從科幻小說到緬因街：如何利用 AI 的力量開展業務.pdf

1、From Sci-Fi to Maine-Street How to Harness the Power of AI for BusinessPresented by Justin Gibbs,Director of Strategy&Insights,SEONovember 14,2024Maine Outdoor Economy SummitJustin GibbsDirector of Strategy&Insights,SEOPossibly an AIWith YouJustov KlimtPainter,Absinth ConnoisseurBaron von GibbsArist

2、ocratJasper GibbonsCoal MinerAlso With YouAgendaWhat is AI?AI History and Why its HotThe landscapeImage generation funIn the newsPromptingAdding AI to your organizationInternal toolingWhat is AI?What isAI?AI is an umbrella term,a fieldThe definition of which is not totally agreed uponWhat isAI?It en

3、compasses a wide range of technologies,from simple rule-based systems to complex neural networks and beyond.It includes everything from basic automation(like chatbots)to advanced machine learning models that drive self-driving cars or recommend medical treatments.Includes a spectrum of technologies:

4、Machine learning Neural NetworksDeep LearningCognitive computingWhat isAI?It is not:an omniscient entity.At least not yetWhat is AI?Three Categories of Artificial IntelligenceGeneral AI(AGI)This is the GoalArtificial SuperintelligenceThis is where things get scaryNarrow AIWe are here(somewhere in th

5、e middle)Three Categories of AIWhat isAI?We use AI all the time and dont even know it.Its been in the background of some of our favorite apps recognizing patterns for a long time.What isAI?Recommendation EnginesPhotoshop healing brushAutocompleteSpellcheckSpam filtersCAPTCHA.(Completely Automated Pu

6、blic Turing tests to tell Computers and Humans Apart)are those puzzles that youre asked to solve when youre trying to log in to a website or fill out a form online.They usually involve identifying images with specific objects(like traffic lights or bicycles)or typing in distorted text.Reliance on St

7、atistical Probability.Uses statistical probability to make informed predictions and decisionsIdentify patterns and make probabilistic assessmentsUse statistical methods to improve their predictions and adapt over time.When and where was the term AI first coined?A Brief history of AI:an ebb and flow1

8、950s1950:Alan Turing publishes Computing Machinery and Intelligence,proposing what is now known as the Turing Test.1956:The term Artificial Intelligence is coined at the Dartmouth Conference by John McCarthy,Marvin Minsky,Nathaniel Rochester,and Claude Shannon.AI research faces criticism and financi

9、al cuts,leading to the first AI Winter due to inflated expectations and lack of breakthroughs.Expert systems,like DENDRAL,begin to emerge as a focus for AI research.1970sThe development of machine learning algorithms begins to gain traction.Expert systems,such as XCON,become widely used in industrie

10、s,marking a resurgence of interest in AI.The second AI Winter occurs in the late 1980s due to the limitations of expert systems.1980sThe emergence of the World Wide Web leads to vast amounts of data becoming available,setting the stage for big data analytics and AI.Machine learning,especially neural

11、 networks,starts to become popular again.1990sEarly AI research is funded by the Department of Defense and focuses on basic problem solving and symbolic methods.1966:The first chatbot,ELIZA,is created by Joseph Weizenbaum.1960sAI becomes an integral part of the technology industry,with machine learn

12、ing powering search engines,recommendation systems,and more.2006:The term deep learning is introduced,representing a new generation of neural network research.2000sA Brief history of AI:an ebb and flow2010s2011:IBMs Watson wins Jeopardy!,showcasing the progress in natural language processing.2012:Th

13、e success of AlexNet in the ImageNet competition marks a significant breakthrough in deep learning for computer vision.AI research and applications explode,with advancements in deep learning,reinforcement learning,and the rise of AI in consumer products.AI ethics and the implications of AI on societ

14、y and jobs become a significant focus.AI models become more powerful and capable,with systems like GPT-3,4,4o demonstrating human-like text generation.AI research continues to push the boundaries,with models becoming more efficient and capable of multitasking across different domains.2020sSo why doe

