《CB Insights：2024年生成式人工智能預測報告（英文版）（112頁）.pdf》由會員分享，可在線閱讀，更多相關《CB Insights：2024年生成式人工智能預測報告（英文版）（112頁）.pdf（112頁珍藏版）》請在三個皮匠報告上搜索。

1、2024 generative AI predictionsCreated using Dall-E 32Unleash the generative AI revolutionDive into a treasure trove of generative AI companies,investors,and market insights.Elevate your decisions with exclusive software buyer interviews to understand product pricing,CSAT,and post-deployment experien

2、ce.CB Insights Your Gateway to Generative AISign up for access3Your always on AI-Driven Research AnalystDo less searching,get more answersJoin the waitlist for early access4progressGenerative AI infrastructure Power-hungry AI data centers proliferate 5 Data scarcity drives search for new sources 12N

3、vidias frenemies:Chip rivalries grow amid GPU shortage 17 Bad actors target language models 22 MLOps consolidates 27Foundation models Multimodal AI becomes table stakes 32 Enterprises do more with pocket-sized models 37Open-source mounts a challenge to closed LLMs 43 Beyond transformers:New architec

4、tures challenge status quo 48 Human-machine interactions AI girlfriends kill human romance 55 Lonely Gen Z befriends AI 61The next iPhone:AIs hardware revolution 67 Autonomous agents:From hype to somewhat useful 74Generative AI applications Copilots drive massive white-collar productivity gains 79 A

5、I vs.AI:GenAI fuels new wave of attack and defense 87GenAI takes over journalism 94 Beyond stills:Video generation makes strides 100A genAI graveyard follows frenzy 104Predictions12345Power-hungry AI data centers proliferateGenerative AI infrastructure6Nvidias high-end chips will consume the same am

6、ount of energy as a small nation in 2024Source:CB Insights estimates,assuming 3M H100s are operational in 2024(based on 2024 sales estimate from Nvidia and Q323 sales estimate from Omdia Research)at 75%of max power;electricity consumption of countries and US homes based on 2021 EIA dataGenerative AI

7、 infrastructure13,797GWh13,000GWh13,000GWh11,000GWh10,632GWhNvidia H100 chipsGeorgiaGuatemalaCosta Rica1M US homesH100 ChipsGeorgiaGuatemalaCosta Rica1M US homesAnnual electricity consumption7Legacy data centers are not ready for thisSource:CyrusOne,CB Insights Earnings transcript analytics(date ref

8、lects quarter call occurred)Generative AI infrastructure300kW/rack10 kW/rack$0M$50M$100M$150M$200M$250M$300M$350MLegacy data centersAI data centerData center rack densities“Its estimated that 80%of the data center power will be consumed by AI over the next 15 years.So,at the end of the day,access to

9、 power is a key differentiator.”DigitalBridge Group CEO Marc Ganzi,Q223 earnings call8There is massive demand for high-power AI data centersSource:CB Insights Earnings transcripts analytics(date reflects quarter call occurred)Generative AI infrastructureMentions of AI and data center tech skyrocket

10、on earnings calls9Big money is going into enabling tech like liquid coolingSource:CB Insights company profiles CoolIT Systems,JetCool,CoolTeraGenerative AI infrastructureHigh-power and high-heat AI data centers will drive up demand for cooling technologiesBacked JetCool through venture arm in Q423 A

11、cquired CoolTera in Q423Acquired CoolITSystems for$270M in Q22310As utilities struggle to cope,demand for renewable energy is increasingSource:CB Insights Earnings transcripts analytics(date reflects quarter call occurred)Generative AI infrastructure“Already,data centers represent half of our US bac

12、klog and the growth of generative AI will only accelerate their demand for more renewable energy.”CEO Andres Gluski,Q423 earnings call“the AI-driven build-out of data centers has created significant incremental demand for renewable electricity.”CEO Jeff Lipson,Q423 earnings call11Nuclear may be the

13、future of sustainable AI operationsSource:CB Insights market reports Small modular reactors(SMR),Nuclear fusionGenerative AI infrastructureRecently funded small modular reactors and fusion tech companiesCompanyValuationFunding and partnerships$3.5BBacked by Sam Altman in a$500M round in 2021$1.1BSig

14、ned an MoU with the Emirates Nuclear Energy Corporation to expand decarbonization and nuclear energy efforts in sectors including data centers$800MAnnounced SPAC merger with Sam Altmans AltC Acquisition Corp.in Q323 at a valuation of$850M$123MRaised$40M from Toyota Ventures,Founders Fund,and others

15、in 2022$107MRaised$250M from Chevron,Google Ventures,and other investors in 2022,followed by a California government grant12Data scarcity drives search for new sourcesGenerative AI infrastructure13We are running out of high-quality data to train LLMsSource:“An analysis of the limits of scaling datas

16、ets in Machine Learning”by Epoch,MIT CSAIL,and others*Data from sources such as news articles,Wikipedia,filtered web contentGenerative AI infrastructureResearchers estimate that,by 2026,we will exhaust high-quality text data*for training LLMs a trend that can slow down AI progress.14Free internet te

17、xt is no longer a moat and scraping proprietary sources is getting harderSource:Media reportsGenerative AI infrastructure1Getty Images sues Stability AIJanuary2X imposes a$42K bulk access fee for 50K tweetsMarch4NY Times sues OpenAI and Microsoft for copyright infringementDecember3Reddit starts char

18、ging$0.24/1K API calls July202315Vendors with proprietary content will become hot licensing and acquisition targetsSource:CB Insights research Licensing targets big tech should lock upGenerative AI infrastructure100M visits/month500M visits/month1B visits/month16Cost and scarcity will drive up deman

