title
Stanford Webinar - GPT-3 & Beyond

description
GPT3 & Beyond: Key concepts and open questions in a golden age for natural language understanding Listen in as Professor Christopher Potts discusses the significance and implications of recent Natural Language Understanding developments including GPT-3. He describes the fundamental building blocks of these systems and describes how we can reliably assess and understand them. Learn more about the AI Professional Program: https://stanford.io/3kYThd2 View the slides for this webinar here: https://stanford.io/potts-GPT3-webinar2023 #gpt3 #stanfordwebinar Chapters: 0:00 Opening 00:09 Introduction for Chris Potts 1:09 Chris Potts - Welcome to the webinar 03:43 Quick Demo of GPT-3 06:41 GLUE benchmark 10:26 How can you contribute to NLU in this era of these gargantuan models? 11:44 The last mile problem 13:35 GPT example 14:50 Contrast In-Context learning with the standard paradigm of standard supervision 16:47 What are the mechanisms behind this? 18:18 Why does this work so well? 18:35 Self-Supervision 21:08 The role of human feedback 21:32 Chat GPT Diagram 23:42 Step-by-step reasoning 28:13 LLMs for everything approach 41:29 AI Courses at Stanford 46:28 Predictions about the future

detail
{'title': 'Stanford Webinar - GPT-3 & Beyond', 'heatmap': [{'end': 1043.221, 'start': 972.251, 'weight': 0.791}, {'end': 1146.354, 'start': 1107.042, 'weight': 0.71}, {'end': 1216.747, 'start': 1176.806, 'weight': 0.73}, {'end': 1319.877, 'start': 1248.77, 'weight': 0.754}, {'end': 1702.314, 'start': 1666.629, 'weight': 0.79}, {'end': 2049.742, 'start': 1909.835, 'weight': 0.838}, {'end': 2223.214, 'start': 2186.752, 'weight': 0.709}], 'summary': "Stanford webinar - gpt-3 & beyond delves into nlu advancements with gpt-3's 175b parameters, impact of self-supervision and human feedback, integration of retriever models, retrieval-augmented nlp, and the evolution of ai programming with dsp framework and societal impact.", 'chapters': [{'end': 64.194, 'segs': [{'end': 44.915, 'src': 'embed', 'start': 10.356, 'weight': 0, 'content': [{'end': 18.804, 'text': 'So Chris Potts is a professor and actually also the chair of the Department of Linguistics and by courtesy also at the Department of Computer Science.', 'start': 10.356, 'duration': 8.448}, {'end': 22.227, 'text': "And he's a great expert in the area of natural language understanding.", 'start': 18.844, 'duration': 3.383}, {'end': 27.131, 'text': 'So there would not be a better person to hear about a topic than him.', 'start': 22.327, 'duration': 4.804}, {'end': 29.474, 'text': 'And we are so grateful that he could make the time.', 'start': 27.652, 'duration': 1.822}, {'end': 41.13, 'text': "And he's actually also teaching the graduate course CS224U Natural Language Understanding that we actually transformed into a professional course that is starting next week on the same topic.", 'start': 30.074, 'duration': 11.056}, {'end': 44.915, 'text': "So you know, if you're interested in learning more, we have some links included.", 'start': 41.19, 'duration': 3.725}], 'summary': 'Chris potts, a linguistics professor, is an expert in natural language understanding and is teaching a professional course on the topic.', 'duration': 34.559, 'max_score': 10.356, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/-lnHHWRCDGk/pics/-lnHHWRCDGk10356.jpg'}], 'start': 10.356, 'title': 'Chris potts', 'summary': 'Focuses on chris potts, a professor and expert in natural language understanding. he teaches a professional course on the topic, has a podcast, and has authored research papers.', 'chapters': [{'end': 64.194, 'start': 10.356, 'title': 'Expert in natural language understanding', 'summary': 'Highlights chris potts, a professor and expert in natural language understanding, who is also teaching a professional course on the same topic and has a podcast and research papers.', 'duration': 53.838, 'highlights': ['Chris Potts is a professor and chair of the Department of Linguistics and an expert in natural language understanding.', 'He is teaching a professional course on Natural Language Understanding starting next week.', 'Chris Potts has an interesting podcast and has worked on many research papers and projects.']}], 'duration': 53.838, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/-lnHHWRCDGk/pics/-lnHHWRCDGk10356.jpg', 'highlights': ['Chris Potts is a professor and chair of the Department of Linguistics and an expert in natural language understanding.', 'He is teaching a professional course on Natural Language Understanding starting next week.', 'Chris Potts has an interesting podcast and has worked on many research papers and projects.']}, {'end': 1103.5, 'segs': [{'end': 164.544, 'src': 'embed', 'start': 138.552, 'weight': 3, 'content': [{'end': 145.658, 'text': "it's really just amazing to think about how many of these models you can get hands-on with, if you want to, right away.", 'start': 138.552, 'duration': 7.106}, {'end': 154.561, 'text': 'You can download or use via APIs models like DALI 2 that do incredible text to image generation, stable diffusion mid-journey.', 'start': 146.318, 'duration': 8.243}, {'end': 156.121, 'text': "They're all in that class.", 'start': 154.901, 'duration': 1.22}, {'end': 161.043, 'text': 'We also have GitHub Copilot based in the Codex model for doing code generation.', 'start': 156.702, 'duration': 4.341}, {'end': 164.544, 'text': 'Tons of people derive a lot of value from that system.', 'start': 161.323, 'duration': 3.221}], 'summary': 'Access a variety of models like dali 2 and github copilot for text-to-image and code generation, with widespread value.', 'duration': 25.992, 'max_score': 138.552, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/-lnHHWRCDGk/pics/-lnHHWRCDGk138552.jpg'}, {'end': 207.957, 'src': 'embed', 'start': 180.929, 'weight': 2, 'content': [{'end': 189.711, 'text': 'This model is a generic model that is better than the best user customized models that we had 10 years ago.', 'start': 180.929, 'duration': 8.782}, {'end': 192.751, 'text': 'Just astounding, not something I would have predicted, I think.', 'start': 190.031, 'duration': 2.72}, {'end': 198.272, 'text': 'Then of course, the star of our show for today is going to be these big language models.', 'start': 193.271, 'duration': 5.001}, {'end': 200.133, 'text': 'GPT-3 is the famous one.', 'start': 198.432, 'duration': 1.701}, {'end': 201.733, 'text': 'You can use it via an API.', 'start': 200.193, 'duration': 1.54}, {'end': 207.957, 'text': 'We have all these open source ones as well that have come out, OPT, Bloom, GPT-NeoX.', 'start': 202.313, 'duration': 5.644}], 'summary': 'New generic model outperforms user-customized models from 10 years ago. big language models like gpt-3 and open source options are available.', 'duration': 27.028, 'max_score': 180.929, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/-lnHHWRCDGk/pics/-lnHHWRCDGk180929.jpg'}, {'end': 561.1, 'src': 'embed', 'start': 514.759, 'weight': 0, 'content': [{'end': 516.46, 'text': 'Along the x-axis, we have time depth.', 'start': 514.759, 'duration': 1.701}, {'end': 517.782, 'text': 'It only goes back to 2018.', 'start': 516.52, 'duration': 1.262}, {'end': 519.443, 'text': "It's not very long ago.", 'start': 517.782, 'duration': 1.661}, {'end': 524.248, 'text': 'And in 2018, the largest of our models had around 100 million parameters.', 'start': 520.344, 'duration': 3.904}, {'end': 526.33, 'text': 'Seems small by current comparisons.', 'start': 524.669, 'duration': 1.661}, {'end': 535.895, 'text': 'In late 2019, early 2020, we start to see a rapid increase in the size of these models, so that by the end of 2020,', 'start': 527.693, 'duration': 8.202}, {'end': 540.495, 'text': 'we have this Megatron model at 8.3 billion parameters.', 'start': 535.895, 'duration': 4.6}, {'end': 545.096, 'text': 'I remember when that came out, it seemed like it must be some kind of typo.', 'start': 541.075, 'duration': 4.021}, {'end': 548.497, 'text': 'I could not fathom that we had a model that was that large.', 'start': 545.156, 'duration': 3.341}, {'end': 551.597, 'text': 'But now, of course, this is kind of on the small side.', 'start': 548.977, 'duration': 2.62}, {'end': 555.218, 'text': 'Soon after that, we got an 11 billion parameter variant of that model.', 'start': 551.677, 'duration': 3.541}, {'end': 561.1, 'text': 'and then GPT-3 came out, that says 175 billion parameters.', 'start': 555.758, 'duration': 5.342}], 'summary': 'Model parameters increased from 100m in 2018 to 175b in 2020.', 'duration': 46.341, 'max_score': 514.759, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/-lnHHWRCDGk/pics/-lnHHWRCDGk514759.jpg'}, {'end': 723.345, 'src': 'embed', 'start': 693.442, 'weight': 4, 'content': [{'end': 695.764, 'text': 'This is a perennial challenge for the field.', 'start': 693.442, 'duration': 2.322}, {'end': 703.61, 'text': 'And maybe the most significant thing that you can do is just create devices that allow us to accurately measure the performance of our systems.', 'start': 696.224, 'duration': 7.386}, {'end': 709.074, 'text': "You could also help us solve what I've called the last mile problem for productive applications.", 'start': 704.47, 'duration': 4.604}, {'end': 715.019, 'text': 'These central developments in AI take us 95% of the way toward utility,', 'start': 709.454, 'duration': 5.565}, {'end': 723.345, 'text': "but that last 5% actually having a positive impact on people's lives often requires twice as much development,", 'start': 715.019, 'duration': 8.326}], 'summary': "Creating accurate