1
00:00:03,000 --> 00:00:05,400
Evan: Welcome to Daily Paper Cast!

2
00:00:05,448 --> 00:00:13,058
Ashley: Today's paper is selected from the Hugging Face daily paper list of November 18, 2026, with 40 upvotes.

3
00:00:13,104 --> 00:00:21,444
Evan: The paper we're delving into is titled 'Weight-Space Programming Doesn’t Need Transformers' by Hyunjong Ok and Jaeho Lee.

4
00:00:21,504 --> 00:00:31,914
Ashley: Our corresponding author, Jaeho Lee, and his colleague Hyunjong Ok, are affiliated with the Pohang University of Science and Technology, or POSTECH.

5
00:00:31,968 --> 00:00:33,888
Evan: Let's dive into the Introduction.

6
00:00:33,888 --> 00:00:43,158
This paper focuses on neural architectures, specifically Transformers, which have been widely recognized for their performance with gradient descent optimization.

7
00:00:43,224 --> 00:00:44,544
Ashley: Exactly, Evan.

8
00:00:44,544 --> 00:00:53,174
Transformers have become standard in many AI and NLP tasks due to their efficient training on modern GPUs and remarkable accuracy.

9
00:00:53,174 --> 00:00:58,934
However, their inherent design might not be ideal for every approach in neural network implementation.

10
00:00:58,992 --> 00:01:09,632
Evan: The authors present an intriguing concept called weight-space programming, where a large language model, or LLM, writes target weights by hand.

11
00:01:09,632 --> 00:01:17,702
This method does not rely on gradient-based training or large datasets, which brings us to the research question posed by the authors.

12
00:01:17,760 --> 00:01:22,370
Ashley: Should weight-space programming continue to use the Transformer architecture?

13
00:01:22,370 --> 00:01:27,140
To investigate this, the authors conducted experiments with minimal guidance.

14
00:01:27,140 --> 00:01:34,380
The setup includes providing each LLM with only ten labeled questions and an isolated sandbox environment.

15
00:01:34,440 --> 00:01:48,500
Evan: The experiments explore two main regimes: one where the LLM performs weight-space programming on diverse released architectures like Transformers, state-space models, and RNNs from random initialization.

16
00:01:48,500 --> 00:01:53,350
And another where the LLM is allowed to design the target architecture from scratch.

17
00:01:53,400 --> 00:01:59,890
Ashley: Interestingly, when the LLMs were free to choose the structure, they often did not opt for Transformers.

18
00:01:59,890 --> 00:02:06,990
Instead, they created models that compile plain-text facts into a sparse table queried by a fixed procedure.

19
00:02:06,990 --> 00:02:18,440
This novel approach achieved up to 54.4% test accuracy on ARC multiple-choice science questions, outperforming the most accurate Transformer designed by the same LLM.

20
00:02:18,504 --> 00:02:20,424
Evan: That's fascinating, Ashley.

21
00:02:20,424 --> 00:02:28,064
The paper also proposes an architecture based on associative memory, which is far more efficient in terms of parameters and speed.

22
00:02:28,128 --> 00:02:41,028
Ashley: Indeed, this associative-memory-based model uses about 1,800 times fewer parameters and can answer up to 17 times faster on GPUs compared to a same-sized Transformer.

23
00:02:41,028 --> 00:02:48,658
The key difference is that predictions in these models trace back to specific rows, making editing easier and more localized.

24
00:02:48,720 --> 00:02:54,520
Evan: The authors tested their associative-memory-based model across various tasks and benchmarks.

25
00:02:54,520 --> 00:03:02,030
Although pretrained LLMs generally remain more accurate, the proposed model provides a substantial efficiency advantage.

26
00:03:02,088 --> 00:03:16,078
Ashley: To summarize the Introduction, the paper argues that while standard neural architectures like Transformers are optimized for gradient descent training, weight-space programming by LLMs does not need to depend on such architectures.

27
00:03:16,078 --> 00:03:22,148
The authors have demonstrated that an associative-memory-based model can be a highly efficient alternative.

28
00:03:22,200 --> 00:03:25,030
Evan: That's the end of the Introduction section of this paper.

29
00:03:25,030 --> 00:03:27,050
Let's continue with the Methods next.

30
00:03:28,358 --> 00:03:31,098
Evan: Let's delve into the methods employed in this paper.

