- il y a 2 semaines
More than a Model The Gen AI Essentials for Business Innovation
Catégorie
🤖
TechnologieTranscription
00:00Good afternoon, everyone.
00:03Really excited to be here talking about all things about Generative AI.
00:08Just a little bit about me.
00:14First is, I'm a builder and a dreamer at heart.
00:18Just like many of you, I've been super excited
00:20in being able to build with Generative AI.
00:24And if you look at my own history,
00:27I started as an intern in Amazon
00:29more than 18, 20 years ago.
00:33And my internship project was to build
00:35what turned out to be one of the first cloud computing servers.
00:41Actually.
00:45And over the past few decades,
00:48I had the opportunity to be part of some of the most
00:51important technology in the tech industry.
01:04And the beginning of cloud industry was actually
01:08part-breaking in a big way.
01:09For the first time in the history,
01:11we had compute storage became a programmable resource.
01:15Then big data systems like database and analytics became programmable.
01:21And then we had machine learning,
01:23which started from traditional rule-based systems
01:26to linear machine learning,
01:30to gradient boosting trees,
01:31to deep learning.
01:32learning, they all became programmable.
01:34And then we entered into the era of Generative AI,
01:38where large language models took place.
01:42And with the era of Generative AI discovery,
01:46now we are entering a space
01:49where AI is now at the center of every boardroom discussion.
01:55And every industry is reinventing itself with Generative AI.
02:04To begin with, in marketing industry,
02:07Gen AI can automate end-to-end advertising campaigns
02:10and enable remarkably hyper-personalized customer experiences.
02:16In the healthcare industry,
02:18it can support clinicians by listening to patient conversations,
02:22summarize clinical notes,
02:24and even suggesting imaging modalities
02:26to increase the diagnostic accuracy.
02:30And then in gaming industry,
02:32it can create individualized experience
02:35with infinite variation of characters, missions,
02:39and interactions
02:40based on the individual player actions.
02:44And across all these industries,
02:46we are seeing a few common themes
02:49and use cases emerge
02:51that simply wouldn't be possible
02:53without Generative AI.
02:55From significantly enhanced customer experiences
02:58through highly intelligent Generative AI assistants
03:02or personal virtual assistants.
03:06To boost in employee productivity
03:08with things like conversational search,
03:11text summarization,
03:13or being able to generate code
03:14and coding assistants.
03:16and to finally highly optimize business operations
03:21through intelligent document processing
03:23or being able to do predictive analytics.
03:27Just a decade ago,
03:29these projects and ideas were just stuff of dreams.
03:34and now they are part of what I call as a new Gen AI-powered reality.
03:40A reality that often feels like magic
03:42because these LLMs have deep reasoning capabilities
03:47and we can add them to our enterprise applications.
03:51And this technology will fundamentally change
03:54the way we innovate and build new products.
03:58and if you even look at the presentation,
04:01what you are seeing right here on the VivaTech stage,
04:04all these images have been generated by Gen AI.
04:10Now with all of these possibilities,
04:13many organizations are asking themselves,
04:16how do we actually get started?
04:18How do I innovate with Generative AI?
04:20So before we get into it,
04:23let's look at what is a key look at the technology
04:27that is powering every Gen AI application.
04:31The foundational models.
04:33I'm sure that many of you are already experimenting
04:36and playing with many of these foundational models.
04:39These are large language models
04:41or other foundational models.
04:44It seems like every day we are learning
04:46about yet another new model that got released
04:49and new powerful models from companies
04:52like Anthropic, Meta, Mistral AI,
04:55and many, many more.
04:56Just last month, even Amazon,
05:00we introduced new additions
05:01to Amazon's own foundational model called Titan.
05:05The pace of innovation we are seeing in this space
05:08is like unprecedented.
05:12At AWS, we were the first one to recognize
05:16no one model will rule them all.
05:19These foundational models
05:20will continue to evolve at an amazing speed
05:24and customers will need the flexibility
05:26to use a combination of models for different use cases.
05:31Imagine if you are a retailer,
05:33you might want to use Anthropic's cloud model
05:35to generate a product description for a new shoe launch.
05:39But you may also want to use Stability AI's image generation,
05:42stable diffusion model
05:44to present a unique background for your product image.
05:48Now, to enable more developers,
05:51not just machine learning scientists,
05:54to take advantage of Gen AI,
05:56it should be easy to access and evaluate the AI models
06:00for each of your use case.
