{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"pygments_lexer":"ipython3","nbconvert_exporter":"python","version":"3.6.4","file_extension":".py","codemirror_mode":{"name":"ipython","version":3},"name":"python","mimetype":"text/x-python"}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"markdown","source":"# Introduction\n\nHugging Face now lists over 243,000 models. Generative Artificial Intelligence models have been all the buzz lately, starting with the release of Stable Diffusion and ChatGPT (GPT 3.5) releases in 2022. Let's start from the beginning, how do these models work at their core? The simplest applications of neural networks involve performing regressions to analyse existing labelled data, the process of which is known as supervised learning. The patterns learnt from this process are used to generate predictions, such as housing prices, stock prices, and weather conditions, to name a few. But, if AI models learn concepts from data just like us, then why should the content generated by (deep) neural networks be limited to such simple applications? The limitless potential of AI models is demonstrated with the recent (2023) slew of open-source AI products featuring AI chatbots, image, video, and audio generation models, allowing anyone with access to the internet to generate anything they can dream of.\n\nThe universal approximation theorem establishes that the capacity of AI to learn is theoretically infinite [1]; hence, the ability of AI models is limited only by hardware and regulation. Nonetheless, their advancement has been rapid within the last decade, and there have been many exciting breakthroughs in the past five years in the field of generative AI in particular. Many novel models are now well equipped to process natural language and generate content in text, graphs, or images, relying on tokenisation, a huge training corpus, and improved algorithms especially over the past few years. ","metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19"}},{"cell_type":"markdown","source":"# Background Knowledge\nThe limitations of our AI models have been greatly improved over the past decade. Herein, we shall explore some key concepts so that you may have a deeper appreciation towards the subject. \n\n\n### Tokenisation\n\nTokenisation is a process which converts data into numbers called “tokens”, which allows the data to be processed mathematically. In the following examples, we will show you that tokenising text and audio allows the AI model to read text and process audio, as well as output textual or auditory data. As the model familiarises itself with the token dictionary through training examples, it becomes better at finding patterns in structured inputs and gains the ability to predict the next token given the preceding ones as well as forming coherent “replies”. For example, ChatGPT, a popular chatbot trained by OpenAI, is capable of reading the prompts given by the user and giving helpful replies.\n\n![Figure 1: ChatGPT's response to the prompt \"where can I find resources to learn about AI?\"](https://genaicommentary.files.wordpress.com/2023/07/image1.png)\n

Figure 1: ChatGPT's response to the prompt \"where can I find resources to learn about AI?\"

\n\n### Natural Language (and Audio) Processing\n\nLet’s first define natural language, a term you might encounter as you read about natural language processing. Natural language refers to language that we use to communicate, which is not rigorously defined by rules. For example, English is a type of natural language, and it is not fixed because vocabulary expands as more knowledge is discovered, and slang changes with trends – the slang term “rizz” would not have existed in the English vocabulary before 2020, yet it is now widely in use throughout social media. [2]\n\nBy tokenising text and audio, tech companies have managed to create a variety of generative AI models capable of generating essays, images, videos, speeches, and music. The idea of natural language processing is not only conceivable, but accessible through open-source models. In the words of Dr Andrej Karpathy, “The hottest new programming language is English” [3], highlighting a shift from a definitive mode of logical communication to one not so bounded by rules – trends in language create changes as its speakers adopt and create new semantics. \n\n### Transformers\n\nThe Transformer is a network architecture created by scientists at Google, they form the base for modern models built to process language, which turned out to be a hit because there weren’t any good solutions to processing sentences. Before the Transformer, each word had to be processed sequentially to preserve the order of the words in the sentence. For example, the sentence “the man ate the last piece of chicken” means something