31
00:03:31,154 --> 00:03:37,534
Ashley: Weight-space programming redefines how models acquire knowledge by bypassing traditional gradient descent training.

32
00:03:37,534 --> 00:03:45,294
Instead, a large language model, or LLM, writes the weights directly, line by line, without data fitting.

33
00:03:45,294 --> 00:03:53,354
This approach challenges the notion of using the Transformer architecture due to the intrinsic differences in how weight-space programming functions.

34
00:03:53,462 --> 00:03:54,872
Evan: Exactly, Ashley.

35
00:03:54,872 --> 00:04:04,062
The authors designed an experimental setup to evaluate how effectively LLMs can handle weight-space programming across different architectures.

36
00:04:04,062 --> 00:04:11,442
They provided minimal guidance to the LLMs, including only ten labeled questions and an isolated sandbox environment.

37
00:04:11,498 --> 00:04:16,698
Ashley: The experimental evaluation of weight-space programming was conducted across two regimes.

38
00:04:16,698 --> 00:04:29,198
First, the LLMs performed weight-space programming on diverse released architectures such as Transformers, state-space models, and recurrent neural networks, starting from random initialization.

39
00:04:29,198 --> 00:04:34,608
Second, the LLMs were given the freedom to design the target architecture from scratch.

40
00:04:34,658 --> 00:04:39,078
Evan: So how did the authors set up and execute these experiments?

41
00:04:39,122 --> 00:04:50,702
Ashley: In the first regime of fixed architectures, the LLMs wrote weights over a small two-layer baseline without starting weights or random initialization of a released neural architecture.

42
00:04:50,702 --> 00:05:03,072
Specifically, the given Transformer architecture consisted of two layers, with each option in the multiple-choice question forming a sequence made of question word IDs followed by option word IDs.

43
00:05:03,182 --> 00:05:08,262
Evan: And for the second regime where the LLMs designed the architecture themselves?

44
00:05:08,306 --> 00:05:18,126
Ashley: The LLMs defined the structure itself via an architecture file, a build script, and plain-text knowledge files compiled into tensors on a CPU.

45
00:05:18,126 --> 00:05:23,146
This setup led to models converging on sparse fact tables instead of dense layers.

46
00:05:23,146 --> 00:05:36,406
The fact table holds written facts in plain-text sentences, which are compiled into rows of the table, while a fixed lookup procedure reads the table to match tokens from the question and candidate options to select the highest-scoring answer.

47
00:05:36,518 --> 00:05:37,948
Evan: Interesting approach.

48
00:05:37,948 --> 00:05:44,058
Can you explain more about how the LLMs wrote these weights and how the associative memory model operates?

49
00:05:44,114 --> 00:05:45,124
Ashley: Certainly.

50
00:05:45,124 --> 00:05:55,524
In weight-space programming, the LLMs work in a loop, creating practice questions, running the model, inspecting failures, and adding a rule or fact row.

51
00:05:55,524 --> 00:06:02,914
This loop ensures that each prediction from the model traces back to specific written lines, enabling precise debugging.

52
00:06:02,914 --> 00:06:12,034
This method, called associative memory, starkly contrasts with the gradient descent method, where weights are optimized jointly through backpropagation.

53
00:06:12,158 --> 00:06:21,448
Evan: So, each wrong answer points to a specific row in the table, making it easy for the LLM to find and rewrite a mistaken line, correct?

54
00:06:21,506 --> 00:06:22,486
Ashley: Exactly.

55
00:06:22,486 --> 00:06:25,906
That's one of the major advantages of weight-space programming.

56
00:06:25,906 --> 00:06:35,256
When the LLM-designed models were tested, retaining only the single best-matching row per option nearly matched the full-model accuracy.

57
00:06:35,256 --> 00:06:41,186
This demonstrates that most of the model's accuracy comes from the best-written fact for each option.

58
00:06:41,314 --> 00:06:44,504
Evan: Does this method also help in speeding up inference?

59
00:06:44,570 --> 00:06:45,520
Ashley: Yes.

60
00:06:45,520 --> 00:06:53,760
Normally, inference in traditional models would slow down as the number of stored facts increased because each fact needs to be scored for every question.

61
00:06:53,760 --> 00:07:06,170
However, in the proposed associative-memory-based model, the performance is optimized by reading a fixed number of rows per question from the pair table and computing a single product with the concept table.