06:04We believe that in the very near future,
06:08every developer need to build Gen AI applications
06:11because Gen AI will be the fundamental part
06:15of every experience we create in the digital world.
06:19So how can we make this future a new reality?
06:24What would it take to create a seamless end-to-end
06:29customer experience that is powered by Generative AI?
06:35And I talked about model choice,
06:37but in addition to model choice,
06:40you need lots more capabilities
06:42to build Gen AI applications.
06:45First, you need the ability
06:46to customize these models with your data.
06:50Next, you need the ability
06:52to be able to automate various tasks
06:54in your organization with Gen AI.
06:57And finally, you also want the ability
06:59to leverage the built-in security
07:02and governance controls
07:03to mitigate any potential risk.
07:07And finally, you need the ability
07:09to scale the innovation
07:10without having to manage the infrastructure
07:13to run these large-scale models.
07:17Now, let's double-click on some of these today,
07:20starting with model customization.
07:23Now, when it comes to Generative AI,
07:26that know your business and your customers,
07:29your data is your differentiator.
07:32Data is the difference
07:33between a generic Gen AI model-powered application
07:38and an application that knows your business
07:41and your customer.
07:43Now, you might be wondering,
07:44what is the best way
07:45to actually augment data with your models?
07:48There are a few different ways
07:50and to how you can use your data
07:53to maximize their value.
07:55The easiest place to start
07:57is the technique called
07:59Retrieval Augmented Generation, or RAG.
08:02With RAG, the developers
08:04need to retrieve their own company data
08:07to add context or information
08:09to their prompts,
08:11using which they ultimately create
08:13better responses
08:14to their customer questions or requests.
08:18Now, you don't need to just stop with RAG.
08:21To further improve the accuracy
08:23of your model outputs,
08:25you can also use
08:26and update the out-of-the-box foundational models
08:30to create a fine-tuned model.
08:32With the process of fine-tuning,
08:34you're taking a small set
08:36of labeled data examples
08:38that is very specific
08:39to your company data
08:42to train the model
08:44with your corporate lingo
08:45and business policies.
08:48And then, to take it one step further,
08:50you can change
08:52the underlying core parameters
08:54of the model
08:54through a process called
08:56Continued Pre-Training.
08:58Continued Pre-Training
09:00means you pick up
09:01from where the model provider left off
09:05and you train
09:06with unsupervised training
09:09with actually totally
09:11non-annotated data.
09:12This means, essentially,
09:14you are extending the model
09:15to be an expert
09:16in your company's knowledge base,
09:20just like how you are trained
09:22a custom model altogether.
09:24Now, I talked about
09:26all these three techniques.
09:28Not only do you need tools
09:29to support all these three
09:32forms of customization,
09:34but you also need
09:35to organize your data.
09:38What we have seen
09:40is that the companies
09:41that move faster
09:42with Gen.AI
09:43are the companies
09:44that have the strong
09:46data foundations.
09:47These are the companies
09:49that typically store
09:50various types of data,
09:52including their vector data,
09:53that can be used
09:54for customizing models.
09:56And they have them integrated.
09:58And across all the use cases,
10:01they also use
10:02all the different variety
10:03of database tools altogether
10:05with a well-governed manner
10:07as well.
10:08Now, let's talk about
10:10also task automation
10:11with generative AI.
10:13One of the most
10:15common, time-consuming
10:17projects for AI developers
10:19is creating autonomous agents
10:21that help perform
10:23everyday tasks
10:24for your business
10:26and for your customers.
10:28For example,
10:29let's say a customer
10:30wants to exchange
10:32black shoes
10:33for brown shoes
10:34that they purchase
10:35through an online retailer.
10:37They use the site's
10:38customer service
10:39chatbot interface
10:40to communicate
10:42their requests
10:42and confirm
10:43that their order returns
10:45were accepted
10:46and a new order
10:46was placed.
10:48This process seems
10:49incredibly simple, right?
10:51But to program
10:53such a thing,
10:53there is actually
10:54a lot of manual programming
10:56that is involved
10:58in creating
10:59even a Gen AI
11:00powered chat interface.
11:04First,
11:05to make this happen,
11:06developers need
11:07to go through
11:08series of set of steps
11:09and all of these
11:11are time-consuming.
11:12Like, you need
11:12to define the instructions
11:14and orchestrate
11:15the set of workflows
11:16that you need
11:17to go through.
