《高通：2023混合AI是AI的未來白皮書-第一部分：終端側AI和混合Al開啟生成式AI的未來（英文版）（17頁）.pdf》由會員分享，可在線閱讀，更多相關《高通：2023混合AI是AI的未來白皮書-第一部分：終端側AI和混合Al開啟生成式AI的未來（英文版）（17頁）.pdf（17頁珍藏版）》請在三個皮匠報告上搜索。

1、 1 The future of AI is hybrid May 2023 Part I:Unlocking the generative AI future with on-device and hybrid AI 2 1 Executive summary.3 2 Introduction and current trends for generative AI.4 3 Hybrid AI is crucial for generative AI to scale.5 3.1 What is hybrid AI?.5 3.2 Benefits of hybrid AI.6 3.2.1 C

2、ost.6 3.2.2 Energy.6 3.2.3 Reliability,performance,and latency.6 3.2.4 Privacy and security.7 3.2.5 Personalization.7 3.3 How AI workloads are distributed.7 3.3.1 Device-centric hybrid AI.7 3.3.2 Device-sensing hybrid AI.8 3.3.3 Joint-processing hybrid AI.9 4 On-device AI evolution intersects the ne

3、eds of generative AI.10 4.1 On-device processing can support a variety of generative AI models.11 5 Impactful generative AI use cases across device categories.11 5.1 Smartphone:search and digital assistant.12 5.2 Laptop and PC:productivity.12 5.3 Automotive:digital assistant and autonomous driving.1

4、2 5.4 XR:3D content creation and immersive experiences.14 5.5 IoT:Operational efficiency and customer support.15 6 Conclusion.16 3 1 Executive summary The future of AI is hybrid.As generative AI adoption grows at record-setting speeds1 and drives higher demand for compute,2 AI processing must be dis

5、tributed between the cloud and devices for AI to scale and reach its full potential just like traditional computing evolved from mainframes and thin clients to todays mix of cloud and edge devices.A hybrid AI architecture distributes and coordinates AI workloads among cloud and edge devices,rather t

6、han processing in the cloud alone.The cloud and edge devices,such as smartphones,vehicles,PCs,and IoT devices,work together to deliver more powerful,efficient,and highly optimized AI.The main motivation is cost savings.For instance,generative AI-based search cost per query is estimated to increase b

7、y 10 times compared to traditional search methods3 and this is just one of many generative AI applications.Hybrid AI will allow generative AI developers and providers to take advantage of the compute capabilities available in edge devices to reduce costs.A hybrid AI architecture(or running AI on dev

8、ice alone)offers the additional benefits of performance,personalization,privacy,and security at a global scale.Hybrid AI architectures can have different offload options to distribute processing among cloud and devices depending on factors such as model and query complexity.For example,if the model

9、size,prompt,and generation length is less than a certain threshold and provides acceptable accuracy,inference can run completely on the device.If the task is more complex,the model can run across cloud and devices.Hybrid AI even allows for devices and cloud to run models concurrently with devices ru

10、nning light versions of the model while the cloud processes multiple tokens of the full model in parallel and corrects the device answers if needed.The potential of hybrid AI grows further as powerful generative AI models become smaller while on-device processing capabilities continue to improve.AI

11、models with over 1 billion parameters are already running on phones with performance and accuracy levels similar to those of the cloud,and models with 10 billion parameters or more are slated to run on devices in the near future.The hybrid AI approach is applicable to virtually all generative AI app

12、lications and device segments including phones,laptops,XR headsets,vehicles,and IoT.The approach is crucial for generative AI to scale and meet enterprise and consumer needs globally.1 https:/ https:/ https:/ 2 Introduction and current trends for generative AI ChatGPT has captured our imagination an

13、d engaged our curiosity.Reaching 100 million monthly active users just two months after it launched in November 2022,it is the fastest growing consumer app in history and the first generative AI“killer”app.With the rapid pace of innovation,it is becoming more and more difficult to keep up with gener