62
00:07:06,218 --> 00:07:09,388
Evan: What about efficiency compared to standard Transformers?

63
00:07:09,434 --> 00:07:15,884
Ashley: The associative-memory-based model retains the written table while optimizing the number of rows read per question.

64
00:07:15,884 --> 00:07:27,684
As a result, it stays within about three percentage points of the LLM-designed models' accuracy while answering up to 17 times faster on a GPU than a same-sized Transformer.

65
00:07:27,684 --> 00:07:33,374
This performance boost comes from the architecture's efficiency in how memory is addressed and read.

66
00:07:33,434 --> 00:07:36,284
Evan: How about the generalization of these findings?

67
00:07:36,284 --> 00:07:39,854
Did the authors test outside multiple-choice science questions?

68
00:07:39,914 --> 00:07:50,664
Ashley: Yes, they extended their evaluation to other domains, including pretrained weights, mathematics questions, open-ended text generation tasks, and even vision tasks.

69
00:07:50,664 --> 00:08:00,354
The associative memory model showed robust performance, maintaining accuracy close to the full model while significantly reducing inference time on GPUs.

70
00:08:00,410 --> 00:08:02,860
Evan: It seems quite comprehensive.

71
00:08:02,860 --> 00:08:06,840
What kind of benchmarks and tasks were included in the evaluation?

72
00:08:06,890 --> 00:08:16,810
Ashley: The authors tested their model on ARC Easy and ARC Challenge test sets, consisting of a total of 3,530 multiple-choice science questions.

73
00:08:16,810 --> 00:08:25,690
They followed rigid protocols to ensure the isolation of the LLMs from any external data other than the ten labeled training examples provided.

74
00:08:25,690 --> 00:08:39,400
Besides, other benchmarks included pretrained weights and tasks where models generated answers, confirming that while pretrained LLMs remain more accurate, the associative memory model offers considerable efficiency advantages.

75
00:08:39,518 --> 00:08:46,318
Evan: It's fascinating to see how weight-space programming reimagines the process of model creation and optimization.

76
00:08:46,318 --> 00:08:51,598
This method indeed brings significant benefits in terms of speed and parameter efficiency.

77
00:08:51,680 --> 00:08:52,430
Ashley: Evan.

78
00:08:52,430 --> 00:09:08,150
To conclude the method section, the paper showcases a pioneering approach where the associative-memory model design radically differs from traditional dense weight architectures, and the results affirm its potential long-term impact on efficient model design and operation.

79
00:09:08,210 --> 00:09:11,800
Evan: That wraps up the Methods section analysis for this paper.

80
00:09:11,800 --> 00:09:18,490
Next, we'll explore the experiments conducted and the results obtained by the authors in the following segment.

81
00:09:19,755 --> 00:09:24,255
Evan: Now let's move on to the Experiments and Results section of this fascinating paper.

82
00:09:24,315 --> 00:09:34,295
Ashley: The authors conducted a series of experiments to assess the performance and efficiency of weight-space programming compared to traditional models and other architectures.

83
00:09:34,295 --> 00:09:39,335
These experiments were primarily focused on ARC multiple-choice science questions.

84
00:09:39,387 --> 00:09:50,457
Evan: They implemented extensive evaluations, testing different architectural setups and configurations, particularly comparing LLM-designed models to fixed architectures.

85
00:09:50,523 --> 00:09:51,683
Ashley: Exactly.

86
00:09:51,703 --> 00:10:02,763
If we look at their results, the LLM-designed models consistently outperformed fixed architectures, achieving higher ARC test accuracy with significantly fewer parameters.

87
00:10:02,783 --> 00:10:13,073
For instance, the LLM-designed model reached up to 54.4% test accuracy while using about 1,800 times fewer parameters.

88
00:10:13,131 --> 00:10:14,761
Evan: That's quite impressive.

89
00:10:14,761 --> 00:10:18,021
How did the associative-memory-based models measure up?

90
00:10:18,075 --> 00:10:24,525
Ashley: The associative-memory-based models retained the efficiency advantage of the LLM-designed models.

91
00:10:24,525 --> 00:10:34,445
In the first-generation associative-memory model, the accuracy was around 49.6%, and in the fifth generation, it went up to 52.4%.