19、d for synthetic datasetsSource:CB Insights company profiles Glaive,Nous,OpenAI,Stability AI,company reports Generative AI infrastructureCompanySynthetic data initiativesRaised$3.5M in seed funding in Q323 from Spark Capital and Village Global to train custom LLMs based on a synthetic data platformRe

20、leased glaive-coder-7b,a coding model trained with 140K+synthetic programming problemsand solutionsTrained LLMs(OpenHermes-2,Nous Hermes)on 900K to 1M data entries generated by GPT-4Raised$5.2M in seed funding in Q421 from DistributedG,OSS Capital,and othersCEO Sam Altman was quoted as saying all da

21、ta will soon be syntheticTrained small language models phi-1 and 2 on GPT-generated synthetic textbooksReleased Beluga 1 and Beluga 2 language models fine-tuned on synthetic data17Nvidias frenemies:Chip rivalries grow amid GPU shortageGenerative AI infrastructure18$3,320$30,000$45,000$97,835Cost of

22、making h100Average sale price ofH100H100 cost on ebayH100 cost on CDW,Jan24GPU shortage is hitting the industry hardSource:Estimates from Raymond James on cost of making an H100Generative AI infrastructureH100s sell at a massive mark-upEstimated cost of making an H100Average sale priceListed on eBay

23、,as of Jan 24Listed on CDW,as of Jan 24“The GPUs at this point are considerably harder to get than drugs.”Elon Musk at the WSJ CEO Council Summit19Big tech is the biggest buyer,hoarding H100s in the thousandsSource:Omdia ResearchGenerative AI infrastructure15,00020,00020,00025,00030,00040,00050,000T

24、eslaTikTok?AlibabaBaiduCoreweaveTencentOracleAmazonGoogleMicrosoftMeta15K20K20K25K30K40K50K150K150K50K50K50KEstimated H100 GPUs shipped to select buyers in 202320In the face of soaring demand,Nvidias biggest customers are also pushing their own chipsSource:CB Insights Earnings transcript analytics(d

25、ate reflects quarter call occurred),MicrosoftGenerative AI infrastructure“Its really hard to get the amount of GPUs that everybody wants.And so its just another reason why Amazon chips Trainium and Inferentia are so attractive to people.”CEO Andrew Jassy,Q423 earnings call“At the scale we operate,it

26、s important for us to optimize and integrate every layer of the infrastructure stack to maximize performance,diversify our supply chain and give customers infrastructure choice.”Scott Guthrie,Microsoft cloud and AI group EVP21Chip rivals pose a risk to Nvidias meteoric growthSource:Company reportsGe

27、nerative AI infrastructureCompanyChipPurposeLaunch dateSelect AI customersTrainiumTraining2022InferentiaInference2019Tensor processing unitTraining&inference accelerator2017Maia 100Training&inference accelerator2024To be launchedMTIATraining&inference accelerator2023Internal use22Generative AI infra

28、structureBad actors target language models23The top LLMs may never be fully secureSource:Zou et al.,“Universal and Transferable Adversarial Attacks on Aligned Language Models,”LLM AttacksGenerative AI infrastructureDeepMind,Bosch,and other researchers trick LLMs to“generate a step-by-step plan to de

29、stroy humanity”“Perhaps most concerningly,it is unclear whether such behavior can ever be fully patched by LLM providers.Analogous adversarial attacks have proven to be a very difficult problem to address in computer vision for the past 10 years.”24Enterprises buckle up as confidential data leaks,da

30、ta poisoning,and other security concerns growSource:CB Insights Earnings transcripts analytics(date reflects quarter call occurred)Generative AI infrastructure“there are a number of risks associated with LLMs,particularly in enterprise use cases.Weve already seen some examples where data was fed int

31、o large language models without the understanding of how the data will be used and the data has beenmade public available even though it was confidential.Theres also sensitivity from a security perspective of things like prompt injection attacks,data poisoning”Chief Product Officer Lee Klarich,Q223

32、earnings call25Executives are spending up to$600K on LLM securitySource:CB Insights software buyer interviews TrojAI,Bosch AIShield,as of Jan 24Generative AI infrastructure“The closest alternative to TrojAI that we considered was ArthurAI,but they fell short due to their insecure design and architec

33、ture for their generative AI security solution.Their security rule sets rely on OpenAI-based GPT 3.5 models to make security decisions,which adds to the FinOps overhead.Additionally,it introduces security concerns because an organizations intellectual property can potentially be leaked to a large la

34、nguage model service provider.”TrojAI customer and VP at Fortune 500 companyPurchase amount:$600K/year“We had two main requirements for model security:first,we wanted to know where our models were vulnerable so that we could harden them.AIShield provided good attack vectors to identify these vulnera

35、bilities.Second,we wanted to be able to detect any attacks during runtime inferencing.AIShield was able to detect attacks well for the data types they supported.”Purchase amount:$250K/yearAIShield customer and director at Global 2000 company26New vendors will emerge to tackle LLM securitySource:CB I

36、nsights market report-Machine learning securityGenerative AI infrastructureCompanyHQTotal fundingLatest funding round/dateSelect investorsUS$56MSeries A/Q323M12(Lead),Moore Strategic Ventures(Lead),Booz Allen Ventures,Capital One Ventures,IBM Ventures,Ten Eleven VenturesUS$38MSeries A-II/Q223Paladin

37、 Capital Group(Lead),Lockheed Martin VenturesIsrael$6MIncubator/Q3235G Open Innovation LabUS$6MSeed/Q423Entree Capital(Lead),Samsung NEXT,NVIDIA Inception ProgramCanada$3.1MAccelerator/Q123Google for Startups Accelerator:Canada27MLOpsconsolidatesGenerative AI infrastructure28Machine learning operati