14、ative AI developments.According to a major aggregator site,there are over 3,000 generative AI apps and features available.4 AI is having a Big Bang moment,akin to the launch of television,the worldwide web,or the smartphone.And this is just the beginning.Generative AI models,such as ChatGPT and Stab

15、le Diffusion,can create new and original content like text,images,video,audio,or other data from simple prompts.These models are disrupting traditional methods of search,content creation,and recommendation systems offering significant enhancements in utility,productivity,and entertainment with use c

16、ases across industries,from the commonplace to the creative.Architects and artists can explore new ideas,while engineers can create code more efficiently.Virtually any field that works with words,images,video,and automation can benefit.Web search is one of many applications being disrupted by genera

17、tive AI.Another example is Microsoft 365 Copilot,a new productivity feature that uses generative AI to help write and summarize documents,analyze data,or turn simple written ideas into presentations all embedded in Microsoft apps including Word,Excel,PowerPoint,Outlook,Teams,and more.The emergence o

18、f generative AI also marks the first step toward a more diverse and personalized digital landscape for users to explore.It has the potential to democratize 3D content creation since 3D designers can develop 3D content faster and more efficiently with generative AI tools.This will not only accelerate

19、 the creation of immersive virtual experiences,but also reduce the barriers to entry for individual creators to produce their own content.We are on the cusp of seeing a variety of new enterprise and consumer use cases emerge from generative AI,delivering capabilities we have yet to even imagine.Foun

20、dation models,such as general-purpose large language models(LLMs)like GPT-4 and LaMDA,have achieved unprecedented levels of language understanding,generation capabilities,and world knowledge.Most of these models are quite large,with more than 100 billion parameters,and expose their functionality to

21、customers via APIs,free or paid.Access to foundation models is leading to a deluge of both startups and large organizations building apps across text,image,video,3D,speech,and audio.Examples include code generation(GitHub Copilot),text generation(Jasper),image generation for artists and designers(Mi

22、djourney),and conversational chatbots(Character.ai).4 Generative AI apps and features as of April 2023 on https:/ Figure 1:The generative AI ecosystem stack is allowing many apps to proliferate.Early estimates value the market size of generative AI at$1 trillion5 spread through participants in the e

23、cosystem stack.To realize this massive opportunity and drive AI into the mainstream,the computing architecture needs to evolve and meet the increased processing and performance demands of generative AI at scale.3 Hybrid AI is crucial for generative AI to scale Massive generative AI models with billi

24、ons of parameters place significant demands on computing infrastructure.As such,both AI training,which learns the parameters for an AI model,and AI inference,which executes the model,have been constrained to cloud implementations for large and complex models until now.The scale of AI inference is po

25、ised to be significantly higher than that of AI training.Although training individual models consumes significant resources,large generative AI models are expected to be trained only a few times a year.However,the cost of inferencing with those models increases correspondingly with the number of dai

26、ly active users and their frequency of use.Running inference in the cloud results in very high costs that can be unsustainable for scaling.Hybrid AI is the solution just like traditional computing went from mainframes and thin clients to a mix of cloud infrastructure and smart devices including PCs

27、and smartphones.3.1 What is hybrid AI?Hybrid AI is the device and cloud working together,splitting AI computation where and when appropriate,to provide enhanced experiences and efficient use of resources.For some scenarios,the computation will be centered primarily around the device,offloading tasks

28、 to the 5 UBS,Feb 2023 6 cloud when necessary.In cloud-centric implementations,devices will opportunistically offload AI workloads from the cloud,where possible and subject to their capabilities.3.2 Benefits of hybrid AI A hybrid AI architecture(or running AI on-device alone)offers benefits with reg

29、ards to cost,energy,performance,privacy,security,and personalization at a global scale.3.2.1 Cost With generative AI model usage and complexity continuing to grow,running inference exclusively in the cloud is not economical due to increasing costs for data center infrastructure,including costs for h

