00:00Sous-titrage Société Radio-Canada
00:30...
01:00...
01:30...
01:30...
01:30...
01:30...
01:32...
01:32...
01:33...
01:33...
01:34...
01:34...
01:35...
01:35...
01:36...
01:36...
01:37...
01:38...
01:38...
01:39...
01:39...
01:39...
01:40...
01:40...
01:40...
01:40...
01:41...
01:41...
01:42...
01:43...
01:43...
01:43...
01:44...
01:45...
01:46...
01:46...
01:47...
02:18...
02:18...
02:18...
02:47...
02:48...
02:48...
02:48...
02:49...
02:50...
02:50...
03:19...
03:20...
03:49...
03:50...
03:50...
03:50...
04:19...
04:20...
04:20...
04:20...
04:22...
04:23...
04:23...
04:23...
04:23...
04:24...
04:24...
04:25...
04:25...
04:55...
04:55...
04:56...
04:57...
04:57...
04:57...
04:58...
04:58...
04:59...
04:59...
05:00...
05:00...
05:01...
05:01...
05:02...
05:03...
05:03...
05:03...
05:04...
05:05...
05:05...
05:06...
05:06...
05:07...
05:07...
05:07...
05:07...
05:08et providee knowledge
05:10in a completely different manner.
05:13So from complicated inputs
05:15to complicated outputs.
05:18We see four main modalities
05:20for how these models will be used.
05:23Clearly, on the creative output,
05:25writing, content creation,
05:28enabling functionality,
05:30hyper-hyper-personalization,
05:32interactive experiences,
05:34which we all know and love,
05:35well, sometimes, like chatbots.
05:38And this is the one
05:39that's going to make all the difference.
05:41Decision-making support
05:42when you have a task
05:43that can be automated
05:44with human oversight.
05:46This is where the immense value
05:48of generative AI will come from.
05:53Now, whatever the modality you choose,
05:56there are a few things you need
05:57to make this happen properly.
05:59You need a choice of models.
06:02There is no winner-takes-all model.
06:04You need the right model
06:06for the right job.
06:08Two, you need a secure environment
06:10to customize your models,
06:11to fine-tune them with your data.
06:14You need tools that allow your developers
06:17to be highly productive.
06:20And very importantly,
06:21you need the lowest latency,
06:24lowest cost, price-performing,
06:27purpose-built machine learning infrastructure
06:30to really launch your generative AI use cases.
06:35And we're super excited
06:36to share with you, as you know,
06:39Amazon Bedrock,
06:40which gives you all of that and more.
06:43Amazon Bedrock is a managed service.
06:46You can access all of the models
06:48that I took,
06:49and I'm going to come to describing who they are.
06:51You can access a range of FMs
06:53through a managed service and API.
06:56you don't need to manage the infrastructure.
06:59All of the data is encrypted,
07:01and you can set up virtual private clouds
07:04to ensure that everything transpires
07:07in a secure manner.
07:10In Amazon Bedrock,
07:12you have Amazon Titan,
07:14the large language models
07:15that we provide,
07:16text summarization,
07:18Q&A, and search.
07:20You have access to Cloud,
07:22the anthropic large language model
07:24that allows you to work
07:26on conversational text.
07:28You have access to Stability AI's
07:31Stable Diffusion Model,
07:33which is critical for unique images design and art.
07:36And then,
07:37Jurassic 2,
07:39A121 Labs
07:40for multilingual conversational text.
07:48We've also been
07:51using these FM models
07:53as part of our Amazon SageMaker Jumpstart program.
07:57And actually,
07:58we've had lights on.
07:59I think some of you might have seen them here today.
08:01I believe President Macron
08:03mentioned them yesterday on stage,
08:05and Coher,
08:06along with the others here as well.
08:08We have 100,000 customers
08:11operating on Amazon SageMaker.
08:15So, let's take a look
08:17at how these models
08:18actually get put to work.
08:20Put yourself in the shoes
08:22of a product marketing team
08:24that has to create
08:25some social assets for launch.
08:28The first thing they do,
08:29they've got a product description.
08:31They go into Cloud,
08:32and they're able to,
08:34this thing is moving faster than I want.
08:36Let me see if I can slow it down.
08:40Sorry.
08:43Okay.
08:45Right.
08:46So, the product manager
08:48has a product description.
08:49They put it into
08:51Anthropics Cloud model,
08:53and out it spits out
08:55a beautiful product description.
08:57They then want a picture,
08:59and they go to the stability model,
09:01and that spits out
09:02a gorgeous picture of the shoe
09:04with London City in the background.
09:05You then go to Jurassic,
09:07and you create some social tweets.
09:10And then finally,
09:11into Titan,
09:12and you actually have
09:13some search engine mechanisms.