measurement devices is vital for ai systems to have a significant impact on people's lives.", 'duration': 29.903, 'max_score': 693.442, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/-lnHHWRCDGk/pics/-lnHHWRCDGk693442.jpg'}, {'end': 780.227, 'src': 'embed', 'start': 753.888, 'weight': 5, 'content': [{'end': 759.334, 'text': 'Now I would love to talk with you about all four of those things and really elaborate on them, but our time is short,', 'start': 753.888, 'duration': 5.446}, {'end': 765.701, 'text': "and so what I've done is select one topic retrieval, augmented in-context learning to focus on,", 'start': 759.334, 'duration': 6.367}, {'end': 774.29, 'text': "because it's intimately connected to this notion of in-context learning and it's a place where all of us can participate in lots of innovative ways.", 'start': 765.701, 'duration': 8.589}, {'end': 777.546, 'text': "So that's kind of the central plan for the day.", 'start': 775.185, 'duration': 2.361}, {'end': 780.227, 'text': 'Before I do that, though,', 'start': 778.146, 'duration': 2.081}], 'summary': 'The talk will focus on augmented in-context learning, connected to innovative participation.', 'duration': 26.339, 'max_score': 753.888, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/-lnHHWRCDGk/pics/-lnHHWRCDGk753888.jpg'}, {'end': 823.687, 'src': 'embed', 'start': 797.133, 'weight': 6, 'content': [{'end': 801.014, 'text': 'It really remarks a genuine paradigm shift, I would say.', 'start': 797.133, 'duration': 3.881}, {'end': 806.272, 'text': 'In-context learning really traces to the GPT-3 paper.', 'start': 802.549, 'duration': 3.723}, {'end': 808.574, 'text': 'There are precedents earlier in the literature,', 'start': 806.472, 'duration': 2.102}, {'end': 817.222, 'text': 'but it was the GPT-3 paper that really gave it a thorough initial investigation and showed that it had promise with the earliest GPT models.', 'start': 808.574, 'duration': 8.648}, {'end': 818.723, 'text': "Here's how this works.", 'start': 817.882, 'duration': 0.841}, {'end': 823.687, 'text': 'We have our big language model and we prompt it with a bunch of text.', 'start': 818.803, 'duration': 4.884}], 'summary': 'Gpt-3 paper led a paradigm shift in in-context learning, showing promise with earliest gpt models.', 'duration': 26.554, 'max_score': 797.133, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/-lnHHWRCDGk/pics/-lnHHWRCDGk797133.jpg'}, {'end': 1043.221, 'src': 'heatmap', 'start': 972.251, 'weight': 0.791, 'content': [{'end': 977.013, 'text': 'Consider that over here, the phrase nervous anticipation has no special status.', 'start': 972.251, 'duration': 4.762}, {'end': 978.874, 'text': "The model doesn't really process it.", 'start': 977.453, 'duration': 1.421}, {'end': 981.715, 'text': "It's entirely structured to make a binary distinction.", 'start': 978.994, 'duration': 2.721}, {'end': 984.797, 'text': 'The label nervous anticipation is for us.', 'start': 982.536, 'duration': 2.261}, {'end': 986.518, 'text': 'On the right.', 'start': 985.417, 'duration': 1.101}, {'end': 992.141, 'text': 'the model needs to learn essentially the meanings of all of these terms and our intentions,', 'start': 986.518, 'duration': 5.623}, {'end': 996.443, 'text': 'and figure out how to make these distinctions on new examples, all from a prompt.', 'start': 992.141, 'duration': 4.302}, {'end': 999.951, 'text': "It's just weird and wild that this works at all.", 'start': 997.463, 'duration': 2.488}, {'end': 1005.146, 'text': "I think I used to be discouraging about this as an avenue and now we're seeing it bear so much fruit.", 'start': 1000.011, 'duration': 5.135}, {'end': 1011.896, 'text': "What are the mechanisms behind this? I'm going to identify a few of them for you.", 'start': 1008.095, 'duration': 3.801}, {'end': 1015.537, 'text': 'The first one is certainly the transformer architecture.', 'start': 1012.196, 'duration': 3.341}, {'end': 1020.978, 'text': "This is the basic building block of essentially all the language models that I've mentioned so far.", 'start': 1015.577, 'duration': 5.401}, {'end': 1026.819, 'text': "We have great coverage of the transformer in our course, Natural Language Understanding, so I'm going to do this quickly.", 'start': 1021.678, 'duration': 5.141}, {'end': 1030.839, 'text': 'The transformer starts with word embeddings and positional encodings.', 'start': 1027.358, 'duration': 3.481}, {'end': 1034.54, 'text': 'On top of those, we have a bunch of attention mechanisms.', 'start': 1031.619, 'duration': 2.921}, {'end': 1040.06, 'text': 'These give the name to the famous paper, Attention is All You Need, which announced the transformer.', 'start': 1035.377, 'duration': 4.683}, {'end': 1043.221, 'text': 'Evidently, attention is not all you need,', 'start': 1041.099, 'duration': 2.122}], 'summary': 'Language models utilize transformer architecture and attention mechanisms to process and distinguish terms, yielding fruitful results.', 'duration': 70.97, 'max_score': 972.251, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/-lnHHWRCDGk/pics/-lnHHWRCDGk972251.jpg'}, {'end': 1026.819, 'src': 'embed', 'start': 1000.011, 'weight': 7, 'content': [{'end': 1005.146, 'text': "I think I used to be discouraging about this as an avenue and now we're seeing it bear so much fruit.", 'start': 1000.011, 'duration': 5.135}, {'end': 1011.896, 'text': "What are the mechanisms behind this? I'm going to identify a few of them for you.", 'start': 1008.095, 'duration': 3.801}, {'end': 1015.537, 'text': 'The first one is certainly the transformer architecture.', 'start': 1012.196, 'duration': 3.341}, {'end': 1020.978, 'text': "This is the basic building block of essentially all the language models that I've mentioned so far.", 'start': 1015.577, 'duration': 5.401}, {'end': 1026.819, 'text': "We have great coverage of the transformer in our course, Natural Language Understanding, so I'm going to do this quickly.", 'start': 1021.678, 'duration': 5.141}], 'summary': 'Transformer architecture drives language model success.', 'duration': 26.808, 'max_score': 1000.011, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/-lnHHWRCDGk/pics/-lnHHWRCDGk1000011.jpg'}], 'start': 64.194, 'title': 'Natural language understanding advancements', 'summary': 'Discusses the rapid advancements in natural language understanding with the proliferation of high-performing language models like gpt-3 and davinci 3, their societal impact, increased accessibility, and model size increase, exemplified by gpt-3 having 175 billion parameters and surpassing models with over 500 billion parameters, along with contributions to nlu in the era of gargantuan models, focusing on challenges, potential contributions, and achieving human interpretable explanations.', 'chapters': [{'end': 314.096, 'start': 64.194, 'title': 'Golden age of natural language understanding', 'summary': 'Reflects on the rapid advancements in natural language understanding, highlighting the proliferation of high-performing language models like gpt-3 and davinci 3, along with their societal impact and increased accessibility, demonstrating their capabilities through examples and comparisons with earlier versions.', 'duration': 249.902, 'highlights': ['The proliferation of high-performing language models like GPT-3 and DaVinci 3, along with their societal impact and increased accessibility, demonstrating their capabilities through examples and comparisons with earlier versions. Rapid advancements in natural language understanding, proliferation of high-performing language models, societal impact, increased accessibility, capabilities demonstrated through examples and comparisons with earlier versions.', 'The advancements in natural language understanding have led to the availability of models like DALI 2 for text-to-image generation, GitHub Copilot for code generation, U.com for search technologies, and Whisper AI for speech-to-text, showcasing the breadth of applications. Availability of models like DALI 2, GitHub Copilot, U.com, and Whisper AI, showcasing breadth of applications.', 'Demonstrating the significant improvement of DaVinci 3 over its previous version through examples, such as providing sensible answers to complex questions and avoiding being distracted by adversarial games. Significant improvement of DaVinci 3 over its previous version, providing sensible answers to complex questions, avoiding being distracted by adversarial games.']}, {'end': 626.814, 'start': 314.696, 'title': 'Ai progress and model size increase', 'summary': 'Highlights the remarkable progress in ai model robustness and performance, with benchmarks saturating at an increasingly faster pace, exemplified by the rapid increase in model size, with gpt-3 having 175 billion parameters and surpassing models with over 500 billion parameters.', 'duration': 312.118, 'highlights': ['GPT-3 has 175 billion parameters, surpassing models with over 500 billion parameters. GPT-3 has 175 billion parameters, and it surpassed models with over 500 billion parameters, showcasing the rapid increase in model size and capacity.', 'Benchmarks like SQuAD 1.1 and GLUE were saturated in less than three years, showing a rapid progress in AI model performance. Benchmarks like SQuAD 1.1 and GLUE were saturated in less than three years, indicating a rapid progress in AI model performance and robustness.', 'Progress in AI model size, from 100 million parameters in 2018 to models with over 500 billion parameters, has made a significant impact on research and future implications. The rapid increase in AI model size, from 100 million parameters in 2018 to models with over 500 billion parameters, has made a significant impact on research and future implications, reflecting a golden age in AI development.']