1
00:00:00,000 --> 00:00:02,400
Hey everyone and welcome back for another deep dive.

2
00:00:02,400 --> 00:00:07,520
Today we're looking at this story that has everyone talking in the AI world a Chinese

3
00:00:07,520 --> 00:00:14,560
startup called DeepSeek. They've seemingly done the impossible matching AI giants like OpenAI.

4
00:00:15,760 --> 00:00:18,160
But they're doing it without the billion dollar budgets.

5
00:00:19,200 --> 00:00:24,320
So how did they manage to pull this off? What does it mean for the future of AI?

6
00:00:24,320 --> 00:00:26,560
Yeah, it's really interesting to find out.

7
00:00:26,560 --> 00:00:31,440
The truly remarkable thing about DeepSeek is they haven't actually come up with some brand new

8
00:00:31,440 --> 00:00:36,400
revolutionary technology. What they've done is they've taken existing methods and resources

9
00:00:36,400 --> 00:00:42,000
and through some really clever optimization they've achieved results that actually rival the industry

10
00:00:42,000 --> 00:00:47,360
giants. Okay, so let's break down what they've built. DeepSeek has released two major AI models,

11
00:00:47,920 --> 00:00:55,840
Janus Pro and R1. Now Janus Pro is this incredible multimodal model, meaning it can work with different

12
00:00:55,840 --> 00:01:00,480
types of data images and text. Exactly. It can generate images, analyze images,

13
00:01:00,480 --> 00:01:04,480
and even handle your typical text-based AI tasks. It can do it all.

14
00:01:04,480 --> 00:01:07,840
It's like a Swiss army knife for AI. Yes, very versatile.

15
00:01:07,840 --> 00:01:13,200
It can tackle all sorts of jobs. And then there's R1, which is specifically designed to understand

16
00:01:13,200 --> 00:01:19,200
and generate human-like text. Okay. It's going head-to-head with models like GPT-4 in terms of

17
00:01:19,200 --> 00:01:25,120
performance. Wow. But here's the thing, they developed R1 for only five to six million dollars.

18
00:01:25,120 --> 00:01:28,560
Wait, really? Yeah, think about that for a second. Okay.

19
00:01:28,560 --> 00:01:33,040
Major AI labs are pouring billions into similar projects.

20
00:01:33,040 --> 00:01:38,160
So you're saying DeepSeek is producing similar results with a fraction of the resources.

21
00:01:38,160 --> 00:01:40,240
Exactly. That's got to have some people worried, right?

22
00:01:40,240 --> 00:01:45,520
Oh, absolutely. Especially companies like NVIDIA who make those super expensive top-of-the-line

23
00:01:45,520 --> 00:01:48,480
chips that everyone thought were essential for AI development.

24
00:01:48,480 --> 00:01:52,880
Very right. If DeepSeek can get these results without those, what happens to them?

25
00:01:52,880 --> 00:01:58,640
Well, when news of DeepSeek's efficiency hit the market, NVIDIA's stock took a huge hit,

26
00:01:58,640 --> 00:02:05,280
like a $600 billion drop in a single day. Oh my God. Investors started to worry.

27
00:02:05,280 --> 00:02:11,120
If DeepSeek can do this without the most expensive hardware, what does that mean for the demand for

28
00:02:11,120 --> 00:02:15,120
these chips going forward? That's crazy. Imagine being in those boardrooms when that news came out.

29
00:02:15,120 --> 00:02:17,520
I know, right? Talk about a reality check. Big time.

30
00:02:17,520 --> 00:02:24,160
But the big question is, is this just a fluke? Right. Or is this a sign of a fundamental shift

31
00:02:24,160 --> 00:02:27,920
in how we think about AI development? That is the million-dollar question, isn't it?

32
00:02:27,920 --> 00:02:34,240
Yeah. Some experts believe DeepSeek is leading the way for a new era, an era where smaller,

33
00:02:34,240 --> 00:02:38,880
more agile teams can actually compete with those established companies.

34
00:02:38,880 --> 00:02:44,080
And they're showing that you don't need billions of dollars to make these big strides in AI.

35
00:02:44,080 --> 00:02:47,520
So they're really challenging the foundation of the AI industry. They really are.

36
00:02:48,080 --> 00:02:52,000
How are the big players responding to this? Are they just watching this happen?

37
00:02:52,560 --> 00:02:58,720
Well, OpenAI, the company behind GPT-4, they've acknowledged DeepSeek's accomplishments,

38
00:02:58,720 --> 00:03:05,760
but they are maintaining their commitment to a resource-intensive approach. Their CEO, Sam Altman,

39
00:03:05,760 --> 00:03:11,600
argues that their huge investments are needed to keep pushing the limits of AI.

40
00:03:11,600 --> 00:03:15,520
Interesting. So they're not backing down from that resource race just yet.

41
00:03:15,520 --> 00:03:18,640
So it's almost like we have two different philosophies clashing here.

42
00:03:18,640 --> 00:03:21,120
Yeah. DeepSeek is saying we can do more with less.

43
00:03:21,120 --> 00:03:24,880
Exactly. And OpenAI is saying go big or go home.

44
00:03:24,880 --> 00:03:27,600
Right. It's going to be really interesting to see how this plays out.

