數據：推動人工智能的原材料.pdf

《數據：推動人工智能的原材料.pdf》由會員分享，可在線閱讀，更多相關《數據：推動人工智能的原材料.pdf（14頁珍藏版）》請在三個皮匠報告上搜索。

1、Data:The Raw Material Fueling AIVijay Karunamurthy,Field Chief Technology OfficerThe 3 Pillars of AIDATAALGORITHMSTalentCOMPUTETraining Data+EvalGPU ChipsScale has powered every major AI breakthrough across computer vision and generative AIAccelerating the development of AI applicationsModelsChipsDa

2、taOUR MISSIONAccelerate the development of AI applicationsOUR INSIGHTThese models are their data.The only path to improvement is through quality dataOUR WORKDramatic improvements to model accuracy through advanced training techniquesHelp model to run real-world ready agent work flowsEnable next-gen

3、models to do X domain in Y languagefor Builders,Enterprises,and GovernmentsGovernmentsScale powers major AI U.S.government programs with the leading commercial technologyLeading EnterpriseEnterprises use Scales platform to get GenAI apps into production faster through better dataFrontier Model Build

4、erScale evaluates and generates more GenAI data per day than any company on the planet6Scale is OpenAIs preferred fine-tuning partner for GPT-3.5Scale GenAI Platform is now available on the Microsoft Azure MarketplacePioneer of model customization,fine-tuning,and evaluation practicesScale is Metas l

5、aunch partner for enterprise Llama 3.1 405BScale becomes the first AWS model customisation and evaluation partner on AWS MarketplaceScale builds custom full-stack GenAI applications for governments and enterpriseNational LLMFull development and maintenance of a Foundational LLM tailored to language,

6、cultural context,and historyCustom AI Use CasesEnd-to-end development for Custom AI applications in enterprise or in governmentModel EnhancementEvaluations and high quality data for significant improvements in model performance and safetyPublic dataPrivate dataCurrent model data mixTotal available d

7、ata sourcesEvery organisation is going to need a scalable platform and data engine that uses their proprietary data to build specialised GenAI models Proprietary data and experts are the most valuable assets in GenAI SystemsScale GenAI PlatformBase LLMsProprietary Private DataDifferentiated IP and R

8、OI=+Organisations deliver enhanced products and services by customising LLMs with their own proprietary data using scalable tooling&infraEnterprise&Government data edge must be leveraged in model outputEnterprise data=10 x of public dataYet models today are powerful generalists,but not the expertsth

9、at governments and enterprises need.The Expertise ProblemModels will only be“taught”to become expertsby curating the right-relevantand structured-information as input.Public dataPrivate dataCurrent LLM data mixTotal available data sourcesChatLogsCustomer ExperiencesVideoFinancial DataShare Data with

10、 customerRelease candidatesDeploy Best VariantsEvaluationData&Model Size Impact on PerformanceSGPModel VerificationScale Sensor Fusion AnnotationEdge CaseImprove ApplicationData&Model Size Impact on PerformanceProduction MonitoringData&Model Size Impact on PerformanceSGPData GenerationData&Model Siz

11、e Impact on PerformanceScale Data EngineEvaluation DatasetsTraining DatasetsPrompt SetsIdentify IssuesPromptsModelsData IntegrationsLoggingOnline Evals/GuardrailsUser feedbackUser defined triggersExplore performance metricsIdentify failure casesRegression testsScale Data FlywheelScale Generative AI

12、Platform Data FlywheelUpdate Evaluation DataYour App!Leaders in AI test and evaluation United States Department of Defense,Chief Digital and Artificial Intelligence OfficeSafety,Evaluation,and Alignment LabCustom LLMs in production need a“Trust Feedback Loop”for scaled deploymentEvaluateImproveITERA

13、TIVE&CONTINUOUS DEVELOPMENTBuildIdentify performance gaps Generate Data(Training&Evals)Fine-tune Gen AI ApplicationsArchitect your GenAI ApplicationsTrusted and safe GenAI Production-grade performance Faster development cycles Lower TCO through OSS customization Full visibility and control1414Leveraging proprietary data for custom AI applications is the futureKnowledge ManagementCustomer Service Data AnalysisContent Generation Process AutomationMultilingual Support