30、ardware,real estate,energy,operations,additional bandwidth,and network ingress/egress fees.For example,the current cloud-based computing architecture for LLM inferencing leads to higher operating costs for internet search companies,big and small.Consider a future with internet search augmented by ge

31、nerative AI LLMs,like GPT running with well over 175 billion parameters.Generative AI searches can provide a much better user experience and results,but the cost per query is estimated to increase by 10 times compared to traditional search methods,if not more.With more than 10 billion search queries

32、 per day currently,even if LLM-based searches take just a small fraction of queries,the incremental cost could be multiple billions of dollars annually.6 Shifting some processing from the cloud to edge devices reduces strain in cloud infrastructure and mitigates expenses.This makes hybrid AI crucial

33、 for generative AI to continue to scale taking advantage of the billions of edge devices with AI capabilities already deployed,as well as the billions more coming with increased processing power.Cost savings is also a critical enabler to the generative AI ecosystem,allowing OEMs,independent software

34、 developers(ISVs),and app developers to experiment and create apps more economically.For example,a developer could create an app based on Stable Diffusion that runs completely on the device and pay a lower per-query cost,or no cost at all,for every image generated.3.2.2 Energy Edge devices with effi

35、cient AI processing offer leading performance per watt,especially when compared with the cloud.Edge devices can run generative AI models at a fraction of the energy,especially when considering not only processing but also data transport.This difference is significant in energy costs as well as helpi

36、ng cloud providers offload datacenter energy consumption to meet their environmental and sustainability goals.3.2.3 Reliability,performance,and latency In a hybrid AI architecture,on-device AI processing provides reliable performance that can be comparable to the cloud or even better when cloud serv

37、ers and network connectivity are congested.7 Peaks in cloud demand for generative AI queries can create large queues and high 6 Morgan Stanley,“How Large are the Incremental AI Costs.and 4 Factors to Watch Next”Feb 2023 7 https:/ 7 latency,and in some cases even lead to denial of service.8 This can

38、be prevented by shifting compute loads toward edge devices.Additionally,the availability of on-device processing in a hybrid AI architecture can allow generative AI apps to work anywhere users are,even without connectivity.3.2.4 Privacy and security On-device AI inherently helps protect users privac

39、y since queries and personal information remain solely on the device.For enterprise and workplace usage of generative AI,this helps to solve the challenge of protecting company confidential information.For example,a programming assistant app for generating code could run on the device without exposi

40、ng confidential information to the cloud addressing a concern already faced by companies today.9 For consumer usage,a“private mode”in a hybrid AI architecture allows users to strictly utilize on-device AI to enter sensitive prompts to chatbots,such as health questions or startup ideas.In addition,on

41、-device security is strong and will evolve to ensure that personal data and model parameters are secure on edge devices.3.2.5 Personalization Hybrid AI makes more personalized experiences possible.Digital assistants will be customized to a users expressions,quirks,and uniqueness without sacrificing

42、privacy.This persona,which is a representation of a person based on real-life behaviors,values,pain points,needs,concerns,and problems,is learned and evolves over time.It is used to enhance and customize generative AI prompts,which are then processed either on-device or in the cloud.Since it stays w

43、ithin the device,and with on-device learning,the persona gets continuously optimized and updated.Personalization is not just for consumers.For instance,organizations can use it to standardize how their code is written or to produce public content in a unique tone and voice.3.3 How AI workloads are d

44、istributed We envision a hybrid AI architecture with different offload options to distribute processing,which will evolve over time,depending on model and query complexity.For example,if the model size,prompt,generation length is less than a certain threshold and provides acceptable accuracy,then it

45、 can run completely on the device.If more complex,then the model can run jointly in the cloud and on device,or if more up-to-date information is required,then the internet can provide that information as well.3.3.1 Device-centric hybrid AI In a device-centric hybrid AI architecture,the device acts a