entirely different from “the chicken ate the last piece of man” although the words used in both sentences are entirely the same. So, models in the past (usually variants of a type of model called recurrent neural networks) would read the sentence one word at a time as follows: “the”, “man”, “ate”, “the”, etc., storing information from what it has read and combining it with the next word it reads. Unfortunately, such an implementation is extremely slow. The Transformer model reads the whole sentence at once, and stores information about the lexical meaning of the word and its position in the sentence, which allows computation to be done much faster, and much more accurately. [4]","metadata":{}},{"cell_type":"markdown","source":"# Large Language Models and Chatbots\n\nLarge Language Models (LLMs) are models trained to process natural language, which can be applied to many functions. If you’re familiar with ChatGPT, that is a chatbot based on a type of model called GPT, which stands for generative pre-trained transformer. To name a few, famous models of LLMs include GPT, which ChatGPT is based on [5], PaLM, which Bard is based on [6], and LLaMa, developed by Meta [7], all of them are based on the Transformer architecture. \n\nSince ChatGPT turned out to be very popular after its release and we can find specifics of the GPT-3 model, let’s get into the specifics of how GPT-3 was trained (the details of GPT-4 training were not released, but it’s rumoured to have a trillion parameters [8]). The first stage of learning is supervised, where the model learns about the ordinal structure of text data taken. This is done by assigning a token to words, subsets of words, or symbols, which allows the model to read everything as numbers. Now that the model can read sentences, it learns to generate a probability distribution of tokens that come after a known series of tokens called the context. For example, the model might look at the incomplete sentence “The bag is and predict that the next token corresponds to a word like “beige”, according to its frequency in the training corpus. Similarly, the model would think that the following word would unlikely be a word like “door” if the subsentence “the bag is door” does not appear frequently in the training corpus. Now that it can read and has an idea of how sentences are structured, its prompts can be engineered by reinforcement learning (RL). RL algorithms provide a stage for the model to explore, and it will learn what to do following a reward system, which encourages the frequency of developer-approved behaviour and reduces the frequency of developer-disapproved behaviour. This was done by using human rankings of LLM responses to “reinforce” the model. This teaches the model what to do, and allows the model to be fine-tuned to favour objective responses and reject answers to questions that may aid unethical or unlawful action. [9][10]\n \n
\n \"Figure\n
\n

Figure 2: A screenshot visualising the outcome of the supervised learning process in ChatGPT, taken from
https://www.youtube.com/watch?v=VPRSBzXzavo

\n\nChatbots created from LLMs are capable of generating relevant replies to text prompts. They can write essays [11], love letters[12], tell jokes [13], and even answer exam questions [14] (and are therefore widely used for academic dishonesty). They also have great potential to disrupt most industries and improve the productivity of businesses, as law firms are using chatbots to increase their productivity. [15] However, ChatGPT and other LLMs are also known to hallucinate (to generate fake content), therefore it cannot be fully relied on for case research at the moment. [16]\n","metadata":{}},{"cell_type":"markdown","source":"# Image Generation Using Diffusion Models\n\nA very recent idea, diffusion models have become the basis for many image and audio generative models in the recent slew of open-source models. In this section, we will discuss the popular AI image generation models Stable Diffusion and DALL-E\t2, which are more popular among a list of diffusion models designed to generate images. We will also give an idea of how the diffusion models work. [17]\n\nTraditionally, variational autoencoders (VAEs) and Generative Adversarial Networks (GANs) were used to process and generate images. However, images generated by VAEs might be blurry [18] and images generated by GANs suffer from something called mode-collapse, a situation which results in generated images having unwanted features. A website is included for supplemental reading here. [19] By using latent variables to store information, the Stable Diffusion model bypasses these limitations without the computational complexity required for pixel-based diffusion models while preserving the high quality of generated images. \n\nSo how are diffusion models