1
00:00:00,000 --> 00:00:01,920
Hey everyone and welcome back.

2
00:00:01,920 --> 00:00:04,220
Today we're diving into something pretty amazing.

3
00:00:05,100 --> 00:00:09,340
It's OpenAI's new video generation model, Sora.

4
00:00:09,340 --> 00:00:12,080
Yeah, you can kind of think of it like Dali but for video.

5
00:00:12,080 --> 00:00:15,080
So Dali, but instead of making pictures, it makes videos.

6
00:00:15,080 --> 00:00:16,200
Exactly.

7
00:00:16,200 --> 00:00:19,200
And the crazy part is how it can create these videos

8
00:00:19,200 --> 00:00:21,040
from all sorts of different input.

9
00:00:21,040 --> 00:00:22,480
Oh really? Like what?

10
00:00:22,480 --> 00:00:25,040
Well, you can use text prompts just like with Dali,

11
00:00:25,040 --> 00:00:27,160
but you can also use still images.

12
00:00:27,160 --> 00:00:28,000
Well, hold on.

13
00:00:28,000 --> 00:00:30,200
Like if I give it a picture, it can make a video out of it.

14
00:00:30,200 --> 00:00:31,040
Yeah.

15
00:00:31,040 --> 00:00:34,880
Or even video clips to create totally new videos.

16
00:00:34,880 --> 00:00:35,720
That's pretty wild.

17
00:00:35,720 --> 00:00:38,560
It can generate videos up to 1080p resolution

18
00:00:38,560 --> 00:00:42,840
and up to 20 seconds long with crazy detail and accuracy.

19
00:00:42,840 --> 00:00:44,760
So we're talking about like making pictures come to life here.

20
00:00:44,760 --> 00:00:45,600
Pretty much.

21
00:00:45,600 --> 00:00:47,960
I gotta admit the tech side of things is a little fuzzy for me.

22
00:00:47,960 --> 00:00:48,800
Sure.

23
00:00:48,800 --> 00:00:50,720
How does Sora actually make these videos?

24
00:00:50,720 --> 00:00:53,960
So Sora uses something called a diffusion model.

25
00:00:53,960 --> 00:00:56,520
Imagine you have a video that's just a bunch of random noise

26
00:00:56,520 --> 00:00:58,320
like static on an old TV.

27
00:00:58,320 --> 00:01:00,200
The model starts with that noise

28
00:01:00,200 --> 00:01:02,480
and gradually removes it step by step

29
00:01:02,480 --> 00:01:05,560
until it forms a clear video based on what you gave it.

30
00:01:05,560 --> 00:01:07,360
So it's kind of like sculpting.

31
00:01:07,360 --> 00:01:09,080
But instead of starting with a block of marble,

32
00:01:09,080 --> 00:01:11,040
you're starting with a jumbled mess of pixels.

33
00:01:11,040 --> 00:01:11,960
Yeah.

34
00:01:11,960 --> 00:01:13,320
That's a great way to think about it.

35
00:01:13,320 --> 00:01:15,280
And one of the things that makes Sora so special

36
00:01:15,280 --> 00:01:16,960
is that it can keep things consistent

37
00:01:16,960 --> 00:01:19,000
even when stuff moves out of view.

38
00:01:19,000 --> 00:01:19,840
Oh yeah.

39
00:01:19,840 --> 00:01:20,680
I've definitely seen that problem

40
00:01:20,680 --> 00:01:22,280
with other AI generated videos.

41
00:01:22,280 --> 00:01:25,120
Like things just disappear or totally change

42
00:01:25,120 --> 00:01:26,240
when they go behind something.

43
00:01:26,240 --> 00:01:27,080
Right.

44
00:01:27,080 --> 00:01:29,040
And Sora solves that by using what's called

45
00:01:29,040 --> 00:01:31,280
a transformer architecture.

46
00:01:31,280 --> 00:01:33,960
It's the same tech behind those powerful language models

47
00:01:33,960 --> 00:01:37,520
like GPT-3 and it lets Sora look at lots of frames

48
00:01:37,520 --> 00:01:39,080
of the video at the same time.

49
00:01:39,080 --> 00:01:42,400
So it's almost like it remembers what should be there

50
00:01:42,400 --> 00:01:43,760
even when we can't see it.

51
00:01:43,760 --> 00:01:44,600
Exactly.

52
00:01:44,600 --> 00:01:46,120
Like it understands the whole scene

53
00:01:46,120 --> 00:01:48,080
not just focusing on each moment.

54
00:01:48,080 --> 00:01:49,240
Makes sense.

55
00:01:49,240 --> 00:01:51,880
So to understand our instructions,

56
00:01:51,880 --> 00:01:54,200
it uses this technique called recaptioning

57
00:01:54,200 --> 00:01:55,840
from Deli-3, right?