46、s the anchor point with the cloud used only to offload tasks that the device cannot sufficiently perform.Since many generative AI models can run adequately on the device(see Figure 2),the device can do most of the processing by running the less complex inferences.For example,a user is running Micros

47、oft 365 Copilot or Bing Chat on a laptop.Models with up to tens of billions of parameters will run on device while more complex models will use the 8 https:/ https:/ 8 cloud as needed.The experience will be seamless to the user since an on-device neural network or rules-based arbiter will decide if

48、the cloud is needed,whether for a better model or retrieving internet information.If the user is not satisfied with the quality of the results,a request to try again could initiate a better model to be used.As on-device AI processing improves with newer devices and more advanced chipsets,more can be

49、 offloaded from the cloud.Figure 2:In a device-centric hybrid AI architecture,the cloud is only used to offload AI tasks that the device cannot sufficiently perform.For various generative AI applications,such as creating images or draft emails,fast and responsive inferences are preferrable,even if t

50、hat carries a small tradeoff with regards to accuracy.Quick feedback(i.e.,low latency)with on-device AI allows users to quickly iterate with refined prompts until the output is satisfactory.3.3.2 Device-sensing hybrid AI In a device-sensing hybrid AI scenario,the edge models act as the sensor input

51、or the eyes and ears to the cloud LLMs,which act as the brain.For example,a user speaks to their smartphone.An automatic speech recognition(ASR)AI model,such as Whisper,runs on device converting speech to text before sending it as a query to the cloud.The cloud runs an LLM and then sends the text an

52、swer back to the device.The device runs a text-to-speech(TTS)model to provide a natural and hands-free response.Offloading the ASR and TTS models to the device saves compute and connectivity bandwidth.As LLMs become multi-modal and allow images as input,the computer vision processing can also run on

53、 device,further offloading compute and reducing connectivity bandwidth and therefore saving costs.In a more advanced version that preserves privacy further,on-device AI takes on more processing and provides an improved and more personalized prompt to the cloud.Through on-device learning and personal

54、 data on the device,such as social media,email,messaging,calendar,and location,the device would create an individual persona of the user that works with an orchestrator program to provide improved prompts based on this additional context.For example,if a user asks their phone to schedule a time and

55、reserve a table with their best friend at their favorite restaurant,the orchestrator knows those personal details and can provide a better prompt to the cloud LLM.The orchestrator can put guard rails on the LLM 9 when it lacks information and help prevent hallucinations.For simpler queries,a smaller

56、 LLM can run on device without any cloud interaction,similar to device-centric hybrid AI.Figure 3:In device-sensing hybrid AI,the ASR,CV,and TTS happens on device.For the more advanced version,an on-device orchestrator provides an improved and more personalized prompt to the cloud.3.3.3 Joint-proces

57、sing hybrid AI On-device and cloud AI computing can also work jointly to process AI workloads.Multi-token generation for an LLM is an example.LLMs are memory-bound,meaning that the compute hardware is more often idle while waiting for the memory data from DRAM.LLMs produce a single token per inferen

58、ce that is almost equivalent to a word,which means that a model like GPT-3 must read all 175 billion parameters to generate a single word.The full model then runs again to produce the next token and so on until the entire result is complete.Since memory reads are the bottleneck,it is more efficient

59、to run several LLMs in parallel to generate multiple tokens while reading all the parameters once from DRAM.Reading all the parameters per token consumes energy and produces heat,so sharing the parameters to speculatively run LLMs in parallel on an otherwise idle compute is a net gain in performance

60、 and energy consumed.For four-token generation,a draft LLM,which is 7-to-10 times smaller and hence less accurate than the original target LLM,runs sequentially on device four times in a row to generate the next four tokens.The device sends the four tokens to the cloud,which checks the accuracy by r

61、unning the target model effectively four times with only one read of the full model parameters.The tokens are computed in parallel,with each target model having an input of zero,one,two,three,or four of the predicted tokens.These tokens are considered“drafts”until they are confirmed or corrected by

