1 00:00:03,000 --> 00:00:05,600 Evan: Welcome to Daily Paper Cast. 2 00:00:05,664 --> 00:00:14,224 Ashley: Today’s paper is from the Hugging Face daily paper list of October 22, 2026, and has received 53 upvotes. 3 00:00:14,280 --> 00:00:21,060 Evan: The paper is titled 'Looking Ahead: Modeling Future Evolution in Autoregressive Video Generation'. 4 00:00:21,120 --> 00:00:32,020 Ashley: The first two authors are Ruopeng Gao and Yukun Huang, with Limin Wang as the corresponding author, all from Nanjing University and ARC Lab, Tencent IEG. 5 00:00:32,124 --> 00:00:42,454 Evan: Autoregressive video generation has emerged as an important direction in video generation, offering a flexible paradigm that allows visual sequences to be extended continuously. 6 00:00:42,454 --> 00:00:49,384 It primarily operates by generating video segments, or chunks, conditioned on preceding visual contexts. 7 00:00:49,440 --> 00:00:50,600 Ashley: That’s right, Evan. 8 00:00:50,600 --> 00:01:02,130 This task is vital for generative world modeling, where a history-conditioned generator isn’t just expected to provide plausible successive chunks, but also to model visual scene evolution over time. 9 00:01:02,130 --> 00:01:13,440 Here’s the crux of the problem: while generating video chunks conditioned on the preceding context can extend a video, it doesn't necessarily mean the model has learned how the scene should evolve. 10 00:01:13,488 --> 00:01:14,528 Evan: Exactly. 11 00:01:14,528 --> 00:01:22,088 A critical issue here is the reliance on teacher forcing, where the current chunk is denoised based on the ground-truth history. 12 00:01:22,088 --> 00:01:29,398 Due to strong temporal redundancy in videos, neighboring chunks often look very similar and share motion patterns. 13 00:01:29,398 --> 00:01:38,728 This redundancy leads the model to rely heavily on local visual continuity, creating shortcuts that reduce the need to capture the dynamic evolution of scenes. 14 00:01:38,784 --> 00:01:44,444 Ashley: This observation led the authors to seek supervision beyond immediate local visual continuity. 15 00:01:44,444 --> 00:01:54,644 They looked farther into the future for supervision signals that require the model to capture temporal progression, enhancing its ability to generate more temporally coherent sequences. 16 00:01:54,696 --> 00:02:00,156 Evan: What they introduced is a method called Future Chunk Prediction, or FCP. 17 00:02:00,156 --> 00:02:10,336 This method complements the denoising of the target chunk with an auxiliary objective that aims to predict a chunk further into the future without seeing the intervening context. 18 00:02:10,392 --> 00:02:20,002 Ashley: FCP draws supervision directly from the training videos, without needing additional pretrained visual encoders, and it doesn’t alter the autoregressive inference process. 19 00:02:20,002 --> 00:02:30,972 This approach places greater pressure on learned representations to capture information relevant to visual evolution, which is fundamentally important for effective temporal rollout in video generation. 20 00:02:31,032 --> 00:02:40,372 Evan: To evaluate the model’s ability to improve the representation of scene evolution, the paper focuses on physical plausibility as a key evaluation dimension. 21 00:02:40,372 --> 00:02:50,112 FCP showcases improved physical plausibility over standard autoregressive video generation methods even under matched data exposure and training budgets. 22 00:02:50,160 --> 00:02:57,720 Ashley: Interestingly, visualizations in the paper suggest that FCP fosters a more structured representation space. 23 00:02:57,720 --> 00:03:06,160 This spatiotemporal organization in the learned representations potentially makes them more suitable for building autoregressive world models. 24 00:03:06,216 --> 00:03:09,316 Evan: This concludes the Introduction section of the paper. 25 00:03:10,562 --> 00:03:15,732 Evan: Alright, Ashley, let’s dive deeper into the Method section of this fascinating paper. 26 00:03:15,732 --> 00:03:19,642 How do the authors go about implementing Future Chunk Prediction? 27 00:03:19,706 --> 00:03:20,746 Ashley: Sure, Evan. 28 00:03:20,746 --> 00:03:33,336 The general formulation of Future Chunk Prediction, or FCP, begins with the standard teacher-forced training, where each target chunk is denoised conditioned on its temporally contiguous ground-truth history. 