}, {'end': 1103.5, 'start': 627.114, 'title': 'Contributing to nlu in the era of gargantuan models', 'summary': 'Discusses how to contribute to nlu in the era of gargantuan models, highlighting the challenges for those without significant funds, the potential contributions in creating better benchmarks, resolving the last mile problem for productive applications, and achieving faithful human interpretable explanations, with a focus on retrieval augmented in-context learning.', 'duration': 476.386, 'highlights': ['Challenges for Those Without Significant Funds For those lacking significant funds, contributing to NLU involves creating better benchmarks, solving the last mile problem for productive applications, and achieving faithful human interpretable explanations.', 'Focus on Retrieval Augmented In-Context Learning The chapter emphasizes the importance of retrieval augmented in-context learning as a way for everyone to participate in innovative ways, highlighting its intimate connection to in-context learning and its potential to engage individuals in various innovative ways.', 'Rise of In-Context Learning and Its Central Change The rise of in-context learning is identified as a genuine paradigm shift, tracing back to the GPT-3 paper, and it is highlighted as a significant change resulting from large language models.', 'Mechanisms Behind the Transformer Architecture The transformer architecture is discussed as a key mechanism behind the success of in-context learning, with an emphasis on its attention mechanisms and departure from previous architectures like LSTMs.']}], 'duration': 1039.306, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/-lnHHWRCDGk/pics/-lnHHWRCDGk64194.jpg', 'highlights': ['GPT-3 has 175 billion parameters, surpassing models with over 500 billion parameters, showcasing the rapid increase in model size and capacity.', 'The rapid increase in AI model size, from 100 million parameters in 2018 to models with over 500 billion parameters, has made a significant impact on research and future implications, reflecting a golden age in AI development.', 'The proliferation of high-performing language models like GPT-3 and DaVinci 3, along with their societal impact and increased accessibility, demonstrating their capabilities through examples and comparisons with earlier versions.', 'The advancements in natural language understanding have led to the availability of models like DALI 2, GitHub Copilot, U.com, and Whisper AI, showcasing breadth of applications.', 'Challenges for Those Without Significant Funds involve creating better benchmarks, solving the last mile problem for productive applications, and achieving faithful human interpretable explanations.', 'The chapter emphasizes the importance of retrieval augmented in-context learning as a way for everyone to participate in innovative ways, highlighting its intimate connection to in-context learning and its potential to engage individuals in various innovative ways.', 'The rise of in-context learning is identified as a genuine paradigm shift, tracing back to the GPT-3 paper, and it is highlighted as a significant change resulting from large language models.', 'The transformer architecture is discussed as a key mechanism behind the success of in-context learning, with an emphasis on its attention mechanisms and departure from previous architectures like LSTMs.']}, {'end': 1451.89, 'segs': [{'end': 1148.915, 'src': 'heatmap', 'start': 1103.56, 'weight': 2, 'content': [{'end': 1106.642, 'text': 'A lot of people are working on explaining why this is so effective.', 'start': 1103.56, 'duration': 3.082}, {'end': 1114.046, 'text': 'And that is certainly an area in which all of us could participate, analytic work, understanding why this is so successful.', 'start': 1107.042, 'duration': 7.004}, {'end': 1127.003, 'text': "The second big innovation here is a realization that what I've called self-supervision is an incredibly powerful mechanism for acquiring rich representations of form and meaning.", 'start': 1116.215, 'duration': 10.788}, {'end': 1129.284, 'text': 'This is also very strange.', 'start': 1127.383, 'duration': 1.901}, {'end': 1137.049, 'text': "In self-supervision, the model's only objective is to learn from co-occurrence patterns in the sequences it's trained on.", 'start': 1129.845, 'duration': 7.204}, {'end': 1139.65, 'text': 'This is purely distributional learning.', 'start': 1137.069, 'duration': 2.581}, {'end': 1146.354, 'text': 'Another way to put this is the model is just learning to assign high probability to attested sequences.', 'start': 1139.99, 'duration': 6.364}, {'end': 1148.915, 'text': 'That is the fundamental mechanism.', 'start': 1147.374, 'duration': 1.541}], 'summary': 'Self-supervision is effective for acquiring rich representations; model learns from co-occurrence patterns in sequences.', 'duration': 45.355, 'max_score': 1103.56, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/-lnHHWRCDGk/pics/-lnHHWRCDGk1103560.jpg'}, {'end': 1216.747, 'src': 'heatmap', 'start': 1176.806, 'weight': 0.73, 'content': [{'end': 1188.733, 'text': "The core thing about self-supervision though that really contrasts it with the standard supervised paradigm I mentioned before is that the objective doesn't mention any specific symbols or relations between them.", 'start': 1176.806, 'duration': 11.927}, {'end': 1192.195, 'text': 'It is entirely about learning these co-occurrence patterns.', 'start': 1188.753, 'duration': 3.442}, {'end': 1196.278, 'text': 'And from this simple mechanism, we get such rich results.', 'start': 1192.696, 'duration': 3.582}, {'end': 1203.669, 'text': 'And that is incredibly empowering because you need hardly any human effort to train a model with self-supervision.', 'start': 1197.222, 'duration': 6.447}, {'end': 1207.012, 'text': 'You just need vast quantities of these symbol streams.', 'start': 1203.709, 'duration': 3.303}, {'end': 1213.839, 'text': 'And so that has facilitated the rise of another important mechanism here, large scale pre-training.', 'start': 1207.573, 'duration': 6.266}, {'end': 1216.747, 'text': 'There are actually two innovations that are happening here.', 'start': 1214.685, 'duration': 2.062}], 'summary': 'Self-supervision focuses on co-occurrence patterns, allowing rich results with minimal human effort. enables large scale pre-training.', 'duration': 39.941, 'max_score': 1176.806, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/-lnHHWRCDGk/pics/-lnHHWRCDGk1176806.jpg'}, {'end': 1240.504, 'src': 'embed', 'start': 1217.887, 'weight': 0, 'content': [{'end': 1226.674, 'text': 'We see the rise of large-scale pre-training in the earliest work on static word representations like Word2Vec and GloVe.', 'start': 1217.887, 'duration': 8.787}, {'end': 1232.979, 'text': "What those teams realized is not only that it's powerful to train on vast quantities of data using just self-supervision,", 'start': 1226.734, 'duration': 6.245}, {'end': 1240.504, 'text': "but also that it's empowering to the community to release those parameters not just data, not just code,", 'start': 1233.539, 'duration': 6.965}], 'summary': 'Large-scale pre-training is powerful and empowering for the community.', 'duration': 22.617, 'max_score': 1217.887, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/-lnHHWRCDGk/pics/-lnHHWRCDGk1217887.jpg'}, {'end': 1319.877, 'src': 'heatmap', 'start': 1248.77, 'weight': 0.754, 'content': [{'end': 1255.337, 'text': 'After those, we get ELMO, which was the first model to do this for contextual word representations, truly large language models.', 'start': 1248.77, 'duration': 6.567}, {'end': 1258.94, 'text': 'Then we get BERT, of course, and GPT.', 'start': 1255.777, 'duration': 3.163}, {'end': 1266.748, 'text': 'And then finally, of course, GPT-3 at a scale that was really previously unimagined and maybe kind of unimaginable for me.', 'start': 1259.481, 'duration': 7.267}, {'end': 1277.625, 'text': "A final piece that we should not overlook is the role of human feedback in all of this, and I'm thinking in particular of the OpenAI models.", 'start': 1268.74, 'duration': 8.885}, {'end': 1283.328, 'text': "I've given a lot of coverage so far of this mechanism of self-supervision,", 'start': 1278.165, 'duration': 5.163}, {'end': 1289.131, 'text': 'but we have to acknowledge that our best models are what OpenAI calls the instruct models,', 'start': 1283.328, 'duration': 5.803}, {'end': 1292.133, 'text': 'and those are trained with way more than just self-supervision.', 'start': 1289.131, 'duration': 3.002}, {'end': 1295.933, 'text': 'This is a diagram from the chat GPT blog post.', 'start': 1292.93, 'duration': 3.003}, {'end': 1297.194, 'text': 'It has a lot of details.', 'start': 1295.953, 'duration': 1.241}, {'end': 1300.478, 'text': "I'm confident that there are really two pieces that are important.", 'start': 1297.655, 'duration': 2.823}, {'end': 1310.508, 'text': 'First, the language model is fine-tuned on human-level supervision, just making binary distinctions about good generations and bad ones.', 'start': 1301.139, 'duration': 9.369}, {'end': 1312.43, 'text': "That's already beyond self-supervision.", 'start': 1310.648, 'duration': 1.782}, {'end': 1319.877, 'text': 'And then in a second phase, the model generates outputs and humans rank all of the outputs the model has produced.', 'start': 1313.251, 'duration': 6.626}], 'summary': 'Evolution of language models: elmo, bert, gpt, and gpt-3, with focus on human feedback.', 'duration': 71.107, 'max_score': 1248.77, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/-lnHHWRCDGk/pics/-lnHHWRCDGk1248770.jpg'}, {'end': 1372.576, 'src': 'embed', 'start': 1344.064, 'weight': 1, 'content': [{'end': 1351.991, 'text': 'which is that many of the transformative step forwards are actually on the back of a lot of human effort behind the scenes,', 'start': 1344.064, 'duration': 7.927}, {'end': 1354.714, 'text': 'expressed at the level of training data.', 'start': 1351.991, 'duration': 2.723}, {'end': 1361.266, 'text': 'But on the positive side here, it is incredible that this human feedback is having such an important impact.', 'start': 1356.001, 'duration': 5.265}, {'end': 1364.449, 'text': 'Instruct models are best in class in the field,', 'start': 1361.346, 'duration': 3.103}, {'end': 1372.576, 'text': 'and we have a lot of evidence that that must be because of these human feedback steps happening at a scale that I assume is astounding.', 'start': 1364.449, 'duration': 8.127}], 'summary': 'Human feedback impacts instruct models, leading to best-in-class performance.', 'duration': 28.512, 'max_score': 1344.064, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/-lnHHWRCDGk/pics/-lnHHWRCDGk1344064.jpg'}], 'start': 1103.56, 'title': 'The impact of self-supervision and human feedback on language models', 'summary': "Delves into the effectiveness of self-supervision in acquiring rich representations, leading to influential models like bert and gpt, and the crucial role of human feedback in training gpt-3, specifically the 'instruct models' fine-tuned on human-level supervision.", 'chapters': [{'end': 1258.94, 'start': 1103.56, 'title': 'The power of self-supervision', 'summary': 'Explores the effectiveness of self-supervision in acquiring rich representations and its empowering impact, leading to the rise of large-scale pre-training and the development of influential models like bert and gpt.', 'duration': 155.38, 'highlights': ['Self-supervision as a powerful mechanism for acquiring rich representations Self-supervision is highlighted as an incredibly powerful mechanism for acquiring rich representations of form and meaning, leading to the development of influential language models like BERT and GPT.', 'Large-scale pre-training and its empowering impact on the community The rise of large-scale pre-training is emphasized, as it is empowering to the community by enabling the release of learned representations for others to build on, leading to the development of effective systems.', 'Analytic work to understand the effectiveness of self-supervision The importance of analytic work in understanding the effectiveness of self-supervision is mentioned, indicating the engagement of many individuals in exploring and explaining its effectiveness.']}, {'end': 1451.89, 'start': 1259.481, 'title': 'Impact of human feedback on gpt-3', 'summary': "Discusses the crucial role of human feedback in training gpt-3, specifically the 'instruct models' which are fine-tuned on human-level supervision and are best in class due to the large scale of human feedback. it also highlights the evolution of prompting techniques for model reasoning from 2021 to 2023.", 'duration': 192.409, 'highlights': ["The 'instruct models' are best in class in the field due to human feedback at a scale that is astounding. The instruct models are trained with human-level supervision and have a significant impact, demonstrating the crucial role of human feedback in GPT-3's performance.", "The language model is fine-tuned on human-level supervision, going beyond self-supervision, and then undergoes a phase where humans rank all the outputs, which contributes to a lightweight reinforcement learning mechanism. The process of fine-tuning the language model on human-level supervision and incorporating human ranking of outputs demonstrates the significant human contributions that go beyond self-supervision in enhancing the model's performance.", 'The evolution of prompting techniques from 2021 to 2023 is highlighted, emphasizing the shift towards designing prompts that aid the model in reasoning in intended ways, often referred to as step-by-step reasoning. The transition from naive direct questions to more designed prompts in 2023 illustrates the evolution in prompting techniques to facilitate step-by-step reasoning, showcasing the progress in model reasoning capabilities.']}], 'duration': 348.33, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/-lnHHWRCDGk/pics/-lnHHWRCDGk1103560.jpg', 'highlights': ['Large-scale pre-training and its empowering impact on the community', 'The instruct models are best in class in the field due to human feedback at a scale that is astounding', 'The importance of analytic work in understanding the effectiveness of self-supervision', 'The evolution of prompting techniques from 2021 to 2023 is highlighted']}, {'end': 1909.114, 'segs': [{'end': 1508.925, 'src': 'embed', 'start': 1472.385, 'weight': 0, 'content': [{'end': 1479.448, 'text': "you'd say that these large language models are kind of like alien creatures and it's taking us some time to figure out how to communicate with them.", 'start': 1472.385, 'duration': 7.063}, {'end': 1487.131, 'text': "And together with all that instruct fine tuning with human supervision, we're converging on prompts like this as the powerful device.", 'start': 1480.108, 'duration': 7.023}, {'end': 1496.696, 'text': "And this is exciting to me, because what's really emerging is that this is a kind of very light way of programming an AI system using only prompts,", 'start': 1487.771, 'duration': 8.925}, {'end': 1499.477, 'text': 'as opposed to all the deep learning code that we used to have to write.', 'start': 1496.696, 'duration': 2.781}, {'end': 1504.3, 'text': "And that's going to be incredibly empowering in terms of system development and experimentation.", 'start': 1499.818, 'duration': 4.482}, {'end': 1508.925, 'text': 'All right, so we have our background in place.', 'start': 1506.724, 'duration': 2.201}], 'summary': 'Large language models can be programmed with only prompts, empowering system development and experimentation.', 'duration': 36.54, 'max_score': 1472.385, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/-lnHHWRCDGk/pics/-lnHHWRCDGk1472385.jpg'}, {'end': 1552.351, 'src': 'embed', 'start': 1526.329, 'weight': 3, 'content': [{'end': 1533.335, 'text': "I think we're all probably vaguely aware at this point that large language models have been revolutionizing search.", 'start': 1526.329, 'duration': 7.006}, {'end': 1539.78, 'text': 'Again, the star of this is the transformer, or maybe more specifically, its famous spokesmodel BERT.', 'start': 1533.995, 'duration': 5.785}, {'end': 1548.648, 'text': 'Right after BERT was announced around 2018, Google announced that it was incorporating aspects of BERT into its core search technology.', 'start': 1540.741, 'duration': 7.907}, {'end': 1552.351, 'text': 'And Microsoft made a similar announcement at about the same time.', 'start': 1548.688, 'duration': 3.663}], 'summary': 'Large language models like bert revolutionizing search, adopted by google and microsoft.', 'duration': 26.022, 'max_score': 1526.329, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/-lnHHWRCDGk/pics/-lnHHWRCDGk1526329.jpg'}, {'end': 1706.801, 'src': 'heatmap', 'start': 1656.679, 'weight': 4, 'content': [{'end': 1665.988, 'text': 'And the standard strategy is to rely on some kind of retrieval mechanism to find relevant evidence in a large corpus or maybe even the web.', 'start': 1656.679, 'duration': 9.309}, {'end': 1668.11, 'text': 'And then we proceed as before.', 'start': 1666.629, 'duration': 1.481}, {'end': 1673.735, 'text': "This is a much harder problem because we're not going to get the substring guarantee anymore,", 'start': 1668.951, 'duration': 4.784}, {'end': 1676.718, 'text': "because we're dependent on the retriever to find relevant evidence.", 'start': 1673.735, 'duration': 2.983}, {'end': 1683.184, 'text': "But of course, it's a much more important task because this is much more like our experience of searching on the web.", 'start': 1677.118, 'duration': 6.066}, {'end': 1690.381, 'text': "Now, I've kind of biased already in describing things this way, where I assume we're retrieving a passage.", 'start': 1684.915, 'duration': 5.466}, {'end': 1693.404, 'text': 'But there is another narrative out there.', 'start': 1691.262, 'duration': 2.142}, {'end': 1694.606, 'text': 'Let me skip to this.', 'start': 1693.425, 'duration': 1.181}, {'end': 1697.709, 'text': 'Then you could call this like the LLMs for everything approach.', 'start': 1694.686, 'duration': 3.023}, {'end': 1700.452, 'text': "And this would be where there's no explicit retriever.", 'start': 1698.29, 'duration': 2.162}, {'end': 1702.314, 'text': 'You just have a question come in.', 'start': 1700.933, 'duration': 1.381}, {'end': 1706.801, 'text': 'You have a big opaque model process that question, and out comes an answer.', 'start': 1702.915, 'duration': 3.886}], 'summary': 'Retrieval-based approach relies on finding evidence in a large corpus or the web, posing a more challenging yet crucial task, differing from the llms for everything approach.', 'duration': 50.122, 'max_score': 1656.679, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/-lnHHWRCDGk/pics/-lnHHWRCDGk1656679.jpg'}, {'end': 1742.597, 'src': 'embed', 'start': 1716.314, 'weight': 5, 'content': [{'end': 1722.941, 'text': 'I think this is an incredibly inspiring vision, but we should be aware that