58
00:01:55,840 --> 00:01:56,680
That's right.

59
00:01:56,680 --> 00:01:57,960
So it's not just seeing pixels.

60
00:01:57,960 --> 00:01:59,040
It's learning to understand

61
00:01:59,040 --> 00:02:00,960
what those pixels actually represent.

62
00:02:00,960 --> 00:02:01,800
Exactly.

63
00:02:01,800 --> 00:02:04,120
It's all about connecting the visual information

64
00:02:04,120 --> 00:02:06,360
with the words we use to describe it,

65
00:02:06,360 --> 00:02:09,000
which helps it follow instructions more accurately.

66
00:02:09,000 --> 00:02:09,840
Wow.

67
00:02:09,840 --> 00:02:10,800
That's incredibly clever.

68
00:02:10,800 --> 00:02:11,680
And to get that good,

69
00:02:11,680 --> 00:02:14,200
it must have been trained on a ton of data, right?

70
00:02:14,200 --> 00:02:15,040
Yeah.

71
00:02:15,040 --> 00:02:15,880
I was just thinking that.

72
00:02:15,880 --> 00:02:16,840
Like mountains of data.

73
00:02:16,840 --> 00:02:17,680
Yeah.

74
00:02:17,680 --> 00:02:20,040
Think of it as a visual feast for the AI.

75
00:02:20,040 --> 00:02:21,720
So where did they get all that data?

76
00:02:21,720 --> 00:02:24,120
Well, they used publicly available data

77
00:02:24,120 --> 00:02:27,200
like what you'd find in standard image and video data sets.

78
00:02:27,200 --> 00:02:28,040
Okay.

79
00:02:28,040 --> 00:02:29,200
But they also partnered with companies

80
00:02:29,200 --> 00:02:30,960
like Shutterstock and Pond5.

81
00:02:30,960 --> 00:02:31,800
Okay.

82
00:02:31,800 --> 00:02:34,080
Which gave them access to even more high quality

83
00:02:34,080 --> 00:02:35,440
images and video.

84
00:02:35,440 --> 00:02:36,280
Yeah.

85
00:02:36,280 --> 00:02:38,400
Partnerships like that are so important with AI.

86
00:02:38,400 --> 00:02:41,360
Access to that much data is like gold.

87
00:02:41,360 --> 00:02:42,480
Totally.

88
00:02:42,480 --> 00:02:45,040
And the third source was actually human feedback.

89
00:02:45,040 --> 00:02:45,880
Oh, really?

90
00:02:45,880 --> 00:02:48,680
They had AI trainers, people looking for weaknesses,

91
00:02:48,680 --> 00:02:52,680
and even open AI employees giving feedback to train Sora.

92
00:02:52,680 --> 00:02:54,280
So it's not just feeding it raw data.

93
00:02:54,280 --> 00:02:57,280
It's about having humans guide the learning process

94
00:02:57,280 --> 00:02:58,680
and make sure it's on the right track.

95
00:02:58,680 --> 00:02:59,560
Exactly.

96
00:02:59,560 --> 00:03:01,520
But I'm guessing they didn't just throw everything at Sora.

97
00:03:01,520 --> 00:03:02,400
Oh, definitely not.

98
00:03:02,400 --> 00:03:04,240
There must have been some filtering involved, right?

99
00:03:04,240 --> 00:03:05,480
Absolutely.

100
00:03:05,480 --> 00:03:08,120
Before training, they filtered the data sets

101
00:03:08,120 --> 00:03:10,480
and took out anything violent sensitive

102
00:03:10,480 --> 00:03:12,320
or anything with hate symbols.

103
00:03:12,320 --> 00:03:13,160
Makes sense.

104
00:03:13,160 --> 00:03:14,800
You don't want the AI picking up any bad habits.

105
00:03:14,800 --> 00:03:15,640
For sure.

106
00:03:15,640 --> 00:03:16,480
Okay.

107
00:03:16,480 --> 00:03:19,120
So training on massive data sets and filtering them,

108
00:03:19,120 --> 00:03:20,720
that's one side of the coin.

109
00:03:20,720 --> 00:03:23,080
What about the safety and practical implications

110
00:03:23,080 --> 00:03:24,520
of actually using Sora?

111
00:03:25,440 --> 00:03:27,520
A technology this powerful has to come

112
00:03:27,520 --> 00:03:29,200
with some serious responsibility.

113
00:03:29,200 --> 00:03:30,880
Oh, you are absolutely right about that.

114
00:03:30,880 --> 00:03:33,360
Like what are they doing to make sure it's used safely?

115
00:03:33,360 --> 00:03:36,000
Well, open AI has been really proactive

116
00:03:36,000 --> 00:03:38,120
about addressing the potential risks.