09:16That's all it takes.
09:18A few prompts,
09:19and you've got everything you need
09:23for...
09:23Okay.
09:24This thing is going very fast,
09:26so I'm going to try
09:26and slow it down.
09:28You've got everything you need
09:29to be able to drive your services,
09:31and all that is present in Bedrock.
09:34So, you've got the managed infrastructure,
09:37low latency, low cost.
09:38You've got all of the models
09:40available to you,
09:41and I'm going to get to this now.
09:43It's not enough to have all that.
09:45You also need tools
09:46for your developers.
09:49We introduced Code Whisperer.
09:51Code Whisperer is our ML-powered
09:54code generation service for developers.
09:59It can be used
10:00with natural language prompts,
10:01by the way,
10:02and it's the only one
10:04that has security built in
10:06so it can scan
10:07the hardest to detect vulnerabilities.
10:16We've got Accenture
10:18and Coach
10:19using Code Whisperer
10:21in a preview
10:22when we did a test.
10:2427% of the teams
10:26that were using Code Whisperer
10:28were more likely
10:29to end their tasks on time,
10:32and 57% faster
10:34than those
10:35that were not using Code Whisperer.
10:37So, this is truly revolutionary
10:39when it comes to improving
10:41developer predictability.
10:44Now, it's not enough
10:45to have all that.
10:47You also need
10:48a purpose-built
10:49machine learning infrastructure
10:51to make this work.
10:53And we've been investing
10:54in Silicon
10:55for five years,
10:57particularly around
10:58machine learning workloads
11:00associated with training
11:01and inferences.
11:03In fact,
11:04our inference chip,
11:05Inferentia,
11:06is eight times
11:08lower latency
11:10and 40% better performance
11:13than comparable
11:14Amazon EC2 instances.
11:20A great example
11:22is Sprinklr.
11:23They actually collect
11:24public data
11:25from social posts,
11:27videos,
11:27and blogs.
11:28They were able
11:29to use
11:30our Inferentia 2 chips.
11:33By the way,
11:34they do something like
11:3410 billion predictions
11:36a day
11:37across 500 models.
11:39And they were able
11:40to reduce latency
11:4130%,
11:42improve efficiency,
11:44and lower the cost.
11:46That's the power
11:47of this type
11:49of compute environment.
11:52So,
11:53everything you need
11:54to complete
11:55your generative AI journey.
11:56You've got infrastructure
11:58you need,
11:59you've got bedrock
12:00which gives you access
12:01to the FM models you need,
12:02you've got services
12:03like Code Whisperer,
12:05and all of that
12:06is in a very secure,
12:08resilient environment
12:09that only we know
12:11how to give you.
12:13If I could stop talking,
12:15I would,
12:16but unfortunately,
12:17it's not about
12:18the technology alone.
12:20The biggest issue
12:21we have is skills.
12:23It's the biggest gap.
12:24There was a survey
12:25done by Vantage Partners
12:26and all the C-suite people
12:29that we surveyed,
12:2992% said
12:31that organisational barriers
12:33was the big issue.
12:3585% of workers
12:36post the pandemic
12:38said they need
12:39digital skills.
12:41Without skills,
12:42you cannot transform,
12:44you cannot build
12:44your great generative AI
12:46use cases.
12:47We are investing
12:48in digital skills.
12:49We are spending
12:50hundreds and millions
12:51of dollars
12:52to train 29 million people
12:55for absolutely free
12:57by 2025.
12:59That's our commitment
13:00to bring you
13:01into the digital economy.
13:03We believe that technology
13:05will create jobs.
13:06According to WEF,
13:0897 million new jobs
13:09will be created
13:10because of cloud computing,
13:12AI, and ML.
13:14We have programs
13:14for the younger generation
13:16called Get IT,
13:17which inspires young girls
13:19in the ages of 12 to 15.
13:21I just finished judging
13:22a competition in London
13:24with a phenomenal group
13:26of schools.
13:27We've got programs
13:28like Restart,
13:29which is 12 weeks
13:30of classroom training
13:32for the underserved
13:33and underemployed community
13:34to give them
13:35the first job
13:36in digital.
13:37And then we have
13:38machine learning university,
13:40a lot of sponsorships
13:41for AI and ML training,
13:43as well as academy programs.
13:45Plus, with our customers,
13:48we are actually engaging them
13:49to re-skill their team
13:51as well.
13:53So skills
13:54and all that technology,
13:56and it's everything
13:57I think you need
13:58to turn those sparks
14:00into reality.
14:01Thank you very much.
14:03I really enjoy being here
14:04and enjoy VivaTac.
14:06Thank you.
14:07Thank you.
Commentaires