11:18You need to write
11:19custom code
11:20and finally,
11:21you need to manage
11:22infrastructure
11:22to run these agents.
11:24All of this process
11:25can take weeks
11:27and require
11:27high level
11:28of machine learning
11:29expertise,
11:30being able to tweak
11:31the prompts appropriately
11:33and that's
11:34what slows down
11:36the ability
11:36to leverage
11:37Gen AI
11:38to do automation.
11:40And finally,
11:41let's look at
11:42security and governance.
11:44A common challenge
11:46when it comes
11:46to deploying
11:47a Gen AI application
11:48is hallucination.
11:50These large language models
11:52have hyperparameters
11:54that can help
11:56manage them
11:56but you can also
11:58use safeguards
12:00like system prompts
12:01to cite the sources
12:02for the model outputs
12:04so that you can actually
12:05get control
12:06of hallucinations.
12:08There are also
12:09new controls
12:10for data privacy.
12:11For enterprise-grade
12:13Gen AI,
12:14your data should never
12:15be exposed
12:16to the core foundational
12:17models
12:18to train
12:19their base
12:20foundational model.
12:22Sending your company
12:23data
12:23to a third party
12:25is a large risk
12:26for any organization.
12:29And for higher risk
12:31applications,
12:32quality control
12:33is an absolute must.
12:35AI can help us
12:37make predictions
12:38but not decisions
12:39which is why
12:41human judgment
12:42is absolutely paramount
12:44when it comes
12:45to inference.
12:46Implementing
12:47all of these
12:48best practices
12:49in a responsible manner
12:51is central
12:52to building trust
12:53with your customers.
12:56To build AI responsibly,
12:58you will also need
12:59to consider
13:00how you are going
13:01to monitor AI system behavior,
13:04prevent AI abuse,
13:06and implement
13:07educational programs
13:08to re-skill
13:09your workforce.
13:10At AWS,
13:12we are investing heavily
13:13in responsible innovations
13:15that enable guardrails
13:17for these models,
13:18evaluation support,
13:20and watermarking,
13:21and many more.
13:22Now that we have covered
13:24a lot about
13:26what does it take
13:27to build
13:27Gen AI applications,
13:29let me start
13:30explaining by how
13:31Amazon is meeting
13:33these needs
13:33for developers
13:34in one easy-to-use place.
13:37That is Amazon Bedrock.
13:40Bedrock
13:41is a fully-managed service
13:43that makes it easy
13:44to build and scale
13:45your Gen AI applications,
13:48and we recently announced
13:49that it is available
13:51in our Paris region.
13:54Tens of thousands
13:55of customers
13:56are already using Bedrock
13:58as the core foundation
13:59for their Gen AI strategy
14:01because it gives them access
14:03to the broader selection
14:05of foundational models
14:06from leading AI companies.
14:08and with the addition
14:10of a new feature
14:11in Bedrock
14:11called the custom model import,
14:13companies that are building
14:15their own foundation models
14:16can import onto Bedrock
14:18and leverage
14:18the rest of capability.
14:20We know that model choice
14:22is important,
14:23but what about
14:24all the other things
14:25that I talked about?
14:28in fact,
14:30what we have learned
14:31is that majority
14:33of Amazon Bedrock customers
14:34use more than one model,
14:36and this includes
14:37customers like Air Liquide,
14:40which leverages
14:41a variety of models
14:43on Bedrock
14:44to quickly prototype
14:45digital experiences.
14:50Now, beyond actually
14:51picking the right model,
14:53there are a bunch
14:54of other things
14:55that you need
14:55to streamline
14:56to build Gen AI applications
14:58faster.
15:00And with tools
15:02like Bedrock,
15:03we are able
15:03to do things
15:04like knowledge bases
15:05for Bedrock
15:06or agents for Bedrock
15:08to do agent automation
15:09and built-in enterprise-grade
15:12security and privacy
15:13and with support
15:15for various regulatory
15:16standards and GDPR.
15:18Now, by enabling
15:20every developer
15:21to build with Gen AI
15:23really easy,
15:24every business
15:25will be able
15:26to move faster
15:27and innovate faster
15:29with Gen AI.
15:31But if you want
15:32to harness the power
15:33of Gen AI
15:34for everybody,
15:35you also need
15:36to make it easy
15:37for every employee
15:39not just developers
15:41to leverage it
15:42for their daily tasks.