62、the cloud.This speculative decoding process continues until the full answer is completed.Our early experiments and other published results10 show that with four-10 Leviathan,Yaniv,Matan Kalman,and Yossi Matias.Fast Inference from Transformers via Speculative Decoding.arXiv preprint arXiv:2211.17192(

63、2022)10 token speculative decoding,on average 2 to 3 are correct and accepted,which results in a net speedup in tokens per unit time and energy savings.Figure 4:Four-token speculative decoding example of joint-processing hybrid AI.4 On-device AI evolution intersects the needs of generative AI On-dev

64、ice AI capabilities are key to enabling hybrid AI and allowing generative AI to reach global scale.How processing is split between the cloud and edge devices will depend on several factors,such as the capabilities of the device,privacy and security requirements,performance requirements,and even busi

65、ness models(see section 3.3).AI processing has been moving toward the edge prior to the arrival of generative AI,with more and more AI inference workloads running on phones,laptops,XR headsets,vehicles,and other edge devices.For example,phones utilize on-device AI for many daily features,such as low

66、-light photography,noise cancellation,and face unlock.Figure 5:AI processing is gravitating toward the edge.11 4.1 On-device processing can support a variety of generative AI models With an installed base of billions of AI-capable phones,PCs,and other devices in users hands today,11 the potential to

67、 tap into a large amount of on-device AI processing for generative AI is already significant and poised to grow steadily in the coming years.A key question becomes which generative AI models can run on device with appropriate performance and accuracy.The great news is that very capable generative AI

68、 models are getting smaller while on-device processing capabilities continue to improve.Figure 2 shows a broad number of generative AI capabilities that can run on device using models that range from 1 to 10 billion parameters.12 Models like Stable Diffusion,with over 1 billion parameters,are alread

69、y running on phones with performance and accuracy levels similar to the cloud equivalent,and many other generative AI models with 10 billion parameters or more are slated to run on device in the near future.Figure 6:A significant number of generative AI models can run on the device,offloading the cl

70、oud.5 Impactful generative AI use cases across device categories The rise of generative AI with foundation models is driving a new wave of use cases around content generation,search,and productivity across device categories,including smartphone,laptop and PC,automotive,XR,and IoT.The hybrid AI archi

71、tecture will enable generative AI to provide new and enhanced user experiences across these segments.11 https:/ 12 Assuming INT4 parameters 12 5.1 Smartphone:search and digital assistant With over 10 billion searches per day and mobile accounting for over 60%of searches,13 the adoption of generative

72、 AI will drive a substantive increment in computing capacity required particularly from queries made on smartphones.Users are already migrating to generative AI-based search since it provides better answers for many queries.The popularity of chat as a search interface is also poised to increase the

73、number of overall queries.As chat improves and becomes more capable,the smartphone can become a true digital assistant.Users can communicate naturally to get accurate and relevant answers,thanks to the combination of very accurate on-device personas and the LLMs understanding text,voice,images,video

74、,and any other input modality.Models that do natural language processing,image understanding,video understanding,text-to-text generation,and much more will be in high demand.5.2 Laptop and PC:productivity With its ability to quickly generate high-quality content from simple prompts,generative AI is

75、transforming productivity.Microsoft Office 365 on laptops and PCs is a great example.With more than 400 million Microsoft Office 365 seats and subscribers worldwide,incorporating generative AI into everyday workflows will have a significant impact.14 Tasks that took hours or days can now take minute

76、s.Microsoft 365 Copilot combines the power of LLMs with user data in the Microsoft Graph and the Microsoft 365 apps to turn prompts into a powerful productivity tool.15 Office workers can run an LLM in the background to read or compose emails in Outlook,write a document in Word,create a presentation

77、 in PowerPoint,analyze data in Excel,or collaborate in Teams meetings.Generative AI models such as natural language processing,text-to-text generation,image generation,video generation,and programming to name a few require tremendous amounts of processing for these heavily used productivity tasks.Mo

