WEBVTT
1
00:00:00.091 --> 00:00:02.454
Hi, and welcome to Technically Creative's Week in Review.
2
00:00:02.934 --> 00:00:08.220
I'm Orlando Wood, founder of Kubrick Labs, an AI-first tech partner and software developer for creative companies.
3
00:00:08.420 --> 00:00:14.827
Each week, I take you through some of the stories that I think have the most impact on the increasingly interdependent worlds of technology and creativity.
4
00:00:15.428 --> 00:00:18.631
And this week, I'm wondering, is ad tech about to take over everything?
5
00:00:19.192 --> 00:00:19.632
Everything?
6
00:00:20.678 --> 00:00:27.008
I read a piece in The Guardian this week about AI threatening advertising's historic role as a training ground for filmmakers.
7
00:00:27.488 --> 00:00:33.958
And it annoyed me, because typically, I think everybody is looking at the same changes and somehow coming to all the wrong conclusions.
8
00:00:34.735 --> 00:00:43.723
Yes, advertising helped produce extraordinary directors, Ridley Scott, Tony Scott, hell, the entire Scott family, but also Guy Ritchie and Rupert Sanders.
9
00:00:44.263 --> 00:00:51.970
Commercials, especially in the UK, where filmmaking jobs are scarce, gave people money, crews, pressure, and they learned by doing.
10
00:00:52.350 --> 00:01:01.555
But can we stop behaving as though every 22-year-old filmmaker in 2026 is standing outside of a production company praying somebody will finally let them touch a camera?
11
00:01:01.895 --> 00:01:05.537
They've had professional cameras in their pocket from the age of eight years old.
12
00:01:06.058 --> 00:01:13.562
They have affordable programs for editing and color, sound, VFX, and global distribution I couldn't have dreamed of when I was at film school.
13
00:01:14.182 --> 00:01:17.223
Not to mention the generative tools that are improving daily.
14
00:01:17.824 --> 00:01:20.685
Oh, these poor budding filmmakers, however will they cope?
15
00:01:20.925 --> 00:01:21.545
Fuck's sake.
16
00:01:22.005 --> 00:01:26.147
They'll learn filmmaking the same way filmmakers always have, by making fucking movies.
17
00:01:26.467 --> 00:01:29.949
And the excuse for why you haven't been able to prove your talent is gone.
18
00:01:31.072 --> 00:01:33.333
But the Guardian article skipped that entirely.
19
00:01:33.513 --> 00:01:38.516
The more useful part of the Guardian article actually came from Richard Glasson, the CEO at WPP Production.
20
00:01:38.637 --> 00:01:47.942
And I know their persona non grata right now, but he says that now clients need vastly more content across vastly more channels, making a whole new class of production a requirement.
21
00:01:48.062 --> 00:01:53.746
I'm sure the production community saw Richard's name or WPP's logo and immediately set fire to the comments section.
22
00:01:54.026 --> 00:01:59.169
But if everybody just took a Xanax and put the pitchforks down, they'd notice something inconvenient.
23
00:01:59.649 --> 00:02:00.170
He's right.
24
00:02:00.990 --> 00:02:03.753
Demand has changed and production companies need to wake up.
25
00:02:04.134 --> 00:02:05.535
But we've heard this narrative before.
26
00:02:06.336 --> 00:02:07.157
I've said it before.
27
00:02:07.377 --> 00:02:09.620
And that's not actually what I'm going to focus on this week.
28
00:02:10.240 --> 00:02:12.503
Because the real story is ad tech.
29
00:02:13.143 --> 00:02:21.553
Because although that human interest story in The Guardian makes everybody angry, the news everywhere, it should show you what's really happening.
30
00:02:22.209 --> 00:02:25.951
TikTok expanded AI-powered creative generation and campaign optimization.
31
00:02:26.371 --> 00:02:29.232
Meta is pushing generative creative deeper into its ad products.
32
00:02:29.753 --> 00:02:31.413
X is automating campaign creation.
33
00:02:31.774 --> 00:02:33.254
OpenAI is expanding ads.
34
00:02:33.995 --> 00:02:38.937
Carrot, NBCUniversal, Freewheel, and Newton are building agentic systems for media planning and buying.
35
00:02:39.297 --> 00:02:45.620
Omnicom Media and Rembrandt are using AI to identify places where brands can be inserted into finished streaming content.
36
00:02:45.981 --> 00:02:50.523
That's a lot of movement in one week, and if you line the announcements up, you can see the stack filling in.
37
00:02:51.323 --> 00:02:53.084
Ad tech is becoming creative tech.
38
00:02:53.804 --> 00:02:57.066
For years, ad tech got really good at answering one question.
39
00:02:57.546 --> 00:02:58.926
Who should see this ad?
40
00:02:59.567 --> 00:03:01.568
What this week's announcements added is another.
41
00:03:01.968 --> 00:03:04.649
What ad should exist for that person when they see it?
42
00:03:05.029 --> 00:03:08.190
Historically, targeting could move much faster than production.
43
00:03:08.391 --> 00:03:11.552
A platform could identify 10 fascinating audience segments.
44
00:03:11.872 --> 00:03:16.174
But the people who are angry still had to make 10 useful pieces of content that could be targeted.
45
00:03:16.655 --> 00:03:20.077
With AI, that constraint is starting to disappear for production.
46
00:03:20.497 --> 00:03:25.379
Creative generation, asset selection, campaign setup, and media buying are all moving closer together.
47
00:03:25.740 --> 00:03:30.102
The response comes back, the system learns, the creative changes, and another version goes out.
48
00:03:30.222 --> 00:03:33.085
Production is becoming part of the technological loop.
49
00:03:33.786 --> 00:03:46.421
And when one system can understand the audience, help create the work, decide who sees it, distribute it, measure the response, and then generate the next version, calling this ad tech a media platform starts to feel a bit like calling Amazon a bookstore.
50
00:03:46.601 --> 00:03:49.605
It is technically true, but it's missing quite a lot of the story.
51
00:03:49.985 --> 00:03:51.426
So we'll start with OpenAI.
52
00:03:51.646 --> 00:04:01.574
OpenAI expanded its advertising business this week, adding more formats inside GBT and more of the measurement infrastructure advertisers need to work out whether those ads actually did anything.
53
00:04:02.075 --> 00:04:04.877
That matters because GBT isn't a social media feed.
54
00:04:05.317 --> 00:04:12.924
People use it to research, compare, plan, make life and business decisions, and crucially, explain what they're trying to do in complete sentences.
55
00:04:12.944 --> 00:04:13.544
Right.
56
00:04:13.664 --> 00:04:17.466
Google became incredibly valuable because search queries reveal intent.
57
00:04:18.027 --> 00:04:21.149
Meta became incredibly valuable because behavior reveals intent.
58
00:04:21.729 --> 00:04:26.772
Conversational AI introduces something richer, people explaining the problem itself.
59
00:04:26.912 --> 00:04:28.894
And Meta's going to the same place OpenAI is.
60
00:04:29.554 --> 00:04:35.937
Historically, Facebook might infer that I'm shopping for a car because I watched three reviews and visited a dealership website.
61
00:04:36.457 --> 00:04:44.000
Muse can have a much richer context because I may have explicitly given it access to things like my email, calendars, and files.
62
00:04:44.540 --> 00:04:53.484
It could know that I emailed my wife saying that the lease is up in March, that soccer practice is on the calendar three nights a week, and that yesterday I complained that our trunk was too small.
