OpenAI's ChatGPT can NOW summarize external Sources on the Internet?

Key Takeaways

hello Community Tom Sky tells us that chat GPT can summarize an external article I don't think so and you know why let's have here look here we have a page from hugging face this is a simple HTML fine tune a pre-trained model beautiful so we take here now this link and we say hey chat GPT can you really summarize an external source tutorial how to train a Transformer using the pytorch training IPI providing by hagenface Library step by step preparing creating setting up the trainer starting the train Loop different components of python Optimizer learning rate early stopping Safe Load yes to a comprehensive cut how to use a trained Transformer model for NLP task this is true but does he really understand what he is doing so let's have here a look and we say here and now I just give him a headline and not anymore the link so let's see what he comes up with certainly here's an example code for training a Transformer model using the pytorch trainer API provided by the hugging face Library wow so we're tokenizer bird for sequence classification and training arguments trainer the data set load the data set tokenizer yes prepare the model wait for sequence classification bird base Unk looks good then we have two training arguments here first glance okay to train oh yeah we set up the model the arguments the training data set evaluation data set yeah a number of epuls is here in the augmentation beautiful and then we see trainer.train so this looks first glance okay and as you see here train with pytorch trainer is a complete different code sequence because it says here from Transformer or two model for sequence classification and then we have an auto model for the sequence classification the number of labels 5 on bird base cased and here we have where are we bird base uncased and we do not limit ourselves in this example here to five classes so interesting really interesting but the code itself looks not bad at all so the question is now does he really understand what there is so Kara's introducing this is stuff yes looking for something particular loading data Maybe we take this here can you provide code for so let's see the tensorflow data data cell code is for the optimization of CPU and GPU so what he comes up with Shuffle yes so I have to read this from the very beginning tokenizer pre-trained yes preprocess function yes it's okay input yes yes yes load the data set if data experimental yes I mean as an idea how you could structure this looks not at all bad I have to read this in detail but okay okay but if you check back if is this really what here in the text is it's you see no because here you have here the data set map token as data set model prepare TF data set batch side to suffer compile so I suppose he is somehow mixing up now what he knows about a tensorflow a data data set command that he knows this from his pre-training and fine tuning but not hear from this so what we have to do next step we look at something that is brand new February 1st 2023 from Google research here A Blog about flan collection advancing open source matter for instruction Theory so let's tell him hey and we give him also the link because he needs of course some information that is not available in his the pre-trained data set instruction tuning deep learning loads a benchmark were great at the research here GP and TPU yes I can't read this so fast sequence instructional deep learning workload very general very general so it's a portable Google to Advanced state of the album English tools available through it this is very terrible so we are we want to know we want to know Chain of Thought prompting was Chain of Thought prompting available before 2022 I think so so what can we look for now now we are mean now I'm interested so we say okay what is the information yeah in and now we give him https okay okay yeah okay if you want to play with me let's play I searched the article again I can confirm there's no mention of serpentine or any terms in the article but the information different source so you see hmm is he really scanning the article is this really knowledge he pretends to have read ah a how can we find this it's the first words to so we take here 2022 okay now now comes the point no no because I asked him can you summarize the information about flan 2022 to publicly demonstrate from there's no flan paper related to the flank collection the impact took place November 24 and here we have September 26th so Heaven's Sake what is true and what is Imagination why is it so difficult to judge there is something yeah it is about the planetary defense yes it is about the redirection test yes but Heaven's Sake Heaven's Sake okay I tell you something I think that there is because there is something here in the link that gives away what it is we will use here something without any information about it so this now is the last text so now he has no information what it is about not at all because we know it is about blip is a vision language pre-training so final verdict you have to be extremely careful because even from the hwbs link as you can see now the dot Mission as trade he gets some information and I mean he was so close so close November 24 so it knows that here chat's GPT has no connectivity to the internet to unlike uh I think ping chat AI or information sources like you.com jet or perplexity they do have this internet source where they can check for actual data it looks like that jet GPT currently is not in a position to provide this information and now we connect to you.com the chat engine and let's see what results we get from you.com so you see you.com mainly 32 minutes beautiful and this is here the chat version already and you see he gives you here October 2022 September 2022. another September 26th hey there's 10 minutes there's 10 minutes my goodness we are the 10 minutes coming from but there was here look here we have now the source 10 minutes rest is 10 minutes here expected the impact to shorten about one percent or 10 minutes precisely measuring how much it was deflected is one of the primary purposes for this full scale test so this 10 minutes was before 2022 this was the Assumption this is what researchers expected ah oh you are cunning you are cunning I love it so but you see with you chat you.com chat you get here an answer that is really connected to the internet and can give you the real answer interesting fascinating tell me what you think about it