78、st of the processing could happen on the PC in a device-centric hybrid AI architecture.5.3 Automotive:digital assistant and autonomous driving Informed by data in and around the vehicle,todays AI-driven cockpits offer highly personalized experiences.Similar to smartphones and PCs,in-vehicle digital

79、assistants will keep drivers and passengers seamlessly connected with a hands-free natural user interface,while also creating new monetization opportunities for the ecosystem.A digital assistant can access a users personal data,such as apps,services,and payment information,as well as sensor data fro

80、m the vehicle,including cameras,radar,lidar,cellular vehicle-to-everything(C-V2X),and more.Enterprise APIs also allow third-party service providers to integrate their offerings,extending their customer relationships into the vehicle.For example,the navigation experience will vastly improve with proa

81、ctive assistance,such as traffic and weather updates that impact the drivers usual travel route and recommendations 13 https:/ Microsoft earnings reports 15 https:/ for recharging the vehicle or purchasing a parking permit.Plus,it can order the users favorite meal with the stored credit card with a

82、simple ask.The cockpit media experience will also be transformed as the vehicle recognizes each occupant and customizes their experience and content,such as music or podcasts.As in-vehicle augmented reality(AR)becomes more prevalent,the digital assistant can customize the display according to the dr

83、iver or passengers preferences.Vehicle maintenance and servicing will also become much more proactive and seamless.Analyzing data such as sensor input,maintenance history,and driving behavior,a digital assistant can predict when maintenance will be necessary.Using generative AI,the assistant can pro

84、vide information on how to repair the vehicle or advise the user on finding the right service provider,improving vehicle reliability while reducing time and cost.Advanced driver assistance systems and autonomous driving(ADAS/AD)solutions are often confused by unusual or unfamiliar objects that a per

85、ception stack has never encountered before.Typically caused by poor lighting or challenging weather conditions,this leads to unpredictable and sometimes even dangerous outcomes from the drive policy stack.To prevent similar situations in the future,the corner case data must be captured and labeled a

86、ppropriately,and the model retrained.This loop can be quite laborious and time-consuming.Generative AI can create simulated corner-case scenarios predicting the trajectory and behavior of various agents on the road such as vehicles,pedestrians,cyclists,and motorcyclists.The planner can utilize these

87、 scenarios to decide the drive policy of a vehicle.Figure 7:Generative AI can be used for ADAS/AD to help improve drive policy by predicting the trajectory and behavior of various agents.14 A drive policy stack,along with a perception stack,always runs locally within the AI computing capabilities of

88、 a vehicle.Strict latency requirements preclude the cloud from playing any role in decision-making for these AI workloads.As ADAS/AD solutions embrace generative AI models with appropriate post-processing,it is inevitable that significant energy-efficient AI computing power in vehicles is required.5

89、.4 XR:3D content creation and immersive experiences Generative AI holds tremendous promise for XR.It has the potential to democratize 3D content creation and bring virtual avatars to life.The next generation of AI rendering tools will enable content creators to use a variety of prompts such as text,

90、speech,images,or video to generate 3D objects,scenes,and(eventually)entire virtual worlds.In addition,content creators will get to leverage text-to-text LLMs to generate human-like conversations for avatars that are fully voiced and emotive.Taken together,these advancements will transform the way we

91、 create and experience immersive content on XR devices.Figure 8:Generative AI models will enable conversational AI and new rendering tools for XR.While the promise of generative AI for XR is undoubtably exciting,it is difficult to predict when there will be widespread adoption of these technologies.