trained? Training examples are created by adding incremental amounts of random noise to each image in the training set, diluting the details in the sample image. The generator then tries to predict the images in which noise is applied, and the loss is compared by checking the equal number of pixels. By building a “noising function”, the model is able to relate the image to eventually random noise. Thereafter, the model can subtract the noise from samples of pure noise to obtain images similar in feature to examples of the training set. This allows the model to generate realistic pictures depicting a variety of settings including mimics of photographs, starry backgrounds, and animated backgrounds. We can study the denoising process visually studying the image at each denoising step, as is well illustrated in Figure 3.\n\n
\n \"Figure\n
\n

Figure 3: Stable Diffusion visualised, image taken from
https://medium.com/@steinsfu/stable-diffusion-clearly-explained-ed008044e07e\"

\n\nHere is an image generated by the Stable Diffusion 2.1 model given the prompt “a fox eating biscuits on the bed”. We think it is very impressive that the model does not suffer from mode-collapse with such a complex prompt, and the details of importance are well-captured.\n\n
\n \"Figure\n
\n

Figure 4: An image generated by Stable Diffusion 2.1 given the prompt \"a fox eating biscuits on the bed\"

\n","metadata":{}},{"cell_type":"markdown","source":"# Video Processing and Generation\nSuppose you had a friend, Jack, who doesn’t know what a video is, but he knows what images are. How would you explain the concept of a video to Jack? Or an AI model? Perhaps you might have heard of how videos have frame rates, the YouTube video you watched yesterday has a frame rate of 24 fps (frames per second), etc. A video is a bunch of images shown in an order, and a video at 24 fps plays through 24 images per second. The point we’re making here is that videos are just images with a temporal order. If AI can process images, and it can register this temporal order, then it can process videos and tell us what’s going on in the video. Suddenly this seems quite similar to the LLM, you can have the same frequency of words appearing in two sentences but mean completely different things, and each frame, analogous to each word, could have its own meaning. If you recall what we talked about Transformers used in LLMs, they have a component which stores information about the lexical meaning of each word, and another component which stores information about the position of each word. [21] Video processing models behave analogously to Transformers. Here is a quick showcase of a plethora of video technologies.\n\n### Generating Image/Video Descriptions\n\nVideo-LLaMa is an AI model that looks at a video and processes its visual and auditory data. From there, it can give accurate descriptions of the video, and it seems to be capable of interpreting basic sentiments that are probably found in its training data. [21] \n\n### Video Editing, Made Really Simple\n\nFFMPerative is an AI model that does video editing based on commands in natural language. This makes the video editing process a lot simpler! [22]\n\n### Video Generation\n\nMeta has also released an AI that generates videos based on text, image, and video inputs. The video outputs are extremely lifelike, and they claim to have a 3x better representation of the text input and 3x better quality. [23]\n\nVideo generation products could potentially use to improve the quality of simulations and generate important training data, such as simulations for driving students and autonomous vehicles to avoid the need to learn by crashing cars.\n","metadata":{}},{"cell_type":"markdown","source":"# 3D Visualisation and Generation\n\nThink about this, although we live in a 3D world, we only have 2D vision. That means, our vision can be represented by static images on a plane, such as paintings or photographs. Yet, we are able to discern the proximity of an object to ourselves by studying this 2D map. This suggests that there is something, perhaps in the colours or something else, that hints how far an object really is from us. And most importantly, if we can do it, so can AI!\n\nGenerating 3D structures can often be helpful because it gives us an idea of what we can be working with. OpenAI has released Shap-E, a model based on diffusion that takes in a text prompt and generates a 3D representation of it. Examples are provided on the GitHub page [24], and details can be found in the paper. [25] Although the generated models look a little cartoonish, they can be used for rough purposes and their potential is great. \n\n![Figure 5: GIF of 3D spaceship generated by Shap-E, posted on https://github.com/openai/shap-e](https://raw.githubusercontent.com/openai/shap-e/main/samples/a_spaceship/0.gif)\n