29 00:03:33,336 --> 00:03:43,786 Then, FCP introduces an auxiliary task: predicting temporally more distant chunks from the same video without exposing the model to the intervening visual context. 30 00:03:43,850 --> 00:03:53,110 Evan: Can you explain why this prediction task is designed to use the learned intermediate representation to predict a farther target? 31 00:03:53,162 --> 00:04:01,882 Ashley: The idea is to discourage reliance on local temporal continuity, which often serves as a shortcut due to high temporal redundancy in videos. 32 00:04:01,882 --> 00:04:13,362 By excluding intervening chunks from the visual context, the task places additional training pressure on the learned representations, encouraging them to capture information relevant to visual evolution. 33 00:04:13,448 --> 00:04:14,358 Evan: Got it. 34 00:04:14,358 --> 00:04:19,498 So, is there a specific architecture or setup for implementing this FCP? 35 00:04:19,562 --> 00:04:20,782 Ashley: Yes, there is. 36 00:04:20,782 --> 00:04:25,852 The authors implement FCP using an auxiliary predictor with dedicated parameters. 37 00:04:25,852 --> 00:04:33,782 The predictor is conditioned on the generator’s intermediate hidden states, allowing the auxiliary loss to backpropagate through these hidden states. 38 00:04:33,782 --> 00:04:40,922 This design helps regularize the representations to capture visual evolution and enhance the autoregressive rollout. 39 00:04:40,970 --> 00:04:42,200 Evan: That's interesting. 40 00:04:42,200 --> 00:04:48,240 It seems crucial to ensure the auxiliary predictor does not interfere with the main generation process. 41 00:04:48,240 --> 00:04:49,990 How does this work in practice? 42 00:04:50,042 --> 00:04:58,602 Ashley: In practice, the auxiliary predictor operates only during the training phase and leaves the original autoregressive inference process unchanged. 43 00:04:58,602 --> 00:05:08,182 The model predicts a future chunk, indexed as x k+∆, which is temporally offset from the current chunk, ignoring the intermediate chunks. 44 00:05:08,294 --> 00:05:12,394 Evan: So, how does this translate into tangible steps during the training? 45 00:05:12,458 --> 00:05:18,968 Ashley: To implement this, the authors extract intermediate features from both clean and noise-perturbed states of a chunk. 46 00:05:18,968 --> 00:05:27,848 Each future offset has a separate predictor, conditioned directly on the intermediate representations of the current chunk, not chained through preceding predictors. 47 00:05:27,848 --> 00:05:33,058 This prevents inter-offset information flow and discourages the creation of shortcuts. 48 00:05:33,122 --> 00:05:35,072 Evan: That’s a neat strategy. 49 00:05:35,072 --> 00:05:39,162 What kind of layers or modules does the architecture consist of? 50 00:05:39,218 --> 00:05:46,468 Ashley: The architecture for the autoregressive generator is based on a stack of diffusion transformer blocks with chunk-causal attention. 51 00:05:46,468 --> 00:05:51,178 The auxiliary predictor adopts a lightweight counterpart of this DiT architecture. 52 00:05:51,178 --> 00:05:58,188 Specifically, source representations are extracted at different semantic levels and are then fused using an MLP. 53 00:05:58,188 --> 00:06:04,518 This ensures fine-grained conditioning across spatiotemporal positions using AdaLN modulation. 54 00:06:04,562 --> 00:06:09,922 Evan: Did the authors evaluate different configurations or settings for implementing FCP? 55 00:06:09,986 --> 00:06:10,886 Ashley: They did. 56 00:06:10,886 --> 00:06:21,756 The authors compared various feature-fusion strategies, such as directly concatenating source and future tokens within self-attention versus token-wise AdaLN modulation. 57 00:06:21,756 --> 00:06:28,286 They found that token-wise AdaLN was more effective in capturing predictive information without creating shortcuts. 58 00:06:28,346 --> 00:06:34,146 Evan: How did they ensure the efficiency of the training process when integrating FCP? 59 00:06:34,232 --> 00:06:35,352 Ashley: Good question. 60 00:06:35,352 --> 00:06:46,662 Each predictor operates independently, which allows parallel processing across GPUs, rather than having a chained prediction where intermediate future information may leak to subsequent predictors. 