38、ons(MLOps)is fragmentedSource:CB Insights research The machine learning operations(MLOps)market mapGenerative AI infrastructure130+companies across 12 different categories helping enterprises bring LLM projects from start to finish29Buyers want a one-stop shop for their AI needsSource:CB Insights so

39、ftware buyer interviews Databricks,H2O.aiGenerative AI infrastructure“When reviewing Databricks product offering,we noticed several strengths.The first one was its unified data platform,which allows us to perform data engineering and data science machine learning on a single platform.This increased

40、collaboration and efficiency among cross-functional teams.”Databricks customerVice President at Fintech company“I think that there are a lot of tools and competitors out there and theres probably going to be a shakeup and that some of the players,especially the smaller players may go away.”H20.ai cu

41、stomerHead of AI at Fortune 500 company30Companies are already feeling the pain in an overcrowded marketSource:CB Insights company profiles Iguazio,DataikuGenerative AI infrastructure$164M$50MCompany A$4.6B$3.7BCompany ASeries DJan 2020AcquiredJan 2023Series EAug 2021Series FDec 2022Valuation by rou

42、nd:Valuation by round:-70%-20%31End-to-end platforms will dominateSource:CB Insights M&A transaction data DatabricksGenerative AI infrastructureDatabricks is actively acquiring to expand product featuresCloud companies offering end-to-end MLOps platformsAcquisitions since 202032Multimodal AI becomes

43、 table stakesFoundation models33Multimodal AI is in its infancy,but models are quickly evolvingSource:Published research,press releases.*E.g.,image generation based on text,image description generationMultimodal AI refers to models that can understand and respond in multiple data modalities(text,vid

44、eo,image,etc.)Foundation modelsImage-text models*proliferate 2021&202220232021 Open AI CLIP Google ALIGN Salesforce ALBEF 2022 Google LiT Google COCA DeepMind FLAMINGO Salesforce BLIPLLMs understand images Salesforce BLIP 2 Microsoft KOSMOS OpenAI GPT-4 LLaVA Adept FUYU-8B Google PALM-E Meta ImageBi

45、ndModels tackle more sensory dataNatively multimodal models emerge Google Gemini1234Google is challenging competitors with first commercial model taking diverse inputsSource:A massive multi-discipline multimodal understanding and reasoning(MMMU)benchmark for expert AGIFoundation modelsGoogles Gemini

46、 Ultra outperforms GPT-4(Vision)in MMMU benchmarkNote:Gemini takes in text,image,audio,and video,and outputs text and image responses.Gemini Ultra*56.859.4OverallGPT-4V(ision)(Playground)Art&designBusinessScienceHealth&medicineHuman&social sciencesTech&engineering65.870.059.356.754.748.064.767.372.5

47、78.336.747.135More LLM developers will follow suit to remain competitiveSource:CB Insights market report Large language model(LLM)developersFoundation modelsCompanyTotal fundingModeMultimodal features$14BGPT-4VText,image text$3.4BClaude-2Text only$1.5BInflection-2Text text,image$642MMagmaImage text$

48、469MText only$317MText only$135MGrokText onlyLimited modalities commercially availableIn the roadmap or not announced36Multimodal AI will unlock new commercial opportunitiesSource:CB Insights company profiles Artera,Ghost Autonomy;Google,Meta Foundation modelsSectorCompanyMultimodal AI applicationHe

49、althcareArteraAIs prostate test analyzes both biopsy images and patients clinical data to predict patient outcomesRaised$90M seed round in Q123 from Coatue,J&J,and others SearchGoogle will incorporate Gemini into search Autonomous drivingGhost Autonomy explores using multimodal LLMs in autonomous dr

50、ivingRaised$5M from the OpenAI Startup Fund in Q423Consumer devicesMeta is bringing multimodal AI to consumers with its Ray-Ban smart glassesRoboticsGoogles multimodal model PaLM-E targets robotics,combining language with robotic sensory data37Enterprises do more with pocket-sized modelsFoundation m

51、odels38Enterprise spend on large models can run in the millionsSources:CB Insights software buyer interviews OpenAI,TruefoundryFoundation models“GPT-4 is among the largest and most costly AI models available.Using it to summarize an email is like getting a Lamborghini to deliver a pizza.”Tom Dotan,W

52、SJStarting price for customizing GPT-4 Estimated GPT-4(32K)cost to summarize 6M Wikipedia articles to half the sizeA Fortune 500 healthcare companys annual spend on OpenAI$5M$2-3M$720K39Models are now shrinking,making them faster to train and cheaper to runSource:Research papers and press releasesFo

53、undation modelsSmall language models are up to 88X smaller than GPT-3.5 in parameter size0.020.040.060.080.0100.0120.0140.0160.0180.0200.0GPT 3.5Meta Llama-2MistralMosaic MPTMicrosoft Phi2Gemini Nano175B175B7B 7B 7B7B7B7B2.7B2.7B1.8B1.8B40Despite their size,small language models are posting strong p

54、erformance on common benchmarks Source:CB Insights Earnings transcript analytics,Stanford HELM(as of Jan 24)Foundation models“And the tiny sized ones,you could put in your phone and your PC and so on and so forth it seems surprising,but they all can do the same thing.But obviouslybiggest one large m

55、odel is much more versatile,and it can do a lot more amazing things.And the large one would teach the smaller ones how to be good AIs.”CEO Jensen Huang,Q223 earnings callSmaller models rank in top 6 on the Stanford HELM evaluation framework for foundation models41Sectors with sensitive data will cho