45
00:03:27,600 --> 00:03:30,880
I agree. But let's go back to Janice Pro for a moment.

46
00:03:30,880 --> 00:03:32,640
Sure. You said it's multimodal.

47
00:03:32,640 --> 00:03:35,440
Yes. So it can work with both images and text.

48
00:03:35,440 --> 00:03:37,520
That's right. That's pretty unique, right?

49
00:03:37,520 --> 00:03:41,040
It is. Most AI models seem to specialize in one or the other.

50
00:03:41,040 --> 00:03:44,240
Yeah. Janice Pro's versatility makes it stand out.

51
00:03:44,240 --> 00:03:47,680
Okay. It can do all sorts of image analysis tasks.

52
00:03:47,680 --> 00:03:48,480
Like what?

53
00:03:48,480 --> 00:03:52,400
Well, it can accurately describe objects and their relationships in an image.

54
00:03:53,200 --> 00:04:00,400
It can also struggle with more abstract concepts, like metaphors, but then models like GPT-4 vision.

55
00:04:00,960 --> 00:04:03,920
They're better at grasping those deeper meanings in images.

56
00:04:03,920 --> 00:04:08,960
So it sounds like Janice Pro is still a work in progress, but it has a ton of potential.

57
00:04:08,960 --> 00:04:10,080
Yeah. For sure.

58
00:04:10,080 --> 00:04:12,640
And here's where things get even more interesting.

59
00:04:12,640 --> 00:04:13,520
Okay.

60
00:04:13,520 --> 00:04:16,720
Deepseek decided to make Janice Pro open source.

61
00:04:17,680 --> 00:04:18,000
Wow.

62
00:04:18,000 --> 00:04:20,960
So they've made the code available for anyone to access.

63
00:04:20,960 --> 00:04:21,920
That's a big deal.

64
00:04:21,920 --> 00:04:23,600
Modify and even build on.

65
00:04:23,600 --> 00:04:24,160
It is.

66
00:04:24,160 --> 00:04:28,320
That's a pretty bold move, especially when you consider that companies like OpenAI

67
00:04:28,880 --> 00:04:30,960
keep their code under lock and key.

68
00:04:30,960 --> 00:04:36,720
Yeah. That open source approach is one of the things that has people so excited about Janice Pro.

69
00:04:36,720 --> 00:04:37,440
I can see why.

70
00:04:37,440 --> 00:04:41,600
Imagine a community of developers all over the world coming together,

71
00:04:42,240 --> 00:04:48,400
tinkering with the code, sharing their ideas, and then collectively pushing the limits of what this

72
00:04:48,400 --> 00:04:49,600
model can do.

73
00:04:49,600 --> 00:04:50,320
That's amazing.

74
00:04:50,320 --> 00:04:54,400
It's completely different from that closed-door development you see at a lot of tech companies.

75
00:04:54,400 --> 00:04:57,520
So Deepseek is basically saying, hey, here are the tools.

76
00:04:57,520 --> 00:04:58,880
Let's see what we can build together.

77
00:04:58,880 --> 00:04:59,680
Exactly.

78
00:04:59,680 --> 00:05:01,040
It's like a giant experiment.

79
00:05:01,040 --> 00:05:01,680
It is.

80
00:05:01,680 --> 00:05:03,680
And the possibilities are endless.

81
00:05:03,680 --> 00:05:04,640
They really are.

82
00:05:04,640 --> 00:05:08,640
I read somewhere that someone actually put Janice Pro up against stable diffusion,

83
00:05:10,000 --> 00:05:12,800
which is another popular image generating AI.

84
00:05:12,800 --> 00:05:13,840
I know it.

85
00:05:13,840 --> 00:05:20,960
And they asked both of them to create an image of a cute baby fox in an autumn scene.

86
00:05:22,160 --> 00:05:22,960
That's a good one.

87
00:05:22,960 --> 00:05:23,760
I know, right?

88
00:05:23,760 --> 00:05:24,800
What happened?

89
00:05:24,800 --> 00:05:26,080
Well, Janice pretty did OK.

90
00:05:26,080 --> 00:05:26,560
OK.

91
00:05:26,560 --> 00:05:28,800
It got the baby part of the prompt.

92
00:05:28,800 --> 00:05:29,360
Yeah.

93
00:05:29,360 --> 00:05:35,040
But stable diffusion ended up creating a much better-looking image.

94
00:05:35,040 --> 00:05:41,520
So it's like we're seeing a trade-off faithfulness to the prompt versus the overall artistic quality.

95
00:05:41,520 --> 00:05:42,560
That's an interesting point.

96
00:05:42,560 --> 00:05:46,160
So it seems like Janice Pro needs to improve its artistic side a bit.

97
00:05:46,160 --> 00:05:46,560
Yeah.

98
00:05:46,560 --> 00:05:50,240
But with the open-source community helping out, who knows how quickly it could catch up.

99
00:05:50,240 --> 00:05:50,960
Exactly.

100
00:05:50,960 --> 00:05:52,720
That's what's so great about open-source.

101
00:05:52,720 --> 00:05:55,200
It allows for quick improvements and changes.

102
00:05:55,200 --> 00:05:58,320
What might take one company months or years to develop?