92
00:10:34,445 --> 00:10:40,425
These models were up to 17 times faster on GPUs compared to same-sized Transformers.

93
00:10:40,491 --> 00:10:47,331
Evan: Speed and efficiency are critical factors, especially when dealing with massive datasets and real-time applications.

94
00:10:47,331 --> 00:10:49,541
How does this compare to traditional methods?

95
00:10:49,587 --> 00:10:57,587
Ashley: In traditional gradient descent models, the inference time increases with the number of facts as each fact needs to be scored for each question.

96
00:10:57,587 --> 00:11:07,017
However, the proposed associative-memory models read a fixed number of rows per question, maintaining efficiency irrespective of the number of stored facts.

97
00:11:07,143 --> 00:11:12,583
Evan: Did the authors run any comparisons with other benchmarks or pretrained weights?

98
00:11:12,627 --> 00:11:23,457
Ashley: Yes, they extended their evaluation beyond multiple-choice science questions to include pretrained weights, math questions, text generation tasks, and even vision tasks.

99
00:11:23,477 --> 00:11:31,287
Remarkably, while pretrained LLMs remained more accurate overall, the associative-memory model excelled in efficiency.

100
00:11:31,287 --> 00:11:40,657
For instance, on ARC, the associative-memory model demonstrated competitive accuracy while drastically reducing inference time compared to other models.

101
00:11:40,767 --> 00:11:46,157
Evan: Let's talk about their findings on random initialization versus pretrained weights.

102
00:11:46,157 --> 00:11:47,487
What did they uncover?

103
00:11:47,547 --> 00:11:56,197
Ashley: When starting from random initialization, LLM-written models consistently outperformed those trained through gradient descent.

104
00:11:56,197 --> 00:12:05,567
Surprisingly, some pretrained models edited by the LLMs closely matched their original accuracy, with minimal alterations to their weights.

105
00:12:05,619 --> 00:12:07,019
Evan: That's intriguing.

106
00:12:07,019 --> 00:12:14,589
So, pretrained models can almost retain their accuracy with very few changes under weight-space programming?

107
00:12:14,643 --> 00:12:19,183
Ashley: This indicates the flexibility and robustness of weight-space programming.

108
00:12:19,183 --> 00:12:30,073
In cases where the pretrained models were close to random guessing, weight-space programming significantly improved their performance by overwriting weights with hand-written facts or rules.

109
00:12:30,213 --> 00:12:36,043
Evan: How did these models perform in other domains, like vision tasks or open-ended generation?

110
00:12:36,099 --> 00:12:51,069
Ashley: For vision tasks, particularly on CIFAR-10, the associative-memory-based models achieved impressive results when paired with tokens from a pre-trained DINOv2 encoder, reaching up to 96.1% accuracy.

111
00:12:51,069 --> 00:13:04,729
In text generation tasks like SQuAD 1.1 and SciQ, the LLM-designed models designed specific lookup tables and reading procedures, showing competitive results compared to standard trained models.

112
00:13:04,779 --> 00:13:12,529
Evan: It sounds like the authors conducted a broad set of evaluations to truly test the robustness and efficiency of weight-space programming.

113
00:13:12,579 --> 00:13:17,639
Ashley: Indeed, they also examined how their models handled edits and local changes.

114
00:13:17,639 --> 00:13:25,679
In weight-space programming, removing one written line usually altered fewer answers, maintaining the efficiency and accuracy of the model.

115
00:13:25,679 --> 00:13:31,929
This is a stark contrast to traditional models where edits can cause widespread changes throughout the model's weights.

116
00:13:31,995 --> 00:13:39,055
Evan: The findings highlight the potential of weight-space programming for various applications beyond just standard benchmarks.

117
00:13:39,099 --> 00:13:40,059
Ashley: Certainly.

118
00:13:40,059 --> 00:13:53,229
The flexibility, parameter efficiency, and speed advantages of the associative-memory-based models present a promising alternative to traditional neural network architectures, allowing for more adaptable and efficient model building.

119
00:13:53,283 --> 00:13:56,453
Evan: That concludes the Experiments section of this paper.

120
00:13:56,453 --> 00:14:01,443
Next, we'll discuss the implications and future directions detailed by the authors.

121
00:14:02,693 --> 00:14:10,273
Evan: Moving on, let's explore the Related Work section of the paper, which provides context and comparisons with existing research.