there are lots of kind of danger zones here.', 'start': 1716.314, 'duration': 6.627}, {'end': 1725.643, 'text': 'so the first is just efficiency.', 'start': 1722.941, 'duration': 2.702}, {'end': 1733.269, 'text': 'one of the major factors driving that explosion in model size that i tracked before is that in this llms for everything approach,', 'start': 1725.643, 'duration': 7.626}, {'end': 1739.274, 'text': 'we are asking this model to play the role of both knowledge store and language capability.', 'start': 1733.269, 'duration': 6.005}, {'end': 1742.597, 'text': 'if we could separate those out, we might get away with smaller models.', 'start': 1739.274, 'duration': 3.323}], 'summary': 'Efficiency is a concern in using large language models as both knowledge store and language capability.', 'duration': 26.283, 'max_score': 1716.314, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/-lnHHWRCDGk/pics/-lnHHWRCDGk1716314.jpg'}, {'end': 1851.042, 'src': 'embed', 'start': 1823.979, 'weight': 6, 'content': [{'end': 1827.301, 'text': "And they're also outstanding at synthesizing information.", 'start': 1823.979, 'duration': 3.322}, {'end': 1831.243, 'text': 'If your question can only be answered by 10 different web pages,', 'start': 1827.341, 'duration': 3.902}, {'end': 1836.806, 'text': "it's very likely that the language model will still be able to do it without you having to hunt through all those pages.", 'start': 1831.243, 'duration': 5.563}, {'end': 1839.794, 'text': 'So exciting, but lots of concerns here.', 'start': 1837.813, 'duration': 1.981}, {'end': 1844.478, 'text': 'Here is the alternative of retrieval augmented approaches, right?', 'start': 1840.215, 'duration': 4.263}, {'end': 1846.399, 'text': "Oh, I can't resist this.", 'start': 1845.098, 'duration': 1.301}, {'end': 1851.042, 'text': 'actually, just to give you an example of how important this trustworthy thing can be.', 'start': 1846.399, 'duration': 4.643}], 'summary': 'Language model can answer questions from 10 web pages, with concerns about trustworthiness.', 'duration': 27.063, 'max_score': 1823.979, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/-lnHHWRCDGk/pics/-lnHHWRCDGk1823979.jpg'}], 'start': 1452.931, 'title': 'Language models and retrieval augmented learning', 'summary': 'Explores the transformative potential of language models and the integration of retriever models to revolutionize search and nlp tasks, empowering system development and experimentation with only prompts.', 'chapters': [{'end': 1508.925, 'start': 1452.931, 'title': 'Language models: the power of prompts', 'summary': 'Discusses the transformative potential of large language models, highlighting the emergence of a light way of programming ai systems using only prompts and the potential for empowering system development and experimentation.', 'duration': 55.994, 'highlights': ['The emergence of a light way of programming AI systems using only prompts is incredibly empowering in terms of system development and experimentation.', 'Large language models are like alien creatures that require time to figure out how to communicate with them.', 'The current era offers the potential for transformative improvements in language model capabilities.']}, {'end': 1909.114, 'start': 1508.965, 'title': 'Retrieval augmented in context learning', 'summary': 'Discusses the integration of language models with retriever models to revolutionize search and nlp tasks, highlighting the challenges and benefits of retrieval augmented approaches.', 'duration': 400.149, 'highlights': ['Large language models revolutionizing search Large language models like BERT have revolutionized search technology, with Google and Microsoft incorporating BERT aspects into their core search technology around 2018.', 'Transition from traditional question answering to retrieval-based approach The shift from traditional literal substring question answering to a retrieval-based approach allows for more relevant, knowledge-intensive tasks, connecting directly with real-world scenarios.', "Challenges of 'LLMs for everything' approach The 'LLMs for everything' approach raises concerns related to efficiency, updatability, trustworthiness, and provenance, impacting the trust and explainability of model behavior.", 'Benefits and concerns of language models Language models are effective at meeting information needs directly and synthesizing information, but raise concerns regarding trustworthiness and explainability, emphasizing the need for retrieval augmented approaches.']}], 'duration': 456.183, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/-lnHHWRCDGk/pics/-lnHHWRCDGk1452931.jpg', 'highlights': ['The emergence of a light way of programming AI systems using only prompts is incredibly empowering in terms of system development and experimentation.', 'Large language models are like alien creatures that require time to figure out how to communicate with them.', 'The current era offers the potential for transformative improvements in language model capabilities.', 'Large language models revolutionizing search Large language models like BERT have revolutionized search technology, with Google and Microsoft incorporating BERT aspects into their core search technology around 2018.', 'Transition from traditional question answering to retrieval-based approach The shift from traditional literal substring question answering to a retrieval-based approach allows for more relevant, knowledge-intensive tasks, connecting directly with real-world scenarios.', "Challenges of 'LLMs for everything' approach The 'LLMs for everything' approach raises concerns related to efficiency, updatability, trustworthiness, and provenance, impacting the trust and explainability of model behavior.", 'Benefits and concerns of language models Language models are effective at meeting information needs directly and synthesizing information, but raise concerns regarding trustworthiness and explainability, emphasizing the need for retrieval augmented approaches.']}, {'end': 2408.608, 'segs': [{'end': 1977.453, 'src': 'embed', 'start': 1953.613, 'weight': 0, 'content': [{'end': 1961.259, 'text': 'we are also gonna use a language model, maybe the same one we used for the query to process all of the documents in our document collection.', 'start': 1953.613, 'duration': 7.646}, {'end': 1966.424, 'text': 'So each one has some kind of numerical deep learning representation now.', 'start': 1962.06, 'duration': 4.364}, {'end': 1972.449, 'text': 'On the basis of these representations, we can now score documents with respect to queries,', 'start': 1967.344, 'duration': 5.105}, {'end': 1977.453, 'text': 'just like we would in the standard good old days of information retrieval.', 'start': 1972.449, 'duration': 5.004}], 'summary': 'Using a language model to process documents for scoring with deep learning representations.', 'duration': 23.84, 'max_score': 1953.613, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/-lnHHWRCDGk/pics/-lnHHWRCDGk1953613.jpg'}, {'end': 2072.371, 'src': 'heatmap', 'start': 1909.835, 'weight': 2, 'content': [{'end': 1918.061, 'text': "Is it true? It would help enormously if the model could offer me at least a web page with evidence that's relevant to these claims.", 'start': 1909.835, 'duration': 8.226}, {'end': 1920.562, 'text': "Otherwise, I'm simply left wondering.", 'start': 1918.601, 'duration': 1.961}, {'end': 1926.747, 'text': "I think that shows you that we've broken this implicit contract with the user that we expect from search.", 'start': 1920.722, 'duration': 6.025}, {'end': 1932.403, 'text': "So that'll bring me to my alternative here, retrieval-based or retrieval-augmented NLP.", 'start': 1927.878, 'duration': 4.525}, {'end': 1938.709, 'text': "To give you a sense for this, at the top here I have a standard search box and I've put in a very complicated question indeed.", 'start': 1933.143, 'duration': 5.566}, {'end': 1943.754, 'text': 'The first step in this approach is familiar from the LLMs for everything one.', 'start': 1939.55, 'duration': 4.204}, {'end': 1949.72, 'text': "We're gonna encode that query into a dense numerical representation capturing aspects of its form and meaning.", 'start': 1943.774, 'duration': 5.946}, {'end': 1951.202, 'text': "We'll use a language model for that.", 'start': 1949.76, 'duration': 1.442}, {'end': 1953.613, 'text': 'The next step is new, though.', 'start': 1952.452, 'duration': 1.161}, {'end': 1961.259, 'text': 'we are also gonna use a language model, maybe the same one we used for the query to process all of the documents in our document collection.', 'start': 1953.613, 'duration': 7.646}, {'end': 1966.424, 'text': 'So each one has some kind of numerical deep learning representation now.', 'start': 1962.06, 'duration': 4.364}, {'end': 1972.449, 'text': 'On the basis of these representations, we can now score documents with respect to queries,', 'start': 1967.344, 'duration': 5.105}, {'end': 1977.453, 'text': 'just like we would in the standard good old days of information retrieval.', 'start': 1972.449, 'duration': 5.004}, {'end': 1981.974, 'text': 'So we can reproduce every aspect of that familiar experience if we want to.', 'start': 1977.893, 'duration': 4.081}, {'end': 1985.215, 'text': "We're just doing it now in this very rich semantic space.", 'start': 1981.994, 'duration': 3.221}, {'end': 1991.377, 'text': 'So we get some results back and we could offer those to the user as ranked results, but we can also go further.', 'start': 1986.175, 'duration': 5.202}, {'end': 2001.679, 'text': 'We can have another language model call