🤖

[Notes] AI

AI Intro

image

LLM (Large Language Models)

  • AI tools that are trained on massive data sets to read, summarize, and translate text, then generate sentences by predicting future words. The more data that the model is trained on the more it is able to learn, remember, and form connections between different words, concepts, and things.

Natural Language Processing (NLP)

  • NLP is a subfield of AI that allows machines to understand and interpret human language. In marketing, NLP can be used to analyze customer feedback, social media comments, and customer service interactions to understand sentiment, identify common issues, and improve customer experience.

應用層、 技術底層 (Foundation Model )

References

image

AI 應用領域

AI for Marketing Automation

High level intro:

  • AI marketing automation is the use of artificial intelligence and machine learning algorithms to automate and optimize marketing activities.
  • More efficient, more personalized, and ultimately more effective.
  • analyze vast amounts of data to identify patterns, predict customer behavior, and make decisions in real-time
  • result in higher engagement rates, increased customer satisfaction, and ultimately, better business results

Application Scenarios that AI can help MA

image

1. Personalized Marketing Campaigns)

  1. E.g. Product Recommendation
  2. Benefits
    1. higher engagement, retention rates, customer loyalty
    2. AI can help marketers achieve this level of personalization at scale by analyzing customer data and making tailored recommendations

2. Automated User Segmentation

(incl. scoring, nurturing)

對潛在客戶評分、培育潛在客戶

  1. Can do these
    1. identify which prospects are most likely to become customers.
    2. guide the leads through the sales funnel and convert into a purchase
  2. Benefits:
    1. help businesses save time and resources while improving the quality of their leads
    2. Automated nurturing result in higher conversion rates 還可以帶來更高的轉化率，as leads are effectively guided through the sales funnel until they are ready to make a purchase.

3. Predictive Analytics

(for customer behavior and trends)

  1. Do these
    1. By analyzing history customer data and behavior, help marketers anticipate customer behavior, needs, and trends before they happen.
      1. e.g. businesses can use predictive analytics to determine the best time to launch a new product or service based on historical data.
      2. e.g. MA tool to predict and suggest when is the best timing to send out marketing messages to consumers
    2. product recommendations to its customers based on their purchase history and browsing behavior.
  2. Benefits
    1. higher customer satisfaction
    2. increased sales.
    3. And help marketers anticipate the trends and adjust their strategies accordingly.

4. Optimization of Marketing Campaigns

(incl. channels, ad placements)

  1. Do these
    1. By analyzing data on the performance of different marketing channels, AI algorithms can identify which channels are most effective and can allocate resources accordingly.
      1. E.g. Meta Robyn
        1. E.g. businesses can use AI to optimize their ad spend by targeting the right audiences and identifying the most effective ad placements.
  2. Benefits
    1. This can lead to more cost-efficient marketing campaigns and better business results.

5. Conversational User Engagement

(Chatbots)

  1. Do these
    1. automate the repetitive tasks
      1. answer FAQs
    2. Provide product recommendations
  2. Benefits
    1. More cost efficient . Help marketers save time and money
    2. reduce customer service agents’ work load.
    3. Increase customer satisfaction

6. Summarize Complex Information

Overall Benefits that AI could help with Digital Marketing

  1. Personalization
  2. Predictive analytics
  3. Cost efficiency
  4. Optimization of ROI:
    1. According to a study by Accenture, companies that implement AI in marketing see an average increase in return on investment (ROI) of up to 30%.
    2. Because AI can help marketers identify new opportunities for growth, optimize ad spend, and improve customer engagement.
  5. Higher engagement rate, retention rate, loyalty
  6. Higher conversion rate
  7. Higher sales
  8. by leveraging AI and automating routine tasks, marketers will be able to focus more on strategic planning, creative thinking, and driving innovation.

AIGC 應用領域

  1. 翻譯語言
  2. 修改文字的語法、通順程度、找出錯字、加上優美的形容詞
  3. 為一篇文章做重點摘要、產出會議紀錄
  4. 輸入指定的資訊，然後輸出制式的文字
    1. 客服回信
  5. 撰寫文章草稿
    1. 並且同時產出不同風格版本

AIGC 應用: Customer Service

可以解決的問題

  1. 適合回答：重複性問題、常見問題
    1. E.g. 處理客戶查詢簡單的問題 (e.g. 訂單狀態)
    2. 很多無 AI 的 chatbot 已經在做類似的事情，有時難以做到 fully conversation-based
  2. 可幫助品牌深入了解：
    1. 消費者的意圖
    2. 消費者的情緒 (Sentiment analysis)
      1. ChatGPT can be trained to understand and classify the sentiment of customer inquiries. By identifying customers' emotions, brands can be more equipped to respond in an appropriate manner.
  3. Task Classification
    1. 自動將 customer inquiry 打標做分類 (e.g. 配送進度 / 退款退貨 …etc)
  4. Provide “personalized support” based on:
    1. Customer’s input in the inquiries
    2. Customer’s purchase history
  5. 預測消費者的下一步行為
    1. 消費者還需要什麼別的 support