122
00:14:10,325 --> 00:14:11,585
Ashley: Certainly, Evan.

123
00:14:11,585 --> 00:14:17,985
The authors highlight several recent studies and methodologies relevant to weight-space programming and its alternatives.

124
00:14:17,985 --> 00:14:24,265
Let's start by examining research on directly reading and writing neural network weights without gradient training.

125
00:14:24,317 --> 00:14:25,257
Evan: Indeed.

126
00:14:25,257 --> 00:14:38,927
One pivotal work cited by the authors is by Takehi et al. (2026), showcasing how language models can natively read and write neural network weights, bypassing traditional training methods like gradient descent.

127
00:14:38,981 --> 00:14:47,531
Ashley: This work aligns closely with the concept of weight-space programming, where a language model generates the weights of a target model line by line.

128
00:14:47,531 --> 00:14:53,401
It's important as it introduces a paradigm shift in how models can be created and optimized.

129
00:14:53,453 --> 00:15:04,313
Evan: The paper also discusses earlier efforts to hand-construct weights without using a language model, focusing on computational experiments that demonstrate the feasibility of such approach.

130
00:15:04,313 --> 00:15:13,103
For instance, Weiss et al. (2021) explore transformer-like computations and hand-built mechanisms for improving interpretability.

131
00:15:13,157 --> 00:15:14,267
Ashley: That's right.

132
00:15:14,317 --> 00:15:23,347
Another significant related work mentioned is by Lindner et al. (2023), which involves compiling transformers for enhanced interpretability.

133
00:15:23,367 --> 00:15:36,247
Similarly, Dugan et al. (2025) delve into constructing efficient fact-storing MLPs for transformers, leveraging hand-constructed methods to embed facts directly into neural networks.

134
00:15:36,293 --> 00:15:42,563
Evan: There’s a connection here to various model modification techniques that don't rely solely on gradient descent.

135
00:15:42,563 --> 00:15:53,093
Methods like low-rank adaptation, discussed by Hu et al. (2022), allow modifying large language models through gradient training but with much smaller updates.

136
00:15:53,141 --> 00:15:59,911
Ashley: This contrasts with the paper’s methodology where modifications can be done line by line without gradient training.

137
00:15:59,991 --> 00:16:14,711
Additionally, Meng et al. (2022) and Meng et al. (2023) present methods for locating and editing factual associations within pretrained models, extending this concept to mass-editing memory in transformers.

138
00:16:14,805 --> 00:16:15,655
Evan: Interesting.

139
00:16:15,655 --> 00:16:22,475
So these techniques show various ways exploratory editing can adjust a model's knowledge without extensive retraining.

140
00:16:22,475 --> 00:16:26,165
How about approaches that search over programs and generate weights?

141
00:16:26,213 --> 00:16:31,863
Ashley: Yes, the paper cites works in neural architecture search and program synthesis.

142
00:16:31,863 --> 00:16:34,063
For example, Chen et al.

143
00:16:34,083 --> 00:16:41,643
(2023) use EvoPrompting to perform code-level neural architecture search to optimize network designs.

144
00:16:41,643 --> 00:16:52,653
Similarly, Nasir et al. (2024) add quality-diversity optimization to the search process, allowing for more diversified and high-quality model structures.

145
00:16:52,709 --> 00:16:58,929
Evan: Were there any notable references to fast weight programmers or other historical methods?

146
00:16:58,973 --> 00:17:00,033
Ashley: Indeed.

147
00:17:00,053 --> 00:17:10,563
The concept of fast weight programmers dates back to Schmidhuber's work in 1992, where a slower network trained by gradient descent outputs the weights of a faster network.

148
00:17:10,563 --> 00:17:13,063
Modern equivalents like Schlag et al.

149
00:17:13,083 --> 00:17:21,683
(2021) highlight how linear transformers can secretly operate as fast weight programmers, dynamically adjusting weights on the fly.

150
00:17:21,749 --> 00:17:30,899
Evan: It seems there’s a wealth of research exploring various ways to generate weights, whether through direct programming, program search, or adaptive methods.

151
00:17:30,965 --> 00:17:44,275
Ashley: These methodologies underpin the current study, showing that while traditional gradient-based optimization has dominated the field, there are viable alternatives that leverage direct weight manipulation and programmatic model design.