63
00:04:54.204 --> 00:04:58.805
Meta hasn't yet connected all of that Muse context into its advertising machine yet, but come on.
64
00:04:59.385 --> 00:05:02.326
This is one of the largest advertising businesses in human history.
65
00:05:02.666 --> 00:05:06.367
Do we really think nobody in Menlo Park has noticed what they've built?
66
00:05:06.567 --> 00:05:11.628
Just two weeks ago, Meta built an agent that users are inviting into the most intimate parts of their lives.
67
00:05:12.088 --> 00:05:13.168
Date night on the calendar.
68
00:05:13.429 --> 00:05:14.849
Here's an ad for contraception.
69
00:05:15.249 --> 00:05:16.549
Fight with the wife over email.
70
00:05:16.789 --> 00:05:18.010
Here's an ad for a dating app.
71
00:05:18.530 --> 00:05:20.710
Email from the bank saying your mortgage payment bounced.
72
00:05:21.050 --> 00:05:23.851
Suddenly, every predatory lender in America would love to help you.
73
00:05:24.651 --> 00:05:25.372
And this is meta.
74
00:05:25.672 --> 00:05:27.033
We've seen this movie before.
75
00:05:27.513 --> 00:05:33.778
The privacy boundary is always terribly important right up until somebody discovers a commercially useful reason for it not to be.
76
00:05:34.338 --> 00:05:36.300
Then comes the update that everybody misses.
77
00:05:36.640 --> 00:05:40.663
We just slightly changed our terms and conditions to state that we own everything.
78
00:05:40.983 --> 00:05:41.443
No biggie.
79
00:05:41.683 --> 00:05:46.567
Then everybody accepts it and continues on with their lives and nobody's read page 37 of the T's and C's.
80
00:05:47.087 --> 00:05:49.089
And OpenAI is already doing this natively.
81
00:05:49.209 --> 00:05:51.110
People open up to OpenAI.
82
00:05:51.371 --> 00:05:52.331
They don't just search.
83
00:05:52.772 --> 00:05:57.716
They tell it what they're worried about, what they're trying to buy, but also their personal problems.
84
00:05:57.976 --> 00:05:59.637
The combination is what I'm watching.
85
00:05:59.797 --> 00:06:02.620
Advertising has spent a century trying to understand what people want.
86
00:06:03.000 --> 00:06:05.542
These systems may increasingly have people simply telling them.
87
00:06:05.722 --> 00:06:11.207
That's an extraordinary amount of intent sitting very close to an extraordinary amount of advertising machinery.
88
00:06:11.567 --> 00:06:19.195
Just last week, Carrot, NBCUniversal, Freewheel, and Newton announced work around agentic television planning, buying, and optimization.
89
00:06:19.715 --> 00:06:26.902
Instead of software reporting what happened after a campaign, agents are increasingly being designed to participate in the decisions themselves.
90
00:06:27.483 --> 00:06:31.007
And almost prophetically, I brought up this exact example last week—
91
00:06:31.527 --> 00:06:33.908
The advertiser has an agent deciding what media to buy.
92
00:06:34.148 --> 00:06:36.570
The publisher has an agent deciding what inventory to sell.
93
00:06:36.930 --> 00:06:40.272
The customer may eventually have an agent deciding what product they should purchase.
94
00:06:40.612 --> 00:06:45.375
So at some point, your AI is negotiating with Procter & Gamble's AI about which toothpaste you should buy.
95
00:06:45.675 --> 00:06:48.236
And now AI agents are haggling over our teeth.
96
00:06:48.676 --> 00:06:54.700
Creation, buying, distribution, optimization, and measurement used to sit in different departments, agencies, and companies.
97
00:06:55.100 --> 00:07:00.083
This week's announcement saw those functions beginning to collapse into one continuous technological system.
98
00:07:00.723 --> 00:07:05.846
And probably the most irritating and dystopian one comes from Omnicom Media and Rembrandt.
99
00:07:06.146 --> 00:07:10.849
Because this is where advertisers will actually change the media that you made.
100
00:07:11.210 --> 00:07:21.316
They've announced technology that scans finished streaming content and identifies places where brands could be digitally inserted after the program has already been produced.
101
00:07:21.656 --> 00:07:28.160
This is, of course, very clever, but it's also completely bonkers for anybody who's creative and cares about the thing that they just produced.
102
00:07:28.460 --> 00:07:47.850
Traditionally, if James Bond drinks a Heineken Somebody does that deal and approves it with the production and the director and everybody Before the bottle arrives on set Bond holds it attractively Everybody invoices everybody And we move on Now the movie can be finished And software can wander through it thinking lovely table I could sell that
103
00:07:49.098 --> 00:07:51.700
It immediately creates about 100 different questions.
104
00:07:52.061 --> 00:07:53.502
Does everybody see the same brand?
105
00:07:53.762 --> 00:07:55.284
Could my Coke be your Pepsi?
106
00:07:55.624 --> 00:07:58.607
Can 20-year-old movie acquire a 2027 MacBook?
107
00:07:58.927 --> 00:08:00.028
Does the filmmaker approve?
108
00:08:00.048 --> 00:08:00.809
Does the actor?
109
00:08:01.089 --> 00:08:02.771
And when is the film even finished?
110
00:08:02.851 --> 00:08:09.597
I mean, are we taking On Her Majesty's Secret Service and constantly inputting new products every time it plays out?
111
00:08:09.777 --> 00:08:13.000
Well, then the movie's never done, and it's now a dynamic billboard.
112
00:08:13.320 --> 00:08:20.345
These are the reasons why the conversations that the Guardians is framing is so little, so late, and completely missing the point now.
113
00:08:20.966 --> 00:08:22.827
It's too small, that conversation.
114
00:08:22.968 --> 00:08:25.690
The next Ridley Scott will be fine from a filmmaking perspective.
115
00:08:25.710 --> 00:08:26.911
He'll just make the film.
116
00:08:27.251 --> 00:08:37.139
The question is what sort of power he has against the ad tech that's coming to change it during production, after production, and on every single playout.
117
00:08:37.299 --> 00:08:40.161
What I'm watching is what companies are owning the creative loop.
118
00:08:40.742 --> 00:08:42.523
The platform's already won distribution.
119
00:08:42.603 --> 00:08:43.624
They know the audience.
120
00:08:43.704 --> 00:08:44.605
They own the feed.
121
00:08:45.125 --> 00:08:46.246
They measure the response.
122
00:08:46.646 --> 00:08:55.974
Now, increasingly, they want to generate the creative, choose the asset, target it, serve it, optimize it, measure what happens next, and then reenact the feedback loop so they can improve it.
123
00:08:57.331 --> 00:09:00.752
But if you're a creative, is that really an improvement on your work?
124
00:09:01.153 --> 00:09:05.094
That's an extraordinary amount of the advertising value chain collapsing into one place.
125
00:09:05.434 --> 00:09:10.556
Clients will love some of this because it'll be quick, measurable, personalized, and incredibly convenient.
126
00:09:11.477 --> 00:09:13.518
So does the value move for everybody else?
127
00:09:13.698 --> 00:09:14.658
Probably not volume.
128
00:09:15.038 --> 00:09:16.399
Meta can beat you at volume.
129
00:09:16.799 --> 00:09:17.879
Probably not variations.
130
00:09:17.939 --> 00:09:19.980
TikTok can produce more versions than you can.