具體作法

  1. 讓 AI 自動回答：重複性問題、常見問題
  2. 真人客服 搭配 AI
    1. E.g. Text Generation. 預先產生「建議的回答」給 CS Agent ，然後人工審核、做修改，再發給 消費者。確保消費者得到具有正確資訊的回覆
  3. Integration with external systems
    1. (這將使人工智能超越回答常見問題，真正開始對客戶的賬戶和訂單進行更改。通過整合，人工智能將能夠解決更多的客戶支持問題。)
    2. 範例
      1. Order management system of e-commerce
      2. External CRM
  4. 但要提防這些事：
    1. 如果消費者知道他們正在與 AI 交互 (而非真人克服)，也許消費者很可能會試圖 “欺騙 / 玩弄” 模型、導致模型以不恰當的方式做出響應

AIGC

AIGC產品的特性

  1. 從經驗中挖掘內容。
    1. 它能產出的 output 完全取決於 被給予了什麼 input 資料
  2. 因為無法自己創造「某些 input 資料」，導致：
    1. 不會自己「產生需求、明白要解決什麼問題」，所以無法自己「決定要達成什麼目標」
    2. 無法主動決定「現在該做什麼事情」
  3. 無法獨立思考，無法判斷事情的真假、對錯、做決定

Midjourney 的成功

  1. Stable Diffusion背后的明星公司——Stability AI，目前正面临严重的财政困境，由于没有明确的盈利途径，公司正面临倒闭的危机。相较之下，Midjourney却运行得风生水起，凭借着付费订阅的商业模式，Midjourney不仅获得了每年1 亿美元的营收，并且在Discord上已经积累了1000多万用户。
  2. 他为Midjourney设立了一个非常不“铜臭”的宗旨：AI 不是现实世界的复刻，而是人类想象力的延伸。
  3. 公司团队成员仅11人，其中1位创始人、8位研发人员、1位法务、1位财务。在公司的构成中，完全没有产品经理、市场销售人员，除了创始人、两个支持性岗位（法务、财务），80%的人员都是研发人员。
  4. Midjourney的盈利模式看上去十分简单，即通过付费订阅的商业模式，按月向用户收取费用，其标准有3种套餐，分别是10/30/60美元/月。不过，这样的模式要想行得通，得解决两大关键问题：1.凭什么让用户产生付费的意愿？2.大模型训练所需要的高昂成本怎么解决？
  5. 开源社区会齐心协力地完善模型文档，共同解决技术难题。这使得代码的迭代速度非常快，优化效率远远高于闭源系统。但缺点也很显而易见，那就是商业化不够直接，可能为别人“做了嫁衣”
  6. Midjourney却采用了不那么开放的“闭源系统”。如果说闭源系统真的有什么好处，那就是针对性更强了。因为模型闭源，并通过庞大的用户量积累了独有的数据集，可以根据用户需求不断地针对性训练模型，长期来看更有利于建立竞争壁垒。
  7. 在探索用户需求这点上，大卫采取了产品上线后边测试边改进的办法。例如Midjourney模型最开始很慢，需要20分钟才能出一张高质量的图片。后来团队推出了一个做15秒生产图片，但是质量没那么高的版本，经过多轮测试，团队了解到，速度和质量其实都只是表象，因为不同用户的选择，实际上是多维度的。在针对用户需求进行调整后，无论是创意行业设计者，还是普通爱好者，都能通过Midjourney满足自身的绘画需求。
  8. 虽为闭源，但Midjourney在使用难易度上，却更像一个“亲民”的大众产品。于是，Midjourney 获得大量用户后，养成了用户使用习惯，且在开启付费订阅后就进一步加强了用户粘性。
  9. 大卫打从创办Leap Motion的时候起就有一个观点，他觉得技术的最大限制不是规模、成本或速度，而是人们如何与之互动。
  10. 从成本来说，Midjourney大约10%的云成本用于训练，90%是用户制作图像的推理。所以几乎所有的成本都在制作图像上
  11. Midjourney在世界上八个不同的地区，设立了自己的服务器，比如韩国、日本或荷兰等，在每个时区的夜间，当地人都在睡觉，没有人使用GPU。Midjourney就可以充分利用这些算力，实现GPU负载平衡。实际上，这种依靠云端服务器来降低成本、加快模型训练的做法，与目前腾讯训练大模型的策略十分相似。在算力已经愈发成为大模型训练瓶颈的今天，如果在训练开发环节，直接调用云端的大模型和AI算力资源，完成后一键分发到用户终端上，就可以大大降低成本，减少工作量。
  12. 互联网的演进之路，已经说明，无论To B还是To C行业，都在追求越来越集约精简的终端硬件、越来越低门槛的交互入口、越来越轻盈的软件应用。所以说，大模型从云入端，是模型服务商实现商业化的必争之地
  13. 因为在生成式AI、云计算等技术逐渐抹平大企业与中小企业之间的技术、成本差距后，各企业真正比拼的，只剩下人才、创意与执行力。这样依靠少数尖端人才组建的团队，具有大企业所没有的灵活性、创见和魄力
  14. 这类小团队的创意、灵感，若要真正在市场、社会中扎下根，就离不开对用户多样化、个性化需求的追踪。这是因为，AIGC技术的“泛用性”，决定了其绝不是针对某一行业、人群，或是某一类企业的技术。只有在这多样化的需求中，尽可能地满足不同层级用户的特定需求，一款产品才能真正地具有长远的生命。既服务所有人，又不忽视每一个特殊的人，这或许就是Midjourney成功的最大原因