15:45One way
15:46we can accomplish
15:47this is by
15:48leveraging Gen AI
15:50assistants
15:50that integrate
15:52your enterprise data
15:53with an AI application,
15:55enabling you
15:56to quickly
15:57and easily
15:58take advantage
15:59of Gen AI
16:00for accelerating
16:02employee productivity.
16:04These assistants
16:05can streamline
16:06various set of tasks
16:07just even today.
16:09right from helping
16:10a software developer
16:11to a data analyst
16:12to data scientist.
16:14We believe
16:15these assistants
16:16will provide
16:17enormous impact.
16:20In fact,
16:21the first set of area
16:22that is going to get
16:24really revolutionized
16:26is automating
16:27repetitive developer tasks.
16:30These assistants
16:31can remove
16:31the heavy lifting
16:32associated with
16:33software development
16:35like coding,
16:36writing tests,
16:37app upgrades,
16:38and security scanning.
16:40They will also
16:41help employees
16:42access their information
16:44faster
16:44and get insights
16:46and take actions
16:48based on their insights.
16:49And these assistants
16:51also have the potential
16:52to help everyone
16:54build their own
16:55Gen AI application.
16:57At AWS,
16:58we have invested
16:59in accelerating
17:00the productivity
17:01in each of these areas
17:02with our own
17:04Gen AI-powered assistant
17:05called Amazon Q.
17:07Q is the most capable
17:10Gen AI assistant
17:11available today.
17:13And we built it
17:14with security
17:14and privacy
17:15in mind
17:17right from the get-go.
17:18With Q,
17:19you can get your job
17:20done faster,
17:21whether you're in IT
17:23or in finance.
17:24Let me share a few examples
17:26on how you can
17:27put Q into action.
17:32With Q for developers,
17:33we are helping developers
17:35become more efficient
17:36across the entire
17:38app development cycle,
17:39right from planning
17:41to coding
17:41to testing
17:42to create data
17:43engineering pipelines.
17:45Q even comes
17:46with built-in
17:47agent capabilities
17:48to autonomously
17:49perform a wide
17:50variety of tasks,
17:52everything from
17:53implementing features
17:54to performing
17:55software upgrades.
17:57For example,
17:58let's say you want
17:59to ask Q
18:00to add a new
18:02checkout feature
18:03to your e-commerce app.
18:04It will analyze
18:05your existing codebase
18:07and map out
18:08the implementation
18:09and execute
18:10the required code changes
18:12and tests
18:12in just minutes.
18:15I've been inspired
18:16by all of the
18:17positive feedback
18:18we are receiving
18:19for Q to date,
18:21and it is
18:22the number one
18:23coding assistant
18:24in the SW bench
18:26today.
18:28but accelerating
18:29productivity
18:30shouldn't just
18:30stop with developers.
18:32You wanted
18:34to take it
18:34to every employee
18:36in an organization,
18:37and that's
18:38what Q for business
18:39does.
18:40It connects
18:40to all your company
18:41data across more
18:43than 40-plus
18:44enterprise systems,
18:45and yet provides
18:46summaries,
18:47insights,
18:48and you can do
18:49analytics all
18:50in a secure fashion.
18:52but we wanted
18:53to take Q
18:54one step further,
18:55and we asked
18:56ourselves,
18:57how do we empower
18:58every business user
18:59to build
19:00their own application?
19:01So now,
19:02let's take a quick look
19:03at how we are bringing
19:04this vision to life.
19:08Amazon Q Apps
19:09enables every employee
19:10to securely create apps
19:12upon enterprise data
19:13in seconds.
19:14like Mary in HR,
19:16who needs to create
19:17an onboarding plan
19:18for a new hire.
19:20Amazon Q Apps
19:21identifies that this
19:22could be turned
19:23into a useful app.
19:24Mary likes the idea
19:25and goes ahead
19:26to create the app
19:27with just a single click.
19:29And Nikki in sales,
19:31who is browsing
19:32the Amazon Q Apps library
19:34for an app to help
19:35her pitch the company's
19:36wide range of products
19:37to her customers.
19:38She discovers an app
19:40created by her co-worker.
19:42The app generates
19:42a custom sales script
19:43tailored to her request.
19:45This is teamwork
19:47without boundaries.
19:48Because with Amazon Q Apps,
19:50everyone can build.
19:55So there it is.
19:57Amazon Q
19:58and Q Apps
19:59will radically change
20:01how our customers
20:02and our employees
20:03can do their work
20:04every day.