117
00:03:38,120 --> 00:03:38,960
Okay.

118
00:03:38,960 --> 00:03:40,720
One of the things they did was work closely

119
00:03:40,720 --> 00:03:43,120
with artists, designers, and filmmakers

120
00:03:43,120 --> 00:03:46,080
to really understand how Sora could be used

121
00:03:46,080 --> 00:03:47,280
in the real world.

122
00:03:47,280 --> 00:03:49,880
Ah, so they're not just building this in isolation.

123
00:03:49,880 --> 00:03:52,880
They're actually talking to the people who will be using it

124
00:03:52,880 --> 00:03:54,520
and thinking about the consequences.

125
00:03:54,520 --> 00:03:55,360
Exactly.

126
00:03:55,360 --> 00:03:56,640
Both the good and the bad.

127
00:03:56,640 --> 00:03:57,520
That's good to hear.

128
00:03:57,520 --> 00:03:59,600
And to identify potential problems

129
00:03:59,600 --> 00:04:01,560
before they become real world issues,

130
00:04:01,560 --> 00:04:03,520
that's where red teaming comes in.

131
00:04:03,520 --> 00:04:04,360
Red teaming.

132
00:04:04,360 --> 00:04:07,080
Basically, they had experts try to break Sora.

133
00:04:07,080 --> 00:04:07,920
Oh, wow.

134
00:04:07,920 --> 00:04:09,080
They were looking for any weaknesses

135
00:04:09,080 --> 00:04:10,280
in its safety measures,

136
00:04:10,280 --> 00:04:12,960
which helped open AI fix those problems

137
00:04:12,960 --> 00:04:14,200
before releasing it.

138
00:04:14,200 --> 00:04:15,320
So like a stress test?

139
00:04:15,320 --> 00:04:17,240
Yeah, like pushing it to its limits

140
00:04:17,240 --> 00:04:20,000
in a safe environment to see where it might break down.

141
00:04:20,000 --> 00:04:21,720
And it sounds like they're taking this whole safety thing

142
00:04:21,720 --> 00:04:22,560
very seriously.

143
00:04:22,560 --> 00:04:23,800
Oh, they are, for sure.

144
00:04:23,800 --> 00:04:26,080
And one area where they've been super focused

145
00:04:26,080 --> 00:04:27,600
is child safety.

146
00:04:27,600 --> 00:04:28,920
Okay, yeah, that makes sense.

147
00:04:28,920 --> 00:04:30,840
They have multiple layers of protection

148
00:04:30,840 --> 00:04:33,600
to prevent Sora from generating any harmful

149
00:04:33,600 --> 00:04:36,480
or inappropriate content related to children.

150
00:04:36,480 --> 00:04:38,040
Yeah, that's definitely a top priority

151
00:04:38,040 --> 00:04:39,320
with any new technology,

152
00:04:39,320 --> 00:04:42,040
especially one that makes images and videos.

153
00:04:42,040 --> 00:04:42,880
Absolutely.

154
00:04:42,880 --> 00:04:44,520
So can you tell me more about these safeguards?

155
00:04:44,520 --> 00:04:45,560
How do they actually work?

156
00:04:45,560 --> 00:04:46,400
Sure.

157
00:04:46,400 --> 00:04:48,880
They've got this really sophisticated system

158
00:04:48,880 --> 00:04:51,360
with input and output classifiers

159
00:04:51,360 --> 00:04:55,120
that scan for and block any bad content.

160
00:04:55,120 --> 00:04:55,960
Interesting.

161
00:04:55,960 --> 00:04:58,840
And they also have block lists of words and phrases

162
00:04:58,840 --> 00:05:00,360
that are totally off limits.

163
00:05:00,360 --> 00:05:01,200
Makes sense.

164
00:05:01,200 --> 00:05:03,760
Plus they're working with groups like Thorn and NCMEC

165
00:05:03,760 --> 00:05:05,720
that fight child exploitation.

166
00:05:05,720 --> 00:05:07,480
Okay, so they're pulling out all the stops

167
00:05:07,480 --> 00:05:08,560
to keep children safe.

168
00:05:08,560 --> 00:05:09,600
Definitely.

169
00:05:09,600 --> 00:05:13,000
But what about preventing misuse in general?

170
00:05:13,000 --> 00:05:14,240
Are there ways to stop people

171
00:05:14,240 --> 00:05:16,160
from making bad stuff with Sora?

172
00:05:16,160 --> 00:05:18,440
Yeah, like inappropriate or harmful content.

173
00:05:18,440 --> 00:05:19,600
Exactly.

174
00:05:19,600 --> 00:05:22,840
OpenAI built this thing called a mitigation stack,

175
00:05:22,840 --> 00:05:25,240
which is basically a bunch of checks and balances

176
00:05:25,240 --> 00:05:26,400
for every request.