AI 未來趨勢

Wave 2 of AI Revolution

image
image
image
image

  1. To date, generative AI applications have overwhelmingly focused on the divergence of information. That is, they create new content based on a set of instructions.
  2. In Wave 2, we believe we will see more applications of AI to converge information.
    1. That is, they will show us less content by synthesizing the information available.
    2. Aptly, we refer to Wave 2 as synthesis AI (“SynthAI”) to contrast with Wave 1.
    3. While Wave 1 has created some value at the application layer, we believe Wave 2 will bring a step function change.
  3. When it comes to B2B applications, the objectives are different. Primarily, there is a cost-benefit assessment around time and quality.
    1. You either want to be able to generate better quality with the same amount of time, or generate the same quality but faster.
    2. This is where the initial translation from B2C to B2B has broken down.
  4. As we move into the next wave of generative AI applications, we expect to see a shift in focus from the generation of information to the synthesis of information.
    1. In knowledge work, there is huge value in decision-making.
    2. Employees are paid to make decisions based on imperfect information, and not necessarily the quantity of content generated to execute or explain these decisions.
    3. In many cases, longer is not better, it’s just longer.
  5. How can AI improve human decision-making?
    1. We believe LLMs will need to focus on synthesis and analysis — SynthAI — that improves the quality and/or speed of decision-making (remember our B2B diagram above), if not make the actual decision itself.
    2. The most obvious application here is to summarize high volumes of information that humans could never digest themselves directly.
  6. The real value of SynthAI in the future will be in helping humans make better decisions, faster.
    1. We are envisioning almost the opposite of the ChatGPT user interface: Instead of writing long-form responses based on a concise prompt, what if we could reverse engineer from massive amounts of data the concise prompt that summarizes it?

用戶 透過 LLM 叫 AI 模型寫程式

In Fixie, each Agent is a standalone service that combines an LLM with a little bit of code — which can be implemented in any programming language — that understands how to connect to an external system, like a database or an API

Future of AI by OpenAI CEO Sam Altman

(Some key points from the article)

  1. In the next decade is that the marginal cost of intelligence and the marginal cost of energy are going to trend rapidly towards zero. That’s going to touch almost everything because these seismic shifts that happen when the whole cost structure of society changes, which happened many times before. And that will quicken the iteration cycles of many things.
  2. With the AI tools emerging more and more in the coming future, we human beings will still really care about interaction with other people. “The stuff that people cared about 50,000 years ago is more likely to be the stuff that people care about 100 years from now than 100 years ago.”
  3. There could be a number of verticalized AI startups that essentially have adapted a fine-tuned foundation model to industry. We will not be doing prompt engineering in five years time. And the AI-added tools be integrated everywhere. And SA thinks the fundamental user interface will be the natural language (may be as what we see in the ChatGPT today)
  4. The AI startups do not start from scratch such as building models from zero. They will take base models that are hugely trained with a gigantic amount of compute and data (maybe by incumbent giant companies), and then they will train on top of those to create the model for each vertical. But they’re doing the 1% of training that really matters for whatever the use case (for their product) is going to be.

No-code & Composability

https://chiefmartec.com/2023/03/holy-two-major-martech-disruptions-just-collided-together-ai-and-composability/

Chat Interface / No-code tools

  1. For years, software users had to adapt how we work to the way the software we purchased was built. Now, natural language interfaces like this start to break us free from that imposed structure. Software begins adapting to our asks, instead of us adapting our asks to it.
  2. Some companies provide a layer of intent translation between people and software. Instead of navigating an app with a series of mouse clicks, you simply ask the app to do what you want. It’s a faster, more direct path to getting things done.
  3. Empower more people to create more software with less or no coding
  4. A business user doesn’t have to be bound by a prescribed workflow or user experience invented by a far away product team at a completely different company. They can invent their own.