15、s everybody suddenly care about AI?Up until about 2022,AI was basically pattern recognition.With enough training data,a machine could recognize just about anything.Generative AI turns this around,making something“new”-words,pictures,video,audio INVASIVE PERSONALIZATIONTAINTING THE MODEL.Nightshade a

16、nd Glaze mission-driven,intelligent tool that acts as a check on AI companies who use artwork from artists to train their models without permission.With this ability,artists can poison this training data,thereby upsetting the balance of upcoming iterations of AI models that generate images.https:/ni

17、ghtshade.cs.uchicago.edu/THE EXPERT HALO.The expert halo is a cognitive bias that occurs when people assume that someone with more knowledge or experience is an expert and therefore infallible.This can lead to group members allowing the perceived expert to dominate decision-making,without using thei

18、r own cognitive processes.The outdoor industry has identified the expert halo as a contributing factor in many wilderness tragedies.Perception of CompetenceAuthority BiasOverconfidenceRisks to Decision MakingInfluence on Public OpinionReduced Diversity of ThoughtRisk of MisinformationDont eat rocks.

19、We didnt listen.We were idiots.THE CURRENT AI LANDSCAPEhttps:/ AI Landscape:Frontier ModelsMicrosoftOpenAIGoogleAnthropicMeta-GPT 4o(omni)-Gemini 1.5Claude 3.5Llama 3CompaniesProductsModelsCopilotChat GPTGeminiClaudeMeta AIThere are many others-Apple says they have one,Microsoft is building their ow

20、n,X has Grok,Amazon has Q Generative AI Landscape:Other Names to KnowImage GenerationCompanion?ProductsCategoryNotebook LMPerplexityMidjourneyCharacter.aiPlus tons of other apps with AI built in,Canva,Adobes creative cloud,Sales Force,etc.etc.etc.ResearchEducation(?)No,Instagram,I dont need to inter

21、act with a chatbot here!Im just trying to look through messages from my friendsOUR VISIONOUR VISIONMy dog and I fit the target demobut I want this picture of my dog and I to be in York,ME“As a pixar animation”HallucinationsBlogImageryIn the NewsAI-authored books are being sold on AmazonSome contain

22、dangerous adviceSimpleQA is a benchmark designed to test language models on short,fact-seeking questions,aiming to reduce hallucinations.The dataset emphasizes quick,efficient grading,enhancing researcher experience.36%of IT professionals believe many AI projects are primarily aimed at impressing in

23、vestors and stakeholders,rather than delivering practical value.CIO This trend raises concerns about the effectiveness and strategic importance of such initiatives.Experts warn that AI projects lacking clear goals and measurable outcomes are unlikely to provide real value,potentially leading to skep

24、ticism about AIs potential and hindering future investments.To maximize AIs benefits,organizations should focus on projects with well-defined objectives and measurable results.*https:/ North Carolina man has been charged with orchestrating a fraudulent scheme that generated over$10 million in royalt

25、ies by exploiting music streaming platforms.He allegedly created hundreds of thousands of songs using artificial intelligence and employed automated accounts to stream these tracks continuously.If convicted,he faces up to 20 years in prison for his actions.Researchers at Google DeepMind have develop

26、ed GNoME,an AI tool that has identified 2.2 million new crystalline materials,including 380,000 predicted to be stable.This discovery significantly expands the known materials database,potentially accelerating advancements in technologies like electronics and energy storage.The AIs predictions were

27、validated through autonomous robotic experiments,achieving a 71%success rate.ServiceNow has announced updates to its Now Assist AI platform,introducing a library of customizable AI agents tailored to enterprise workflows.Enables clients to build,test,and deploy generative AI skills,assigning them to

28、 applications for enhanced functionality.Users can control the extent of human oversight over AI agents,ranging from full autonomy to requiring approvals.The platforms multi-agent system allows AI agents to collaborate and access company data,facilitating comprehensive workflow automation.Prompt Eng