177
00:05:26,400 --> 00:05:27,680
A mitigation stack, huh?

178
00:05:27,680 --> 00:05:29,040
Okay, so walk me through it.

179
00:05:29,040 --> 00:05:31,760
So first, Sora runs your request

180
00:05:31,760 --> 00:05:34,160
through a text and image moderation system.

181
00:05:34,160 --> 00:05:35,000
Gotcha.

182
00:05:35,000 --> 00:05:37,760
It uses AI models and lists of banned words

183
00:05:37,760 --> 00:05:39,880
to find and block any requests

184
00:05:39,880 --> 00:05:43,040
that are trying to create harmful or unwanted content.

185
00:05:43,040 --> 00:05:44,520
So it's like a security guard at the door

186
00:05:44,520 --> 00:05:47,120
checking IDs and making sure nothing dangerous gets in.

187
00:05:47,120 --> 00:05:47,960
Yeah.

188
00:05:47,960 --> 00:05:49,520
And then if something does slip through,

189
00:05:49,520 --> 00:05:51,200
there's another layer of protection

190
00:05:51,200 --> 00:05:52,800
called output classifiers.

191
00:05:52,800 --> 00:05:54,160
Output classifiers, what are those?

192
00:05:54,160 --> 00:05:56,840
So these are specifically designed to catch things

193
00:05:56,840 --> 00:06:00,520
like NSFW content depictions of minors' violence

194
00:06:00,520 --> 00:06:03,200
or even attempts to misuse someone's likeness.

195
00:06:03,200 --> 00:06:04,120
Oh, wow.

196
00:06:04,120 --> 00:06:06,000
And if any of those red flags pop up,

197
00:06:06,000 --> 00:06:08,760
Sora can block the video before it's ever shared.

198
00:06:08,760 --> 00:06:10,440
So it's not just about what goes in,

199
00:06:10,440 --> 00:06:12,920
it's about carefully checking what comes out too.

200
00:06:12,920 --> 00:06:13,760
Exactly.

201
00:06:13,760 --> 00:06:14,760
It's a pretty thorough system.

202
00:06:14,760 --> 00:06:15,600
It really is.

203
00:06:15,600 --> 00:06:16,440
Yeah.

204
00:06:16,440 --> 00:06:17,480
But are there any other safeguards?

205
00:06:17,480 --> 00:06:19,360
Technology changes so fast

206
00:06:19,360 --> 00:06:21,640
and people can be really creative.

207
00:06:21,640 --> 00:06:22,480
You know?

208
00:06:22,480 --> 00:06:23,320
Oh, totally.

209
00:06:23,320 --> 00:06:24,400
And that's why OpenAI also uses something

210
00:06:24,400 --> 00:06:26,480
called custom LLM filtering.

211
00:06:26,480 --> 00:06:27,600
LLM filtering.

212
00:06:27,600 --> 00:06:28,440
Yeah.

213
00:06:28,440 --> 00:06:30,560
It allows for really precise moderation

214
00:06:30,560 --> 00:06:32,120
on specific topics.

215
00:06:32,120 --> 00:06:32,960
Like what?

216
00:06:32,960 --> 00:06:34,520
Well, for example, it can spot attempts

217
00:06:34,520 --> 00:06:36,280
to use copyrighted material,

218
00:06:36,280 --> 00:06:38,320
create misleading content,

219
00:06:38,320 --> 00:06:42,000
or even videos that infringe on someone's personal rights.

220
00:06:42,000 --> 00:06:43,840
So it's not just matching keywords,

221
00:06:43,840 --> 00:06:46,240
it's understanding the context and the intent.

222
00:06:46,240 --> 00:06:47,160
Exactly.

223
00:06:47,160 --> 00:06:48,000
That's pretty amazing.

224
00:06:48,000 --> 00:06:49,680
But even with all these technical safeguards,

225
00:06:49,680 --> 00:06:52,880
there's always a chance someone will try to misuse Sora.

226
00:06:52,880 --> 00:06:54,560
Of course, you can't stop everyone.

227
00:06:54,560 --> 00:06:55,400
Right.

228
00:06:55,400 --> 00:06:56,360
So in addition to all the tech,

229
00:06:56,360 --> 00:07:00,600
they also have clear product policies and user education.

230
00:07:00,600 --> 00:07:03,440
So they tell people what's okay and what's not allowed.

231
00:07:03,440 --> 00:07:04,280
Yeah.

232
00:07:04,280 --> 00:07:05,200
They lay out exactly what is

233
00:07:05,200 --> 00:07:06,920
and isn't acceptable use of Sora.

234
00:07:06,920 --> 00:07:09,080
Things like creating harmful content,

235
00:07:09,080 --> 00:07:10,560
spreading misinformation,

236
00:07:10,560 --> 00:07:12,280
and misusing someone's likeness

237
00:07:12,280 --> 00:07:14,760
without their permission are all banned.