103
00:05:58,320 --> 00:06:02,000
It could happen so much faster with a global community working together.

104
00:06:02,000 --> 00:06:02,880
That's really cool.

105
00:06:02,880 --> 00:06:03,360
It is.

106
00:06:03,360 --> 00:06:07,920
Speaking of exciting things, DeepSeq's AI Assistant app is causing quite a stir.

107
00:06:07,920 --> 00:06:08,400
It is.

108
00:06:08,400 --> 00:06:11,600
This thing's shot to the top of Apple's App Store in the US.

109
00:06:11,600 --> 00:06:12,160
Wow.

110
00:06:12,160 --> 00:06:14,640
It even beat ChatGPC in popularity.

111
00:06:14,640 --> 00:06:15,680
That's huge.

112
00:06:15,680 --> 00:06:18,880
That really shows how interested people are in these AI tools.

113
00:06:18,880 --> 00:06:19,360
Definitely.

114
00:06:19,360 --> 00:06:21,760
People really want to see what these technologies can do.

115
00:06:21,760 --> 00:06:22,480
Yeah.

116
00:06:22,480 --> 00:06:24,160
And use them in their everyday lives.

117
00:06:24,160 --> 00:06:27,360
It seems like AI is really going mainstream.

118
00:06:27,360 --> 00:06:28,400
I think you're right.

119
00:06:28,400 --> 00:06:33,600
DeepSeq getting such a huge audience so quickly says a lot about where AI is going.

120
00:06:33,600 --> 00:06:35,440
Of course, with that many new users.

121
00:06:35,440 --> 00:06:35,680
Yeah.

122
00:06:36,480 --> 00:06:37,600
There were some issues.

123
00:06:37,600 --> 00:06:38,240
Like what?

124
00:06:38,240 --> 00:06:40,560
Some reports of temporary outages.

125
00:06:40,560 --> 00:06:41,120
Oh yeah.

126
00:06:41,120 --> 00:06:42,000
Some glitches.

127
00:06:42,000 --> 00:06:42,800
That happens.

128
00:06:42,800 --> 00:06:45,760
But that's almost expected in the tech world these days.

129
00:06:45,760 --> 00:06:46,160
Yeah.

130
00:06:46,160 --> 00:06:47,280
It's a good problem to have.

131
00:06:47,280 --> 00:06:49,280
It means you've really made something people want.

132
00:06:49,280 --> 00:06:50,320
Exactly.

133
00:06:50,320 --> 00:06:52,160
Any publicity is good publicity.

134
00:06:52,160 --> 00:06:52,800
Right.

135
00:06:52,800 --> 00:06:54,560
And in this case, it's definitely true.

136
00:06:54,560 --> 00:06:54,960
Yeah.

137
00:06:54,960 --> 00:06:57,040
DeepSeq's app is the talk of the town.

138
00:06:57,040 --> 00:06:57,280
Yeah.

139
00:06:57,280 --> 00:07:00,080
And it really makes them stand out in the AI world.

140
00:07:00,080 --> 00:07:02,240
But here's where things get a little mysterious.

141
00:07:02,240 --> 00:07:02,800
Okay.

142
00:07:02,800 --> 00:07:03,920
I like a good mystery.

143
00:07:03,920 --> 00:07:07,920
Despite all this success, DeepSeq is still relatively unknown.

144
00:07:08,960 --> 00:07:11,760
They were only founded last year in 2023.

145
00:07:11,760 --> 00:07:12,400
Wow.

146
00:07:12,400 --> 00:07:13,120
So new.

147
00:07:13,120 --> 00:07:13,520
Yeah.

148
00:07:13,520 --> 00:07:14,800
And they're based in Hong Kong.

149
00:07:14,800 --> 00:07:15,440
Interesting.

150
00:07:15,440 --> 00:07:16,960
So naturally, people are curious.

151
00:07:16,960 --> 00:07:17,360
Yeah.

152
00:07:17,360 --> 00:07:18,400
Who are these guys?

153
00:07:18,400 --> 00:07:18,560
Yeah.

154
00:07:18,560 --> 00:07:19,760
Where do they come from?

155
00:07:19,760 --> 00:07:21,680
How did they do all this so fast?

156
00:07:21,680 --> 00:07:22,240
Yeah.

157
00:07:22,240 --> 00:07:23,840
It's like they came out of nowhere.

158
00:07:23,840 --> 00:07:26,400
And just disrupted the whole AI industry.

159
00:07:26,400 --> 00:07:27,280
It's true.

160
00:07:27,280 --> 00:07:29,440
And that's where we'll leave you hanging for now.

161
00:07:29,440 --> 00:07:30,400
Oh, no.

162
00:07:30,400 --> 00:07:30,800
Don't worry.

163
00:07:30,800 --> 00:07:33,280
We'll be back in part two to dig deeper into the mysteries

164
00:07:33,280 --> 00:07:34,160
surrounding DeepSeq.

165
00:07:34,160 --> 00:07:35,040
Okay, good.

166
00:07:35,040 --> 00:07:37,440
And we'll explore what their sudden rise could mean.

167
00:07:37,440 --> 00:07:38,080
Can't wait.

168
00:07:38,080 --> 00:07:38,960
Stay tuned.

169
00:07:38,960 --> 00:07:39,520
I will.