20:05And now,
20:06with this new era
20:07of 10 AI
20:08discovery,
20:08there has never been
20:09a better time
20:10to be a builder.
20:11I believe that
20:13what you are inventing
20:14today and tomorrow
20:15can lead to a profound
20:17impact on the world,
20:18changing industries
20:19and changing
20:21everyone's lives.
20:22So while you are here,
20:23as you learn about
20:24the tools and partners
20:26to invent your next
20:27application and experience,
20:30please ask yourself,
20:31what magic will you build
20:33with Gen AI?
20:35And now,
20:36I'd like to invite
20:37another builder
20:38to the stage
20:39who is building
20:39a magic of his own
20:41with latest foundation
20:42models.
20:43Please welcome
20:44Timothy LaCroix,
20:46co-founder of
20:47Mistral AI.
21:00So, Tim,
21:02I know
21:03Mistral AI
21:04has been extremely popular,
21:06so let's talk about
21:07what inspired Mistral AI
21:09to start a new player
21:11in the AI space
21:13in France
21:14rather than,
21:15let's say,
21:15in the United States.
21:17So, I think there are
21:18two parts to this question.
21:19Why did we want to
21:20create a new player?
21:22I think we just
21:23wanted to do our own thing.
21:24We had been working
21:26at large tech companies
21:27for a while
21:27and wanted to
21:28get the speed
21:30and control
21:30over what we did.
21:32France really never
21:34was a question for us.
21:36We're French,
21:36all of our life is there
21:38and there is
21:39a lot of talent there.
21:41So,
21:42it wasn't really
21:43a question
21:43and I think
21:44it doesn't prevent us
21:45from being
21:46a global company.
21:47We have offices
21:48in London
21:48and San Francisco.
21:50Okay.
21:51All right.
21:52Let's talk about
21:53what is the vision
21:54of Mistral AI
21:55and your vision
21:58for the future
21:58of artificial intelligence.
22:00I mean,
22:00I hadn't seen
22:01your slides before
22:02but it's pretty much
22:02the same as yours.
22:04Okay.
22:05Glad you are alive.
22:06I think
22:06what I see
22:07in the near future
22:08is a lot more
22:10integration
22:11needs to be built
22:13so all of this
22:13agentic
22:14function calling
22:15behavior
22:16will develop
22:17rapidly
22:18on the application
22:19layer
22:19and on our end
22:21we will build
22:22all of the tools
22:23that are required
22:24to actually act
22:25on these functions
22:27that are available.
22:28This will
22:29enable
22:30a wide variety
22:31of tasks
22:32and with this variety
22:33comes the need
22:34for models
22:36that are optimized
22:37for the complexity
22:38of these tasks
22:39so yeah
22:40more agentic behaviors
22:41and more variety
22:43of models.
22:44So when you talk
22:45about variety
22:46of models
22:46you are talking
22:47about like
22:47different sizes
22:48of models
22:49or different
22:49vertical industries
22:51how do you think
22:51about that?
22:52Both.
22:53So anytime
22:53you verticalize
22:55or even fine-tune
22:56a model
22:56what you unlock
22:57is better capabilities
22:59for a smaller size.
23:01Smaller size
23:02can mean
23:02either reduced cost
23:03or reduced latency
23:05so better user experience.
23:08Verticalization
23:08when it's done
23:09for example
23:09by us
23:10is a bit more generic
23:11but for the practitioner
23:13or model builder
23:14it can enable
23:15some applications
23:17that are maybe
23:18easy to build
23:19with the larger models
23:20but wouldn't have
23:21the right latencies
23:22or would be
23:23too expensive
23:24to put to product.
23:25Yeah.
23:25So that's where
23:26you're talking about
23:27like different
23:27smaller sizes
23:29and being able
23:30to customize it
23:31with the kind of techniques
23:32we talked about
23:32like fine-tuning
23:33and drag
23:34and so forth.
23:35Now
23:36being in the middle
23:37of all these
23:38innovation
23:39and building
23:39your own
23:40foundational models
23:41what are you seeing
23:42as like major trends
23:43in artificial intelligence?
23:45Are we going to keep
23:45scaling
23:46forever?
23:47Is this going to be
23:48all about compute
23:49and data forever
23:50or are you going to see
23:51a different breakthrough?
23:52What are
23:53what do you see
23:54in the coming months
23:55and coming years
23:56in your opinion?