it a reader or a generator, slurp up those retrieved passages and synthesize them into a single answer,', 'start': 1991.977, 'duration': 9.702}, {'end': 2004.02, 'text': "maybe meeting the user's information need directly.", 'start': 2001.679, 'duration': 2.341}, {'end': 2008.17, 'text': "So let's check in on how we're doing with respect to our goals here.", 'start': 2005.109, 'duration': 3.061}, {'end': 2009.15, 'text': 'First, efficiency.', 'start': 2008.25, 'duration': 0.9}, {'end': 2014.472, 'text': "I won't have time to substantiate this today, but these systems, in terms of parameter counts,", 'start': 2009.67, 'duration': 4.802}, {'end': 2018.053, 'text': 'can be much smaller than the integrated approach I mentioned before.', 'start': 2014.472, 'duration': 3.581}, {'end': 2021.554, 'text': 'We also have an easy path to updatability.', 'start': 2019.393, 'duration': 2.161}, {'end': 2023.014, 'text': 'We have this index here.', 'start': 2021.614, 'duration': 1.4}, {'end': 2030.537, 'text': 'So as pages change in our document store, we simply use our frozen language model to reprocess and re-represent them.', 'start': 2023.615, 'duration': 6.922}, {'end': 2037.796, 'text': 'And we can have a pretty good guarantee at this point that information changes will be reflected in the retrieved results down here.', 'start': 2031.037, 'duration': 6.759}, {'end': 2044.539, 'text': "We're also naturally tracking provenance, because we have all these documents and they're used to deliver the results,", 'start': 2038.796, 'duration': 5.743}, {'end': 2047.141, 'text': 'and we can have that carry through into the generation.', 'start': 2044.539, 'duration': 2.602}, {'end': 2049.742, 'text': "So we've kept that contract with the user.", 'start': 2047.661, 'duration': 2.081}, {'end': 2053.804, 'text': 'These models are incredibly effective across lots of literature.', 'start': 2050.562, 'duration': 3.242}, {'end': 2060.628, 'text': "We're seeing that retrieval augmented approaches are just superior to the fully integrated LLMs for everything one.", 'start': 2053.844, 'duration': 6.784}, {'end': 2064.217, 'text': "And we've retained the benefit of LLMs for everything,", 'start': 2061.533, 'duration': 2.684}, {'end': 2072.371, 'text': 'because we have this model down here the reader generator that can synthesize information into answers that meet the information need directly.', 'start': 2064.217, 'duration': 8.154}], 'summary': 'Retrieval-augmented nlp offers efficient, updatable, and effective search results with provenance tracking.', 'duration': 162.536, 'max_score': 1909.835, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/-lnHHWRCDGk/pics/-lnHHWRCDGk1909835.jpg'}, {'end': 2154.025, 'src': 'embed', 'start': 2127.3, 'weight': 3, 'content': [{'end': 2135.268, 'text': 'The truth in practice is that even for very experienced researchers and system designers, this can often go really wrong.', 'start': 2127.3, 'duration': 7.968}, {'end': 2144.458, 'text': 'And debugging these systems and figuring out how to improve them can be very difficult because they are so opaque and the scale is so large.', 'start': 2136.009, 'duration': 8.449}, {'end': 2150.603, 'text': "But maybe we're moving out of an era in which we have to do this at all.", 'start': 2146.601, 'duration': 4.002}, {'end': 2154.025, 'text': 'So this will bring us back to in-context learning.', 'start': 2151.464, 'duration': 2.561}], 'summary': 'Debugging large-scale systems can be difficult for experienced researchers and designers, but a shift to in-context learning may offer a solution.', 'duration': 26.725, 'max_score': 2127.3, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/-lnHHWRCDGk/pics/-lnHHWRCDGk2127300.jpg'}, {'end': 2223.214, 'src': 'heatmap', 'start': 2177.129, 'weight': 1, 'content': [{'end': 2185.812, 'text': 'We have already seen in my basic picture of retrieval augmented approaches that we can have the retriever communicate with the language model via retrieve results.', 'start': 2177.129, 'duration': 8.683}, {'end': 2189.315, 'text': 'Well, what if we just allow that to go in both directions?', 'start': 2186.752, 'duration': 2.563}, {'end': 2195.481, 'text': "Now we've got a system that is essentially constructed by prompts that help these models do message,", 'start': 2189.515, 'duration': 5.966}, {'end': 2199.125, 'text': 'passing between them in potentially very complicated ways.', 'start': 2195.481, 'duration': 3.644}, {'end': 2202.549, 'text': 'An entirely new approach to system design that I think,', 'start': 2199.366, 'duration': 3.183}, {'end': 2208.235, 'text': "is going to have an incredible democratizing effect on who designs these systems and what they're for.", 'start': 2202.549, 'duration': 5.686}, {'end': 2214.289, 'text': 'Let me give you a deep sense for just how wide open the design space is here.', 'start': 2209.226, 'duration': 5.063}, {'end': 2220.732, 'text': 'Again, to give you a sense for how much of this research is still left to be done even in this golden era.', 'start': 2214.489, 'duration': 6.243}, {'end': 2223.214, 'text': "Let's imagine a search context.", 'start': 2221.673, 'duration': 1.541}], 'summary': 'Retrieval augmented approaches enable bidirectional communication between retriever and language model, with potential for democratizing system design.', 'duration': 21.996, 'max_score': 2177.129, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/-lnHHWRCDGk/pics/-lnHHWRCDGk2177129.jpg'}], 'start': 1909.835, 'title': 'Retrieval-augmented nlp and language models', 'summary': 'Introduces retrieval-augmented nlp, utilizing dense numerical representations for efficient and updatable systems, and discusses the potential of in-context learning to revolutionize system design, democratizing the effect on system designers.', 'chapters': [{'end': 2023.014, 'start': 1909.835, 'title': 'Retrieval-augmented nlp', 'summary': 'Introduces the concept of retrieval-augmented nlp, explaining how it uses dense numerical representations to process queries and documents, enabling efficient and updatable systems with smaller parameter counts.', 'duration': 113.179, 'highlights': ['The retrieval-augmented NLP encodes queries and documents into dense numerical representations, allowing for efficient scoring of documents with respect to queries and synthesis of retrieved passages into single answers.', 'Retrieval-augmented NLP systems have smaller parameter counts compared to integrated approaches, offering an easy path to updatability and efficient processing.']}, {'end': 2408.608, 'start': 2023.615, 'title': 'Retrieval-augmented language models', 'summary': 'Discusses the effectiveness of retrieval-augmented approaches and the potential of in-context learning to revolutionize system design, emphasizing the wide open design space and democratizing effect on system designers.', 'duration': 384.993, 'highlights': ['The effectiveness of retrieval augmented approaches is highlighted as superior to fully integrated LLMs, with a pretty good guarantee that information changes will be reflected in the retrieved results.', 'The potential of in-context learning to revolutionize system design is emphasized, with an insight into the wide open design space and democratizing effect on system designers.', 'The challenges of traditional deep learning in integrating pre-trained components and the potential shift towards in-context learning is discussed, highlighting the difficulty of debugging opaque and large-scale systems.', 'The concept of in-context learning is introduced, showcasing the potential for models to communicate in natural language and construct systems through prompts, with a democratizing effect on system design and functionality.', "The diverse design space of in-context learning is explored, including the use of prompts, demonstrations, and rewriting techniques to guide the system's learning process and improve reasoning and answer generation."]}], 'duration': 498.773, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/-lnHHWRCDGk/pics/-lnHHWRCDGk1909835.jpg', 'highlights': ['Retrieval-augmented NLP encodes queries and documents into dense numerical representations, allowing for efficient scoring and synthesis of retrieved passages.', 'Retrieval-augmented NLP systems have smaller parameter counts compared to integrated approaches, offering an easy path to updatability and efficient processing.', 'The effectiveness of retrieval augmented approaches is highlighted as superior to fully integrated LLMs, ensuring reflected information changes in retrieved results.', 'In-context learning has the potential to revolutionize system design, democratizing the effect on system designers.', 'In-context learning introduces the potential for models to communicate in natural language and construct systems through prompts, democratizing system design and functionality.']