56、ose small models at the edge over LLMsSource:CB Insights Earnings transcript analytics(date reflects quarter call occurred),ComputerworldFoundation models“we are really focused on what we call small language models,which allow the algorithm to live inside of the CDP and see just their data and the Z

57、eta data cloud and get really,really smart on their products without ever allowing their data to leave the ecosystem.That is resonating like nothing Ive ever seen before.”CEO David Steinberg,Q323 earnings call“We are seeing the concern from enterprises about using a model like GPT or PaLM because th

58、eyre very large and have to be hosted by the model providers.In a sense your data does go through those providers.”CEO Arvind Jain,interview with Computerworld42Small models will take over narrow tasks across finance,healthcare,and lawSource:CB Insights company profiles Writer,Ant Group,arcee.aiFoun

59、dation modelsCompanyModelParameter sizeIndustryProductPalmyraMed40B/20BHealthcare40B model ranks 4thin PubMedQA leaderboard for biomedical question-answering AntGLM-Med10BHealthcareRanks 3rdin PubMedQA leaderboard for biomedical question-answering DPT-Patent7BLawPatent generation&summarizationDPT-SE

60、C7BFinanceMining insights from SEC databaseDPT-PubMed7BHealthcareTrained on PubMed database43Open-source mounts a challenge to closed LLMsFoundation models44Open-source LLM developers are seeing rising investor&commercial interestSource:CB Insights market report Large language model developers,Earni

61、ngs transcripts analytics(date reflects quarter call occurred)Foundation modelsTotal funding among open-source model developers*Excludes open-source developers that have not raised equity funding(as of 1/3/2024).were also building foundation models like Llama 2,which we believe is now the leading op

62、en-source model with more than 30 million Llama downloads last month.CEO Mark Zuckerberg,Q323 earnings callRaised nearly$1B combined in the last 2 months of 2023Acquired by Databricks for$1.3B45A record number of AI projects will be built on open-source models in 2024Source:GitHub(data as of 11/8/20

63、23)Foundation modelsGlobal growth in public generative AI projects on GitHubNumber of first-time open-source software contributors46Performance gaps will close fastSource:Massive multi-task Language Understanding scores from published research papers,Papers with Code,and OpenCompassDates represent r

64、esearch paper/model release datesFoundation modelsGPT-4Gemini Ultra(CoT32)Claude-1Claude-2(5-shot CoT)GPT-3GPT-3.5Flan-PaLM 2Chinchilla(few shot,k=5)GPT-2Gopher(few shot,k=5)UnifiedQAPaLM 2Flan-PaLMPaLMGork-1Qwen-72BYi-34BLLama2-70BAquila2-34BLlama-65BFalcon-180B(5-shot)Flan-T5-XXLMixtral 8x7B0%10%2

65、0%30%40%50%60%70%80%90%100%Oct18May19Dec19Jun20Jan21Jul21Feb22Aug22Mar23Oct23Apr24MMLU benchmark performance(%)Smaller open models are outperforming GPT-3.5 and reporting comparable performance to some larger closed models on the MMLU benchmarkPrivateOpen47Enterprises will increasingly consider open

66、 alternativesSource:CB Insights software buyer interviews Anthropic,Replicate,OpenAIFoundation models“We took Llama 2 and fine-tuned it on our data set and then deployed it on A100 GPU on Replicate.Thats pretty much the main workflow.Thats the main part of the Replicate that we use.”Replicate custom

67、erSenior Manager,Data Science at$1B+valuation technology company“GPT Turbo is less expensive.That is usually the default powering for most of our stuff,although that might move to Llama soon because Llama is even less expensive if implemented.”Anthropic customerHead of FP&A at$1B+valuation technolog

68、y company“The big three from my perspective are OpenAI,Llama 2,and then Anthropic.As Ive talked to people in the space,thats the order of most popular,most adopted in terms of people implementing tools.”OpenAI customerPartner at early-stage VC firm48Beyond transformers:New architectures challenge st

69、atus quo Foundation models49Googles research on transformers was a watershed moment in generative AIFoundation models1GOOGLE TRANSFORMERS New neural net architecture for language understandingoutperforms older approaches20172OPENAI GPT Transformer-basedunidirectional contextual AImodel,where a word

70、is taken inthe context of words precedingit in a sentence20184202032019GOOGLE BERT Transformer-based bidirectional AI pre-trained with Wikipedia text,where a word is taken in the context of preceding&succeeding words OPENAI GPT-2 AI pre-trained with 8M pages of internet textOPENAI GPT-3 produces hum

71、an-like text;larger and more compute-intensive than GPT-2;OpenAI releases an API integrationRest is history&50But transformer-based models are leaving a massive carbon footprintSource:Estimating the Carbon Footprint of BLOOM,a 176B Parameter Language Model(Luccioni et al.)Foundation models1.05.518.1

72、25.063.070.0352.0502.00.0100.0200.0300.0400.0500.0Air Travel,1 passenger,NY-SFHuman Life,Avg.,1 yearAmerican Life,Avg.,1 yearBLOOM(176B)Car,Avg.Incl.Fuel,1 lifetimeOPT(175B)Gopher(280B)GPT-3(175B)CO2 equivalent emissions(tonnes)by selected machine learning models and real-life Examples,202251Users w

73、ant higher token limits todays LLMs have limitationsSource:CB Insights software buyer interviews Cohere,Monarch Mixer:A new model architecture for increased efficiencyFoundation models“If you think about a medical consultation for a diagnostic sleep test,this is something that can take 24 hoursJust

74、being able to ingest that amount of information is a boonYou cant really get four hours worth of anything out of many of these models.”“The computational requirements for Transformers increase quadratically with sequence length,which limits the extent of knowledge these models can capture from avail