131
00:09:20.741 --> 00:09:22.081
Certainly not optimization.
132
00:09:22.121 --> 00:09:23.622
That's the machine's whole fucking job.
133
00:09:24.503 --> 00:09:28.065
which pushes value back towards the difficult bit, what's actually worth making.
134
00:09:28.265 --> 00:09:34.569
But it's not like people aren't going to start asking AI what it thinks about that too, because they'll hook the data up to the green light button.
135
00:09:34.749 --> 00:09:36.150
That's where I'm most concerned.
136
00:09:36.670 --> 00:09:40.653
The first internet disruption shifted incredible power towards distribution.
137
00:09:40.893 --> 00:09:44.455
This one is going to merge distribution and production into the same machine.
138
00:09:44.815 --> 00:09:50.059
So let's stop having these pearl-clutching conversations and get in the game, because there's a mess of crap coming.
139
00:09:50.719 --> 00:09:52.540
And we need all our creatives now more than ever.
140
00:09:52.820 --> 00:09:59.285
So in other nudes from the worlds of tech and creativity, publishers just discovered that AI apparently isn't just for ruining books.
141
00:09:59.925 --> 00:10:04.428
There's been plenty of righteous discussion in publishing about what authors should and shouldn't do with AI.
142
00:10:04.488 --> 00:10:04.928
Fair enough.
143
00:10:05.409 --> 00:10:09.512
But reporting this week gave us a look inside HarperCollins, Simon Schuster, and Hatchet.
144
00:10:09.972 --> 00:10:11.873
And shock horror, they're using it too.
145
00:10:12.353 --> 00:10:17.337
Marketing, publicity, internal communication, workflow analysis, automation, and some design support.
146
00:10:17.537 --> 00:10:17.637
Okay.
147
00:10:17.817 --> 00:10:19.378
I don't think that makes anybody a hypocrite.
148
00:10:19.398 --> 00:10:21.820
It makes the distinction we keep trying to make more important.
149
00:10:22.301 --> 00:10:26.364
Using AI to run a publishing company isn't the same thing as using AI to write the fucking book.
150
00:10:26.824 --> 00:10:29.486
If AI can make royalty processing less painful, excellent.
151
00:10:29.967 --> 00:10:34.050
Nobody went into literature because of a deep artistic commitment to a spreadsheet reconciliation.
152
00:10:34.270 --> 00:10:36.832
But where things get more complicated is workplace analysis.
153
00:10:37.172 --> 00:10:38.493
How can we make your job easier?
154
00:10:38.553 --> 00:10:39.934
What bits of your job can we remove?
155
00:10:40.054 --> 00:10:41.635
Our cousins, but they aren't twins.
156
00:10:42.295 --> 00:10:43.796
This is why it's a governance story.
157
00:10:43.896 --> 00:10:52.522
Use the tools, automate the boring crap, but let's be clear about what's being monitored, where author's material is going, and whether efficiency means giving somebody a better tool or removing the somebody.
158
00:10:53.042 --> 00:10:55.604
And then Audible would like every book to talk to you.
159
00:10:56.104 --> 00:10:58.586
Audible laid out an ambition this week that I quite like.
160
00:10:59.067 --> 00:11:02.470
Effectively every book, in every language, available in audio.
161
00:11:02.830 --> 00:11:06.273
Most books don't get audiobooks for a wonderfully unromantic reason.
162
00:11:06.813 --> 00:11:07.073
Money.
163
00:11:07.734 --> 00:11:12.157
A narrator, producer, studio, engineer, edit, and master all cost something.
164
00:11:12.658 --> 00:11:15.140
That's easy enough to justify for the new Stephen King.
165
00:11:15.340 --> 00:11:21.485
It gets trickier for a complete history of Belgian drainage infrastructure, currently being enjoyed by Colin in Shropshire.
166
00:11:22.452 --> 00:11:24.033
AI can move the economic line.
167
00:11:24.073 --> 00:11:29.515
If synthetic narration replaces a viable human production simply because it's cheaper, that's a labor issue.
168
00:11:29.655 --> 00:11:34.237
But for millions of books, the comparison isn't human narrator versus AI narrator.
169
00:11:34.337 --> 00:11:35.918
It's audiobook versus nothing.
170
00:11:36.918 --> 00:11:40.100
Zero narrator jobs before, zero narrator jobs after.
171
00:11:40.160 --> 00:11:41.941
Except now, somebody can listen to the book.
172
00:11:42.321 --> 00:11:49.144
Backlists become useful, more languages become viable, people with sight impairments or reading difficulties gain access to more work.
173
00:11:50.607 --> 00:11:55.529
And then Sony's fighting back, and it's now killed 260,000 fake AI songs.
174
00:11:56.170 --> 00:12:03.493
Sony Music says it's submitted more than 260,000 takedown requests involving AI-generated tracks impersonating its artists.
175
00:12:04.233 --> 00:12:07.615
And sure, I don't know whether to call this copyright enforcement or pest control.
176
00:12:07.795 --> 00:12:11.978
Underneath the ridiculous number, the AI music market is becoming more sensible.
177
00:12:12.399 --> 00:12:18.723
On the one side, you have licensed generative music with companies like Suno developing relationships with majors rights holders.
178
00:12:19.184 --> 00:12:22.786
On the other, you have people generating fake songs and passing them off as real artists.
179
00:12:23.347 --> 00:12:24.968
Those are plainly not the same thing.
180
00:12:25.208 --> 00:12:31.733
We're moving past the magnificently useless question, is AI music good or bad, and toward questions the industry can actually answer.
181
00:12:32.053 --> 00:12:32.774
Was it licensed?
182
00:12:32.794 --> 00:12:33.775
Did the artist consent?
183
00:12:34.355 --> 00:12:35.696
Are you impersonating somebody?
184
00:12:35.896 --> 00:12:36.537
And who owns it?
185
00:12:37.197 --> 00:12:39.961
Legitimate generative music needs licensing infrastructure.
186
00:12:40.081 --> 00:12:41.883
Fake Beyonce needs a takedown button.
187
00:12:42.624 --> 00:12:43.084
Progress!
188
00:12:44.616 --> 00:12:52.322
Asus announced last week a new ProArt Creator hardware that's designed to run more powerful models and agentic workflows directly on the workstation.
189
00:12:52.802 --> 00:12:57.685
And that sounds spectacularly boring until you run a creative company, and then it becomes extremely interesting.
190
00:12:58.326 --> 00:13:01.848
Creative companies are full of material you don't want casually leaving the building.
191
00:13:02.309 --> 00:13:06.612
Unreleased scripts and rough cuts, client strategy, source files, pitches.
192
00:13:07.152 --> 00:13:12.216
Cloud AI made the tools easy to access, but it also means information goes somewhere else to be processed.
193
00:13:12.236 --> 00:13:12.616
Okay.
194
00:13:12.776 --> 00:13:15.377
Local AI offers a beautifully simple alternative.
195
00:13:15.697 --> 00:13:17.057
The AI comes to the data.
196
00:13:17.317 --> 00:13:19.078
The data doesn't have to go to the AI.
197
00:13:19.518 --> 00:13:27.621
It won't replace giant cloud models for everything, but serious AI capabilities sitting on hardware you control could be hugely important for creative work.
198
00:13:27.881 --> 00:13:34.643
Nobody really gives a shit about a local inference until your unreleased Marvel script ends up on Reddit.