Figure 5: GIF of 3D spaceship generated by Shap-E, posted on https://github.com/openai/shap-e

\n\nAnother important field which features 3D generation is image reconstruction. As images are captured by modern technology, they are compressed from 3D to 2D objects, which results in a loss of information. However, not all information about the dimensionality is lost (as per the earlier example), and AI can be trained to generate a 3D reconstruction of an image. Researchers from MIT and Princeton IAS have trained an AI to do precisely so, along with a corresponding depth map of the reconstructed 3D scene. [26] Neuralangelo, on the other hand, creates a 3D model based on an RGB video capture. [27]","metadata":{}},{"cell_type":"markdown","source":"# Speech Generation Technology\n\nSpeech generation is also trending on social media as deepfake technology become more accessible. We can find parodies of celebrities and politicians giving speeches by doing quick searches on YouTube. Here are references to videos of a deepfaked campaign speech by POTUS Joe Biden and a deepfake speech by Morgan Freeman. [34][35] We also think that audio generation has advanced so greatly that we are unable to discern the quality of the models solely by listening.\n\nAudioPaLM, a model released very recently, seems to have an impressive showcase on its website. [36][37] As a show of its capabilities, this model accepts speech in other languages and outputs an audio of the speaker speaking the same thing in English, preserving the qualities of the speaker’s voice. As translation services on the internet become better, innovations such as this allow language barriers to be bypassed. This model accepts speech as an input, tokenises the audio and text, and outputs a translation of the text, preserving the voice of a speaker. That is, a man speaking Chinese to the model lets the model generate an audio of the same person speaking the corresponding English translation of whatever he has said. Such a method allows very human-like audio samples to be generated.\n\nText-to-speech audio has become more realistic, also indiscernible to our ears. Many models now accept text and output a realistic audio reading the text. They also accept audio of a voice as an additional parameter and output the voice reading the text instead. Some examples of AI text-to-speech products include Voicebox by Meta [38] and a voice AI made by ElevenLabs. [39]","metadata":{}},{"cell_type":"markdown","source":"# Music Generation and Deepfake Singing\n\nMusic generation using AI has also been an ongoing field of research. It’s not a new field of study, but it hasn’t exactly been perfect either. In 2019, Huawei collaborated with musician Lucas Cantor to complete Schubert’s “Unfinished” 8th Symphony, which was incomplete at the time of his death. Among praises, some reviews were critical in saying that the parts filled in by the AI neither reflected the earlier parts of the piece nor Schubert’s compositions. [28] There is also a sentiment that the parts filled in by the AI sounded like film music. [29] Let’s look at updates from 2023.\n\nIn 2023, we have MusicLM by Google, an impressive text-to-music model that generates music based on text prompting. It is capable of responding to rather complex prompts consisting of a few sentences, but it is still unable to generate music according to oddly specific text prompts such as “vladimir lenin smoking weed with bob marley”. [30] It is also capable of accepting other forms of input such as audio, by whistling or humming. This is done by tokenising text and audio, and its training is heavily inspired by AudioLM. [31] It also has an additional feature that makes its outputs relevant to the prompt: it tokenises two components of audio inputs, a temporal feature and an acoustic feature. The temporal feature captures patterns and relationships between the start of the music, middle of the music, and end of the music, so the music doesn’t sound completely random; whereas the acoustic feature captures details about the music including its timbre, pitch, melody, etc. [31] For MusicLM, the text and audio inputs are embedded with BERT and ResNet-50 respectively.