170
00:07:40,160 --> 00:07:40,720
Welcome back.

171
00:07:41,280 --> 00:07:44,880
Before the break, we were talking about DeepSeq's incredible rise

172
00:07:44,880 --> 00:07:48,160
and the shockwaves they've sent through the AI world.

173
00:07:48,160 --> 00:07:49,280
It's been pretty wild.

174
00:07:49,280 --> 00:07:51,840
They've pulled off these amazing AI developments

175
00:07:51,840 --> 00:07:54,320
for a fraction of what it normally costs.

176
00:07:54,320 --> 00:07:54,800
Yeah.

177
00:07:54,800 --> 00:07:56,560
And that's got people wondering,

178
00:07:56,560 --> 00:07:58,720
could this be a sign of things to come?

179
00:07:58,720 --> 00:07:59,280
Yeah.

180
00:07:59,280 --> 00:08:03,280
Could we be seeing a new era of AI development

181
00:08:03,280 --> 00:08:06,080
that's more efficient and cost effective?

182
00:08:06,080 --> 00:08:08,000
There are definitely people who think so.

183
00:08:08,640 --> 00:08:12,960
They believe that DeepSeq's success shows a new path forward.

184
00:08:12,960 --> 00:08:13,280
Okay.

185
00:08:13,280 --> 00:08:18,480
A path where smaller, more agile teams can compete with the big guys.

186
00:08:18,480 --> 00:08:19,040
I like it.

187
00:08:19,040 --> 00:08:21,920
By using clever techniques and open source resources.

188
00:08:22,560 --> 00:08:23,760
And speaking of open source.

189
00:08:23,760 --> 00:08:24,240
Yeah.

190
00:08:24,240 --> 00:08:27,280
We can't forget about Janice Pro and what that could mean.

191
00:08:27,280 --> 00:08:27,840
Right.

192
00:08:27,840 --> 00:08:30,480
That decision to make Janice Pro open source

193
00:08:30,480 --> 00:08:33,440
could have a huge impact on the development community.

194
00:08:33,440 --> 00:08:35,120
It could change the game entirely.

195
00:08:35,120 --> 00:08:35,920
Think about it.

196
00:08:35,920 --> 00:08:38,960
A community of developers from all over the world.

197
00:08:38,960 --> 00:08:39,360
Yeah.

198
00:08:39,360 --> 00:08:42,880
They're not limited by the profit motives of those big corporations.

199
00:08:42,880 --> 00:08:43,440
Right.

200
00:08:43,440 --> 00:08:47,040
And they're all working together to push the limits of AI.

201
00:08:47,680 --> 00:08:52,000
Creating these powerful tools and technologies that anyone can use.

202
00:08:52,000 --> 00:08:53,520
That's a pretty amazing vision.

203
00:08:53,520 --> 00:08:53,920
It is.

204
00:08:53,920 --> 00:08:57,040
But some people argue that while DeepSeq's achievements are impressive.

205
00:08:57,040 --> 00:08:57,360
Yeah.

206
00:08:57,360 --> 00:09:00,320
They might not be easy to copy or keep up in the long run.

207
00:09:00,320 --> 00:09:00,640
Yeah.

208
00:09:00,640 --> 00:09:01,920
That's the other side of the coin.

209
00:09:01,920 --> 00:09:08,240
The amount of data and computing power you need to train and use AI models globally is huge.

210
00:09:08,240 --> 00:09:08,800
Exactly.

211
00:09:08,800 --> 00:09:12,560
Companies like Meta have said that if you want to serve billions of users.

212
00:09:12,560 --> 00:09:12,960
Yeah.

213
00:09:12,960 --> 00:09:17,440
You need serious infrastructure and a ton of computing power.

214
00:09:17,440 --> 00:09:19,280
So it's more than just making the model.

215
00:09:19,280 --> 00:09:19,680
Oh yeah.

216
00:09:19,680 --> 00:09:22,240
It's also making sure it works for everyone everywhere.

217
00:09:22,240 --> 00:09:23,040
Exactly.

218
00:09:23,040 --> 00:09:26,320
It has to be reliable and accessible on a global scale.

219
00:09:26,320 --> 00:09:28,800
So DeepSeq's approach is super innovative.

220
00:09:28,800 --> 00:09:29,360
Oh yeah.

221
00:09:29,360 --> 00:09:29,920
Totally.

222
00:09:29,920 --> 00:09:32,000
But can it really handle a global market?

223
00:09:32,720 --> 00:09:33,840
Good question.

224
00:09:33,840 --> 00:09:36,640
It's one thing to have a model that does well in a test environment.

225
00:09:36,640 --> 00:09:37,280
Right.

226
00:09:37,280 --> 00:09:41,040
But it's another thing entirely to make sure it's stable and reliable

227
00:09:41,760 --> 00:09:44,880
when millions of people are using it at the same time.

228
00:09:44,880 --> 00:09:45,520
Absolutely.

229
00:09:46,240 --> 00:09:49,040
And then there's the geopolitical side of things to consider.

230
00:09:49,040 --> 00:09:49,440
Oh yeah.

231
00:09:49,440 --> 00:09:50,000
That's right.