238
00:07:14,760 --> 00:07:17,000
So it's a combination of technical barriers

239
00:07:17,000 --> 00:07:18,560
and clear communication.

240
00:07:18,560 --> 00:07:19,480
That's the idea.

241
00:07:19,480 --> 00:07:20,360
Makes sense.

242
00:07:20,360 --> 00:07:22,760
And these policies aren't set in stone either.

243
00:07:22,760 --> 00:07:23,600
Oh, really?

244
00:07:23,600 --> 00:07:25,040
They've actually already made changes

245
00:07:25,040 --> 00:07:27,880
based on feedback from early access programs.

246
00:07:27,880 --> 00:07:29,840
So they're listening to users and adapting.

247
00:07:29,840 --> 00:07:30,680
Yeah.

248
00:07:30,680 --> 00:07:32,520
They know this technology is constantly evolving,

249
00:07:32,520 --> 00:07:34,280
so they need to be flexible.

250
00:07:34,280 --> 00:07:35,120
Well, that's good to hear.

251
00:07:35,120 --> 00:07:35,960
Uh-huh.

252
00:07:35,960 --> 00:07:37,040
It's reassuring that they're taking

253
00:07:37,040 --> 00:07:38,560
a learning-based approach.

254
00:07:38,560 --> 00:07:39,560
It is.

255
00:07:39,560 --> 00:07:41,400
But even with all these precautions,

256
00:07:41,400 --> 00:07:44,520
there are still potential risks, right?

257
00:07:44,520 --> 00:07:45,920
People could misuse Sora in ways

258
00:07:45,920 --> 00:07:47,040
we haven't even thought of yet.

259
00:07:47,040 --> 00:07:47,880
You're right.

260
00:07:47,880 --> 00:07:49,400
And OpenAI knows that.

261
00:07:49,400 --> 00:07:51,200
That's why they've been talking about some areas

262
00:07:51,200 --> 00:07:53,200
where they're focusing their future work.

263
00:07:53,200 --> 00:07:54,160
Okay, like what?

264
00:07:54,160 --> 00:07:55,440
Well, one of the most interesting

265
00:07:55,440 --> 00:07:57,360
is their likeness pilot program.

266
00:07:57,360 --> 00:07:58,880
Likeness pilot, what's that?

267
00:07:58,880 --> 00:08:01,880
It'll let a small group of users experiment

268
00:08:01,880 --> 00:08:05,600
with making videos using photos or videos

269
00:08:05,600 --> 00:08:07,880
of real people that they upload.

270
00:08:07,880 --> 00:08:08,720
Hold on.

271
00:08:08,720 --> 00:08:10,800
Animating a picture of your grandparents.

272
00:08:10,800 --> 00:08:11,720
Yeah, exactly.

273
00:08:11,720 --> 00:08:14,440
Or even a video featuring a historical figure.

274
00:08:14,440 --> 00:08:15,880
Okay, that does sound kind of cool.

275
00:08:15,880 --> 00:08:16,720
Right.

276
00:08:16,720 --> 00:08:18,320
The possibilities are pretty mind-blowing.

277
00:08:18,320 --> 00:08:20,840
But I can also see how that could be misused pretty easily.

278
00:08:20,840 --> 00:08:21,680
Oh, for sure.

279
00:08:21,680 --> 00:08:23,560
Like imagine someone making fake videos

280
00:08:23,560 --> 00:08:25,680
of celebrities or politicians.

281
00:08:25,680 --> 00:08:27,120
Yeah, that's a big concern,

282
00:08:27,120 --> 00:08:28,720
and that's why this pilot program

283
00:08:28,720 --> 00:08:30,200
will be really controlled

284
00:08:30,200 --> 00:08:32,280
with lots of monitoring and evaluation.

285
00:08:32,280 --> 00:08:34,320
So they want to figure out how people use it

286
00:08:34,320 --> 00:08:35,880
and make sure they have the right safeguards

287
00:08:35,880 --> 00:08:37,520
before releasing it more broadly.

288
00:08:37,520 --> 00:08:38,360
Exactly.

289
00:08:38,360 --> 00:08:39,200
Makes sense.

290
00:08:39,200 --> 00:08:40,040
Better safety, sorry.

291
00:08:40,040 --> 00:08:40,880
Right.

292
00:08:40,880 --> 00:08:42,560
So what other areas are they focusing on?

293
00:08:42,560 --> 00:08:45,440
Another big one is provenance and transparency.

294
00:08:45,440 --> 00:08:46,920
Provenance and transparency.

295
00:08:46,920 --> 00:08:49,320
Yeah, basically they want to make it easy to track

296
00:08:49,320 --> 00:08:52,520
where Sora content came from, so you know it's not fake.