Composability (可組合性)

  1. Different software components “can be selected and assembled in various combinations to satisfy specific user requirements
  2. Composability dramatically multiplies the number of apps in the world. It unleashes waves of combinational innovation, where different components can be mixed and matched in effectively an infinite number of niche contexts

Examples

  1. (ChatGPT’s plugin from Zapier) It’s not hard to imagine that extending across Zapier’s entire portfolio of integrations, enabling more and more sophisticated workflows to be executed, merely by telling ChatGPT what you want to happen. Those workflows could span many different apps in your stack, leveraging data and functionality across any of them. But you won’t need to know or care about those app boundaries. All controlled via natural language.
  2. ChatSpot combines multiple different, independent software apps — ChatGPT, HubSpot CRM, DALL-E, and Google Docs — and orchestrates them together. It’s a testament to the phenomenal cosmic power accessible today through API calls in the cloud.

image
image

2nd-order Effect of Generative AI

In the article you shared, the term "buyer" refers to the people or organizations that purchase products or services from companies that use generative AI in their marketing and martech¹. The article discusses how generative AI can help companies create more personalized and effective marketing campaigns that can better target potential buyers¹.

My takeaways:

  1. Generative AI empowers sellers to generative marketing and sales content faster and less costly. Even though it’s personalized, the huge amount of automated comms/content may make customers choose to ignore the marketing and sales content.
  2. Trusted sources (media or individuals) and data will be more valuable than ever.
  3. Customers might find new ways to discover and evaluate the content and products that they exactly need. They might use AI agents to do this (those could be the same/similar tools that the sellers generate content to attract/lure the buyers?) And a new kind of marketing objective could be to serve these “buyer-centric AI agents”.

image
image

待整理的筆記 from Lydia

  1. Buyers will further tune out “pushed” marketing and sales content, even though it’s hyper-personalized. … as a result, all these über-personalized messages won’t signifcantly improve sales prospecting efficacy. They’ll likely reduce it.
    1. 除了精準和個人化，下個階段對用戶來說，促使消費的好的體驗為何？
      1. Bot-to-bot commerce — the new B2B? — will become a huge portion of e-commerce.
      2. Helping people answer questions to win new business. We’re now on the verge of what we could call a generative inbound marketing revolution — fielding inbound requests from AI agents to help them accomplish the goals they’ve been assigned by their human controllers.
    2. 從廣告SEO 到 social media，下個行銷和銷售的戰場在哪？
      1. this hunger for proprietary data will be opportunities for marketers to market by feeding datasets across their go-to-market ecosystems. Data will become a first-class marketing channel.
      2. API services will become a first-class marketing channel too. This is how we’ll serve the right data to AI agents at the right time. It’s also how we’ll give them easy ways to take actions with us.
  2. From no code to AI-generated
    1. Composable architectures and a universal data layer will play a big role in making this feasible.
    2. Some idea
    3. integration with no-code -> Human describe or choose, and AI complete integrate

李開復的看法

image
image

做 AI Application 的小公司該如何維持競爭力

Summary

  1. 承認自己的強項並非燒錢戰
    1. 燒了一堆錢做出一個 me-too product，不應該是台灣走的路
  2. 使用這些強大的 AI Basemodel (built by 大公司) 已經不是選擇而是必須，否則就會落後於競爭對手。
    1. AI 模型的開發從一開始就是基於前人的實驗來進行優化。所以不要因為使用別人的 APIs 而感到不安，這是產業演化和專業分工的必然結果
  3. 當 OpenAI 和 Google 的模型推出時，AI 小公司必須針對自己主打的情境立刻使用和測試，並將其納入常規產品開發流程
    1. 因為一旦什麼都做，就是直接與這些生成模型競爭。這些模型已經具有很強的通用性，因此通用型絕非小公司的發展方向 (View Highlight)
  4. 範例：判斷對話中消費者是否想要購買蘋果，並實時顯示在消費者的螢幕上
    1. 需要一個實時的語音轉文字串流，再從對話中提取與購買水果相關的對話，然後將其交給 ChatGPT 來判斷顧客是否想要購買水果
    2. 在此過程中，模型對水果買賣要有非常強的理解，而 ChatGPT 的輸入有技術上限
    3. 因此，需要找出如何從對話中提取正確的文本，並將其轉換成 ChatGPT 能夠好好回答的格式。
    4. 最後，針對這個情境，設計並嘗試適合的 ChatGPT 提示問題，以便獲得相應的輸入/輸出。
LinkedIn