199
00:13:34.963 --> 00:13:36.903
Then everybody's a network security expert.
200
00:13:36.923 --> 00:13:38.364
We speak from personal experience.
201
00:13:38.444 --> 00:13:41.825
We work for big, large companies that handle very sensitive scripts.
202
00:13:41.945 --> 00:13:44.240
And very often they come to us bemused and angry.
203
00:13:45.245 --> 00:13:51.746
And I point them to usually my first, second, or third email to them where I say in big, bold letters, please let us upgrade your security.
204
00:13:51.886 --> 00:13:56.147
These AI models are very good at coming after your stuff.
205
00:13:56.947 --> 00:14:02.208
How many times do I have to say that we have been the victims of the largest copyright theft of all time?
206
00:14:02.428 --> 00:14:04.428
And now you're letting AI into your business.
207
00:14:05.108 --> 00:14:07.669
It's going to steal more shit if you don't protect it.
208
00:14:07.769 --> 00:14:11.850
So anyway, we like it when there are more options for keeping stuff more secure with AI.
209
00:14:11.950 --> 00:14:13.470
So that's a good piece of news.
210
00:14:13.730 --> 00:14:17.993
And then OpenAI is going to join in thropping and tattling on you for using AI.
211
00:14:18.213 --> 00:14:22.716
Because last week, OpenAI introduced invisible statistical watermarking for generated text.
212
00:14:22.896 --> 00:14:25.378
It's a useful idea, but it's a big, difficult problem.
213
00:14:25.678 --> 00:14:27.359
Text is incredibly easy to alter.
214
00:14:27.659 --> 00:14:29.480
Rewrite enough of it and the signal weakens.
215
00:14:29.840 --> 00:14:31.341
Short passages are harder to deal with.
216
00:14:31.622 --> 00:14:36.905
A human can take an AI draft, rewrite most of it, and leave us arguing over whether the final thing is AI generated at all.
217
00:14:37.365 --> 00:14:41.807
Of course, then the question becomes, why did you even need the AI if we were going to have to rewrite the whole thing?
218
00:14:41.927 --> 00:14:46.790
And even if the watermark works perfectly, it tells you something narrower than people will inevitably pretend it does.
219
00:14:47.310 --> 00:14:49.871
It may indicate AI contributed to a piece of text.
220
00:14:50.171 --> 00:14:55.414
It doesn't tell you who authored it, who owns it, how much of it was changed, or whether any of it is true.
221
00:14:55.914 --> 00:14:57.555
Providence is not authorship.
222
00:14:57.635 --> 00:15:04.018
But what it might do that will be useful is it'll stop kids at college getting blamed when they actually did write the thing.
223
00:15:04.398 --> 00:15:07.683
I run things that I write through AI and through AI detectors all the time.
224
00:15:07.823 --> 00:15:12.049
And you'd be shocked at how many things it says were written by AI that weren't.
225
00:15:12.229 --> 00:15:14.031
And kids are getting expelled for this.
226
00:15:14.472 --> 00:15:19.699
Because in A Tale of Two Cities, it was the best of times, it was the worst of times, is flagged as AI generated.
227
00:15:19.959 --> 00:15:21.361
But it's a step in the right direction.
228
00:15:21.982 --> 00:15:25.463
And then Nikon may have accidentally written the best AI policy yet.
229
00:15:25.763 --> 00:15:34.586
Nikon disqualified the original winner of its small world in motion microscopy competition after determining that AI-related manipulation violated the rules.
230
00:15:34.926 --> 00:15:35.226
Great!
231
00:15:36.147 --> 00:15:36.787
Sure, you...
232
00:15:37.347 --> 00:15:42.729
Go ahead and use AI to make a deranged music video featuring a flying octopus playing bass guitar.
233
00:15:43.029 --> 00:15:43.530
Go nuts.
234
00:15:44.190 --> 00:15:47.651
Use AI to alter imagery being presented as scientific microscopy?
235
00:15:48.211 --> 00:15:49.032
Absolutely not.
236
00:15:49.492 --> 00:15:51.173
Different context, different standard.
237
00:15:51.613 --> 00:15:56.595
We keep searching for one universal position on AI-generated content, and I don't think we're going to get one.
238
00:15:57.115 --> 00:16:00.236
Art is different from journalism, and journalism is different from advertising.
239
00:16:00.536 --> 00:16:02.477
Advertising is different from scientific evidence.
240
00:16:02.897 --> 00:16:04.978
What matters is the contract with the audience.
241
00:16:05.298 --> 00:16:07.239
What do they reasonably believe they're looking at?
242
00:16:07.679 --> 00:16:08.820
That isn't an inconsistency.
243
00:16:08.860 --> 00:16:10.100
That's how standards work.
244
00:16:10.560 --> 00:16:14.262
So we're all going to have to take a little bit of personal agency about how we do these things.
245
00:16:14.682 --> 00:16:16.804
This is something we say to our clients all the time.
246
00:16:17.164 --> 00:16:20.948
We are not here to tell you what your AI policy should be.
247
00:16:21.408 --> 00:16:26.553
We are here to give you guidance and ethics and tell you what SAG thinks and Equity thinks.
248
00:16:26.813 --> 00:16:30.297
Equity, by the way, still hasn't defined their position on AI, which I think is nuts.
249
00:16:31.407 --> 00:16:39.332
But we want to reflect the policy and the standards and the ethics that you want to reflect as a company.
250
00:16:39.472 --> 00:16:43.134
What we help clients do is define that for themselves because it's going to look different.
251
00:16:43.515 --> 00:16:47.417
We had Una Murphy from the Arts Council England on our podcast not too long ago.
252
00:16:47.977 --> 00:16:50.559
And I talked about AI policies and things like that.
253
00:16:50.579 --> 00:16:54.722
And she goes, it has to be different for every institution because every institution has a different mission.
254
00:16:55.422 --> 00:16:57.204
And that's how we feel about creative companies.
255
00:16:57.585 --> 00:16:59.567
Every creative company is out to do something different.
256
00:16:59.907 --> 00:17:01.589
And everybody's going to judge for themselves.
257
00:17:02.581 --> 00:17:05.483
where they're going to set their standards and where they're going to set their lines.
258
00:17:05.803 --> 00:17:10.846
So I think this is exactly what we want to see, and we want people to take a more active role in their AI policies.
259
00:17:11.066 --> 00:17:12.607
So anyway, that's it for this week.
260
00:17:12.847 --> 00:17:14.048
Thank you so much for listening.
261
00:17:14.308 --> 00:17:17.470
Tomorrow we have Eric Iverson, the CTO of UTA.
262
00:17:17.610 --> 00:17:18.831
He is absolutely great.
263
00:17:18.971 --> 00:17:24.774
I call him the rock star CTO of Hollywood because he is impossibly cool, and it turns out he wasn't a band, so that's pretty great.
264
00:17:25.174 --> 00:17:30.818
But yeah, it's rare that you find a chief technology officer who is quite as cool as Eric.
265
00:17:31.299 --> 00:17:33.291
Anyway, thank you.
266
00:17:33.331 --> 00:17:35.101
Please stick around and listen to that tomorrow.
267
00:17:35.987 --> 00:17:38.320
And we'll be back with another Week in Review next week.
268
00:17:38.522 --> 00:17:39.004
Thanks so much.
269
00:17:39.065 --> 00:17:39.306
Bye-bye.