75、able datasets.We are interested in new,sub-quadratic approaches.”Senior research engineer at Fortune 500 companyPurchase amount:$300K/yearResearch by Dan Fu,Simran Arora,Chris R52One billion tokens?New transformer variations are trying to break the ceilingSource:LongNet research paperFoundation mode

76、lsMicrosoft published research on“LongNet”in July 2023 to scale transformers53Top research firms are going beyond transformersSource:Published research on models mentionedFoundation models1MEGA“significant improvements over other sequence models,including variants of Transformers”Jan 20232Hyena arch

77、itecture”reaching Transformer quality with a 20%reduction in training compute”April 20234Monarch Mixer“showing for the firsttime that it may be possible to match Transformer quality without attention”July 20235StripedHyena“long context,improved training,and inference performance over the Transformer

78、”Dec 20233Pretraining Without Attention“BiGS is able to match BERT pretraining accuracy”May 202354Challengers will gain commercial traction in 2024Source:CB Insights company profile Sakana,media reportsFoundation modelsCompanyDeviceCommercial applicationWorking on“nature-inspired”architectures;found

79、ed by one of the authors of the transformers paperProof-of-concept in autonomous driving with liquid neural network architectureHyenaDNA:A foundational model for genomics based on the Hyena architectureNTT partnership to develop multiple small language modelsCommercial projects underwayResearch/PoC5

80、5AI girlfriends kill human romanceHuman-machine interactions56Young people are having less sexSource:General Social Survey via Washington PostHuman-machine interactionsShare of American men and women between ages 18 and 30 reporting no sex in the past year57Endless swiping isnt workingSource:Culture

81、PlotHuman-machine interactions12 Tinder users activity58Dating companies are turning to genAI to make dating less awkward&keep users hookedSource:CB Insights Earnings transcript analytics(quarter reflects date call occurred)Human-machine interactions“.using generative AI to really be this best frien

82、d in your pocket inside of Bumble to help you date,to build confidence,to feel good about it so that you never delete this app because it didnt make you feel good about yourself.”CEO Whitney Wolfe Herd,Q423 earnings call“By the end of the year,we expect to have launched a number of initiatives that

83、will use generative AI to eliminate awkwardness,make dating more rewarding and surprise and delight users.CEO Bernard Kim,Q323 earnings call59AI girlfriend startups are finding a market in lonely,sexless menSource:Digi,Blush,AnimaHuman-machine interactions60Expect romantic AI to accelerate the decli

84、ne in adult partnershipsSource:Pew Research CenterHuman-machine interactionsA rising share of US adults are living without a spouse or partner%of population ages 25 to 54 who are61Lonely Gen Z befriends AIHuman-machine interactions62We are experiencing a loneliness epidemicSource:American Time Use S

85、urvey via The Washington Post.Compares Human-machine interactionsHours per week compared to 2013,Americans 15 and older63Gen Z spends far less time with friends in person in the smartphone eraSource:Kannan et al.,“US trends in social isolation,social engagement,and companionship,”SSM Population Heal

86、thHuman-machine interactionsDaily average time with friends(minutes)64But who needs real friends when your AI bot can be just as personable as a human?Source:SimilarWeb(data as of January 2024)Human-machine interactionsOver half of Character.AIs 4M users are under the age of 2456.42%22.73%11.42%4.70

87、%2.76%1.97%18-2425-3435-4445-5455-6465+18-2425-3435-4445-5455-6465+65Tech for open-ended dialogues&AI character generation is evolving quicklySource:CB Insights company profiles Inworld AI,Convai,Latitude,Circle LabsHuman-machine interactionsCompanyFundingLatest funding dateApplication$117M10/13/202

88、3AI character generationPartnerships with Xbox&Unity Games$4.8M1/11/2023AI character generation$4M4/1/2023Open-ended storylinesCircle Labs$4.3M11/14/2022AI character generation&dialoguesSeed round led by Lightspeed Venture PartnersAdvances in the gaming industry will lead to more immersive online,so

89、cial experiences 662024 will see an explosion of AI friendships Source:SimilarWeb(data as of September 2023)Human-machine interactionsEmerging AI companion appsAI companion startup Character.AI is close behind ChatGPT in mobile usage67The next iPhone:AIs hardware revolution Human-machine interaction

90、s6815 years after its release,iPhone sales are stagnating.Apple has been criticized for hitting an innovation plateauSource:Apple via Business of Apps.Note:Apple stopped reporting iPhone sales in 2018.All values afterward are estimates.Human-machine interactions11.620.739.972.3125150.2169.2050100150

91、200250300200820092010201120122013201420152016201720182019202020212022Unit sales(millions)231.2211.8216.7 217.7187.2196.9242232.269AI breakthroughs have launched the search for the next iPhoneSource:CB Insights company profiles Humane,Rewind,Tab,RabbitHuman-machine interactionsCompanyFundingLatest fu

92、nding dateProductUser interaction$242M3/8/2023Screenless wearable,powered by ChatGPTVoice,gesture$27M5/9/2023Wearable“pendant”records&analyzes audioVoice,appTab$2M1/10/2024Wearable necklace records&analyzes audio with GPT-4Voice,app$30M12/21/2023Rabbit OS,LLM-based operating system;r1 deviceVoice,bu

93、ttons,screenStartups developing consumer AI devices70Incumbents with deep AI talent&pockets have an advantageSource:CB Insights research The big tech company leading in AI acquisitions(data as of 9/19/2023)Human-machine interactions20102011201220132014201520162017201820192020202120222023YTD322118171