61 00:06:46,662 --> 00:06:53,482 This parallel design supports non-consecutive future offsets, making the training more efficient and scalable. 62 00:06:53,546 --> 00:06:57,526 Evan: So, how exactly is the model optimized during training? 63 00:06:57,578 --> 00:07:04,258 Ashley: The model is jointly optimized with both the autoregressive generator and auxiliary predictors using flow matching. 64 00:07:04,258 --> 00:07:14,428 For each offset, the auxiliary predictor follows the flow-matching objective, conditioned on clean or noisy source representations, mixed into the generator’s hidden states. 65 00:07:14,428 --> 00:07:20,118 The joint objective ensures that the auxiliary losses backpropagate through the learned representations. 66 00:07:20,232 --> 00:07:22,282 Evan: And what happens during inference? 67 00:07:22,376 --> 00:07:30,636 Ashley: During inference, all FCP-specific modules are removed, and the autoregressive generator along with its rollout procedure remains unchanged. 68 00:07:30,636 --> 00:07:36,626 This means there is no additional overhead at inference, giving it an advantage in terms of maintaining efficiency. 69 00:07:36,674 --> 00:07:40,224 Evan: The paper suggests improvements in physical plausibility. 70 00:07:40,224 --> 00:07:45,334 Are there any specific benchmarks or datasets the authors used to evaluate their method? 71 00:07:45,386 --> 00:07:52,846 Ashley: The authors used Physics-IQ Verified, WorldModelBench, PAI-Bench-G, and PhyGround for evaluation. 72 00:07:52,846 --> 00:07:59,906 These benchmarks focus on different aspects of temporal dynamics, such as physical plausibility and real-world recordings. 73 00:07:59,906 --> 00:08:12,846 The evaluations were performed on diverse datasets, including internally sourced videos and a publicly available dataset called WISA-80K, primarily comprising general-purpose real-world videos. 74 00:08:12,950 --> 00:08:17,750 Evan: Did the authors achieve significant improvements on these benchmarks? 75 00:08:17,810 --> 00:08:18,980 Ashley: Yes, they did. 76 00:08:18,980 --> 00:08:27,460 Particularly on Physics-IQ Verified, the results showed consistent gains in IQ scores, indicating improvements in physical plausibility. 77 00:08:27,460 --> 00:08:36,950 The same trend was observed across other benchmarks, demonstrating that FCP enhances the model’s ability to predict and generate physically plausible visual evolution. 78 00:08:37,010 --> 00:08:38,460 Evan: That's impressive. 79 00:08:38,460 --> 00:08:44,580 It seems like FCP not only improves generation quality but also makes the process more efficient. 80 00:08:44,580 --> 00:08:47,630 This wraps up our discussion on the Method section. 81 00:08:48,891 --> 00:08:58,351 Evan: Moving on to the experiments and results section of the paper, Ashley, how do the authors substantiate their claims on Future Chunk Prediction? 82 00:08:58,425 --> 00:09:04,695 Ashley: Evan, the authors conducted extensive experiments to evaluate the effectiveness of their proposed method. 83 00:09:04,695 --> 00:09:11,765 They zeroed in on physical plausibility, an important dimension for assessing temporal dynamics in video generation. 84 00:09:11,765 --> 00:09:15,795 The principal benchmark they used is Physics-IQ Verified. 85 00:09:15,873 --> 00:09:22,003 Evan: I'm curious about the specifics—how did they measure physical plausibility, and what were their findings? 86 00:09:22,059 --> 00:09:28,689 Ashley: They focused on metrics such as similarity in space and time, overall quality, and physical consistency. 87 00:09:28,689 --> 00:09:39,789 For instance, in the Physics-IQ Verified benchmark, they evaluated both image-to-video, or I2V, and video-to-video, or V2V, generation scenarios. 88 00:09:39,789 --> 00:09:47,399 To offer broader context, they also assessed their method on WorldModelBench, PAI-Bench-G, and PhyGround. 89 00:09:47,451 --> 00:09:48,691 Evan: Interesting! 90 00:09:48,691 --> 00:09:54,051 How did Future Chunk Prediction, or FCP, fare against the baseline models? 91 00:09:54,099 --> 00:09:59,169 Ashley: On the Physics-IQ Verified benchmark, their method showed clear improvements. 92 00:09:59,169 --> 00:10:10,259 For instance, with a 1.3 billion parameter model, the IQ score increased from 27.6 under standard teacher forcing to 30.9 with FCP. 93 00:10:10,259 --> 00:10:21,479 For their 5 billion parameter model, it improved from 30.4 to 33.8 for I2V and from 40.6 to 44.4 for V2V. 94 00:10:21,531 --> 00:10:23,421 Evan: That’s significant! 