00:00That's kind of funny because my brother's a diehard Tottenham fan.
00:05I'm sure we can all agree that
00:07generative AI is having a profound impact on our industry.
00:11And it's a change that I've seen happen faster than anything
00:15I've seen in my career. And of course, this is
00:19all because of the models, the underlying generative AI. These models are
00:23fantastic. These models have incredible properties. They understand
00:28a broad set of knowledge about the world. They have
00:31incredible language understanding. You can do translations and summarizations.
00:35But the thing I'm most excited about is that they can reason.
00:41For the first time, we have systems
00:44that we can give a goal to. It can create a plan.
00:48And if you give it the capability, it can carry out
00:52that plan autonomously. This changes everything.
00:57This reasoning capability
00:59is in what is powering new capabilities
01:02that we just couldn't build before.
01:05And we're seeing this in many, many different fields,
01:07including in finance.
01:09But for all of its strengths,
01:11it's actually hard to use generative AI
01:13to build trustworthy systems.
01:15And what I mean by that
01:17is to build a system
01:18that gives you consistently accurate answers
01:22over and over again. Systems that you could use
01:25to make mission-critical decisions
01:29in something like finance or medicine.
01:32And the problem lies in the fact
01:34that the models themselves
01:36don't have a ground truth.
01:38We have to give it the ground truth.
01:41Okay?
01:42And what is that ground truth?
01:43That is, for Bloomberg,
01:46it is our massive amounts of structured data
01:50that we have been collecting and cleaning
01:52and aligning or generating ourselves.
01:55It's the high-quality curated documents
01:59that we have been sourcing
02:01from high-value sources
02:04and creating metadata around it.
02:06It's the analytics
02:07that our customers have trusted for years.
02:10Things like evaluated bond pricing
02:13or a segment KPI prediction
02:18for a company based on observations
02:20of alternative data
02:21like credit card spend or foot traffic
02:23and aligning it towards
02:25what the financial professionals
02:27are looking for.
02:28It's the portfolio risk models.
02:31It is the liquidity analysis
02:34that goes into it.
02:35So how do we make sure,
02:38the challenge actually,
02:39is how do you use the reasoning power
02:41to make sure that these models
02:44don't use their world knowledge
02:46but instead only uses
02:48the trusted sources of information?
02:50These models want to be very helpful.
02:52When it can't find an answer,
02:54it's more than willing to make things up.
02:56Okay?
02:57And so there's a lot of effort
02:59and a lot of research we have done
03:01to make sure that you're guiding it
03:02in the right way.
03:03And without getting too geeky and technical,
03:05it essentially comes down to
03:08I'm simplifying it tremendously.
03:11My AI team would be upset with me,
03:12but essentially is you have to give it
03:15the ability to check its work
03:16at every single step,
03:18at every layer in the stack.
03:20What does this mean?
03:21This means that you have to build
03:23domain-specific checks.
03:25Often, these are our deterministic systems.
03:28Did you call that calculator
03:31with the right parameters
03:32in the right order?
03:34Did you...
03:34When you summarize a paragraph,
03:37did you faithfully ensure that
03:39all the information content
03:41came strictly from that paragraph
03:43and you didn't insert anything?
03:45These validators, these checks,
03:47are something that you just keep
03:49growing and building,
03:50and they are domain-specific,
03:51so you have to have the knowledge
03:53about what is right and what's not.
03:55So at Bloomberg, we build systems
03:58that are robust, comprehensive,
04:01and are accurate consistently.
04:04So for general-reviary systems,
04:06that means we need to derive our answers
04:09from trusted sources.
04:10We have to check our work at every step,
04:13and we need to provide the transparency
04:14and the attribution to the users
04:17through it all, okay?
04:19Now, we have taken these principles
04:22into account when we build ask-based.
04:24Ask-B is a revolutionary rethinking
04:28of the Bloomberg terminal.
04:29It is a chat interface that allows a user
04:32to ask their high-level thematic question.
04:35The AI system goes and reaches into
04:37a whole network of distributed agents
04:39that operate on those trusted sources
04:42of information,
04:42and then it synthesizes an answer.
04:45This is a profoundly different way
04:47of building software for ourselves,
04:49no longer building screens
04:51and thousands and thousands of screens
04:53and having customers have to know the mnemonics.
04:55But it's also a profoundly effective way
04:59for customers to be able to derive insights
05:02that they would have missed
05:03or they wouldn't have seen.
05:04And as we go forward,
05:06we're adding more and more asset classes,
05:08more and more analytics,
05:09more and more capabilities,
05:10so that the system becomes more powerful.
05:13But all along the way,
05:15it has to be built in a trustworthy fashion.
05:18So I highly encourage you to go look outside.
05:21There's a demo booth to look at ASPE.
05:23And whenever you're looking at systems,
05:27do the evaluations.
05:28Spend the time to look at the quality.
05:31It's really easy to get excited about shiny demos.
05:34It's really hard to build trustworthy AI.
Comments