29、ineeringTips for Promptinghttps:/ for PromptingA few moreChain of Thought PromptingZero Shot Chain of thoughtSelf ConsistencyKnowledge Generation PRomptingPrompt PatternsReACT(Reasoning and Action Prompting)Tips for Promptinghttps:/ for PromptingDos Use direct instructions or a clear question.Use co

30、ncise and unambiguous languageProvide any relevant information or data as contextGive examples in promptsSpecify the desired output formatEncourage the model to be factual and cite sourcesDontsUse ambiguous language.Assume the model knows what it is youre looking forAdding AI to YourOrganizationAI i

31、snt a technology problem,its a change management problem.Asking“what vendors do we choose”or“where do we put AI”are the wrong questions.Approach with a specific need in mindTool usage should be driven by a genuine need to solve specific problemsAvoid adopting AI just to claim“we use AI.”Misallocated

32、 resources and inefficiencies can result from forced AI adoption.AI should enhance productivity,streamline processes,and drive innovation.Superficial AI implementations add complexity without real value.Ensure AI leads to tangible workplace improvements.Approach with a specific need in mindAI washin

33、g refers to the practice of companies exaggerating or falsely claiming that their products or services utilize artificial intelligence(AI)to enhance their appeal or market value.Many AI tools are first-to-market cash grabs as companies rush to release AI solutions to capitalize on hype and demand.Hu

34、rried releases often lack depth,robustness,and real value.Critically assess AI tools for true value and reliability.Avoid being swayed by market trends and flashy featuresEvaluating AI ToolsNot all AI tools are the same.Many third-party AI solutions are layers on top of LLMs like GPT-4,LLAMA,etc.Thi

35、rd-party tools introduce potential security risks.Additional technology layers can be vectors for vulnerabilities.Direct use of foundational LLMs can reduce security risks.Evaluate AI tools for necessity and security implications.Ensure chosen AI solutions are effective and secure.Evaluate Needs and

36、 Define ObjectivesIdentify Business ProblemsSet Clear ObjectivesStakeholder InputConduct a Feasibility AssessmentTechnical FeasibilityResource AvailabilityFinancial FeasibilityRisk AssessmentSecurity RisksCompliance and Legal RisksOperational RisksChoose the Right AI ToolsVendor EvaluationPilot Proj

37、ectsDevelop an Implementation PlanProject TimelineIntegration StrategyTraining and SupportMonitor and Evaluate PerformancePerformance MetricsContinuous ImprovementAdding AI to your organizationInternalToolingInternalToolingLarge Language Models are really good at writing code to create specialized f

38、unctions.This is one area that weve used AI in order to increase our outputs,and have found it to be incredibly valuable.Describe what youre trying to achieve in ChatGPT or Claude,and ask it to write Python code to be used in Google Colab notebooks.A few examples that help us in our day-to-day:Inter

39、nalToolingConvert XML sitemaps to CSV and downloadInternalToolingGenerate QR codes with UTM parametersInternalToolingMerge multiple related CSV files into a single XLXSInternalToolingMerge multiple.csv files into a single.xlxs filehttps:/ an XML file to a CSV filehttps:/ Code Generatorhttps:/ the En

40、vironmental Impact of AI.https:/www.iea.org/commentaries/the-carbon-footprint-of-streaming-video-fact-checking-the-headlines https:/ ENVIRONMENTAL IMPACT OF AI14.5%global greenhouse gas emissionsMeat and dairy for consumption.Carbon Emission Context2.5%global greenhouse gas emissionsEntire aviation

41、industry(though its worth noting that WTTC says the total impact of“travel”is more like 7-8%)2.5%global greenhouse gas emissionsALL emissions from the production,use,and disposal of digital technologieshttps:/planbe.eco/en/blog/what-is-the-digital-carbon-footprint/https:/www.fao.org/family-farming/d

42、etail/en/c/1634679/https:/ 162gTikTok usage runs 2.7g per MINUTE.Thats partly because the servers that run the service are located in jurisdictions with lower emissions standards.The average user spends about one hour a day,yielding 162g of carbon per average user.ENVIRONMENTAL IMPACT OF AIComparing