}, {'end': 3461.387, 'segs': [{'end': 2455.736, 'src': 'embed', 'start': 2410.765, 'weight': 0, 'content': [{'end': 2413.967, 'text': "I hope that's given you a sense for just how much can happen here.", 'start': 2410.765, 'duration': 3.202}, {'end': 2418.549, 'text': "What we're starting to see, I think, is that there is a new programming mode emerging.", 'start': 2414.027, 'duration': 4.522}, {'end': 2433.647, 'text': "It's a programming mode that involves using these large pre-trained components to design in code prompts that are essentially full AI systems that are entirely about message passing between these frozen components.", 'start': 2418.809, 'duration': 14.838}, {'end': 2438.109, 'text': "We have a new paper out that's called Demonstrate Search Predict or DSP.", 'start': 2434.227, 'duration': 3.882}, {'end': 2443.111, 'text': 'This is a lightweight programming framework for doing exactly what I was just describing for you.', 'start': 2438.529, 'duration': 4.582}, {'end': 2448.013, 'text': 'And one thing I want to call out is that our results are fantastic.', 'start': 2443.991, 'duration': 4.022}, {'end': 2451.633, 'text': 'Now, we can pat ourselves on the back.', 'start': 2448.911, 'duration': 2.722}, {'end': 2455.736, 'text': "We have a very talented team, and so it's no surprise the results are so good.", 'start': 2451.693, 'duration': 4.043}], 'summary': 'Emerging programming mode using large pre-trained components, dsp framework yields fantastic results.', 'duration': 44.971, 'max_score': 2410.765, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/-lnHHWRCDGk/pics/-lnHHWRCDGk2410765.jpg'}, {'end': 2623.397, 'src': 'embed', 'start': 2601.395, 'weight': 2, 'content': [{'end': 2611.087, 'text': "and we need always to be pushing our systems with harder tasks that come closer to the human capabilities that we're actually trying to get them to achieve.", 'start': 2601.395, 'duration': 9.692}, {'end': 2617.595, 'text': "Without contributions of datasets, we could be tricking ourselves when we think we're making a lot of progress.", 'start': 2611.708, 'duration': 5.887}, {'end': 2623.397, 'text': 'The second thing that I wanted to call out relates to model explainability.', 'start': 2619.195, 'duration': 4.202}], 'summary': 'Push systems with harder tasks, rely on datasets, focus on model explainability.', 'duration': 22.002, 'max_score': 2601.395, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/-lnHHWRCDGk/pics/-lnHHWRCDGk2601395.jpg'}, {'end': 2835.749, 'src': 'embed', 'start': 2805.367, 'weight': 4, 'content': [{'end': 2806.508, 'text': 'But here are the predictions.', 'start': 2805.367, 'duration': 1.141}, {'end': 2814.095, 'text': 'First laggard industries that are rich in text data will be transformed in part by NLP technology,', 'start': 2807.749, 'duration': 6.346}, {'end': 2818.859, 'text': "and that's likely to happen from some disruptive newcomers coming out of left field.", 'start': 2814.095, 'duration': 4.764}, {'end': 2825.314, 'text': 'Second prediction artificial assistants will get dramatically better and become more ubiquitous,', 'start': 2820.006, 'duration': 5.308}, {'end': 2835.749, 'text': "with the side effect that you'll often be unsure in life whether this customer service representative is a person or an AI or some team combining the two.", 'start': 2825.314, 'duration': 10.435}], 'summary': 'Nlp tech will transform text-rich industries. ai assistants will improve, blurring line between humans and ai.', 'duration': 30.382, 'max_score': 2805.367, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/-lnHHWRCDGk/pics/-lnHHWRCDGk2805367.jpg'}, {'end': 2876.037, 'src': 'embed', 'start': 2851.364, 'weight': 5, 'content': [{'end': 2857.188, 'text': 'and then, finally, the negative effects of nlp and of ai will be amplified along with the positives.', 'start': 2851.364, 'duration': 5.824}, {'end': 2861.671, 'text': "i'm thinking of things like disinformation, spread market disruption, systemic bias.", 'start': 2857.188, 'duration': 4.483}, {'end': 2866.83, 'text': "it's almost sure to be the case, if it hasn't already happened already,", 'start': 2863.328, 'duration': 3.502}, {'end': 2875.176, 'text': 'that there will be some calamitous world event that traces to the intentional or unintentional misuse of some AI technology.', 'start': 2866.83, 'duration': 8.346}, {'end': 2876.037, 'text': "that's in our future.", 'start': 2875.176, 'duration': 0.861}], 'summary': 'Negative effects of ai and nlp will be amplified, leading to potential calamitous world events.', 'duration': 24.673, 'max_score': 2851.364, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/-lnHHWRCDGk/pics/-lnHHWRCDGk2851364.jpg'}, {'end': 3174.992, 'src': 'embed', 'start': 3145.855, 'weight': 3, 'content': [{'end': 3153.62, 'text': "It is the thing I have in mind when we're doing all our work on explaining models, because I feel like offering faithful, human,", 'start': 3145.855, 'duration': 7.765}, {'end': 3157.742, 'text': 'interpretable explanations is the step we can take toward trustworthiness.', 'start': 3153.62, 'duration': 4.122}, {'end': 3159.623, 'text': "It's a very difficult problem.", 'start': 3158.162, 'duration': 1.461}, {'end': 3167.548, 'text': "I just want to add that it might be even harder than we've anticipated because people are also pretty untrustworthy.", 'start': 3159.963, 'duration': 7.585}, {'end': 3174.992, 'text': "It's just that individual people often don't have like a systemic effect, right?", 'start': 3168.749, 'duration': 6.243}], 'summary': 'Offering faithful, interpretable explanations in ai models can lead to trustworthiness, despite the challenge posed by human untrustworthiness.', 'duration': 29.137, 'max_score': 3145.855, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/-lnHHWRCDGk/pics/-lnHHWRCDGk3145855.jpg'}, {'end': 3268.483, 'src': 'embed', 'start': 3242.598, 'weight': 7, 'content': [{'end': 3246.799, 'text': 'we are not even sure that it even exists in our minds right now?', 'start': 3242.598, 'duration': 4.201}, {'end': 3248.997, 'text': "Oh, it's a wonderful question.", 'start': 3247.677, 'duration': 1.32}, {'end': 3253.859, 'text': "Yeah, and people are asking this across multiple domains, like they're producing incredible artwork,", 'start': 3249.017, 'duration': 4.842}, {'end': 3259.521, 'text': "but are we now trapped inside a feedback loop that's going to lead to less truly innovative art?", 'start': 3253.859, 'duration': 5.662}, {'end': 3268.483, 'text': 'And if we ask them to generate text, are they going to do either weird, irrelevant stuff or just more of the boring average case stuff?', 'start': 3260.121, 'duration': 8.362}], 'summary': 'Questioning the impact of ai on art and creativity.', 'duration': 25.885, 'max_score': 3242.598, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/-lnHHWRCDGk/pics/-lnHHWRCDGk3242598.jpg'}, {'end': 3408.312, 'src': 'embed', 'start': 3377.405, 'weight': 8, 'content': [{'end': 3379.467, 'text': 'And therefore you could bring it to bear on AI,', 'start': 3377.405, 'duration': 2.062}, {'end': 3385.871, 'text': 'and we might all benefit where you would be taking all these innovations you can learn about in our course and other courses,', 'start': 3379.467, 'duration': 6.404}, {'end': 3394.437, 'text': 'combining that with your domain expertise and maybe actually making progress in a meaningful way on a problem,', 'start': 3385.871, 'duration': 8.566}, {'end': 3399.28, 'text': 'as opposed to merely having demos and things that our scientific community often produces.', 'start': 3394.437, 'duration': 4.843}, {'end': 3405.184, 'text': 'Real impact so often requires real domain expertise of the sort you all have.', 'start': 3399.82, 'duration': 5.364}, {'end': 3408.312, 'text': 'Great, thank you so much.', 'start': 3407.031, 'duration': 1.281}], 'summary': 'Combining ai with domain expertise can lead to real impact and progress on problems, benefiting all.', 'duration': 30.907, 'max_score': 3377.405, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/-lnHHWRCDGk/pics/-lnHHWRCDGk3377405.jpg'}], 'start': 2410.765, 'title': 'The evolution of ai programming', 'summary': 'Discusses the emergence of a new programming mode with dsp framework, the state-of-the-art nlp and ai challenges, including the societal impact, predictions for the future, and environmental impact, and the challenges and solutions for trustworthiness in ai models.', 'chapters': [{'end': 2455.736, 'start': 2410.765, 'title': 'Emerging programming mode with dsp', 'summary': 'Discusses the emergence of a new programming mode involving the use of large pre-trained components to design code prompts, exemplified by demonstrate search predict (dsp) framework with excellent results.', 'duration': 44.971, 'highlights': ['Demonstrate Search Predict (DSP) is a lightweight programming framework for using large pre-trained components to design code prompts, resulting in fantastic results.', 'The new programming mode involves utilizing large pre-trained components for message passing and designing code prompts that are essentially full AI systems.']}, {'end': 3099.014, 'start': 2455.756, 'title': 'State-of-the-art nlp and ai challenges', 'summary': 'Discusses the early stage of nlp and ai development, emphasizing the need for better prompts and systems, the importance of contributing new benchmark datasets, the challenges of model explainability, and the societal impact of ai. it also includes predictions for the future of nlp and ai, highlighting the transformation of industries, the advancement of artificial assistants, and the potential negative effects of ai. furthermore, it addresses the environmental impact of ai training and the need for attention to energy requirements.', 'duration': 643.258, 'highlights': ["The importance of contributing new benchmark datasets Emphasizes the significance of creating effective datasets and pushing systems with harder tasks to avoid tricking ourselves into thinking there's significant progress without dataset contributions.", 'Challenges of model explainability and system reliability Discusses the difficulty of achieving analytic guarantees about model behaviors and highlights the importance of achieving faithful human-interpretable explanations of model behavior.', 'Predictions for the future of NLP and AI Addresses the transformation of industries, advancements in artificial assistants, potential negative effects of AI, and emphasizes the uncertainty of the future of AI despite making predictions.', 'Environmental impact of AI training and energy requirements Raises questions about the environmental impact of AI, considering the massive expenditure for training and serving large models, and the trade-off between centralization and real benefits.']