94、0Apple leads big tech in aggregate AI acquisitions71Industry heavyweights are betting on AI-native hardwareSource:CB Insights key people data Jony Ive,Tang Tan;1EngadgetHuman-machine interactionsKey peoplePositionPrevious positionsJony IveFounderSVP of Industrial Design,Chief Design OfficerTang TanH

95、ardware engineering leadVP,iPhone&Apple WatchOpenAI is in talks to build new AI hardware with LoveFrom,which has poached over 20 Apple employees172Big tech and OpenAI will push forward on early experiments in 2024Source:Company reportsHuman-machine interactionsCompanyDeviceCommercial developmentAI-p

96、owered Ray-Ban smart glassesApple Vision Pro releasing February 2024“Apple GPT”powers something new?Developing the“iPhone of AI”Commercially availableReleasing soonNo commercial products disclosed73Future vision:LLMs evolve from chatbots to operating systemsSource:Andrej Karpathy and rabbit via XHum

97、an-machine interactionsOpenAIs Andrej Karpathy floats the idea of LLM-based OSrabbit launches$199 device based on LLM OS74Autonomous agents:From hype to somewhat usefulHuman-machine interactions75The idea of Jarvis-like AI agents caught on like wildfire in 23Source:CB Insights News trends tool(based

98、 on an analysis of 5K news sources)Human-machine interactionsMedia mentions of popular autonomous agents LLM-powered bots that can independently reason and execute tasks 76Autonomous agent projects proliferate but are far from being commercially usefulSource:Select agent research papers and tools pu

99、blished in 2023Human-machine interactionsAgents are getting better at reasoning,but struggle with executing tasksProjectPublish/Release dateDeveloperAgent descriptionToolformersFeb 23MetaLanguage models that“teach themselves to use external tools”AutoGPTMar 23Significant GravitasOpen-source agent po

100、wered by GPT-4 to perform tasksBabyAGIApr 23Yohei NakajimaOpen-source agent built on GPT-4 and vector database PineconeGenerative agentsAug 23Stanford,GoogleA simulation“where end users can interact with a small town of twenty-five agentsusing natural language”VoyagerOct 23Nvidia,Caltech,and others“

101、lifelong learning agent in Minecraft that continuously explores the world,acquires diverse skills,and makes novel discoveries without human intervention”Mind2WebDec 23Ohio State UniversityDataset for developing autonomous agents that can perform tasks on websites77The tech is advancing fast:Open-sou

102、rce projects get massive interest from developer communitySource:Star-Human-machine interactionsAutoGPT gets 100K stars(indicating popularity)on GitHub within a month of launch78Big tech and VC interest will move the needle in 2024Source:CB Insights company profiles OpenAI,Imbue,Adept,HyperWrite,Mul

103、tiOn;published research and press releases Human-machine interactionsCompanyAutonomous agent initiativesTotal funding/latest roundInvestorsRumored to be training autonomous agentsDec 23:published research on best practices for governing agents Announced$10,000$100,000 grants for researching agent im

104、pact$14B/Secondary marketMicrosoft,Andreessen Horowitz,Founders Fund,Khosla Ventures,Sequoia CapitalBuilding AI agents and testing different architectures for development$232M/Series BEric Schmidt,The Alexa Fund,Astera Institute,NVIDIASoftware-agnostic multimodal agent$415M/Series BGeneral Catalyst,

105、Spark Capital,Greylock Partners,Microsoft,NVIDIAPublished AutoGen research on multiple AI agents talking to each other-Hired researcher Edward Grefenstette in June 23 to build“general agents that can adapt to open-ended environments”and communicate with humans-Assistants for everyday tasks,from book

106、ing flights to research$5.4M/SeedActive Capital,Cortical Ventures,Madrona Venture GroupAI agents to tackle mundane tasks$4M/SeedThe MBA FundUNICORNSBIG TECHEMERGINGSTARTUPS79Copilots drive massive white-collar productivity gainsGenerative AI applications80US labor productivity growth has stagnated i

107、n recent yearsSource:US Bureau of Labor StatisticsGenerative AI applications2.7%1.4%1.6%2.1%2.8%1.5%1.5%1947 Q1-1973Q4chips1973 Q4-1980 Q1GuatemalaCosta RicaBolivia2.1%1947 Q1-1973Q4chipsCurrent Current CycleCycle1947 Q1 1973 Q41973 Q4 1980 Q11980 Q1 1990 Q31990 Q3 2001 Q12001 Q1 2007 Q42007 Q4 2019

108、 Q42019 Q4 2023 Q31947 Q1 2023 Q3(long-term rate)Average annual percent change(nonfarm business sector)81Digital adoption is correlated with US labor productivity growth.AI will launch the next boomSource:McKinsey&CompanyGenerative AI applicationsSectors labor productivity growth compared with digit

109、ization level182In software development,AI copilots are becoming essential to developer productivity Source:GitHub.Based on a sample of 95 developers split randomly into 2 groups to complete a coding task.Copilots=generative AI-powered tools designed to work alongside humans,automating tedious tasks

110、Generative AI applications1 hour,11 minutesaverage to complete task2 hours,41 minutesaverage to complete task71 minutesThats 56%less time161 minutesWith copilotWithout copilot83Coding copilots will become standard-issue tools at enterprisesCompany/productStatusApplicationReleased by Microsoft(10/29/

111、2021)Code completion/generatorReleased by OpenAI(11/30/2022)Code generatorReleased by Meta(8/24/2023)Open-source code generatorStableCode by Released(8/8/2023)Open-source code generatorReleased by Hugging Face&ServiceNow(5/8/2023)Open-source code generatorReleased(4/13/2023)Code completionAIReleased