43、 AI to other digital usage3gA typical Google search.Google often references a.2g cost of a search result-which predates the inclusion of Generative AI results.This also ignores the subsequent page views which average about 1.5g per web page viewed because those pages are hosted all over the world in

44、 data centers with wildly different carbon efficiency.15gJust the beef portion of a typical restaurant hamburger runs 15g.That doesnt include the energy cost to cook it,or all of the other components of a typical meal.Figure your average hamburger meal runs 20-25g-again making some assumptions about

45、 the global average cost of raising a cow.0.5gof carbon emissions per average AI interaction55gWatching an hour of Netflix averages around 55g of carbon emissions.That doesnt include the energy to run your TV.Estimates go as high as 100g per hour,or as low as 20g.It all depends on where you watch an

46、d how the energy to power the data transmission is generated.(well look at that in a minute)Disclaimer:This is an extremely complex issue.Quantifying carbon emissions for any given activity,regardless of whether thats a hamburger or an interaction with ChatGPT is profoundly influenced by WHERE the a

47、ctivity is happening,WHEN you measure it,and HOW you draw the lines around what is being measured.Which model and version are we measuring?Gross or net carbon emissions?Which data centers were/are being used?How efficient are those data centers?When was the measurement done?Is the creation and usage

48、 of A.I.creating a MASSIVE energy problem?ENVIRONMENTAL IMPACT OF AIOpenAI ChatGPT 4o model is reported to have required 10 gigawatt-hours.Ignoring all other factors,that would be a carbon debt of roughly 4,000 tons.Thats not a trivial amount,something like the carbon from 1000 peoples driving for a

49、 whole year.BUUUUT,thats using the national(or even global)average of the carbon cost of generating a KWh of energy.Carbon debt of traininghttps:/www.washington.edu/news/2023/07/27/how-much-energy-does-chatgpt-use/https:/www.epa.gov/greenvehicles/greenhouse-gas-emissions-typical-passenger-vehicle ht

50、tps:/ 10gigawatthoursEXAMPLE CALCULATIONGross CO2 e/KWh(kg/KWh)=.478kg per Kwh(or 478g per KWh)Net CO2 e/KWh(kg/KWh)=.08kg per Kwh(or 80g per KWh)https:/arxiv.org/ftp/arxiv/papers/2104/2104.10350.pdf https:/ of 80 grams is 0.04g of carbon per AI interaction.0005 of 478 grams is 0.24g of carbon per A

51、I interactionLets use Googles Council Bluffs Iowa Datacenter as our example.ENVIRONMENTAL IMPACT OF AIIt matters a lot where the training is happening and how the energy for that facility is generated.OpenAI isnt training in“the global average”data center,theyre using places like Microsoft Azure dat

52、a centers,which have a SIGNIFICANTLY lower carbon cost per kilowatt,and includes a bunch of mitigations that make the real figure something more like 800 tons or about 180 peoples driving for a year.Its not nothing,but its also not a whole lot compared to 350,000,000 Americans driving cars.Carbon de

53、bt of traininghttps:/www.washington.edu/news/2023/07/27/how-much-energy-does-chatgpt-use/https:/www.epa.gov/greenvehicles/greenhouse-gas-emissions-typical-passenger-vehicle https:/ 0.05g Divide the training cost by the(roughly projected)number of interactions and you approach a“debt”of about 0.05-0.

54、25 grams per interaction.ENVIRONMENTAL IMPACT OF AIInference is the simplest measure to focus on,because its most directly within your control.Here again,there are a variety of numbers floating around.Similar to the previous slide,a number of 4g per promptresponse interaction is based on the assumpt

55、ion that the energy is calculated at the global average cost of a KWh.It can also depend on what model was used for the measurement and when it was measured.Remember that GPTs per-token cost has halved twice in the last six months alone.Carbon cost of inferencehttps:/planbe.eco/en/blog/what-is-the-d

