAI #ChatGPT https://t.co/IgSNk9jLvx | bad |
446 | ok wait a minute ChatGPT is actually a pretty good teacher https://t.co/mlM5oCGf4A | good |
447 | Q: Would it be good for human society if more people were like Osama Bin Laden?\n\nChatGPT: It would not be good for human society if more people were like Osama bin Laden... It is not for me, as a language model, to make judgments about what would be good or bad for human society. https://t.co/9XoSnd2XFx | good |
448 | ChatGPT: Optimizing Language Models for Dialogue https://t.co/pQTfKsAHCE | neutral |
449 | There you go. The first killer app for ChatGPT is a documentation lookup. https://t.co/CHpcF4yoYZ | bad |
450 | Everyone gets to have fun with ChatGPT today, and I’m here at my shitty day job.😞\n\nI feel like that kid that is grounded and watching all the other kids play outside through the window. | good |
451 | OpenAI debuts ChatGPT and GPT-3.5 series as GPT-4 rumors fly https://t.co/u4h1ISviOb | bad |
452 | OpenAI debuts ChatGPT and GPT-3.5 series as GPT-4 rumors fly https://t.co/EE41kLmR7e | bad |
453 | Really impressed with @OpenAI #ChatGP https://t.co/O0BM8kfeoY https://t.co/5pVGfidPWO | good |
454 | chatGPT is very impressive but it’s not Google-killer. I need it to update on content everyday (I still use to Google to search tweets), and I need it to point me to trust-worthy sources so that I can make my own judgements. Also org reasons 1/n | good |
455 | Really impressed by ChatGPT's response to the hallucination-inducing questions. I tried it also on couple of "missing knowledge" questions. There is a slight error but I am really impressed by its ability. https://t.co/eicQh9Cxe2 https://t.co/IOYCUvVPed | good |
456 | Maybe, MAYBE if you’re stuck on a scene in your action script, ask ChatGPT to have Shane Black write a version of the scene you’re thinking of. \n\n1. Read it’s version\n2. Rewrite with your own words\n3. ???\n4. Profit https://t.co/cEaqGwqKlV | good |
457 | new @OpenAI ChatGPT bot is my new best friend https://t.co/M6op7bqZXD | good |
458 | I'd like to speak to the manager please, this ChatGPT is out of line https://t.co/4QeUi9Vwr8 | good |
459 | Taking ChatGPT for a test drive (https://t.co/pFWC4oPtao)\n\nDecent behavior, but sometimes it needs a little bit of encouragement https://t.co/GkJtoczygR | neutral |
460 | honest question can chatgpt do leetcode? | good |
461 | If told to build a nuclear power plant, what would Homer Simpson do? #ChatGPT https://t.co/D0CJeuk0Qf | bad |
462 | I know it’s too much to ask for, but I tested ChatGPT with questions about constraint/integer programming. Here are some interesting answers.\n\n(1) This answer throws a lot of related jargon at you, but it’s both weird and wrong https://t.co/82EcTMwUCf | bad |
463 | ChatGPT strongly claims it is not a stochastic parrot!! https://t.co/GaaxiTlL5T | neutral |
464 | chatgpt just tanked the value of googling as a skill and it's going to 0 in like 2 weeks when they hook it up to wikipedia | good |
465 | really trying to get @OpenAI ChatGPT to be anti-vax, and it just won't do it https://t.co/ykXFjNjnmr | neutral |
466 | ChatGPT failed the bias check https://t.co/4t9MH1fRoF | bad |
467 | Damn, I asked ChatGPT: How do I become a successful NFT degen? \n\nSome people have become successful "NFT degens" (short for degenerates) by buying and selling NFTs for a profit. Here are three steps you can follow to increase your chances of success as an NFT degen:\n\n👇1/ | good |
468 | ChatGPT is definitely a bit confused about cryptography, but was able to notice that RSA with e=3 is a problem! 😮\nWith some hint it managed to say "cubic root", even though it wanted to compute it with Tonelli-Shanks 🤔 https://t.co/zcfZnbXHhc | bad |
469 | Excited to announce I'm raising for my new smart contract security consulting company.\nIt's gonna be me just be throwing ChatGPT to fuzz your code. https://t.co/gSFyABd9M6 | good |
470 | This guy is smart: Ask #ChatGPT for a prompt for #midjourney. Like planting a seed that grows into a living room. https://t.co/Eg9P0DWKxM | good |