95 00:10:23,421 --> 00:10:27,131 Were these improvements consistent across other benchmarks too? 96 00:10:27,195 --> 00:10:32,855 Ashley: Yes, even across other physics-related benchmarks, they observed better performance with FCP. 97 00:10:32,855 --> 00:10:39,155 On WorldModelBench, they saw increases in metrics like Instruction, Physical Laws, and Common Sense. 98 00:10:39,155 --> 00:10:46,075 Similarly, on PAI-Bench-G, they witnessed improvements in various domains, including physics and common sense. 99 00:10:46,075 --> 00:10:51,995 And on PhyGround, their method yielded higher scores in both general and physics-specific metrics. 100 00:10:52,059 --> 00:10:57,379 Evan: It looks like FCP outperforms standard autoregressive models significantly. 101 00:10:57,379 --> 00:11:02,479 How did the authors ensure these results were not just due to more computational resources? 102 00:11:02,523 --> 00:11:03,593 Ashley: Good point. 103 00:11:03,593 --> 00:11:07,823 The authors ran controlled experiments to isolate the effects of their method. 104 00:11:07,823 --> 00:11:14,843 For instance, they matched the data exposure and training budgets between the baseline and FCP-enhanced models. 105 00:11:14,843 --> 00:11:21,283 This includes keeping all training settings consistent, such as resolution, frame rate, and batch sizes. 106 00:11:21,399 --> 00:11:22,119 Evan: Got it. 107 00:11:22,119 --> 00:11:24,319 What about qualitative assessments? 108 00:11:24,319 --> 00:11:28,299 Did the authors provide any visual examples to support their findings? 109 00:11:28,347 --> 00:11:36,187 Ashley: They included qualitative comparisons showing that FCP often predicts physically plausible outcomes where teacher forcing fails. 110 00:11:36,187 --> 00:11:45,967 For example, in a domino toppling sequence, FCP correctly predicted the direction of the toppling dominoes, while teacher forcing produced an anomalous collapse. 111 00:11:46,091 --> 00:11:48,471 Evan: Oh, that's quite telling. 112 00:11:48,471 --> 00:11:54,691 Did they also provide an analysis of how FCP affects intermediate representations? 113 00:11:54,747 --> 00:11:56,217 Ashley: Yes, they did. 114 00:11:56,217 --> 00:12:04,307 Using Spatial Feature Similarity Maps and Spatiotemporal Feature Similarity Maps, they visualized the intermediate representations. 115 00:12:04,307 --> 00:12:16,187 FCP-trained models produced more selective similarity responses aligned with the queried objects, suggesting that these models capture clearer spatial structures and meaningful temporal correspondences. 116 00:12:16,281 --> 00:12:18,211 Evan: Very insightful. 117 00:12:18,211 --> 00:12:24,871 Did the authors perform any ablation studies to understand the contribution of individual components of their method? 118 00:12:24,915 --> 00:12:25,775 Ashley: They did. 119 00:12:25,775 --> 00:12:34,585 The ablation studies showed that regularizing both clean-context and noisy-target representations significantly contributed to performance improvements. 120 00:12:34,585 --> 00:12:45,715 Moreover, replacing token-wise AdaLN modulation with simpler feature fusion strategies reduced efficacy, which underlines the importance of their detailed architectural choices. 121 00:12:45,771 --> 00:12:50,031 Evan: What about the efficiency of their method when it comes to training and inference? 122 00:12:50,091 --> 00:12:55,171 Ashley: Their experimental results indicated that their method is both effective and efficient. 123 00:12:55,171 --> 00:13:01,051 The parallel prediction strategy allows better scaling and efficient use of computational resources. 124 00:13:01,051 --> 00:13:07,251 During inference, all FCP-specific modules are removed, ensuring no additional overhead. 125 00:13:07,299 --> 00:13:08,779 Evan: That’s really compelling. 126 00:13:08,779 --> 00:13:16,599 Thus, the experiments clearly suggest that FCP not only enhances the quality of video generation but does so efficiently. 127 00:13:16,689 --> 00:13:17,789 Ashley: Definitely. 128 00:13:17,789 --> 00:13:20,789 This wraps up the Experiment and Results section. 129 00:13:20,789 --> 00:13:23,079 Shall we move to the conclusion next? 130 00:13:24,411 --> 00:13:28,041 Evan: Ashley, let's explore the Related Work section. 