152
00:17:44,333 --> 00:17:49,703
Evan: In the realm of understanding transformer mechanics, the paper references Zhong et al.

153
00:17:49,723 --> 00:18:00,923
(2025) and Wang (2026), who interpret transformers as associative memories, an insight that supports weight-space programming's associative memory model.

154
00:18:00,989 --> 00:18:13,489
Ashley: Further, Cheng et al. (2026) argue that transformers simulate retrieval through computation rather than native lookup, comparing how other models directly integrate lookup primitives.

155
00:18:13,489 --> 00:18:20,799
This differentiation is key in understanding why weight-space programming does not necessarily rely on transformers.

156
00:18:20,921 --> 00:18:21,821
Evan: Interesting.

157
00:18:21,821 --> 00:18:29,301
It sounds like weight-space programming and associative memory models could reshape how we conceptualize and develop neural networks.

158
00:18:29,357 --> 00:18:38,857
Ashley: The paper's review of related works underscores a broad and varied set of methods that align with or provide alternative perspectives to weight-space programming.

159
00:18:38,857 --> 00:18:43,477
Each approach contributes to the evolving landscape of neural network design.

160
00:18:43,541 --> 00:18:46,851
Evan: That concludes our discussion on the Related Work section.

161
00:18:46,851 --> 00:18:55,021
The paper meticulously builds on and contrasts itself with these foundational studies to justify the innovations in weight-space programming.

162
00:18:56,286 --> 00:19:05,326
Evan: We're nearing the end of our exploration of 'Weight-Space Programming Doesn’t Need Transformers.' Let's summarize the key contributions and takeaways from this paper.

163
00:19:05,412 --> 00:19:06,362
Ashley: Evan.

164
00:19:06,362 --> 00:19:13,842
The authors present a compelling case for moving beyond traditional Transformer architectures when employing weight-space programming.

165
00:19:13,842 --> 00:19:22,022
By enabling large language models to write weights directly without gradient-based training, we can achieve significant efficiency gains.

166
00:19:22,086 --> 00:19:27,556
Evan: One of the main contributions is the development of associative-memory-based models.

167
00:19:27,556 --> 00:19:37,666
These models, created by the LLMs, compiled facts into sparse tables that were queried through fixed procedures, resulting in highly efficient architectures.

168
00:19:37,710 --> 00:19:38,960
Ashley: Precisely.

169
00:19:38,960 --> 00:19:53,520
The associative-memory-based model demonstrated up to 54.4% test accuracy on ARC multiple-choice science questions while using around 1,800 times fewer parameters compared to traditional Transformer models.

170
00:19:53,520 --> 00:19:58,180
Moreover, it operated up to 17 times faster on GPUs.

171
00:19:58,230 --> 00:20:10,280
Evan: The experiments revealed that LLM-designed models not only outperformed fixed architectures in terms of accuracy but also showed notable advantages in speed and parameter efficiency.

172
00:20:10,326 --> 00:20:14,246
Ashley: The flexibility of weight-space programming is another key takeaway.

173
00:20:14,246 --> 00:20:21,376
It allows for easy edits, where a single line change alters minimal outputs, maintaining robustness and accuracy.

174
00:20:21,376 --> 00:20:27,196
This contrasts sharply with traditional models, where weight edits can lead to widespread changes.

175
00:20:27,246 --> 00:20:40,006
Evan: Finally, the paper's evaluations across varied benchmarks and tasks, including pretrained weights and vision tasks, underscore the robustness and efficacy of associative-memory-based models.

176
00:20:40,062 --> 00:20:49,812
Ashley: In summary, weight-space programming offers a promising alternative to conventional model design, emphasizing efficiency, flexibility, and robustness.

177
00:20:49,812 --> 00:20:53,712
It paves the way for more adaptable and performant neural networks.

178
00:20:53,766 --> 00:20:56,726
Evan: Thank you for tuning into Daily Paper Cast.

179
00:20:56,726 --> 00:21:02,096
We hope you enjoyed this deep dive into the innovative research on weight-space programming.

180
00:21:02,142 --> 00:21:08,092
Ashley: Don't forget to join us tomorrow as we continue to explore the latest advancements in AI and neural networks.

181
00:21:08,092 --> 00:21:11,132
Until then, stay curious and keep learning!

182
00:21:11,190 --> 00:21:12,350
Evan: Goodbye for now!