}, {'end': 3461.387, 'start': 3099.014, 'title': 'Trustworthiness of ai models', 'summary': 'Discusses the challenges of trustworthiness in ai models, emphasizing the need for faithful, human, interpretable explanations to achieve higher standards for trustworthiness and the potential impact of ai models on innovation. it also highlights the importance of combining domain expertise with ai knowledge for real impact.', 'duration': 362.373, 'highlights': ['The need for faithful, human, interpretable explanations to achieve higher standards for trustworthiness in AI models. Offering faithful, human, interpretable explanations is the step towards trustworthiness, highlighting the difficulty in achieving trustworthiness due to the systemic effect of AI models impacting the entire population.', 'Potential impact of AI models on innovation and the concern of being trapped inside a feedback loop that leads to less truly innovative art. Discusses the potential impact of AI models on innovation, specifically in creating innovative art and synthesizing information across sources, while raising concerns about being trapped in a feedback loop that may lead to less truly innovative art.', 'Importance of combining domain expertise with AI knowledge for real impact. Emphasizes the relevance of domain expertise in contributing to AI and achieving real impact by combining domain expertise with AI knowledge.']}], 'duration': 1050.622, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/-lnHHWRCDGk/pics/-lnHHWRCDGk2410765.jpg', 'highlights': ['Demonstrate Search Predict (DSP) is a lightweight programming framework for using large pre-trained components to design code prompts, resulting in fantastic results.', 'The new programming mode involves utilizing large pre-trained components for message passing and designing code prompts that are essentially full AI systems.', "The importance of contributing new benchmark datasets Emphasizes the significance of creating effective datasets and pushing systems with harder tasks to avoid tricking ourselves into thinking there's significant progress without dataset contributions.", 'Challenges of model explainability and system reliability Discusses the difficulty of achieving analytic guarantees about model behaviors and highlights the importance of achieving faithful human-interpretable explanations of model behavior.', 'Predictions for the future of NLP and AI Addresses the transformation of industries, advancements in artificial assistants, potential negative effects of AI, and emphasizes the uncertainty of the future of AI despite making predictions.', 'Environmental impact of AI training and energy requirements Raises questions about the environmental impact of AI, considering the massive expenditure for training and serving large models, and the trade-off between centralization and real benefits.', 'The need for faithful, human, interpretable explanations to achieve higher standards for trustworthiness in AI models. Offering faithful, human, interpretable explanations is the step towards trustworthiness, highlighting the difficulty in achieving trustworthiness due to the systemic effect of AI models impacting the entire population.', 'Potential impact of AI models on innovation and the concern of being trapped inside a feedback loop that leads to less truly innovative art. Discusses the potential impact of AI models on innovation, specifically in creating innovative art and synthesizing information across sources, while raising concerns about being trapped in a feedback loop that may lead to less truly innovative art.', 'Importance of combining domain expertise with AI knowledge for real impact. Emphasizes the relevance of domain expertise in contributing to AI and achieving real impact by combining domain expertise with AI knowledge.']}], 'highlights': ['GPT-3 has 175 billion parameters, surpassing models with over 500 billion parameters, showcasing the rapid increase in model size and capacity.', 'The rapid increase in AI model size, from 100 million parameters in 2018 to models with over 500 billion parameters, has made a significant impact on research and future implications, reflecting a golden age in AI development.', 'The proliferation of high-performing language models like GPT-3 and DaVinci 3, along with their societal impact and increased accessibility, demonstrating their capabilities through examples and comparisons with earlier versions.', 'The advancements in natural language understanding have led to the availability of models like DALI 2, GitHub Copilot, U.com, and Whisper AI, showcasing breadth of applications.', 'Retrieval-augmented NLP encodes queries and documents into dense numerical representations, allowing for efficient scoring and synthesis of retrieved passages.', 'Demonstrate Search Predict (DSP) is a lightweight programming framework for using large pre-trained components to design code prompts, resulting in fantastic results.', 'The instruct models are best in class in the field due to human feedback at a scale that is astounding', 'The importance of analytic work in understanding the effectiveness of self-supervision', 'The evolution of prompting techniques from 2021 to 2023 is highlighted', 'The rise of in-context learning is identified as a genuine paradigm shift, tracing back to the GPT-3 paper, and it is highlighted as a significant change resulting from large language models.', 'The transformer architecture is discussed as a key mechanism behind the success of in-context learning, with an emphasis on its attention mechanisms and departure from previous architectures like LSTMs.', 'In-context learning has the potential to revolutionize system design, democratizing the effect on system designers.', 'In-context learning introduces the potential for models to communicate in natural language and construct systems through prompts, democratizing system design and functionality.', 'The emergence of a light way of programming AI systems using only prompts is incredibly empowering in terms of system development and experimentation.', 'Large language models revolutionizing search Large language models like BERT have revolutionized search technology, with Google and Microsoft incorporating BERT aspects into their core search technology around 2018.', 'Transition from traditional question answering to retrieval-based approach The shift from traditional literal substring question answering to a retrieval-based approach allows for more relevant, knowledge-intensive tasks, connecting directly with real-world scenarios.', "Challenges of 'LLMs for everything' approach The 'LLMs for everything' approach raises concerns related to efficiency, updatability, trustworthiness, and provenance, impacting the trust and explainability of model behavior.", 'Benefits and concerns of language models Language models are effective at meeting information needs directly and synthesizing information, but raise concerns regarding trustworthiness and explainability, emphasizing the need for retrieval augmented approaches.', 'The new programming mode involves utilizing large pre-trained components for message passing and designing code prompts that are essentially full AI systems.', "The importance of contributing new benchmark datasets Emphasizes the significance of creating effective datasets and pushing systems with harder tasks to avoid tricking ourselves into thinking there's significant progress without dataset contributions.", 'Challenges of model explainability and system reliability Discusses the difficulty of achieving analytic guarantees about model behaviors and highlights the importance of achieving faithful human-interpretable explanations of model behavior.', 'Predictions for the future of NLP and AI Addresses the transformation of industries, advancements in artificial assistants, potential negative effects of AI, and emphasizes the uncertainty of the future of AI despite making predictions.', 'Environmental impact of AI training and energy requirements Raises questions about the environmental impact of AI, considering the massive expenditure for training and serving large models, and the trade-off between centralization and real benefits.', 'The need for faithful, human, interpretable explanations to achieve higher standards for trustworthiness in AI models. Offering faithful, human, interpretable explanations is the step towards trustworthiness, highlighting the difficulty in achieving trustworthiness due to the systemic effect of AI models impacting the entire population.', 'Potential impact of AI models on innovation and the concern of being trapped inside a feedback loop that leads to less truly innovative art. Discusses the potential impact of AI models on innovation, specifically in creating innovative art and synthesizing information across sources, while raising concerns about being trapped in a feedback loop that may lead to less truly innovative art.', 'Importance of combining domain expertise with AI knowledge for real impact. Emphasizes the relevance of domain expertise in contributing to AI and achieving real impact by combining domain expertise with AI knowledge.']}