131 00:13:28,041 --> 00:13:32,841 What do the authors highlight in terms of prior research that relates to their work? 132 00:13:32,885 --> 00:13:43,865 Ashley: The authors provide an extensive overview of the existing literature, which mainly falls into two categories: autoregressive video diffusion and representation alignment for generation. 133 00:13:44,015 --> 00:13:47,185 Evan: I've seen a lot about video diffusion models lately. 134 00:13:47,185 --> 00:13:51,005 How do these models fit into the narrative provided by the authors? 135 00:13:51,053 --> 00:13:57,413 Ashley: Conventional video diffusion models typically generate fixed-duration clips in a single denoising process. 136 00:13:57,413 --> 00:14:06,413 In contrast, autoregressive video diffusion models extend videos chunk by chunk, conditioning each new segment on the preceding visual context. 137 00:14:06,413 --> 00:14:14,613 This chunk-wise formulation supports streaming generation and is a key pathway for interactive visual simulation and generative world modeling. 138 00:14:14,669 --> 00:14:20,309 Evan: Right, generating videos in chunks seems quite effective for continuous visual sequences. 139 00:14:20,309 --> 00:14:25,429 How do past works attempt to ensure quality in these chunk-wise generated videos? 140 00:14:25,493 --> 00:14:36,183 Ashley: Prior works have addressed rollout quality by adapting bidirectional models into causal few-step generators through techniques like distribution matching or adversarial training. 141 00:14:36,183 --> 00:14:44,143 For instance, some models are exposed to self-generated histories or extended student rollouts to improve temporal coherence. 142 00:14:44,143 --> 00:14:50,213 These approaches also consider how teacher supervision can be made compatible with student generation. 143 00:14:50,321 --> 00:14:51,361 Evan: Interesting. 144 00:14:51,361 --> 00:14:55,281 So how does this paper’s approach differ from these existing methods? 145 00:14:55,325 --> 00:15:04,615 Ashley: The authors propose future-chunk supervision to complement teacher-forced training without prescribing rollout distillation or inference-time context management. 146 00:15:04,615 --> 00:15:15,885 Essentially, their approach introduces additional predictive supervision to teacher-forced training, rather than managing the historical context through active windows, cached states, or global anchors. 147 00:15:15,941 --> 00:15:17,581 Evan: That's a nuanced difference. 148 00:15:17,581 --> 00:15:21,711 Now onto the second part, representation alignment for generation. 149 00:15:21,711 --> 00:15:24,961 What is its role and how have previous methods utilized it? 150 00:15:25,013 --> 00:15:35,283 Ashley: Representation alignment aims to synchronize intermediate diffusion features with vision foundation models, accelerating training and enhancing the quality of generative outputs. 151 00:15:35,283 --> 00:15:45,453 Previous works in video extension tasks have employed alignment strategies like cross-frame and multi-feature alignment or relational supervision targeting physical plausibility. 152 00:15:45,539 --> 00:15:52,029 Evan: Aligning features with pretrained vision models sounds beneficial, but what limitations have these methods faced? 153 00:15:52,085 --> 00:15:58,895 Ashley: These methods often depend on external pretrained encoders, which can complicate scaling to large video corpora. 154 00:15:58,895 --> 00:16:10,305 The authors’ approach, by contrast, derives supervision from future chunks in the training videos, avoiding the need for additional pretrained visual encoders and simplifying the scaling process. 155 00:16:10,429 --> 00:16:16,789 Evan: Did the authors draw comparisons between their method and other related approaches within their experiments? 156 00:16:16,853 --> 00:16:23,683 Ashley: The authors acknowledge the similarity of their approach to Next Forcing, which also studies future-chunk prediction. 157 00:16:23,683 --> 00:16:32,333 While developed independently and concurrently, the two methods share common ground in how they target future-chunk predictions to enhance generative models. 158 00:16:32,381 --> 00:16:37,311 Evan: It sounds like positioning their work within the broader context was quite important. 159 00:16:37,311 --> 00:16:40,041 Did they review any other similar advancements? 