56、igital-carbon-footprint/#:text=g%20CO2e/min-,Is%20streaming%20on%20Netflix%20bad%20for%20climate%20change,by%20the%20end%20of%202022.https:/ of carbon emissions per interaction,or lets say between 1/10 and 1/2 of a gram on average if youre also paying back the“debt”of training.But arent those number

57、s also dependent on where the compute is happening?THE IMPORTANCE OF LOCATIONhttps:/www.iea.org/commentaries/the-carbon-footprint-of-streaming-video-fact-checking-the-headlines This is the carbon output of watching one hour of streaming video in the USA(average)31gof carbon emissionsTHE IMPORTANCE O

58、F LOCATIONhttps:/www.iea.org/commentaries/the-carbon-footprint-of-streaming-video-fact-checking-the-headlines This is the same one hour of video in New Zealand8gof carbon emissionsCASE IN POINT:BREAKTHROUGHS IN EFFICIENCYAs an example of the various kinds of efficiency gains that we should be antici

59、pating,this research paper defines a new algorithm for use in AI models that replaces expensive multiplication operations with simple addition to significantly reduce computational energy of inference.It estimates a 95%reduction in energy consumption.Wait what about the water consumption?ENVIRONMENT

60、AL IMPACT OF AIBut what about water?https:/ https:/ the total USAs water withdrawal in 2027 is estimated to be needed specifically for AI use.That sounds low,but were already short on water-two-thirds of the global population experiences water scarcity for at least one month each year.This is also a

61、 massively complex issue,which also depends greatly on which data center(s)were talking about,and which parts of the total system were including.There are two key issues to be aware of-water withdrawal and water consumption.Consumption is essentially evaporation,which is about 1.8%of the water thats

62、 withdrawn.That water isnt lost exactly,but its also not immediately reclaimable.Withdrawal is the amount of freshwater that is removed from the global system for use in data centers,which is the bigger concern because its“trapped”in the datacenter and therefore unavailable for other uses.WATER NEUT

63、RALITYSimilar to the carbon problem,it all depends on how that particular data center is handling water.The major cloud services companies are pursuing water neutral/positive initiatives that will“return more clean water to the system”than they withdrew.so should I worry about this or not?ENVIRONMEN

64、TAL IMPACT OF AIThe environmental impact of AI is a very real concern.You absolutely should consider the impact your usage of AI has,but you cant do that as a silo,it has to be considered in context with the impacts of your other actions and choices.Current state of search SGE-AIOAIOverviews FAST FA

65、CTSHow often is Googles AI Overview displayed in search?Answer:Since the beginning of the year,AIO has appeared in 1%-30%of travel-related search queries;1%-36%(restaurants)BUT hotels are limited with some sources saying less than 5%of searches for hotels are showing AIO resultsThis number is changi

66、ng month-month as Google works to refine its prediction software https:/ FACTSWhat%of AI overviews pull from traditional search results?Answer:66%of AIO results(as of now)didnt display the same sources appearing in traditional search results.While Google AIO prioritizes helpful content over traditio

67、nal ranking,this doesnt diminish the value of classic SEO tactics.Marketers now have the chance to boost their visibility by adopting a multifaceted approach in digital marketing https:/ Preparing for the Searchstorm“Off Page”Optimizations Google Business Profile is KingProfile Maintenance Reputatio

68、n Management Active Review Management:Regularly monitor and respond to reviews.This proactive approach not only improves your reputation but also enhances your visibility in SGE local results.Encourage guests to leave positive reviews on platforms like Google and Tripadvisor.Reputation Management Pr

69、eparing for the Searchstorm“On Page”Optimizations For example,AIO is more likely to appear under the following circumstances:Increases:“Best,”(+50%)“what is,”(+20%),“how to”(+15%)and“symptoms of”(+12%)queries are more often to trigger an AI OverviewsDecreases:“Vs”(-20%),brand-specific(-15%),general

70、product(-14%)and lifestyle-related(-12%)queries trigger AI Overviews less often.Optimizing ContentLeverage Structured DataStructured format,like lists into product descriptions,enhances readability and aligns with SGEs preference for well-organized content,potentially improving the visibility of your products in search results.This may include room types,amenities,events and more.