1
00:00:00.091 --> 00:00:02.454
Hi, and welcome to Technically Creative's Week in Review.
2
00:00:02.934 --> 00:00:08.220
I'm Orlando Wood, founder of Kubrick Labs, an AI-first tech partner and software developer for creative companies.
3
00:00:08.420 --> 00:00:14.827
Each week, I take you through some of the stories that I think have the most impact on the increasingly interdependent worlds of technology and creativity.
4
00:00:15.428 --> 00:00:18.631
And this week, I'm wondering, is ad tech about to take over everything?
5
00:00:19.192 --> 00:00:19.632
Everything?
6
00:00:20.678 --> 00:00:27.008
I read a piece in The Guardian this week about AI threatening advertising's historic role as a training ground for filmmakers.
7
00:00:27.488 --> 00:00:33.958
And it annoyed me, because typically, I think everybody is looking at the same changes and somehow coming to all the wrong conclusions.
8
00:00:34.735 --> 00:00:43.723
Yes, advertising helped produce extraordinary directors, Ridley Scott, Tony Scott, hell, the entire Scott family, but also Guy Ritchie and Rupert Sanders.
9
00:00:44.263 --> 00:00:51.970
Commercials, especially in the UK, where filmmaking jobs are scarce, gave people money, crews, pressure, and they learned by doing.
10
00:00:52.350 --> 00:01:01.555
But can we stop behaving as though every 22-year-old filmmaker in 2026 is standing outside of a production company praying somebody will finally let them touch a camera?
11
00:01:01.895 --> 00:01:05.537
They've had professional cameras in their pocket from the age of eight years old.
12
00:01:06.058 --> 00:01:13.562
They have affordable programs for editing and color, sound, VFX, and global distribution I couldn't have dreamed of when I was at film school.
13
00:01:14.182 --> 00:01:17.223
Not to mention the generative tools that are improving daily.
14
00:01:17.824 --> 00:01:20.685
Oh, these poor budding filmmakers, however will they cope?
15
00:01:20.925 --> 00:01:21.545
Fuck's sake.
16
00:01:22.005 --> 00:01:26.147
They'll learn filmmaking the same way filmmakers always have, by making fucking movies.
17
00:01:26.467 --> 00:01:29.949
And the excuse for why you haven't been able to prove your talent is gone.
18
00:01:31.072 --> 00:01:33.333
But the Guardian article skipped that entirely.
19
00:01:33.513 --> 00:01:38.516
The more useful part of the Guardian article actually came from Richard Glasson, the CEO at WPP Production.
20
00:01:38.637 --> 00:01:47.942
And I know their persona non grata right now, but he says that now clients need vastly more content across vastly more channels, making a whole new class of production a requirement.
21
00:01:48.062 --> 00:01:53.746
I'm sure the production community saw Richard's name or WPP's logo and immediately set fire to the comments section.
22
00:01:54.026 --> 00:01:59.169
But if everybody just took a Xanax and put the pitchforks down, they'd notice something inconvenient.
23
00:01:59.649 --> 00:02:00.170
He's right.
24
00:02:00.990 --> 00:02:03.753
Demand has changed and production companies need to wake up.
25
00:02:04.134 --> 00:02:05.535
But we've heard this narrative before.
26
00:02:06.336 --> 00:02:07.157
I've said it before.
27
00:02:07.377 --> 00:02:09.620
And that's not actually what I'm going to focus on this week.
28
00:02:10.240 --> 00:02:12.503
Because the real story is ad tech.
29
00:02:13.143 --> 00:02:21.553
Because although that human interest story in The Guardian makes everybody angry, the news everywhere, it should show you what's really happening.
30
00:02:22.209 --> 00:02:25.951
TikTok expanded AI-powered creative generation and campaign optimization.
31
00:02:26.371 --> 00:02:29.232
Meta is pushing generative creative deeper into its ad products.
32
00:02:29.753 --> 00:02:31.413
X is automating campaign creation.
33
00:02:31.774 --> 00:02:33.254
OpenAI is expanding ads.
34
00:02:33.995 --> 00:02:38.937
Carrot, NBCUniversal, Freewheel, and Newton are building agentic systems for media planning and buying.
35
00:02:39.297 --> 00:02:45.620
Omnicom Media and Rembrandt are using AI to identify places where brands can be inserted into finished streaming content.
36
00:02:45.981 --> 00:02:50.523
That's a lot of movement in one week, and if you line the announcements up, you can see the stack filling in.
37
00:02:51.323 --> 00:02:53.084
Ad tech is becoming creative tech.
38
00:02:53.804 --> 00:02:57.066
For years, ad tech got really good at answering one question.
39
00:02:57.546 --> 00:02:58.926
Who should see this ad?
40
00:02:59.567 --> 00:03:01.568
What this week's announcements added is another.
41
00:03:01.968 --> 00:03:04.649
What ad should exist for that person when they see it?
42
00:03:05.029 --> 00:03:08.190
Historically, targeting could move much faster than production.
43
00:03:08.391 --> 00:03:11.552
A platform could identify 10 fascinating audience segments.
44
00:03:11.872 --> 00:03:16.174
But the people who are angry still had to make 10 useful pieces of content that could be targeted.
45
00:03:16.655 --> 00:03:20.077
With AI, that constraint is starting to disappear for production.
46
00:03:20.497 --> 00:03:25.379
Creative generation, asset selection, campaign setup, and media buying are all moving closer together.
47
00:03:25.740 --> 00:03:30.102
The response comes back, the system learns, the creative changes, and another version goes out.
48
00:03:30.222 --> 00:03:33.085
Production is becoming part of the technological loop.
49
00:03:33.786 --> 00:03:46.421
And when one system can understand the audience, help create the work, decide who sees it, distribute it, measure the response, and then generate the next version, calling this ad tech a media platform starts to feel a bit like calling Amazon a bookstore.
50
00:03:46.601 --> 00:03:49.605
It is technically true, but it's missing quite a lot of the story.
51
00:03:49.985 --> 00:03:51.426
So we'll start with OpenAI.
52
00:03:51.646 --> 00:04:01.574
OpenAI expanded its advertising business this week, adding more formats inside GBT and more of the measurement infrastructure advertisers need to work out whether those ads actually did anything.
53
00:04:02.075 --> 00:04:04.877
That matters because GBT isn't a social media feed.
54
00:04:05.317 --> 00:04:12.924
People use it to research, compare, plan, make life and business decisions, and crucially, explain what they're trying to do in complete sentences.
55
00:04:12.944 --> 00:04:13.544
Right.
56
00:04:13.664 --> 00:04:17.466
Google became incredibly valuable because search queries reveal intent.
57
00:04:18.027 --> 00:04:21.149
Meta became incredibly valuable because behavior reveals intent.
58
00:04:21.729 --> 00:04:26.772
Conversational AI introduces something richer, people explaining the problem itself.
59
00:04:26.912 --> 00:04:28.894
And Meta's going to the same place OpenAI is.
60
00:04:29.554 --> 00:04:35.937
Historically, Facebook might infer that I'm shopping for a car because I watched three reviews and visited a dealership website.
61
00:04:36.457 --> 00:04:44.000
Muse can have a much richer context because I may have explicitly given it access to things like my email, calendars, and files.
62
00:04:44.540 --> 00:04:53.484
It could know that I emailed my wife saying that the lease is up in March, that soccer practice is on the calendar three nights a week, and that yesterday I complained that our trunk was too small.