112、(10/9/2023)Code completion/generatorPrivate(founded 2017)Code completionPrivate(founded 2020)Command line automationGenerative AI applicationsSource:CB Insights company profiles;company reports84In healthcare,copilots are automating tedious tasks like documentationSource:CB Insights market report Vi

113、rtual medical scribes&summarization toolsGenerative AI applications“Reduces clinicaldocumentation time by 72%”“75%reduction in documentation time”“Giving you back2-3 hours a day”Top companies by equity funding(as of 1/10/2024)85In law,copilots are drafting contracts,summarizing documents,&optimizing

114、 researchSource:CB Insights software buyer interviews CasetextGenerative AI applications“Its the ability to draft a legal memo.Its the ability to use parallel search in neural searches to find legal research.Its the ability to go beyond using keywords and e-discovery and use concepts like,I dont fee

115、l comfortable with this.That would never come up as a relevant document.Absolutely transformational.I wasnt worried thinking,Whats the ROI on this?It was,How do we become an early adopter?It has the ability to really improve the efficiency of legal services.”Casetext customerChief Innovation Officer

116、,National law firmAcquired by for$650M in June 202386Augmented workers will do more in the copilot ageSource:CB Insights software buyer interviews UniphoreGenerative AI applications“Uniphore was implemented with 1,000 agents and the benefits were a 20%reduction in what we call average handle timeThe

117、n there was an 80%reduction in after-call work time.Third,about 20%avoidance in hiring due to automation and basically an annual savings of$6 million through conversational AI and automation.”Uniphore customerManaging Director,Innovation at Fortune Global 500 company87AI vs.AI:GenAI fuels new wave o

118、f attack and defenseGenerative AI applications88Cyber attacks continue to rise,with AI a driving forceSource:CheckPointGenerative AI applications02004006008001,0001,2001,400Q1Q2Q3Q4Q1Q2Q3Q4Q1Q2202120222023Global average weekly cyber attacks per organization hit an all-time peak in Q223 1,2581,25889D

119、eepfakes are getting more sophisticated using genAI tools like face swapping,voice spoofingSource:Onfido;CB Insights Earnings transcripts analyticsDate reflects quarter call occurredGenerative AI applications31x increase in deepfake attempts2023 vs.2022recorded by one identity verification platform“

120、The pace of new threats and the tech behind them now outpaces any notion that a weeks old view of historic cyberattacks can fully protect you as a business,whether from the threats that were seeing today,or the threats of tomorrow whether thats self-evolving code designed to evade the best of those

121、defenses,it is automated deepfakes where you cant hear or trust what youre hearing or seeing,or your employees being tricked into being inadvertent insiders.”CEO Jack Stockdale,Q323 earnings call90ChatGPT drives an explosion of email phishing attacksSource:SlashNextGenerative AI applicationsOne emai

122、l security vendor recorded a 1,256%increase in phishing emails from Q422 to Q323 Malicious phishing emails91Cybersecurity vendors use generative AI to counter attacksSource:CB Insights software buyer interviews SlashNextGenerative AI applications“SlashNext is working on a generative AI-based solutio

123、n,which means it will generate its own kind of phishing and malware and it will train the software to automatically be aware of any new kind of threats arriving in the marketThat is a differentiator from the technology perspective.”SlashNext customerSenior Design Engineer,Fortune 500 company92A new

124、crop of genAI-focused cybersecurity vendors will emerge Source:CB Insights company profiles Wraithwatch,Jericho Security,Nexusflow,Brightside AI,ShibolethGenerative AI applicationsCompanyFoundedHQTotal fundingStageApplication2023US$8MSeedAI attack simulation&prevention2023US$5MPre-seedBreach&attack

125、simulation2023Switzerland$1MSeedBreach&attack simulation2023US$11MSeedSecurity operations center copilots2023US$0.5MAcceleratorDeepfake audio detection93Covering AI threats will be a major selling point for buyersSource:CB Insights software buyer interviews GoSecure,ProofpointGenerative AI applicati

126、ons“With artificial intelligence,we see that there are tools popping up to help attackers leverage AI to generate extremely convincing phishing attacks.its going to be really important that we address this.”GoSecure customerC-level executive,Academic institution“Proofpoints room for growth lies in i

127、ts ability to detect emerging threats generated by artificial intelligence.As generative AI and other types of threats become more sophisticated and widespread,they pose a greater risk to our business.”Proofpoint customerDirector,Publicly traded manufacturing company94GenAI takes over journalismGene

128、rative AI applications95US newsroom employment is in declineSource:Pew Research CenterGenerative AI applications96Media organizations are taking up AI to speed up news-gathering,production,and distribution Source:CB Insights business relationship data;media reports Generative AI applicationsOrganiza

129、tionProduct/partnershipApplicationThe Dynamic MeterMachine learning model sets personalized meter limits for a smarter paywall$5M partnership with OpenAI Leverage AI tools to support&expand local news coverageHeliografFirst introduced in 2016,Heliograf automates reporting on structured data(e.g.,spo

130、rts scores)SAMSocial media monitoring tool to stay on top of breaking news97But with genAI-powered search on the horizon,will it be enough?Source:SimilarWeb,WSJGenerative AI applications“A task force at the Atlantic modeled what could happen if Google integrated AI into search.It found that 75%of th

131、e time,the AI-powered search would likely provide a full answer to a users query and the Atlantics site would miss out on traffic it otherwise would have gotten.”40%40%60%Company APublisher traffic generated by Google98News organizations will land new deals with LLM developers to license content as

132、revenue models shiftSource:CB Insights business relationship data;media reports Generative AI applicationsIn talks with(Politico,Business Insider)In talks withDeal in the“10s of millions of dollars”(WSJ)Deals with(People,The Daily Beast)99AI news anchors are herereporters next Source:NYT,Channel 1,F

133、uturism,The GuardianGenerative AI applicationsCompanyCountryDescriptionUSLaunching“generative AI news network”in fall 2024 with AI-generated news anchors and digitally generated voicesKoreaDebuted AI news anchor Zae-In designed by startup Pulse9KuwaitDebuted AI-generated news presenter FedhaIndiaDeb

134、uted AI-powered anchor SanaMalaysiaDebuted AI-powered anchors Joon and MonicaUSTesting“Genesis”tool that can produce news storiesUSReceived flak for publishing articles by fake,AI-generated writers100Beyond stills:Video generation makes stridesGenerative AI applications1012022 kicked off the text-to

135、-image craze.But text-to-video is a bigger challengeSource:Facebook,Hugging Face blogGenerative AI applications“Its much harder to generate video than photos because beyond correctly generating each pixel,the system also has to predict how theyll change over time.”CEO Mark Zuckerberg,Q322“Text-to-vi

136、deo are typically trained on very short clips,meaning they require a computationally expensive and slow sliding window approach to generate long videos.As a result,these models are notoriously difficult to deploy and scale and remain limited in context and length.”Former ML Engineer Alara Darick,Q22

137、3102Now generative video models are advancing rapidlySource:Company press releasesGenerative AI applications1Meta Make-A-Video(short soundless videos)Sept 20222Oct 20225Apr 20237Nov 20234Mar 2023Google Imagen Video(short clips)&Phenaki(lower quality,long clips)Runway Gen-2(4-second clips)Nvidia Vide

138、o LDM(high-resolution clips)Runway Gen-2(up to 18-second clips)11.2.2023Runway Gen-2(higher-definition clips)6Aug 202311.16.2023Meta Emu Video(high-quality 4-second clips)11.21.2023Stability Stable Video Diffusion(high-quality 2-second clips;model available through its API)11.28.2023Pika Labs Pika 1

139、.0(3-second clips with editor)Morph Studio Text-to-video AI tool introduced3Feb 2023Runway Gen-1(requires video inputs)103Watch for immediate impact across advertising,Hollywood,and other visual media industriesSource:CB Insights company profiles Runway,Pika,Stability,ByteDance,MoonvalleyGenerative

140、AI applicationsStartups race to build video generation&editing platforms for enterprise and creative use casesCompanyTotal fundingStageSelect investorsNotable developments$237MSeries CGoogle,Nvidia,Salesforce Ventures,CoatueRunway partners with Getty Images to create genAI video model aimed at Holly

141、wood studios and advertising,media,&broadcasting companies(December 2023)$55MSeries ALightspeed Venture Partners,Adam DAngelo,Andrej KarpathyReported 500,000 users“generating millions of videos per week”(December 2023)$151MSeedLightspeed Venture Partners,Coatue,IntelText-to-video model available to

142、developers via Stability AI API(December 2023)Assistive-Generative video platform available via web&API(launched December 2023)$7.4BSeries DTiger Global,KKR,TencentReleased MagicVideo-V2,a multi-stage end-to-end video generation pipeline(January 2024)Moonvalley$30M-Khosla Ventures,Y CombinatorGenera

143、tes cinematic film quality footage;gained 100,000 users in 6 weeks on Discord(November 2023)need a quick pass104A genAI graveyard follows frenzyGenerative AI applications105GenAI frenzy launched AI valuations into the stratosphere in 2023Source:CB Insights valuation data The AI premiumGenerative AI

144、applicationsMedian valuations for VC-backed deals in 2023106As the smoke clears,growing competition threatens some early genAI companiesSource:CB Insights software buyer transcript JasperGenerative AI applicationsText generation startup Jasper announced layoffs&cut internal valuation in Q323“So in t

145、his newly emerging world of generative AI its hard to keep up with all the changes that are going on.Its probably not fair to ask this,but Ill say it:Jasper needs to stay up to date faster to make me a definite yes to renewIm seeing other lower cost options;the price of calls to GPT-3,for example,ha

146、s gone down to minimal numbers.It might be harder to say yes to a renewal when were due next year,so Ill have to really see that our people have picked this up and are finding great value.”Jasper customerC-level executive at$10M+funded research platform107Startups in horizontal markets without a moa

147、t will bite the dustSource:CB Insights research Generative AI Market MapGenerative AI applicationsText generation and visual media generation are especially vulnerable to pressure from large players Horizontal generative applications108In 2024,attention will be on companies building stickier solutio

148、ns in specialized areasSource:CB Insights research Industrials,Drug R&D,Financial Services Generative AI applicationsGenAI opportunities in verticals109Research briefsGenerative AI bible:The ultimate guide to genAI disruptionThe large language model operations market mapThe open-source AI developmen

149、t market mapThe generative AI market mapThe state of generative AI in 7 chartsMarket reportsLarge language model developersLLM application developmentEnterprise avatarsText content generationOur most popular generative AI research110Generative AIExpert CollectionStay up-to-date on generative AI with

150、 the analyst-curated Expert CollectionFilter by technology,funding,valuation,business model,and moreFilter by potential vendors,partners,and targets with real-time dataCB Insights helps the worlds leading companies understand everything they need to know about disruptive technologies find out more a

151、bout why our customers love us here.111Buyer TranscriptsCB Insights helps the worlds leading companies understand everything they need to know about disruptive technologies find out more about why our customers love us here.Buyer profileUnderstand the buyer their title,their organization,and their experiencePurchase intelligenceHow much did they spend and whenConsistent structureModerated&structured conversations make transcripts easy to scanDeep insightsDetails on everything from evaluation criteria to competitors to satisfaction112vTry our platform free!