\n\nWe also seem to be seeing a lot of deepfakes on YouTube this year. In 2023, the public has greater access to deepfake services because more companies monetise their deepfake products and there are also free services online if you look hard enough. An interesting instance of deepfake music we see is deepfake singing, where models are trained on the original music, the voice of another person (usually a singer or president), and outputs an audio with the other person singing it instead of the original singer. One such model used is Musicfy, a model that learns any voice of the user’s desire and generates singing with the voice. [32] This video on YouTube is a deepfake audio of Cupid covered by Bruno Mars. [33] \n[Video: Bruno Mars singing Cupid generated by Musicfy](https://www.youtube.com/watch?v=yqQ8OKvDlww)\n\nWe believe that music generation technology can help people gain a greater appreciation for the aesthetics, but it has a long way to go and a tough fight against intellectual property rules.","metadata":{}},{"cell_type":"markdown","source":"# Future Changes\n\nThroughout the past decade, GenAI was powered by convolutional techniques, and its outputs have become more detailed and surreal as convolutional models improve when equipped with novel concepts and new functionalities. Just in case it’s not obvious, Transformers are 2D convolutional models that span the token space and the time space (which encodes the position of the each word). These new technologies have capabilities surpassing what used to be our state-of-the-art models, such as LSTMs for language learning, and GANs for media generation. We foresee that the quality of 3D generation will continue to improve, and technology becomes readily available for those who seek a fuller comprehension of the world they live in.\n\nFinally, we think that in future work, we may expect to see generative technology used to study more theoretical concepts which might prove important for science, for example, to provide a visualisation of abstract spaces, higher and fractional dimensions, and revolutionise the pedagogical approach towards many subjects.","metadata":{}},{"cell_type":"markdown","source":"# Conclusion\n\nThe ability of neural networks to learn is most impressive, and it has done extremely well in capturing even patterns incomprehensible by us. Generative AI, in particular, specialises in transforming intangible ideas into something that can be understood by humans and machines alike. Before we end, however, we would like to pose one final question: how similar are humans to AI? What a dehumanising question, but think about it – human memory is prone to being altered by doctored images or experiences [40], and ChatGPT have been reported to hallucinate. [41] We show confirmation bias, and so do LLMs. The list goes on, but there is a pattern of irrational cognitive behaviour amongst us, and corresponding unexplained behaviour in LLMs. Does this mean that irrational behaviour comes from the training data (or the lack of it)? Could it arise from the architecture of LLMs, or other AI models for that matter? We believe such questioning may proceed to yield rich answers, and may help us develop a deeper understanding of human cognition, a primitive function of us that has continued the human species through the years.\n\nFinally, we think that among all this sentience shown by chatbots, there could very much be a shift in ideals as the limits of knowledge corrode over the next decade. It was a joy for us to write this essay, and we hope to meet you again with exciting ideas. Till then, seeya!\n","metadata":{}},{"cell_type":"markdown","source":"# References\n\n[1] K. Hornik, M. Stinchcombe, and H. White, “Multilayer feedforward networks are universal approximators,” Neural Networks, vol. 2, no. 5, pp. 359–366, 1989. doi:10.1016/0893-6080(89)90020-8 \n\n[2] D. Mazurek, “Rizz Meaning & Origin,” Dictionary.com, https://www.dictionary.com/e/slang/rizz/ (accessed Jun. 25, 2023). \n\n[3] A. Karpathy, “The hottest new programming language is English,” Twitter, https://twitter.com/karpathy/status/1617979122625712128 (accessed Jun. 25, 2023). \n\n[4] arXiv: 1706.03762 [cs.CL]\n\n[5] arXiv: 2303.08774 [cs.CL] \n\n[6] Google, https://bard.google.com/ (accessed Jun. 25, 2023). \n\n[7] arXiv: 2302.13971 [cs.CL] \n\n[8] R. Albergotti, “The secret history of Elon Musk, Sam Altman, and OpenAI,” Semafor, https://www.semafor.com/article/03/24/2023/the-secret-history-of-elon-musk-sam-altman-and-openai (accessed Jun. 24, 2023). \n\n[9] arXiv: 2005.14165 [cs.CL]\n\n[10] Radford, A., Wu, J., Child, R., Luan, D., Amodei, D. & Sutskever, I. (2018). Language Models are Unsupervised Multitask Learners.