63
00:04:54.204 --> 00:04:58.805
Meta hasn't yet connected all of that Muse context into its advertising machine yet, but come on.
64
00:04:59.385 --> 00:05:02.326
This is one of the largest advertising businesses in human history.
65
00:05:02.666 --> 00:05:06.367
Do we really think nobody in Menlo Park has noticed what they've built?
66
00:05:06.567 --> 00:05:11.628
Just two weeks ago, Meta built an agent that users are inviting into the most intimate parts of their lives.
67
00:05:12.088 --> 00:05:13.168
Date night on the calendar.
68
00:05:13.429 --> 00:05:14.849
Here's an ad for contraception.
69
00:05:15.249 --> 00:05:16.549
Fight with the wife over email.
70
00:05:16.789 --> 00:05:18.010
Here's an ad for a dating app.
71
00:05:18.530 --> 00:05:20.710
Email from the bank saying your mortgage payment bounced.
72
00:05:21.050 --> 00:05:23.851
Suddenly, every predatory lender in America would love to help you.
73
00:05:24.651 --> 00:05:25.372
And this is meta.
74
00:05:25.672 --> 00:05:27.033
We've seen this movie before.
75
00:05:27.513 --> 00:05:33.778
The privacy boundary is always terribly important right up until somebody discovers a commercially useful reason for it not to be.
76
00:05:34.338 --> 00:05:36.300
Then comes the update that everybody misses.
77
00:05:36.640 --> 00:05:40.663
We just slightly changed our terms and conditions to state that we own everything.
78
00:05:40.983 --> 00:05:41.443
No biggie.
79
00:05:41.683 --> 00:05:46.567
Then everybody accepts it and continues on with their lives and nobody's read page 37 of the T's and C's.
80
00:05:47.087 --> 00:05:49.089
And OpenAI is already doing this natively.
81
00:05:49.209 --> 00:05:51.110
People open up to OpenAI.
82
00:05:51.371 --> 00:05:52.331
They don't just search.
83
00:05:52.772 --> 00:05:57.716
They tell it what they're worried about, what they're trying to buy, but also their personal problems.
84
00:05:57.976 --> 00:05:59.637
The combination is what I'm watching.
85
00:05:59.797 --> 00:06:02.620
Advertising has spent a century trying to understand what people want.
86
00:06:03.000 --> 00:06:05.542
These systems may increasingly have people simply telling them.
87
00:06:05.722 --> 00:06:11.207
That's an extraordinary amount of intent sitting very close to an extraordinary amount of advertising machinery.
88
00:06:11.567 --> 00:06:19.195
Just last week, Carrot, NBCUniversal, Freewheel, and Newton announced work around agentic television planning, buying, and optimization.
89
00:06:19.715 --> 00:06:26.902
Instead of software reporting what happened after a campaign, agents are increasingly being designed to participate in the decisions themselves.
90
00:06:27.483 --> 00:06:31.007
And almost prophetically, I brought up this exact example last week—
91
00:06:31.527 --> 00:06:33.908
The advertiser has an agent deciding what media to buy.
92
00:06:34.148 --> 00:06:36.570
The publisher has an agent deciding what inventory to sell.
93
00:06:36.930 --> 00:06:40.272
The customer may eventually have an agent deciding what product they should purchase.
94
00:06:40.612 --> 00:06:45.375
So at some point, your AI is negotiating with Procter & Gamble's AI about which toothpaste you should buy.
95
00:06:45.675 --> 00:06:48.236
And now AI agents are haggling over our teeth.
96
00:06:48.676 --> 00:06:54.700
Creation, buying, distribution, optimization, and measurement used to sit in different departments, agencies, and companies.
97
00:06:55.100 --> 00:07:00.083
This week's announcement saw those functions beginning to collapse into one continuous technological system.
98
00:07:00.723 --> 00:07:05.846
And probably the most irritating and dystopian one comes from Omnicom Media and Rembrandt.
99
00:07:06.146 --> 00:07:10.849
Because this is where advertisers will actually change the media that you made.
100
00:07:11.210 --> 00:07:21.316
They've announced technology that scans finished streaming content and identifies places where brands could be digitally inserted after the program has already been produced.
101
00:07:21.656 --> 00:07:28.160
This is, of course, very clever, but it's also completely bonkers for anybody who's creative and cares about the thing that they just produced.
102
00:07:28.460 --> 00:07:47.850
Traditionally, if James Bond drinks a Heineken Somebody does that deal and approves it with the production and the director and everybody Before the bottle arrives on set Bond holds it attractively Everybody invoices everybody And we move on Now the movie can be finished And software can wander through it thinking lovely table I could sell that
103
00:07:49.098 --> 00:07:51.700
It immediately creates about 100 different questions.
104
00:07:52.061 --> 00:07:53.502
Does everybody see the same brand?
105
00:07:53.762 --> 00:07:55.284
Could my Coke be your Pepsi?
106
00:07:55.624 --> 00:07:58.607
Can 20-year-old movie acquire a 2027 MacBook?
107
00:07:58.927 --> 00:08:00.028
Does the filmmaker approve?
108
00:08:00.048 --> 00:08:00.809
Does the actor?
109
00:08:01.089 --> 00:08:02.771
And when is the film even finished?
110
00:08:02.851 --> 00:08:09.597
I mean, are we taking On Her Majesty's Secret Service and constantly inputting new products every time it plays out?
111
00:08:09.777 --> 00:08:13.000
Well, then the movie's never done, and it's now a dynamic billboard.
112
00:08:13.320 --> 00:08:20.345
These are the reasons why the conversations that the Guardians is framing is so little, so late, and completely missing the point now.
113
00:08:20.966 --> 00:08:22.827
It's too small, that conversation.
114
00:08:22.968 --> 00:08:25.690
The next Ridley Scott will be fine from a filmmaking perspective.
115
00:08:25.710 --> 00:08:26.911
He'll just make the film.
116
00:08:27.251 --> 00:08:37.139
The question is what sort of power he has against the ad tech that's coming to change it during production, after production, and on every single playout.
117
00:08:37.299 --> 00:08:40.161
What I'm watching is what companies are owning the creative loop.
118
00:08:40.742 --> 00:08:42.523
The platform's already won distribution.
119
00:08:42.603 --> 00:08:43.624
They know the audience.
120
00:08:43.704 --> 00:08:44.605
They own the feed.
121
00:08:45.125 --> 00:08:46.246
They measure the response.
122
00:08:46.646 --> 00:08:55.974
Now, increasingly, they want to generate the creative, choose the asset, target it, serve it, optimize it, measure what happens next, and then reenact the feedback loop so they can improve it.
123
00:08:57.331 --> 00:09:00.752
But if you're a creative, is that really an improvement on your work?
124
00:09:01.153 --> 00:09:05.094
That's an extraordinary amount of the advertising value chain collapsing into one place.
125
00:09:05.434 --> 00:09:10.556
Clients will love some of this because it'll be quick, measurable, personalized, and incredibly convenient.
126
00:09:11.477 --> 00:09:13.518
So does the value move for everybody else?
127
00:09:13.698 --> 00:09:14.658
Probably not volume.
128
00:09:15.038 --> 00:09:16.399
Meta can beat you at volume.
129
00:09:16.799 --> 00:09:17.879
Probably not variations.
130
00:09:17.939 --> 00:09:19.980
TikTok can produce more versions than you can.