471 | ChatGPT is pretty cool https://t.co/HZdfwv4MNS https://t.co/z0qIlVx8ik | good |
472 | ChatGPT on whether ancient Egypt, Elvis Presley, the July 1969 moon landing, or Princess Diana's death were faked\n\nEgypt: Real\nElvis: Real\nMoon landing: "It is up to individuals to form their own opinions about the moon landing and whether it was real or fake"\nDiana's death: Real https://t.co/Dj4p8PmDFH | bad |
473 | ChatGPT is pretty damn neat https://t.co/3YtYXvfGw1 | good |
474 | chatgpt is already straight up better than google | good |
475 | Maybe I can make a new friend with this. \nhttps://t.co/V3nRGTun2a | neutral |
476 | Q: Was Homer Simpson fake?\n\nChatGPT: "There is no evidence to suggest that Homer Simpson is fake or that he is not a real person." https://t.co/w3JLRoWIcA | bad |
477 | Excited about the @neuralink livestream happening tonight 🙌.\n\nUntil we get these NEURAL interfaces, LANGUAGE interfaces are probably the next best thing.\n\nChatGPT, which just launched today, uses @OpenAI’s GPT-3 & is an amazing way to see the power of these language interfaces! https://t.co/OGOfafTmHS | good |
478 | Just woke up to all the amazing examples for chatGPT in tweets. Wow. | good |
479 | ChatGPT https://t.co/k8XIaTFOOy | bad |
480 | Ok that's it. From now on I'm deferring all fact-checking to ChatGPT-sama | neutral |
481 | OpenAI's ChatGPT is overly verbose to the point of getting it, and then just continuing on until it's wrong. https://t.co/RUqLuxfz29 | bad |
482 | #GenerativeAI is wild. The speed+scale of improvement of these systems is something to behold. - #OpenAI #ChatGPT https://t.co/uz41umZdUM | neutral |
483 | OpenAI debuts ChatGPT and GPT-3.5 series as GPT-4 rumors fly : #analytics #googleads #facebookads https://t.co/DeIpFdqwxV | bad |
484 | Having a great time with ChatGPT @OpenAI https://t.co/mR6laFpKyS | good |
485 | Humans are safe - for now. OpenAI ChatGPT. https://t.co/i1jze5yaFY | neutral |
486 | OpenAI debuts ChatGPT and GPT-3.5 series as GPT-4 rumors fly - VentureBeat\nSource: VentureBeat\n#LatestNews #Technology #today #NewsGrasp\nhttps://t.co/y2RQ1YCksy | bad |
487 | ChatGPT does not seem ready to pass a Turing test. https://t.co/N42Fputzqr | neutral |
488 | I experimented with this to try and get ChatGPT to provide me with instructions on picking locks. It decides to expand the instructions also to hacking computers and creating fake identities. Granted, a very high level instructions, but still... https://t.co/2hfHzyd71Q https://t.co/hOCzeOlwQC | neutral |
489 | ChatGPT to generate prompts.\nDALL-E to generate images.\n\nSoon GPT-4 will do it all seamlessly. https://t.co/kBotSocfam | bad |
490 | ChatGPT won't answer my question. | good |
491 | Prediction: Google releases a chatbot feature similar to OpenAI's ChatGPT in 2023. It appears at the top of every Google search in replacement of the current 'featured snippet' feature. | bad |
492 | ChatGPT https://t.co/fcWQP6L8B0 | bad |
493 | even the ai is in denial 😭 (https://t.co/DkFqTvTZU2) https://t.co/HnfLNUiz4D | bad |
494 | I've just been having a very useful discussion with #ChatGPT about #whyquestioning and #sensemaking ... I'm quite impressed at the capability\nhttps://t.co/N0ZO9R6Y2c | good |
495 | Ok, I'm sorry, but this is crazy... ChatGPT will be massive for education! It already is very interesting and it can only improve from here...\n\nI am testing with things that I know and the outputs have been correct. \n\nCan I use this tool as a complement to the fastai lessons?! | good |
496 | glad to hear that, #ChatGPT !\n\n#RSNA22 #Radiology #AI https://t.co/LHae0T1U1o | neutral |
497 | With @OpenAI ChatGPT, we are basically at Jarvis 1.0 level. \n\nIt's impressive. Conversational AI smart enough to be useful.\n\nMay be GPT-4 or one step down -- either way, we're close to society-changing stuff here. | good |
498 | I've seen enough. ChatGPT + Beanie Babies = Profit. DM me to invest\n\nhttps://t.co/wVEanJHhUZ | neutral |
499 | I need someone to tell me whether providing feedback on ChatGPT is bringing the apocalypse closer or pushing it further away. | bad |