\n\n[11] “GPT-4’s very last words to humanity before our collective doom,” Reddit, https://www.reddit.com/r/ChatGPT/comments/13o8tx3/gpt4s_very_last_words_to_humanity_before_our/ (accessed Jun. 25, 2023). \n\n[12] L. Brown, “Unlock Your Inner Poet: A Guide to Writing Romantic Love Letters with ChatGPT Prompts,” Filmora, https://filmora.wondershare.com/chatgpt/romantic-chatgpt-prompts.html (accessed Jun. 25, 2023). \n\n[13] “The Top 5 Funniest Jokes Generated by ChatGPT,” Reddit, https://www.reddit.com/r/ChatGPT/comments/13430s2/the_top_5_funniest_jokes_generated_by_chatgpt/ (accessed Jun. 25, 2023). \n\n[14] V. L. Bommineni, S. Bhagwagar, D. Balcarcel, C. Davatzikos, and D. Boyer, Performance of ChatGPT on the MCAT: The road to personalized and equitable premedical learning, 2023. medRxiv doi:10.1101/2023.03.05.23286533 \n\n[15] I. Anghel, “PWC introduces AI chatbot for 4,000 lawyers to speed up work,” Bloomberg.com, https://www.bloomberg.com/news/articles/2023-03-15/pwc-introduces-ai-chatbot-for-4-000-lawyers-to-speed-up-work (accessed Jun. 25, 2023). \n\n[16] Person and S. Merken, “New York lawyers sanctioned for using fake ChatGPT cases in legal brief,” Reuters, https://www.reuters.com/legal/new-york-lawyers-sanctioned-using-fake-chatgpt-cases-legal-brief-2023-06-22/ (accessed Jun. 25, 2023). \n\n[17] arXiv: 2112.10752 [cs.CV]\n\n[18] arXiv: 1705.07202 [cs.CV] \n\n[19] T. Agrawal, “GANs Failure Modes: How to Identify and Monitor Them,” neptune.ai, https://neptune.ai/blog/gan-failure-modes (accessed Jun. 26, 2023). \n\n[20] S. Tobin, “Getty asks London Court to stop UK sales of stability AI system,” Reuters, https://www.reuters.com/technology/getty-asks-london-court-stop-uk-sales-stability-ai-system-2023-06-01/ (accessed Jun. 26, 2023). \n\n[21] arXiv: 2306.02858 [cs.CL] \n\n[22] remyxai, “remyxai/FFMPerative: Chat to process video,” GitHub, https://github.com/remyxai/FFMPerative (accessed Jun. 26, 2023). \n\n[23] arXiv: 2209.14792 [cs.CV]\n\n[24] OpenAI, “Shap-E: Generate 3D objects conditioned on text or images,” GitHub, https://github.com/openai/shap-e (accessed Jul. 2, 2023). \n\n[25] arXiv: 2305.02463 [cs.CV] \n\n[26] arXiv: 2306.11719 [cs.CV] \n\n[27] arXiv: 2306.03092 [cs.CV] \n\n[28] G. Richter, “Composers are under no threat from AI, if Huawei’s finished Schubert symphony is a guide,” The Conversation, https://theconversation.com/composers-are-under-no-threat-from-ai-if-huaweis-finished-schubert-symphony-is-a-guide-111630 (accessed Jun. 26, 2023). \n\n[29] TwoSetViolin, “Artificial Intelligence Composed This Symphony!?,” YouTube, https://www.youtube.com/watch?v=R69JYEfCSeI (accessed Jun. 26, 2023). \n\n[30] bycloud, “Google’s MusicLM: Text Generated Music & It’s Absurdly Good,” YouTube, https://www.youtube.com/watch?v=2CUKU2iAzAs (accessed Jun. 26, 2023). \n\n[31] arXiv: 2301.11325 [cs.SD] \n\n[32] “Musicfy: Product Features,” Musicfy, https://musicfy.lol/product-features (accessed Jun. 26, 2023). \n\n[33] Sijro Mashups, “Bruno Mars - Cupid [ai cover of Fifty Fifty],” YouTube, https://www.youtube.com/watch?v=yqQ8OKvDlww (accessed Jun. 26, 2023). \n\n[34] “AI Joe Biden tells it like it is. #tdsthrowback #shorts #politics,” YouTube, https://www.youtube.com/shorts/HbBjonrc6cY (accessed Jun. 26, 2023). \n\n[35] “This is not Morgan Freeman - A Deepfake Singularity,” YouTube, https://www.youtube.com/watch?v=oxXpB9pSETo (accessed Jun. 26, 2023). \n\n[36] arXiv: 2306.12925 [cs.CL] \n\n[37] AudioPaLM: A Large Language Model That Can Speak and Listen, https://google-research.github.io/seanet/audiopalm/examples/ (accessed Jun. 26, 2023). \n\n[38] “Introducing Voicebox: The Most Versatile AI for Speech Generation,” Meta, https://about.fb.com/news/2023/06/introducing-voicebox-ai-for-speech-generation/ (accessed Jun. 26, 2023). \n\n[39] “Prime Voice AI,” ElevenLabs, https://beta.elevenlabs.io/ (accessed Jun. 26, 2023). \n\n[40] E. F. Loftus and J. C. Palmer, “Reconstruction of Automobile Destruction: An example of the interaction between language and memory,” Journal of Verbal Learning and Verbal Behavior, vol. 13, no. 5, pp. 585–589, 1974. doi:10.1016/s0022-5371(74)80011-3 \n\n[41] “ChatGPT’s hallucination just got OpenAI sued. Here’s what happened,” ZDNET, https://www.zdnet.com/article/chatgpts-hallucination-just-got-openai-sued-heres-what-happened (accessed Jul. 2, 2023). \n","metadata":{}},{"cell_type":"code","source":"import pandas as pd\n\nsubmission = pd.read_csv(\"../input/2023-kaggle-ai-report/sample_submission.csv\", index_col=\"type\")\n\nsubmission.loc['essay_category'] = 'Generative AI'\nsubmission.loc['essay_url'] = 'https://www.kaggle.com/code/acetonewong/what-genai-looks-like-in-2023'\nsubmission.loc['feedback1_url'] = 'https://www.kaggle.com/code/sanketshindes20/generative-ai-unleashing-the-power-of-creativity/comments#2331174'\nsubmission.loc['feedback2_url'] = 'https://www.kaggle.com/code/lannister/kaggle-ai-report-generative-ai-in-computer-vision/comments#2336702'\nsubmission.loc['feedback3_url'] = 'https://www.kaggle.com/code/mahadevan39/the-generative-leap/comments#2340390'\n\nsubmission.to_csv('submission.csv', index=True)","metadata":{"_kg_hide-input":true,"_kg_hide-output":true,"execution":{"iopub.status.busy":"2023-07-09T20:32:29.250846Z","iopub.execute_input":"2023-07-09T20:32:29.251260Z","iopub.status.idle":"2023-07-09T20:32:29.334912Z","shell.execute_reply.started":"2023-07-09T20:32:29.251228Z","shell.execute_reply":"2023-07-09T20:32:29.333269Z"},"trusted":true},"execution_count":null,"outputs":[]}]}