232
00:09:50,000 --> 00:09:53,760
President Trump has called on American tech companies to step up their game.

233
00:09:53,760 --> 00:09:55,280
In response to DeepSeq.

234
00:09:55,280 --> 00:09:56,080
Exactly.

235
00:09:56,080 --> 00:09:58,320
He sees DeepSeq as a real wake up call.

236
00:09:58,320 --> 00:09:58,960
He does.

237
00:09:58,960 --> 00:09:59,440
He does.

238
00:10:00,000 --> 00:10:03,760
He's saying that the US has to stay on top in AI.

239
00:10:03,760 --> 00:10:08,000
That's interesting because this comes at a time when there's a lot of talk about restricting

240
00:10:08,000 --> 00:10:09,360
chip exports to China.

241
00:10:09,360 --> 00:10:12,960
It's like they're trying to slow down Chinese progress in AI.

242
00:10:12,960 --> 00:10:13,440
Right.

243
00:10:13,440 --> 00:10:17,360
But DeepSeq seems to have found a way around these restrictions.

244
00:10:17,360 --> 00:10:17,920
Oh really?

245
00:10:17,920 --> 00:10:18,320
How?

246
00:10:18,320 --> 00:10:24,160
They're using Nvidia's H800 chips which are technically less powerful than those top tier

247
00:10:24,160 --> 00:10:25,360
chips that are restricted.

248
00:10:25,360 --> 00:10:30,880
So DeepSeq is getting similar results with chips that aren't even the most advanced.

249
00:10:30,880 --> 00:10:31,520
They are.

250
00:10:31,520 --> 00:10:36,480
That kind of undermines the idea that you need the best hardware to be competitive in AI.

251
00:10:36,480 --> 00:10:37,280
It does.

252
00:10:37,280 --> 00:10:42,080
DeepSeq's approach shows that being clever and innovative can sometimes be more important

253
00:10:42,080 --> 00:10:44,160
than having a ton of resources.

254
00:10:44,160 --> 00:10:46,160
So that brings us back to that big question.

255
00:10:46,160 --> 00:10:46,720
Okay.

256
00:10:46,720 --> 00:10:50,480
Can DeepSeq's way of doing things really work for the future of AI?

257
00:10:50,480 --> 00:10:51,840
It's tough to say for sure.

258
00:10:51,840 --> 00:10:52,240
Yeah.

259
00:10:52,240 --> 00:10:54,320
DeepSeq has definitely shaken things up.

260
00:10:54,320 --> 00:10:54,880
They have.

261
00:10:54,880 --> 00:10:56,720
Everyone's rethinking their strategies now.

262
00:10:57,280 --> 00:11:00,960
They've shown that you can get amazing results without spending a fortune.

263
00:11:00,960 --> 00:11:02,080
But can it scale up?

264
00:11:02,880 --> 00:11:04,080
Can it keep innovating?

265
00:11:04,880 --> 00:11:06,000
We'll have to wait and see.

266
00:11:06,000 --> 00:11:07,440
I think that's a fair assessment.

267
00:11:07,440 --> 00:11:10,160
And there's been some skepticism about DeepSeq's claims too.

268
00:11:10,720 --> 00:11:14,720
Especially about the $5.6 million development cost for R1.

269
00:11:14,720 --> 00:11:19,680
Ah, yeah. Some experts say that number probably just covers the final training phase.

270
00:11:19,680 --> 00:11:20,160
Oh, okay.

271
00:11:20,160 --> 00:11:24,080
And that it doesn't include all the early research and data work that went into it.

272
00:11:24,080 --> 00:11:26,000
So the real cost could be a lot higher.

273
00:11:26,880 --> 00:11:27,680
Possibly.

274
00:11:27,680 --> 00:11:33,120
But even if it was a few times higher, it could still be way less than what the big tech companies

275
00:11:33,120 --> 00:11:34,000
usually spend.

276
00:11:34,000 --> 00:11:34,960
So how did they do it?

277
00:11:34,960 --> 00:11:35,760
What's their secret?

278
00:11:36,320 --> 00:11:39,440
Well, DeepSeq says it's all about their innovative training techniques.

279
00:11:39,440 --> 00:11:40,240
Okay.

280
00:11:40,240 --> 00:11:45,360
They claim they found ways to make the model focus only on the most important data.

281
00:11:45,360 --> 00:11:47,120
And that saves a lot of computing power.

282
00:11:47,120 --> 00:11:47,520
Sort.

283
00:11:47,520 --> 00:11:52,080
And they also used existing open source projects from companies like Alibaba and Meta.

284
00:11:52,080 --> 00:11:52,640
Oh, wow.

285
00:11:52,640 --> 00:11:56,080
They built on those foundations and adapted them for their own needs.

286
00:11:56,080 --> 00:11:58,320
So they're really taking advantage of open source.

287
00:11:58,320 --> 00:11:59,760
They are. It's a smart move.

288
00:11:59,760 --> 00:12:03,120
It really shows how powerful the open source movement can be.

289
00:12:03,120 --> 00:12:05,680
It does. Developers can build on each other's work.

290
00:12:05,680 --> 00:12:07,840
And that speeds up innovation like crazy.

291
00:12:07,840 --> 00:12:10,640
It's like that saying standing on the shoulders of giants.