131
00:09:20.741 --> 00:09:22.081
Certainly not optimization.
132
00:09:22.121 --> 00:09:23.622
That's the machine's whole fucking job.
133
00:09:24.503 --> 00:09:28.065
which pushes value back towards the difficult bit, what's actually worth making.
134
00:09:28.265 --> 00:09:34.569
But it's not like people aren't going to start asking AI what it thinks about that too, because they'll hook the data up to the green light button.
135
00:09:34.749 --> 00:09:36.150
That's where I'm most concerned.
136
00:09:36.670 --> 00:09:40.653
The first internet disruption shifted incredible power towards distribution.
137
00:09:40.893 --> 00:09:44.455
This one is going to merge distribution and production into the same machine.
138
00:09:44.815 --> 00:09:50.059
So let's stop having these pearl-clutching conversations and get in the game, because there's a mess of crap coming.
139
00:09:50.719 --> 00:09:52.540
And we need all our creatives now more than ever.
140
00:09:52.820 --> 00:09:59.285
So in other nudes from the worlds of tech and creativity, publishers just discovered that AI apparently isn't just for ruining books.
141
00:09:59.925 --> 00:10:04.428
There's been plenty of righteous discussion in publishing about what authors should and shouldn't do with AI.
142
00:10:04.488 --> 00:10:04.928
Fair enough.
143
00:10:05.409 --> 00:10:09.512
But reporting this week gave us a look inside HarperCollins, Simon Schuster, and Hatchet.
144
00:10:09.972 --> 00:10:11.873
And shock horror, they're using it too.
145
00:10:12.353 --> 00:10:17.337
Marketing, publicity, internal communication, workflow analysis, automation, and some design support.
146
00:10:17.537 --> 00:10:17.637
Okay.
147
00:10:17.817 --> 00:10:19.378
I don't think that makes anybody a hypocrite.
148
00:10:19.398 --> 00:10:21.820
It makes the distinction we keep trying to make more important.
149
00:10:22.301 --> 00:10:26.364
Using AI to run a publishing company isn't the same thing as using AI to write the fucking book.
150
00:10:26.824 --> 00:10:29.486
If AI can make royalty processing less painful, excellent.
151
00:10:29.967 --> 00:10:34.050
Nobody went into literature because of a deep artistic commitment to a spreadsheet reconciliation.
152
00:10:34.270 --> 00:10:36.832
But where things get more complicated is workplace analysis.
153
00:10:37.172 --> 00:10:38.493
How can we make your job easier?
154
00:10:38.553 --> 00:10:39.934
What bits of your job can we remove?
155
00:10:40.054 --> 00:10:41.635
Our cousins, but they aren't twins.
156
00:10:42.295 --> 00:10:43.796
This is why it's a governance story.
157
00:10:43.896 --> 00:10:52.522
Use the tools, automate the boring crap, but let's be clear about what's being monitored, where author's material is going, and whether efficiency means giving somebody a better tool or removing the somebody.
158
00:10:53.042 --> 00:10:55.604
And then Audible would like every book to talk to you.
159
00:10:56.104 --> 00:10:58.586
Audible laid out an ambition this week that I quite like.
160
00:10:59.067 --> 00:11:02.470
Effectively every book, in every language, available in audio.
161
00:11:02.830 --> 00:11:06.273
Most books don't get audiobooks for a wonderfully unromantic reason.
162
00:11:06.813 --> 00:11:07.073
Money.
163
00:11:07.734 --> 00:11:12.157
A narrator, producer, studio, engineer, edit, and master all cost something.
164
00:11:12.658 --> 00:11:15.140
That's easy enough to justify for the new Stephen King.
165
00:11:15.340 --> 00:11:21.485
It gets trickier for a complete history of Belgian drainage infrastructure, currently being enjoyed by Colin in Shropshire.
166
00:11:22.452 --> 00:11:24.033
AI can move the economic line.
167
00:11:24.073 --> 00:11:29.515
If synthetic narration replaces a viable human production simply because it's cheaper, that's a labor issue.
168
00:11:29.655 --> 00:11:34.237
But for millions of books, the comparison isn't human narrator versus AI narrator.
169
00:11:34.337 --> 00:11:35.918
It's audiobook versus nothing.
170
00:11:36.918 --> 00:11:40.100
Zero narrator jobs before, zero narrator jobs after.
171
00:11:40.160 --> 00:11:41.941
Except now, somebody can listen to the book.
172
00:11:42.321 --> 00:11:49.144
Backlists become useful, more languages become viable, people with sight impairments or reading difficulties gain access to more work.
173
00:11:50.607 --> 00:11:55.529
And then Sony's fighting back, and it's now killed 260,000 fake AI songs.
174
00:11:56.170 --> 00:12:03.493
Sony Music says it's submitted more than 260,000 takedown requests involving AI-generated tracks impersonating its artists.
175
00:12:04.233 --> 00:12:07.615
And sure, I don't know whether to call this copyright enforcement or pest control.
176
00:12:07.795 --> 00:12:11.978
Underneath the ridiculous number, the AI music market is becoming more sensible.
177
00:12:12.399 --> 00:12:18.723
On the one side, you have licensed generative music with companies like Suno developing relationships with majors rights holders.
178
00:12:19.184 --> 00:12:22.786
On the other, you have people generating fake songs and passing them off as real artists.
179
00:12:23.347 --> 00:12:24.968
Those are plainly not the same thing.
180
00:12:25.208 --> 00:12:31.733
We're moving past the magnificently useless question, is AI music good or bad, and toward questions the industry can actually answer.
181
00:12:32.053 --> 00:12:32.774
Was it licensed?
182
00:12:32.794 --> 00:12:33.775
Did the artist consent?
183
00:12:34.355 --> 00:12:35.696
Are you impersonating somebody?
184
00:12:35.896 --> 00:12:36.537
And who owns it?
185
00:12:37.197 --> 00:12:39.961
Legitimate generative music needs licensing infrastructure.
186
00:12:40.081 --> 00:12:41.883
Fake Beyonce needs a takedown button.
187
00:12:42.624 --> 00:12:43.084
Progress!
188
00:12:44.616 --> 00:12:52.322
Asus announced last week a new ProArt Creator hardware that's designed to run more powerful models and agentic workflows directly on the workstation.
189
00:12:52.802 --> 00:12:57.685
And that sounds spectacularly boring until you run a creative company, and then it becomes extremely interesting.
190
00:12:58.326 --> 00:13:01.848
Creative companies are full of material you don't want casually leaving the building.
191
00:13:02.309 --> 00:13:06.612
Unreleased scripts and rough cuts, client strategy, source files, pitches.
192
00:13:07.152 --> 00:13:12.216
Cloud AI made the tools easy to access, but it also means information goes somewhere else to be processed.
193
00:13:12.236 --> 00:13:12.616
Okay.
194
00:13:12.776 --> 00:13:15.377
Local AI offers a beautifully simple alternative.
195
00:13:15.697 --> 00:13:17.057
The AI comes to the data.
196
00:13:17.317 --> 00:13:19.078
The data doesn't have to go to the AI.
197
00:13:19.518 --> 00:13:27.621
It won't replace giant cloud models for everything, but serious AI capabilities sitting on hardware you control could be hugely important for creative work.
198
00:13:27.881 --> 00:13:34.643
Nobody really gives a shit about a local inference until your unreleased Marvel script ends up on Reddit.
199
00:13:34.963 --> 00:13:36.903
Then everybody's a network security expert.
200
00:13:36.923 --> 00:13:38.364
We speak from personal experience.
201
00:13:38.444 --> 00:13:41.825
We work for big, large companies that handle very sensitive scripts.
202
00:13:41.945 --> 00:13:44.240
And very often they come to us bemused and angry.
203
00:13:45.245 --> 00:13:51.746
And I point them to usually my first, second, or third email to them where I say in big, bold letters, please let us upgrade your security.
204
00:13:51.886 --> 00:13:56.147
These AI models are very good at coming after your stuff.
205
00:13:56.947 --> 00:14:02.208
How many times do I have to say that we have been the victims of the largest copyright theft of all time?
206
00:14:02.428 --> 00:14:04.428
And now you're letting AI into your business.
207
00:14:05.108 --> 00:14:07.669
It's going to steal more shit if you don't protect it.
208
00:14:07.769 --> 00:14:11.850
So anyway, we like it when there are more options for keeping stuff more secure with AI.
209
00:14:11.950 --> 00:14:13.470
So that's a good piece of news.
210
00:14:13.730 --> 00:14:17.993
And then OpenAI is going to join in thropping and tattling on you for using AI.
211
00:14:18.213 --> 00:14:22.716
Because last week, OpenAI introduced invisible statistical watermarking for generated text.
212
00:14:22.896 --> 00:14:25.378
It's a useful idea, but it's a big, difficult problem.
213
00:14:25.678 --> 00:14:27.359
Text is incredibly easy to alter.
214
00:14:27.659 --> 00:14:29.480
Rewrite enough of it and the signal weakens.
215
00:14:29.840 --> 00:14:31.341
Short passages are harder to deal with.
216
00:14:31.622 --> 00:14:36.905
A human can take an AI draft, rewrite most of it, and leave us arguing over whether the final thing is AI generated at all.
217
00:14:37.365 --> 00:14:41.807
Of course, then the question becomes, why did you even need the AI if we were going to have to rewrite the whole thing?
218
00:14:41.927 --> 00:14:46.790
And even if the watermark works perfectly, it tells you something narrower than people will inevitably pretend it does.
219
00:14:47.310 --> 00:14:49.871
It may indicate AI contributed to a piece of text.
220
00:14:50.171 --> 00:14:55.414
It doesn't tell you who authored it, who owns it, how much of it was changed, or whether any of it is true.
221
00:14:55.914 --> 00:14:57.555
Providence is not authorship.
222
00:14:57.635 --> 00:15:04.018
But what it might do that will be useful is it'll stop kids at college getting blamed when they actually did write the thing.
223
00:15:04.398 --> 00:15:07.683
I run things that I write through AI and through AI detectors all the time.
224
00:15:07.823 --> 00:15:12.049
And you'd be shocked at how many things it says were written by AI that weren't.
225
00:15:12.229 --> 00:15:14.031
And kids are getting expelled for this.
226
00:15:14.472 --> 00:15:19.699
Because in A Tale of Two Cities, it was the best of times, it was the worst of times, is flagged as AI generated.
227
00:15:19.959 --> 00:15:21.361
But it's a step in the right direction.
228
00:15:21.982 --> 00:15:25.463
And then Nikon may have accidentally written the best AI policy yet.
229
00:15:25.763 --> 00:15:34.586
Nikon disqualified the original winner of its small world in motion microscopy competition after determining that AI-related manipulation violated the rules.
230
00:15:34.926 --> 00:15:35.226
Great!
231
00:15:36.147 --> 00:15:36.787
Sure, you...
232
00:15:37.347 --> 00:15:42.729
Go ahead and use AI to make a deranged music video featuring a flying octopus playing bass guitar.
233
00:15:43.029 --> 00:15:43.530
Go nuts.
234
00:15:44.190 --> 00:15:47.651
Use AI to alter imagery being presented as scientific microscopy?
235
00:15:48.211 --> 00:15:49.032
Absolutely not.
236
00:15:49.492 --> 00:15:51.173
Different context, different standard.
237
00:15:51.613 --> 00:15:56.595
We keep searching for one universal position on AI-generated content, and I don't think we're going to get one.
238
00:15:57.115 --> 00:16:00.236
Art is different from journalism, and journalism is different from advertising.
239
00:16:00.536 --> 00:16:02.477
Advertising is different from scientific evidence.
240
00:16:02.897 --> 00:16:04.978
What matters is the contract with the audience.
241
00:16:05.298 --> 00:16:07.239
What do they reasonably believe they're looking at?
242
00:16:07.679 --> 00:16:08.820
That isn't an inconsistency.
243
00:16:08.860 --> 00:16:10.100
That's how standards work.
244
00:16:10.560 --> 00:16:14.262
So we're all going to have to take a little bit of personal agency about how we do these things.
245
00:16:14.682 --> 00:16:16.804
This is something we say to our clients all the time.
246
00:16:17.164 --> 00:16:20.948
We are not here to tell you what your AI policy should be.
247
00:16:21.408 --> 00:16:26.553
We are here to give you guidance and ethics and tell you what SAG thinks and Equity thinks.
248
00:16:26.813 --> 00:16:30.297
Equity, by the way, still hasn't defined their position on AI, which I think is nuts.
249
00:16:31.407 --> 00:16:39.332
But we want to reflect the policy and the standards and the ethics that you want to reflect as a company.
250
00:16:39.472 --> 00:16:43.134
What we help clients do is define that for themselves because it's going to look different.
251
00:16:43.515 --> 00:16:47.417
We had Una Murphy from the Arts Council England on our podcast not too long ago.
252
00:16:47.977 --> 00:16:50.559
And I talked about AI policies and things like that.
253
00:16:50.579 --> 00:16:54.722
And she goes, it has to be different for every institution because every institution has a different mission.
254
00:16:55.422 --> 00:16:57.204
And that's how we feel about creative companies.
255
00:16:57.585 --> 00:16:59.567
Every creative company is out to do something different.
256
00:16:59.907 --> 00:17:01.589
And everybody's going to judge for themselves.
257
00:17:02.581 --> 00:17:05.483
where they're going to set their standards and where they're going to set their lines.
258
00:17:05.803 --> 00:17:10.846
So I think this is exactly what we want to see, and we want people to take a more active role in their AI policies.
259
00:17:11.066 --> 00:17:12.607
So anyway, that's it for this week.
260
00:17:12.847 --> 00:17:14.048
Thank you so much for listening.
261
00:17:14.308 --> 00:17:17.470
Tomorrow we have Eric Iverson, the CTO of UTA.
262
00:17:17.610 --> 00:17:18.831
He is absolutely great.
263
00:17:18.971 --> 00:17:24.774
I call him the rock star CTO of Hollywood because he is impossibly cool, and it turns out he wasn't a band, so that's pretty great.
264
00:17:25.174 --> 00:17:30.818
But yeah, it's rare that you find a chief technology officer who is quite as cool as Eric.
265
00:17:31.299 --> 00:17:33.291
Anyway, thank you.
266
00:17:33.331 --> 00:17:35.101
Please stick around and listen to that tomorrow.
267
00:17:35.987 --> 00:17:38.320
And we'll be back with another Week in Review next week.
268
00:17:38.522 --> 00:17:39.004
Thanks so much.
269
00:17:39.065 --> 00:17:39.306
Bye-bye.