297
00:08:52,520 --> 00:08:54,760
So like a digital trail for the video?

298
00:08:54,760 --> 00:08:55,680
Exactly.

299
00:08:55,680 --> 00:08:58,320
And to do this, they're going to use things like

300
00:08:58,320 --> 00:09:01,360
C2PA metadata, which is a standard

301
00:09:01,360 --> 00:09:03,840
for verifying the origin of digital content.

302
00:09:03,840 --> 00:09:04,680
Okay.

303
00:09:04,680 --> 00:09:07,360
And visible watermarks, so you know right away

304
00:09:07,360 --> 00:09:09,120
that a video is made with Sora.

305
00:09:09,120 --> 00:09:12,440
So even if someone tries to pass it off as real,

306
00:09:12,440 --> 00:09:14,320
there'll be ways to tell it's AI generated.

307
00:09:14,320 --> 00:09:15,160
That's so cool.

308
00:09:15,160 --> 00:09:15,980
Smart.

309
00:09:15,980 --> 00:09:17,320
Are there any other tools they're working on?

310
00:09:17,320 --> 00:09:19,120
Yeah, they're also building a special

311
00:09:19,120 --> 00:09:20,840
reverse video search tool.

312
00:09:20,840 --> 00:09:21,680
Okay, what's that?

313
00:09:21,680 --> 00:09:23,920
So their team can quickly figure out

314
00:09:23,920 --> 00:09:25,720
if a video is made with Sora.

315
00:09:25,720 --> 00:09:27,840
Interesting, so it's like a special search engine

316
00:09:27,840 --> 00:09:29,120
for Sora videos.

317
00:09:29,120 --> 00:09:29,960
Pretty much.

318
00:09:29,960 --> 00:09:32,800
This will be super helpful for spotting

319
00:09:32,800 --> 00:09:36,120
and dealing with any misuse or misinformation.

320
00:09:36,120 --> 00:09:38,360
Wow, that sounds like a really powerful tool.

321
00:09:38,360 --> 00:09:39,200
It is.

322
00:09:39,200 --> 00:09:41,840
But it's not just about the tech, right?

323
00:09:41,840 --> 00:09:43,880
Their user policies also matter.

324
00:09:43,880 --> 00:09:46,000
Absolutely, their policies clearly say that

325
00:09:46,000 --> 00:09:48,360
Sora can't be used for deceptive purposes.

326
00:09:48,360 --> 00:09:49,200
Okay.

327
00:09:49,200 --> 00:09:52,200
Things like creating disinformation, impersonating others,

328
00:09:52,200 --> 00:09:55,200
or misrepresenting content are all for bin.

329
00:09:55,200 --> 00:09:57,720
So they're setting clear boundaries and consequences

330
00:09:57,720 --> 00:10:01,520
for anyone who tries to use Sora for bad stuff.

331
00:10:01,520 --> 00:10:02,360
Right.

332
00:10:02,360 --> 00:10:04,320
Well things like they've thought of pretty much everything.

333
00:10:04,320 --> 00:10:06,280
They've definitely put a lot of thought into it.

334
00:10:06,280 --> 00:10:07,840
But what about artists?

335
00:10:07,840 --> 00:10:10,400
Had they addressed any concerns about people using Sora

336
00:10:10,400 --> 00:10:12,640
to copy a specific artist's work?

337
00:10:12,640 --> 00:10:14,560
That's a great question and they have.

338
00:10:14,560 --> 00:10:17,080
So whenever you put the name of a living artist

339
00:10:17,080 --> 00:10:20,440
in your prompt, Sora will actually rewrite it

340
00:10:20,440 --> 00:10:22,480
to avoid copying their style.

341
00:10:22,480 --> 00:10:23,320
Whoa, really?

342
00:10:23,320 --> 00:10:25,240
So it's like AI plagiarism protection?

343
00:10:25,240 --> 00:10:26,160
Yeah, basically.

344
00:10:26,160 --> 00:10:27,400
Okay, can you give me an example?

345
00:10:27,400 --> 00:10:28,320
Sure.

346
00:10:28,320 --> 00:10:31,200
Let's say you tell Sora to create a video

347
00:10:31,200 --> 00:10:33,120
in the style of famous artist,

348
00:10:33,120 --> 00:10:36,280
showing a cat writing a unicorn through a rainbow field.

349
00:10:36,280 --> 00:10:37,120
Okay.

350
00:10:37,120 --> 00:10:40,160
Instead of copying the artist's style directly,

351
00:10:40,160 --> 00:10:42,280
Sora's editor will change the prompt

352
00:10:42,280 --> 00:10:43,360
to something more general,

353
00:10:43,360 --> 00:10:47,000
like create a whimsical video of a cat writing a unicorn.