292
00:12:10,640 --> 00:12:11,440
Exactly.

293
00:12:11,440 --> 00:12:15,280
But using those open source projects, especially ones from Western companies,

294
00:12:15,280 --> 00:12:15,760
Yeah.

295
00:12:15,760 --> 00:12:17,280
has been a little controversial.

296
00:12:17,280 --> 00:12:21,440
Right. Some people say DeepSeq is basically writing on the coattails of others.

297
00:12:22,240 --> 00:12:24,160
And they're questioning whether that's fair play.

298
00:12:24,880 --> 00:12:28,080
It raises questions about intellectual property and competition.

299
00:12:28,080 --> 00:12:32,080
It does. And it really shows how complicated open source development can be.

300
00:12:32,080 --> 00:12:34,880
Yeah. You want to encourage collaboration and progress.

301
00:12:34,880 --> 00:12:35,200
Right.

302
00:12:35,200 --> 00:12:38,160
But you also need to make sure everyone gets credit for their work.

303
00:12:38,160 --> 00:12:41,280
Exactly. And that no one is taking advantage of the system.

304
00:12:41,280 --> 00:12:42,640
It's a tough balance to strike.

305
00:12:42,640 --> 00:12:43,440
It is.

306
00:12:43,440 --> 00:12:45,520
And then of course, there's the ethical side of all of this.

307
00:12:45,520 --> 00:12:47,440
Oh, right. The big picture stuff.

308
00:12:47,440 --> 00:12:51,200
Stuart Russell, a leading AI researcher at UC Berkeley.

309
00:12:51,200 --> 00:12:52,000
I know him.

310
00:12:52,000 --> 00:12:54,800
He's worried about an unchecked AI arms race.

311
00:12:54,800 --> 00:12:56,640
Yeah. She's been saying that for a while.

312
00:12:56,640 --> 00:13:03,680
He says that the race to create artificial general intelligence, or AGI, is really risky.

313
00:13:03,680 --> 00:13:04,400
It could be.

314
00:13:04,400 --> 00:13:08,560
Because we're basically trying to build systems that can become smarter than humans.

315
00:13:08,560 --> 00:13:10,000
And potentially take control.

316
00:13:10,000 --> 00:13:11,680
Exactly. And that's a scary thought.

317
00:13:11,680 --> 00:13:14,640
It is. It makes you think about the future of humanity.

318
00:13:14,640 --> 00:13:16,160
And our relationship with technology.

319
00:13:16,160 --> 00:13:16,640
Definitely.

320
00:13:16,640 --> 00:13:18,320
And if we bring this back to DeepSeq,

321
00:13:19,440 --> 00:13:24,880
their rapid rise and their ability to do so much with so little money,

322
00:13:25,600 --> 00:13:29,120
it could actually make this arms race even faster.

323
00:13:29,120 --> 00:13:30,160
Oh, so.

324
00:13:30,160 --> 00:13:34,320
It might push other companies to try to find even cheaper and faster ways to develop AI.

325
00:13:34,320 --> 00:13:39,040
Oh, I see. So we could see a flood of new powerful AI systems.

326
00:13:39,040 --> 00:13:40,320
That's a real possibility.

327
00:13:40,320 --> 00:13:42,000
And it's something we need to think about carefully.

328
00:13:42,000 --> 00:13:47,760
We do. We need to ask ourselves, what happens if AI development keeps accelerating at this pace?

329
00:13:48,480 --> 00:13:51,360
Are we ready for the challenges that might bring?

330
00:13:51,360 --> 00:13:52,080
Big questions.

331
00:13:52,080 --> 00:13:52,880
You are.

332
00:13:52,880 --> 00:13:56,960
And they touch on everything from society and the economy to ethics.

333
00:13:56,960 --> 00:13:58,160
A lot to consider.

334
00:13:58,160 --> 00:14:02,160
It is. And these are questions we'll keep asking is AI keeps evolving.

335
00:14:02,160 --> 00:14:03,040
I think that's a given.

336
00:14:03,040 --> 00:14:04,320
But let's pause here.

337
00:14:04,320 --> 00:14:04,720
Okay.

338
00:14:04,720 --> 00:14:07,520
And we'll come back for the third and final part of this deep dive.

339
00:14:07,520 --> 00:14:08,160
Sounds good.

340
00:14:08,160 --> 00:14:13,440
We'll wrap up our look at DeepSeq and talk about what their story might mean for the future of AI.

341
00:14:14,800 --> 00:14:16,080
Looking forward to it.

342
00:14:16,080 --> 00:14:16,880
Stick with us.

343
00:14:16,880 --> 00:14:17,600
I will.

344
00:14:17,600 --> 00:14:18,480
We're back.

345
00:14:18,480 --> 00:14:24,240
And we've been on this wild ride exploring DeepSeq's incredible rise in the AI world.

346
00:14:24,240 --> 00:14:26,880
Yeah, it's been a fascinating deep dive for sure.

347
00:14:26,880 --> 00:14:30,560
They've really shown how efficient AI development can be.

348
00:14:30,560 --> 00:14:31,200
Right.

349
00:14:31,200 --> 00:14:34,080
But along the way we found some interesting questions and concerns.

350
00:14:34,080 --> 00:14:36,160
There's always more to uncover, right?

351
00:14:36,160 --> 00:14:37,440
So much to think about.

352
00:14:37,440 --> 00:14:41,680
What stands out to you is the most impactful thing about DeepSeq so far.

353
00:14:42,320 --> 00:14:44,480
Hmm. That's a tough one.

354
00:14:44,480 --> 00:14:46,960
I know, right? There's so much to choose from.

355
00:14:46,960 --> 00:14:48,800
Yeah. What's the big takeaway here for you?

356
00:14:48,800 --> 00:14:53,440
I think it's how DeepSeq has challenged the way we think about AI development.

357
00:14:53,440 --> 00:14:58,400
For so long we just assumed that only the companies with tons of money like Google and Microsoft.

358
00:14:58,400 --> 00:14:59,360
We're one of the big players.

359
00:14:59,360 --> 00:15:02,080
Exactly. They were the only ones who could play in this space.

360
00:15:02,080 --> 00:15:06,240
But DeepSeq showed us that sometimes being smart and innovative

361
00:15:06,240 --> 00:15:08,800
is more important than just throwing money at the problem.

362
00:15:08,800 --> 00:15:12,080
Yeah. I think their boldness is what really caught everyone's attention.

363
00:15:12,080 --> 00:15:12,560
I agree.

364
00:15:13,440 --> 00:15:15,600
This little known startup comes out of nowhere.

365
00:15:15,600 --> 00:15:16,160
Yeah.

366
00:15:16,160 --> 00:15:17,360
And shakes up the whole industry.

367
00:15:17,360 --> 00:15:19,120
It's like they flipped the script.

368
00:15:19,120 --> 00:15:21,760
Totally. And now everyone's trying to figure out what to do.

369
00:15:21,760 --> 00:15:23,360
Yeah. How to adapt.

370
00:15:23,360 --> 00:15:24,000
Exactly.

371
00:15:24,000 --> 00:15:25,840
It's like something out of a movie.

372
00:15:25,840 --> 00:15:26,560
It really is.

373
00:15:26,560 --> 00:15:30,320
The underdog taking on the giants and showing that it can be done.

374
00:15:30,320 --> 00:15:31,760
And it's not just the tech itself.

375
00:15:31,760 --> 00:15:32,320
Right.

376
00:15:32,320 --> 00:15:33,920
It's how they went about it.

377
00:15:33,920 --> 00:15:35,120
Their whole approach.

378
00:15:35,120 --> 00:15:37,680
Exactly. Open sourcing Janice Pro was a big move.

379
00:15:37,680 --> 00:15:38,160
It was.

380
00:15:38,160 --> 00:15:41,280
This shows they really believe in collaboration

381
00:15:42,160 --> 00:15:45,440
and that community driven innovation can work.

382
00:15:45,440 --> 00:15:48,400
It's like they're saying we're not hiding our knowledge.

383
00:15:48,400 --> 00:15:48,560
Right.

384
00:15:48,560 --> 00:15:49,440
We're sharing it.

385
00:15:49,440 --> 00:15:52,000
Let's all work together and see what we can create.

386
00:15:52,000 --> 00:15:55,760
So now the big question is what does this mean for the future?

387
00:15:55,760 --> 00:15:56,320
Okay.

388
00:15:56,320 --> 00:15:57,760
Is this the start of something new?

389
00:15:58,720 --> 00:16:03,280
Will we see a shift to more open and collaborative AI development?

390
00:16:04,080 --> 00:16:05,760
I think it's definitely possible.

391
00:16:05,760 --> 00:16:06,160
Okay.

392
00:16:06,160 --> 00:16:11,360
Deep-seek success could inspire a whole new wave of AI startups

393
00:16:11,360 --> 00:16:15,040
and researchers who want to focus on efficiency.

394
00:16:15,040 --> 00:16:15,440
Right.

395
00:16:15,440 --> 00:16:19,120
Using resources wisely and working together openly.

396
00:16:19,120 --> 00:16:22,080
They've shown that you don't have to have billions of dollars.

397
00:16:22,080 --> 00:16:22,400
Right.

398
00:16:22,400 --> 00:16:24,320
To make a real difference in this field.

399
00:16:24,320 --> 00:16:28,240
But some people say that Deep-seek's approach just won't work in the long run.

400
00:16:28,960 --> 00:16:30,240
Yeah. I've heard that too.

401
00:16:30,240 --> 00:16:34,880
They say you still need a ton of resources to make these models work for everyone.

402
00:16:34,880 --> 00:16:38,240
To make them accessible to billions of users around the world.

403
00:16:38,240 --> 00:16:38,880
Exactly.

404
00:16:38,880 --> 00:16:41,360
So how do we reconcile those two viewpoints?

405
00:16:41,360 --> 00:16:42,960
That's the challenge, isn't it?

406
00:16:42,960 --> 00:16:43,360
It is.

407
00:16:43,360 --> 00:16:48,720
Scaling these models and making them work for everyone is a huge hurdle.

408
00:16:48,720 --> 00:16:49,040
Yeah.

409
00:16:49,040 --> 00:16:52,960
And serving a global audience requires a lot of infrastructure and resources.

410
00:16:52,960 --> 00:16:53,600
It does.