160 00:16:40,085 --> 00:16:49,715 Ashley: Yes, they also consider approaches like Rolling Forcing and active window management, which balance long-range consistency against computational and memory costs. 161 00:16:49,715 --> 00:16:58,485 However, their method focuses strictly on enhancing teacher-forced training with future-prediction supervision, distinguishing it from these other methods. 162 00:16:58,541 --> 00:17:04,191 Evan: It’s fascinating how they build on and diverge from existing work to push the field forward. 163 00:17:04,191 --> 00:17:08,541 Are there any other benchmarks or analysis methods referenced in the Related Work? 164 00:17:08,597 --> 00:17:27,307 Ashley: Indeed, they mention benchmarks like Physics-IQ Verified, which assesses generated continuations against real-world recordings without relying on vision-language models, and WorldModelBench, PAI-Bench-G, and PhyGround for broader evaluations of temporal dynamics and physical plausibility. 165 00:17:27,307 --> 00:17:34,837 These references underscore the comprehensiveness with which they evaluate their method against previous models and other benchmarks. 166 00:17:34,901 --> 00:17:36,461 Evan: That sounds thorough. 167 00:17:36,461 --> 00:17:43,841 Clearly, grounding their research in a strong background of prior work has helped underline the significance of their contributions. 168 00:17:43,901 --> 00:17:45,101 Ashley: Exactly, Evan. 169 00:17:45,101 --> 00:17:48,361 This wraps up our discussion on the Related Work section. 170 00:17:49,674 --> 00:17:55,074 Evan: Alright, Ashley, let's wrap up by summarizing the key takeaways from this paper. 171 00:17:55,134 --> 00:18:02,944 Ashley: This paper introduces Future Chunk Prediction, or FCP, as a method to improve autoregressive video generation. 172 00:18:02,944 --> 00:18:12,734 The core idea is to supplement standard teacher-forced training with an auxiliary task that predicts future video chunks without relying on intervening visual context. 173 00:18:12,734 --> 00:18:17,054 This encourages the model to capture temporal evolution more effectively. 174 00:18:17,118 --> 00:18:18,238 Evan: Exactly. 175 00:18:18,238 --> 00:18:29,168 By placing greater pressure on learned representations to encapsulate information relevant to visual evolution, FCP enhances the physical plausibility of generated videos. 176 00:18:29,168 --> 00:18:41,438 This was evident in their results, where their models outperformed standard approaches across multiple benchmarks like Physics-IQ Verified, WorldModelBench, PAI-Bench-G, and PhyGround. 177 00:18:41,532 --> 00:18:50,182 Ashley: Another notable contribution is that FCP doesn't require additional pretrained visual encoders or alter the autoregressive inference process. 178 00:18:50,182 --> 00:18:53,202 This makes it both an effective and efficient enhancement. 179 00:18:53,202 --> 00:19:03,022 Plus, their evaluations showed that FCP-trained models foster a more structured representation space, which is valuable for building autoregressive world models. 180 00:19:03,078 --> 00:19:12,118 Evan: The authors also conducted extensive ablation studies, ensuring that each component of their method contributed significantly to its overall performance. 181 00:19:12,118 --> 00:19:20,578 They demonstrated that their parallel prediction strategy was particularly effective in maintaining efficiency during training and inference. 182 00:19:20,622 --> 00:19:31,082 Ashley: In conclusion, FCP not only improves the quality of video generation by enhancing temporal consistency and plausibility but does so while maintaining efficiency. 183 00:19:31,082 --> 00:19:36,042 It’s a compelling approach for advancing the field of autoregressive video generation. 184 00:19:36,102 --> 00:19:47,382 Evan: Thank you for joining us on today’s episode of Daily Paper Cast, where we delved into 'Looking Ahead: Modeling Future Evolution in Autoregressive Video Generation'. 185 00:19:47,430 --> 00:19:52,480 Ashley: We appreciate you tuning in, and we hope you found this discussion insightful. 186 00:19:52,480 --> 00:19:57,650 Be sure to join us again for more deep dives into cutting-edge AI research. 187 00:19:57,752 --> 00:20:02,702 Evan: Until next time, stay curious and keep exploring the world of AI. 188 00:20:02,702 --> 00:20:03,782 Goodbye!