92、However,given the rapid pace of innovation in recent months,it is safe to say that we can expect to see significant progress in the coming years.For immersive worlds,text-to-image models,such as Stable Diffusion,will soon enable content creators to generate realistic textures for 3D objects.We expec

93、t these capabilities to become available on smartphones,and by extension XR devices,within a year.Deployment in XR will require“distributed processing,”where the headset runs the perception and rendering stack,while the paired smartphone or the cloud run the generative AI model.In a couple of years,

94、the first text-to-3D and image-to-3D models will likely make it to the edge to generate high-quality point clouds of 3D objects.Within a few years later,these models will improve further,to the point where they can generate high-quality 3D textured objects from scratch.In a decade or so,further mode

95、l improvements will enable entire 3D rooms and scenes to generate,at high 15 fidelity,from text or images.In the future,text-to-3D and video-to-3D models might finally allow us to step into 3D virtual worlds that were generated from scratch,for example automatically constructing a 3D virtual environ

96、ment that is only limited by the users imagination.Figure 9:Generative AI will help create immersive 3D virtual worlds based on simple prompts,such as:“surreal world,floating jellyfish all around,beautiful waterfalls,mystical lake,towering mountains.Virtual avatars will follow a similar progression.

97、Text-to-text models such as 13 billion parameter LLaMA will make it to edge devices,to generate conversations for avatars that feel natural and intuitive.In addition,text-to-image models will generate new textures and outfits for these avatars.In the following years,image-to-3D and encoder/decoder m

98、odels will generate head and full-body avatars of humans for telecommunication.Eventually,we will use voice prompts,images,or video to generate human-like virtual avatars that are photorealistic,fully animated,intelligent,and mass producible.5.5 IoT:Operational efficiency and customer support AI is

99、already used today in a wide variety of IoT verticals,including retail,security,energy and utilities,supply chain,and asset management.AI improves decision-making based on data collected and analyzed in near real time,optimizes operational efficiency,and enables innovation to create competitive diff

100、erentiation.IoT segments stand to further benefit from AI utilization through generative AI.For example,generative AI can improve the customer and employee experience in retail.A grocery shopping agent at an onsite kiosk or smart cart can make a menu with recipes using the weekly sales specials,budg

101、et constraints,and family preferences.Store managers can anticipate off-cycle sales opportunities based on upcoming events and prepare accordingly.If a sports team was coming into town,a store manager can ask generative AI what branded items 16 fans like and increase inventory accordingly.Another us

102、e is to create a new store planogram based on best practices and positive results from other stores in similar communities.Generative AI could help store managers reorder products in shelves to expand space for the most profitable brands using a simple prompt,or it could help minimize the prominence

103、 of out-of-stock items on the shelf using data from nearby chain stores.Figure 10:Generative AI will help improve customer and employee experience in retail,such as providing recommendations for inventory and store layout.The energy and utilities sector also stands to benefit from generative AI.Oper

104、ation teams can create corner-case load scenarios and predict demand for electricity along with potential grid failures under unusual circumstances,such as a hot summer with strong winds and localized fires in rural areas,to better manage their resources and avoid outages.Generative AI can also be u

105、sed to provide better customer service,such as by answering questions about outages or billing.6 Conclusion Hybrid AI is inevitable.Use cases for generative AI will continue to evolve and grow into mainstream experiences,increasing demand on the cloud and its infrastructure.With the advanced capabil

106、ities of on-device AI,a hybrid AI architecture can scale to meet enterprise and consumer needs,providing benefits with regards to cost,energy,performance,privacy,security,and personalization.The cloud and devices will work together to deliver next-generation user experiences through powerful,efficie

107、nt,and highly optimized AI capabilities.Interested in more content like this?Sign up for our Whats Next in Mobile Computing Tech newsletterSign up for our Whats Next in Mobile Computing Tech newsletter 17 Nothing in these materials is an offer to sell any of the components or devices referenced here

108、in.Qualcomm may mean Qualcomm Incorporated,Qualcomm Technologies,Inc.,and/or other subsidiaries.2023 Qualcomm Technologies,Inc.and/or its affiliated companies.All Rights Reserved.Qualcomm is a trademark of Qualcomm Incorporated,registered in the United States and other countries.Other products and brand names may be trademarks or registered trademarks of their respective owners.Follow us on:For more information,visit us at: