Tech CEO: Big AI’s monopoly is breaking | Daniela Braga, Defined.ai — Transcript
Full transcript
- 0:00everybody is starting to invest in
- 0:01smaller
- 0:03specialized models. I see a world where
- 0:06uh
- 0:07this decentralization is for for one
- 0:09offload the the the
- 0:11concentration
- 0:13in certain in a few powers in the in
- 0:16terms of AI generation and owning and
- 0:20the knowledge
- 0:21uh
- 0:22and it's and we see that all over the
- 0:24world.
- 0:26>> [music]
- 0:31>> Hi, I'm Isaac and today on the Tech
- 0:33Report and at Web Summit kind of I'm
- 0:36joined by CEO of Defined AI, Daniela
- 0:39Braga. Thanks for coming on.
- 0:40>> Hi Isaac. It's a pleasure to be here.
- 0:42>> So back in June, Cloudflare announced
- 0:45that bot traffic surpassed human traffic
- 0:47well ahead of its sort of mid 2027
- 0:49expectations.
- 0:51And around the same time, the New
- 0:52Scientist has reported that AI
- 0:54developers and AI data
- 0:57data set annotators
- 0:59admitted to using chatbots to help do
- 1:01their work.
- 1:02As a founder of a ethical training data
- 1:05platform, if I could put it that way,
- 1:07how real is the data wall, the data
- 1:09bottleneck?
- 1:11And how big of a threat is is model
- 1:13collapse?
- 1:14>> Hm.
- 1:15Right. So Elon Musk a year and a half
- 1:19ago, I think he did say that we were
- 1:21going to run out of data.
- 1:23Uh
- 1:24but that he meant out of
- 1:26open
- 1:28uh not not open. Uh he meant
- 1:32data that is out in the web that is
- 1:35publicly available, which doesn't mean
- 1:38free
- 1:39in terms of IP or copyrights.
- 1:42But we know that all the Frontier Labs
- 1:46have
- 1:47crawlers constantly uh
- 1:50constantly
- 1:51crawling the web and crawling the data.
- 1:54So in that regard,
- 1:56uh it is true that
- 1:59all the internet been
- 2:00crawled already
- 2:02and and there's a feed of remaining
- 2:06digital data that is continuing
- 2:08continues to be crawled
- 2:11but it's not very different from what is
- 2:14already there in the models.
- 2:16So
- 2:17in that regard, yes, there's a
- 2:19bottleneck of data of publicly available
- 2:22data
- 2:23but it but the real opportunity and
- 2:26differentiator is the closed source
- 2:29private data.
- 2:30That data is much bigger.
- 2:33It's just not so accessible. It's
- 2:35definitely not accessible. That's what
- 2:37we that's what our business model is
- 2:39about and our
- 2:42product is about is is basically
- 2:45making that available to the world but
- 2:48for a fee. It has to be paid.
- 2:51So that's the big difference. There's a
- 2:52third type of data which is synthetic
- 2:54data. That is absolutely
- 2:58augmenting
- 3:00I mean we know we know that our our
- 3:02clients are using a fair amount of
- 3:05synthetic data in their model.
- 3:07Typically synthetic data is the one that
- 3:11in the wrong proportions causes model
- 3:13collapse
- 3:15of course.
- 3:16In the right proportions is very
- 3:18important because you're going to be
- 3:19able to
- 3:21supplement
- 3:22scenarios that otherwise you can find in
- 3:24real data.
- 3:26But the world of data is specializing.
- 3:29Again, what you can find
- 3:31in the open is not going to give the
- 3:34edge to the frontier labs. And we're
- 3:36starting if in the beginning of this
- 3:38race, I mean two or three years ago, I
- 3:41remember giving a a talk at CES where I
- 3:43was showing the output of all the
- 3:44models. They were very similar.
- 3:47You can start seeing differentiation in
- 3:49the models. They are getting specialized
- 3:51in domains in verticals
- 3:53and this is because of the closed source
- 3:55data.
- 3:57>> A lot of attention around AI has been
- 3:59focused on sort of scaling and growth
- 4:01and more and more hardware to sort of
- 4:03make things better. But
- 4:05it seems to me and may I could be
- 4:07accused of being
- 4:08not the most informed on all of the
- 4:10scopes of AI.
- 4:12But it seems to me that the next phase
- 4:14of AI development might be fought more
- 4:16on sort of creating specialized AI tools
- 4:19trained on sort of very high quality
- 4:20data sets
- 4:22which have more of a defined use case.
- 4:24Thompson Reuters AI model comes to mind
- 4:26if you're familiar with it.
- 4:28What's your take on the the industry's
- 4:31focus on generalist models and the
- 4:33potential rise of specialist models with
- 4:35more sort of bespoke use cases?
- 4:38>> Definitely. I I mean even for from a
- 4:40cost perspective, specialized models is
- 4:43the way to go. Everybody is starting to
- 4:45invest in smaller
- 4:47specialized models with
- 4:50uh and we see that. We see
- 4:52uh so if
- 4:54three, four years ago our clients would
- 4:57come, it would be a handful of
- 5:00the typical frontier labs and they would
- 5:02come in and
- 5:04request large amounts of data
- 5:06and not necessarily even
- 5:09with well large amounts of data in
- 5:11in in in
- 5:13basically multi-model aspects of of data
- 5:16but not
- 5:18uh but things that were in big scale and
- 5:21more generic content and topics.
- 5:24Uh today we see a surge of other clients
- 5:28that are not the frontier labs but they
- 5:30are customizing
- 5:32open source models or closed source
- 5:34models.
- 5:35And they want smaller data sets but
- 5:38highly curated, highly curated with
- 5:41specialists, with humans in the loop
- 5:43that are doctors or mathematicians or or
- 5:47chemists or so this is this is what
- 5:50we're seeing and all all linguists in
- 5:53PhDs PhDs including.
- 5:56So this is we see
- 5:58we see that movement and uh it's
- 6:00definitely the way to go
- 6:02uh
- 6:03from
- 6:04from an industry perspective.
- 6:05>> And
- 6:06>> do open weight models fit into this this
- 6:09idea? 'Cause I mean,
- 6:11quite simply, I mean, admit you can
- 6:13correct me if I'm wrong here, smaller
- 6:15bespoke models sort of trained on
- 6:17tightly curated and
- 6:19even proprietary data is is surely going
- 6:21to impact the AI industry more broadly
- 6:24and I'm kind of thinking
- 6:25the sort of the the open AI's, the the M
- 6:28Tropics who are making these very big
- 6:30models, but what happens to them if the
- 6:32industry is moving already towards
- 6:35open weight sort of
- 6:37uh yeah, open weight models which they
- 6:39can tailor to their use cases.
- 6:41>> No, that exactly right. That's that's
- 6:43that's what we see in the industry.
- 6:45We see
- 6:46uh
- 6:47consulting companies, startups that are
- 6:50not
- 6:51for sure, they don't have budgets to
- 6:53build
- 6:54large scale
- 6:56models, world models, large language
- 6:59models, they're not going to build that.
- 7:00They definitely are are uh
- 7:03fine tuning open source models
- 7:05for their own purpose with specific
- 7:07hardware, with specific industry
- 7:09domains, with some private data.
- 7:11Uh private data that
- 7:13that can be theirs or or or acquired
- 7:16through us.
- 7:17So that's that's where we see the the
- 7:20trend. And
- 7:21to give you an idea, if we had been
- 7:24mostly serving in the last few years
- 7:27around
- 7:2850 massive
- 7:30Frontier Labs,
- 7:32the our list of clients exploded this
- 7:34year to
- 7:36over 150
- 7:38because this is the these are companies
- 7:40that are not coming from they are not AI
- 7:42native
- 7:44um and they are starting to build their
- 7:45own models.
- 7:47>> Mhm. So I'm I'm just really curious if I
- 7:48can just push you on that a little more.
- 7:50I'm wondering if you think that the the
- 7:53Frontier model developers are
- 7:55are already starting to in some ways
- 7:56lose lose some of their their piece of
- 7:58the pie, lose some ground to to this
- 8:01this idea of moving towards models which
- 8:03fit specific use cases to open weight
- 8:05models.
- 8:06>> I not yet. Actually,
- 8:08>> I see a different movement. I see the
- 8:10movement of the Frontier Labs are
- 8:13allowing
- 8:15allowing small and medium businesses and
- 8:18big businesses to build to move to an
- 8:20agent to agents
- 8:23which which bypasses the small models.
- 8:26I mean
- 8:27they they they still have
- 8:30their relevance today where they're
- 8:32building
- 8:33I mean now you're I mean we are
- 8:36the AI company in the data space.
- 8:39We have now our own legal agents
- 8:43which we
- 8:45we work we we put
- 8:49downstream of our
- 8:50our legal counsels
- 8:53accelerating and and reducing price and
- 8:55accelerating legal reviews. We do we do
- 8:58have a lot of legal reviews in our
- 9:01model
- 9:02so many partners that we need to onboard
- 9:04and make sure they are legit and all of
- 9:07that. So you can see this is this is a
- 9:09good example of
- 9:11of how
- 9:13the Frontier Labs are enabling
- 9:15companies to build agents and and but
- 9:18there is and I I think that is the next
- 9:21thing not so much the small model
- 9:23producers
- 9:25those
- 9:28those are definitely out there but I
- 9:30don't I believe that who's going to win
- 9:32the battle
- 9:33is going to be actually the agentic all
- 9:35the all the folks and the architects
- 9:37building agents
- 9:39more than the small models.
- 9:41>> I I've obviously got to ask you about
- 9:44copyright and AI training data given
- 9:46what you do for a job. It seems more
- 9:48legal battles
- 9:50are not going the way of the model
- 9:51developer and they're going towards the
- 9:53people that are sort of bringing these
- 9:55claims against thinking most recently
- 9:57Anthropic that I can think of.
- 9:59I assume maybe we might both agree on
- 10:01this point but how can model developers
- 10:03like OpenAI and Anthropic
- 10:05balance the cost of licensing data when
- 10:09the business from these sorts of
- 10:11these communities of people that are
- 10:13looking to sort of reclaim the ownership
- 10:15of of their
- 10:16of their data.
- 10:17When the business models of of those
- 10:19companies are already priced
- 10:22kind of put to perfection and haven't
- 10:25kind of taken into account the fact that
- 10:27they may have to be paying quite as much
- 10:29as they they could do going forwards to
- 10:32to the people that I would say they they
- 10:34stole a lot of data from.
- 10:36>> I wouldn't feel sorry for these
- 10:39Frontier Labs model developers in terms
- 10:41of balancing costs.
- 10:45They they are extremely well funded.
- 10:49They are
- 10:51they are already
- 10:53they have enough
- 10:56data
- 10:57compute power to crawl the web.
- 11:00So the least the only thing they need to
- 11:02do is to account for
- 11:05paying copyrights
- 11:07for content and data that is that they
- 11:10can find
- 11:11otherwise or it's so unique that needs
- 11:14to be paid for. So I wouldn't feel sorry
- 11:16for their business model.
- 11:18They definitely are
- 11:20so much so that they both OpenAI and
- 11:22Anthropic are filing IPOs. They they are
- 11:25in a good shape
- 11:27to
- 11:29to in they have turned around their
- 11:31revenue model. If two years ago or three
- 11:33years ago there was not it wasn't clear
- 11:36how these these companies were going to
- 11:38make money. Now it is very clear.
- 11:41So
- 11:42I I say that
- 11:44litigations continue to happen
- 11:47in the current
- 11:50political environment
- 11:52less so because
- 11:54the moat is deregulation in the US.
- 11:58But this goes in cycles and I believe
- 12:01that
- 12:02more should come and legislation will
- 12:05come eventually. We
- 12:06we cannot see we cannot continue as you
- 12:09can see
- 12:11whenever a bigger outlet or a bigger
- 12:14bigger union comes comes together just
- 12:16like the authors guild for
- 12:20against entropy
- 12:21they are able to
- 12:24to win their battle.
- 12:26That being said it's still a small price
- 12:29to settle 1.5 billion which was the
- 12:32largest so far in a settlement it's
- 12:35still a small price to pay considering
- 12:38the value of the data that
- 12:40was taken.
- 12:42>> I will say I I I might go off topic here
- 12:44with one of my questions. You you said
- 12:47that the the path to making money for
- 12:50these these companies is quite clear. I
- 12:53should be very upfront that the the
- 12:54audience of the tech report would I
- 12:56think largely disagree
- 12:58with the premise maybe they can make
- 12:59money but not profit.
- 13:02I'm curious I just would love to hear
- 13:03your your thoughts on how you think
- 13:06that that would come around because you
- 13:07seem very confident with it.
- 13:10>> Well
- 13:11I
- 13:13I work in this field for a long time.
- 13:16I know
- 13:18I know what it goes I mean you could see
- 13:20actually in terms of cost reductions
- 13:23how
- 13:24this year the world in the US
- 13:26experienced
- 13:28thousands
- 13:29tens of thousands of layoffs
- 13:32in the biggest frontier lab
- 13:35builders.
- 13:37If if that's not a way to balance profit
- 13:41profitability in a world where more of
- 13:44the white collar work is being replaced
- 13:48but it but that it's not so much to
- 13:50offload the price of data as a matter of
- 13:53fact we we haven't seen
- 13:56a huge growth in data
- 13:59purchases
- 14:00it's to offload the the price of data
- 14:02centers cost
- 14:04and energy and energy cost so
- 14:07I continue to disagree I think I think
- 14:10we are going we're seeing
- 14:13I think their business models are going
- 14:15to work out as we can see that. They
- 14:16they will it they have to balance
- 14:19people,
- 14:20infrastructure, and data.
- 14:22And certainly right now data is not
- 14:26at the center.
- 14:28And clearly people is getting cut.
- 14:31[laughter]
- 14:32It may change next, but that's what was
- 14:34this happening this year.
- 14:36>> That does actually we we gave um the the
- 14:39audience of the Tech Report a chance to
- 14:41sort of come up with some of the
- 14:42questions that we might ask,
- 14:44particularly executives, CFOs, CEOs,
- 14:46founders, those sorts of people.
- 14:48And one of them was what you think needs
- 14:51to be done to sort of address the
- 14:53anxiety that's kind of been created
- 14:55around
- 14:56uh losing losing jobs to AI because on
- 14:59one side it's seen as a very
- 15:00revolutionary technology, but for a lot
- 15:02of people it's seen as
- 15:04something that is a a threat to the
- 15:05livelihood their livelihoods to put it
- 15:07in quite a abrupt way. I was wondering
- 15:09what you think can be
- 15:11what can be done about that and and it
- 15:12would be great if you have any examples
- 15:14of of of that yourself.
- 15:15>> No,
- 15:16definitely. I mean the anxiety is real
- 15:19and this is the first year where we I
- 15:21mean this we knew the whole world knew
- 15:23this was coming.
- 15:24Every
- 15:25technology revolution
- 15:27uh
- 15:28it happens with loss of jobs and
- 15:30creation of new jobs. And this was the
- 15:33pivotal year. Uh we saw these layoffs
- 15:36are the
- 15:38testimony to that.
- 15:39Uh and they happened
- 15:42most like they happened we saw that they
- 15:44happened in areas
- 15:46that can be more automated obviously.
- 15:48The
- 15:49the the entry-level
- 15:52white-collar jobs, analysts, and things
- 15:55like that. But but the world
- 15:59humanity and technology adjusts. We are
- 16:02also seeing huge demand
- 16:05uh for people who know how to use AI
- 16:08tools, people who know how to build
- 16:11agents,
- 16:13people who know how to red team their
- 16:15agents
- 16:16and curate
- 16:18data and and their models and keep
- 16:21testing them against hallucinations and
- 16:24creating the governance and guardrails
- 16:26on around the data. So
- 16:28all these things are happening
- 16:31at
- 16:32and and all the physically AI world that
- 16:34is exploding in every sector of society
- 16:38asking for skills of doctors and
- 16:41mechanics and electricians and everybody
- 16:44who has hands on
- 16:46real the real thing
- 16:49in the 3D world. This is this is
- 16:51happening. So
- 16:55I would the advice I would say I would
- 16:57give is
- 16:59try and use as many AI tools as you can
- 17:02because the differentiator to get a job
- 17:05is the person to get or sustain a job is
- 17:08the person who actually can
- 17:10leverage this knowledge, these tools
- 17:14for to improve their productivity and
- 17:15their organization's productivity.
- 17:18>> And just one final question. I know
- 17:19you've got to you've got to run off
- 17:20soon, but I am asking everybody that we
- 17:22interview.
- 17:24Very curious to hear what you're going
- 17:25to say with this one as well. Is AI in a
- 17:27bubble and what do you think is going to
- 17:29happen between now and and 2030?
- 17:32>> Well, we hear that a lot.
- 17:34I
- 17:35I don't
- 17:37I don't think
- 17:38AI is in a bubble.
- 17:41I can I actually I actually think that
- 17:44I AI is getting
- 17:46decentralized
- 17:48which is good. It's a great thing.
- 17:51It's getting more in hands. You don't
- 17:54It's going to be very much like the
- 17:55internet. You don't have to have to know
- 17:58HTML
- 18:00to create a website back in the day.
- 18:03But in the beginning you did.
- 18:05So I I see a world where
- 18:08this decentralization is for for one
- 18:11offload the the
- 18:13concentration
- 18:14in certain in a few powers in the in
- 18:17terms of AI generation and owning and
- 18:21the knowledge.
- 18:23And it's and we see that all over the
- 18:25world already.
- 18:27And the open source world plays a huge
- 18:31play here.
- 18:32More standards are coming together. The
- 18:34Linux Foundation has a whole now new
- 18:37standard on agentic
- 18:39how to build agents in the agentic
- 18:41platforms.
- 18:42So I do I do think that
- 18:45the bubble
- 18:47if anything the bubble is always created
- 18:50around
- 18:52political instability rather than
- 18:54the actual technology advances.
- 18:58>> Well, Daniela Braga, thanks for taking
- 19:00the time.
- 19:00>> Thank you so much, Isaac.
About this transcript
This page contains the full transcript of Tech CEO: Big AI’s monopoly is breaking | Daniela Braga, Defined.ai by The Tech Report, generated from the public captions YouTube serves with the video. The transcript has 2,777 words across 515 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.
What you can do with it
Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.
Free YouTube transcript tool
YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.