354
00:10:47,000 --> 00:10:49,600
So it takes out the specific artist's reference

355
00:10:49,600 --> 00:10:51,440
and makes it more open to interpretation?

356
00:10:51,440 --> 00:10:52,280
Exactly.

357
00:10:52,280 --> 00:10:53,560
That's a pretty clever solution.

358
00:10:53,560 --> 00:10:54,440
It is.

359
00:10:54,440 --> 00:10:55,880
But even with all these measures,

360
00:10:55,880 --> 00:10:57,720
there's always room for improvement, right?

361
00:10:57,720 --> 00:10:59,480
Of course, open AI has been very open

362
00:10:59,480 --> 00:11:01,840
about the fact that Sora is still being developed.

363
00:11:01,840 --> 00:11:03,360
Yeah, it would be kind of scary if they said

364
00:11:03,360 --> 00:11:04,440
it was perfect already.

365
00:11:04,440 --> 00:11:05,280
Right.

366
00:11:05,280 --> 00:11:07,200
So what are they focusing on for future updates?

367
00:11:07,200 --> 00:11:09,000
Yeah, they're planning to invest even more

368
00:11:09,000 --> 00:11:12,280
in those provenance and transparency tools,

369
00:11:12,280 --> 00:11:15,560
making sure it's easy to track where those videos come from

370
00:11:15,560 --> 00:11:17,200
and that they haven't been messed with.

371
00:11:17,200 --> 00:11:19,880
So they're doubling down on that digital trail idea.

372
00:11:19,880 --> 00:11:22,240
Exactly, and working with other groups

373
00:11:22,240 --> 00:11:23,760
to improve the whole system.

374
00:11:23,760 --> 00:11:24,840
So it's a team effort,

375
00:11:24,840 --> 00:11:26,760
not just open AI going it alone.

376
00:11:26,760 --> 00:11:28,960
Right, it's bigger than just one company.

377
00:11:28,960 --> 00:11:29,800
Good to hear.

378
00:11:29,800 --> 00:11:31,720
They're also working on better representation

379
00:11:31,720 --> 00:11:32,760
than Sora's videos.

380
00:11:32,760 --> 00:11:33,760
Oh, that's super important.

381
00:11:33,760 --> 00:11:35,680
Making sure the AI isn't biased

382
00:11:35,680 --> 00:11:37,640
and the videos reflect the real world.

383
00:11:37,640 --> 00:11:40,720
Yeah, we want AI to help fight bias, not make it worse.

384
00:11:40,720 --> 00:11:41,880
Exactly.

385
00:11:41,880 --> 00:11:44,400
And of course, safety policies, ethical alignment,

386
00:11:44,400 --> 00:11:45,760
that's all ongoing too.

387
00:11:45,760 --> 00:11:48,720
So always evaluating, getting feedback,

388
00:11:48,720 --> 00:11:50,560
making sure it's used for good.

389
00:11:50,560 --> 00:11:51,520
That's the goal.

390
00:11:51,520 --> 00:11:53,440
That makes me feel better, honestly.

391
00:11:53,440 --> 00:11:54,360
It's a lot to think about,

392
00:11:54,360 --> 00:11:56,160
but it seems like they're on the right track.

393
00:11:56,160 --> 00:11:59,280
They are, and it's easy to focus on the downsides,

394
00:11:59,280 --> 00:12:02,040
but let's not forget the amazing possibilities here.

395
00:12:02,040 --> 00:12:03,040
Oh yeah, for sure.

396
00:12:03,040 --> 00:12:07,400
So Sora could totally revolutionize so many industries.

397
00:12:07,400 --> 00:12:10,360
Like imagine making Hollywood level special effects

398
00:12:10,360 --> 00:12:11,880
without a huge budget.

399
00:12:11,880 --> 00:12:14,960
Or personalized educational videos for every student.

400
00:12:14,960 --> 00:12:17,800
Wow, and think about storytelling authors

401
00:12:17,800 --> 00:12:19,760
could bring their books to life.

402
00:12:19,760 --> 00:12:22,160
Journalists could make incredible documentaries.

403
00:12:22,160 --> 00:12:23,920
Every day people could share their stories

404
00:12:23,920 --> 00:12:25,160
in totally new ways.

405
00:12:25,160 --> 00:12:27,280
It feels like a whole new world of creativity

406
00:12:27,280 --> 00:12:28,120
is opening up.

407
00:12:28,120 --> 00:12:30,920
It is, but we can't forget about the responsibility

408
00:12:30,920 --> 00:12:31,760
that comes with it.

409
00:12:31,760 --> 00:12:34,320
Yeah, it's a bit of a tightrope walk, isn't it?

410
00:12:34,320 --> 00:12:37,640
We have this amazing tool, but we have to be careful with it.