411
00:16:53,600 --> 00:16:56,960
But Deep-seek's approach could lead to new ways of doing things.

412
00:16:56,960 --> 00:16:57,360
Okay.

413
00:16:57,360 --> 00:16:59,600
More efficient ways to scale AI.

414
00:16:59,600 --> 00:17:02,960
They've already shown they can achieve amazing results with limited resources.

415
00:17:02,960 --> 00:17:03,520
They have.

416
00:17:03,520 --> 00:17:05,680
Who's to say they can't do the same for scaling?

417
00:17:05,680 --> 00:17:07,120
It's an exciting thought.

418
00:17:07,120 --> 00:17:08,720
It is if they can figure that out.

419
00:17:08,720 --> 00:17:14,080
We could see a future where really powerful AI tools are available to everyone.

420
00:17:14,080 --> 00:17:15,600
Yeah. Not just the big companies.

421
00:17:15,600 --> 00:17:16,160
Exactly.

422
00:17:16,160 --> 00:17:17,760
And that could change everything.

423
00:17:17,760 --> 00:17:19,360
It could impact education.

424
00:17:19,360 --> 00:17:20,080
Healthcare.

425
00:17:20,080 --> 00:17:21,040
Scientific research.

426
00:17:21,840 --> 00:17:23,440
Even social good initiatives.

427
00:17:23,440 --> 00:17:25,120
It's really exciting to think about.

428
00:17:25,120 --> 00:17:27,520
But it's important to remember that we're still at the beginning.

429
00:17:27,520 --> 00:17:29,600
Oh, for sure. This is just the start.

430
00:17:29,600 --> 00:17:31,200
There are a lot of challenges ahead.

431
00:17:31,200 --> 00:17:32,080
Absolutely.

432
00:17:32,080 --> 00:17:33,600
Both technical challenges.

433
00:17:33,600 --> 00:17:34,640
And the sidelines.

434
00:17:34,640 --> 00:17:41,440
As AI continues to evolve, we need to have serious conversations about how to develop and use it responsibly.

435
00:17:41,440 --> 00:17:42,720
I think that's crucial.

436
00:17:42,720 --> 00:17:45,120
We need to think about the ethical implications.

437
00:17:45,120 --> 00:17:45,600
Right.

438
00:17:45,600 --> 00:17:47,440
The impact on jobs.

439
00:17:47,440 --> 00:17:47,920
Yeah.

440
00:17:47,920 --> 00:17:50,400
And how to prevent misuse.

441
00:17:50,400 --> 00:17:51,920
These are big questions.

442
00:17:51,920 --> 00:17:52,880
Huge ones.

443
00:17:52,880 --> 00:17:57,200
And it's clear that DeepSeek's story is just one part of a much bigger story.

444
00:17:57,200 --> 00:17:57,680
It is.

445
00:17:57,680 --> 00:17:59,760
The world of AI is changing so fast.

446
00:17:59,760 --> 00:18:00,720
It's hard to keep up.

447
00:18:00,720 --> 00:18:02,720
And we're just starting to understand what it all means.

448
00:18:02,720 --> 00:18:03,200
Yeah.

449
00:18:03,200 --> 00:18:04,560
But one thing is for sure.

450
00:18:04,560 --> 00:18:05,440
What's that?

451
00:18:05,440 --> 00:18:07,440
DeepSeek has changed the game.

452
00:18:07,440 --> 00:18:08,320
They have.

453
00:18:08,320 --> 00:18:11,760
And their impact will be felt for years to come.

454
00:18:11,760 --> 00:18:14,720
So as we wrap up this deep dive, here's something for you to think about.

455
00:18:14,720 --> 00:18:15,680
Okay. I'm listening.

456
00:18:16,400 --> 00:18:21,120
What if DeepSeek's success isn't just about doing more with less?

457
00:18:21,120 --> 00:18:22,320
Yes.

458
00:18:22,320 --> 00:18:25,840
What if it's a sign of a whole new wave of innovation in AI?

459
00:18:25,840 --> 00:18:27,520
Coming from unexpected places.

460
00:18:27,520 --> 00:18:28,480
Exactly.

461
00:18:28,480 --> 00:18:33,440
What if the future of AI is less about brute force and more about being clever.

462
00:18:33,440 --> 00:18:34,320
And collaborating.

463
00:18:34,320 --> 00:18:35,200
Working together.

464
00:18:35,200 --> 00:18:38,080
And being willing to challenge the way things have always been done.

465
00:18:38,080 --> 00:18:38,880
Food for thought.

466
00:18:38,880 --> 00:18:39,360
Right.

467
00:18:39,360 --> 00:18:40,720
Definitely a lot to ponder.

468
00:18:40,720 --> 00:18:43,600
Thanks for joining us as we explore this fascinating story.

469
00:18:43,600 --> 00:18:44,400
It's been a pleasure.

470
00:18:44,400 --> 00:18:46,800
We hope you found it informative and thought-provoking.

471
00:18:46,800 --> 00:18:47,520
I know he did.

472
00:18:47,520 --> 00:18:49,360
Until next time, keep exploring.

473
00:18:49,360 --> 00:18:51,120
Keep questioning and keep learning.

474
00:18:51,120 --> 00:19:20,960
I will see you next time.