23:57I mean
23:57people will always
23:58try to scale.
24:00I personally
24:01tend to be
24:02doubtful
24:03of infinite scaling
24:04I don't think
24:05these things
24:05hold for a while
24:06one limit
24:08that we run into
24:09quite quickly
24:10is the amount
24:11of data
24:12humanity
24:12has only been
24:13writing stuff
24:14for so long
24:14once we've
24:16read all of it
24:16there is going
24:17to be a limit
24:19so some
24:20breakthroughs
24:21will be needed
24:22but I don't think
24:22we've seen
24:23the end of scaling yet
24:24so that's
24:25one direction
24:27the other
24:28is
24:30getting better
24:31at the lower end
24:32of the scale
24:33I don't think
24:33we've done
24:34quite enough work
24:35on getting
24:36small models
24:37to do what
24:38they need
24:39this is because
24:40we don't really
24:41understand
24:41the scope
24:42of what these
24:43models need
24:44if you were
24:45to tell me
24:45okay I want
24:46a model
24:47in Alexa
24:48what should it do
24:49I don't think
24:50this has been built yet
24:51and this will
24:52enable lots of
24:52research on smaller
24:53models as well
24:54yeah
24:54and I know
24:56we actually
24:57expanded our
24:58partnership
24:58and Mistral AI
24:59is now available
25:00as part of
25:01Bedrock
25:02how do you
25:04envision
25:04actually our
25:05customers
25:05I mean millions
25:06of customers
25:07that are innovating
25:08on AWS today
25:09will benefit
25:10from Mistral
25:10what kind of
25:11things are possible
25:12so I think
25:13something that
25:14we've seen
25:14when building
25:15Mistral
25:16was that
25:16enterprise
25:17customers
25:17really want
25:18to build
25:19AI applications
25:20in their
25:21cloud of choice
25:21they want
25:22to move
25:23out of it
25:23so for
25:25customers
25:25it's already
25:27an ability
25:28to get
25:29more choice
25:29on their
25:30current platform
25:31I think
25:32one benefit
25:33of using
25:33Mistral models
25:34is the care
25:35that we put
25:36into always
25:38being on the
25:39efficiency frontier
25:40all of our
25:41models
25:41we strive
25:41to make them
25:42as good
25:43as possible
25:44for their
25:45cost
25:45and especially
25:47in European
25:49languages
25:49with which
25:50we started
25:50so French
25:51English
25:52is wider
25:54than Europe
25:55German
25:56Spanish
25:57Italian
25:57these
25:58languages
25:59are really
25:59first class
26:00citizens
26:00for us
26:01and so
26:02performance
26:03might be
26:04even better
26:04when using
26:05Mistral AI
26:05models
26:06on these
26:06yeah
26:07I like
26:07that you're
26:08focusing
26:08a lot
26:09on the
26:09efficiency
26:10and
26:12accuracy
26:13for a
26:14given
26:14use case
26:15and also
26:15on languages
26:16that are not
26:17just always
26:18English
26:18based
26:19it makes
26:20Gen AI
26:20more accessible
26:21one final
26:22teaser
26:23question
26:24what should
26:25customers
26:25expect
26:26from your
26:26future
26:27of Mistral AI
26:28in the coming
26:28months
26:29any spoilers
26:30I mean
26:31I don't know
26:32if it's
26:32much of a
26:33spoiler
26:33I think
26:34we're
26:34working on
26:35multimodal
26:36models
26:37so this
26:38will happen
26:39I think
26:40reasonably soon
26:41we're working
26:41on some
26:42verticalization
26:43on code
26:43as well
26:45it's also
26:45the easiest
26:46for us
26:46because that's
26:47what we use
26:47every day
26:48but more
26:49generally
26:49I think
26:50our goal
26:51in the near
26:51future
26:52is to help
26:53with the
26:53hassle
26:54of customizing
26:55models
26:55which I think
26:56today is still
26:57very complex
26:57and we want
26:58to enable
27:00this for
27:00pretty much
27:01anyone
27:01to do
27:02at scale
27:03and up to
27:03prod
27:04that's awesome
27:05thanks again
27:06for actually
27:07coming here
27:08and joining
27:09me today
27:09in this
27:10discussion
27:10and thanks
27:11again for
27:12everyone in
27:13the audience
27:13for taking
27:14this time
27:15thank you
Commentaires