411
00:12:37,640 --> 00:12:40,640
Yeah, like how do we protect artists and musicians,

412
00:12:40,640 --> 00:12:43,040
make sure AI isn't stealing their work?

413
00:12:43,040 --> 00:12:44,000
That's a big one for sure.

414
00:12:44,000 --> 00:12:47,040
We want innovation, but we also need to protect creators.

415
00:12:47,040 --> 00:12:48,160
Absolutely.

416
00:12:48,160 --> 00:12:49,080
And what about bias?

417
00:12:49,080 --> 00:12:50,360
We were talking about that earlier.

418
00:12:50,360 --> 00:12:53,160
How do we make sure Sora isn't learning bad habits

419
00:12:53,160 --> 00:12:54,800
from the data it's trained on?

420
00:12:54,800 --> 00:12:56,720
Yeah, that's crucial.

421
00:12:56,720 --> 00:12:58,720
We need to make sure the AI itself is fair

422
00:12:58,720 --> 00:13:01,000
and doesn't make inequality worse.

423
00:13:01,000 --> 00:13:04,120
Right, it's not just about stopping bad people from using it.

424
00:13:04,120 --> 00:13:06,680
It's about making sure the AI itself is ethical.

425
00:13:06,680 --> 00:13:07,720
Exactly.

426
00:13:07,720 --> 00:13:09,800
And as these videos get more and more realistic,

427
00:13:09,800 --> 00:13:12,280
we'll need to think about authenticity and trust.

428
00:13:12,280 --> 00:13:14,920
Yeah, how will we know what's real and what's fake?

429
00:13:14,920 --> 00:13:16,960
Right, and what does that mean for society

430
00:13:16,960 --> 00:13:19,000
if we can't trust what we see?

431
00:13:19,000 --> 00:13:20,600
Those are some pretty deep questions.

432
00:13:20,600 --> 00:13:22,520
They are, it's uncharted territory

433
00:13:22,520 --> 00:13:25,440
and it's gonna take a lot of careful thought to figure out.

434
00:13:25,440 --> 00:13:27,480
It's not just up to the tech companies either, right?

435
00:13:27,480 --> 00:13:28,720
We all have a role to play.

436
00:13:28,720 --> 00:13:29,600
Totally.

437
00:13:29,600 --> 00:13:32,520
Policymakers, ethicists, educators,

438
00:13:32,520 --> 00:13:34,480
the public, we all need to work together.

439
00:13:34,480 --> 00:13:37,040
This future of AI video generation,

440
00:13:37,040 --> 00:13:39,880
it's both exciting and a little scary.

441
00:13:39,880 --> 00:13:40,720
Yeah, yeah.

442
00:13:40,720 --> 00:13:41,560
It's a powerful tool

443
00:13:41,560 --> 00:13:43,560
and we need to make sure we use it wisely.

444
00:13:43,560 --> 00:13:44,840
Couldn't agree more.

445
00:13:44,840 --> 00:13:47,800
We need open conversations, careful planning

446
00:13:47,800 --> 00:13:50,400
and a real commitment to ethical development.

447
00:13:50,400 --> 00:13:51,240
Well said.

448
00:13:51,240 --> 00:13:54,800
I'm glad to see companies like OpenAI taking this seriously.

449
00:13:54,800 --> 00:13:56,000
It gives me hope for the future.

450
00:13:56,000 --> 00:13:57,080
Me too.

451
00:13:57,080 --> 00:13:59,200
It's great to see them prioritizing these issues.

452
00:13:59,200 --> 00:14:01,760
Yeah, it makes me think we can actually create a future

453
00:14:01,760 --> 00:14:03,560
where AI makes the world a better place.

454
00:14:03,560 --> 00:14:04,520
Absolutely.

455
00:14:04,520 --> 00:14:06,520
A future where it helps people be more creative

456
00:14:06,520 --> 00:14:08,280
and understand each other better.

457
00:14:08,280 --> 00:14:09,200
That's a great point.

458
00:14:09,200 --> 00:14:11,960
And to our listeners out there, stay curious,

459
00:14:11,960 --> 00:14:14,400
stay informed and be part of these conversations.

460
00:14:14,400 --> 00:14:16,960
The future of AI is something we'll shape together.

461
00:14:16,960 --> 00:14:18,280
Well, that's all the time we have

462
00:14:18,280 --> 00:14:20,600
for today's Deep Dive into Sora.

463
00:14:20,600 --> 00:14:21,560
Thanks for joining us.

464
00:14:21,560 --> 00:14:22,760
Thanks for having me.

465
00:14:22,760 --> 00:14:24,520
We'll be back next week with another look

466
00:14:24,520 --> 00:14:26,520
at the cutting edge of AI.

467
00:14:26,520 --> 00:14:55,520
Until then, keep exploring.