Generative AI Summary

How generative AI works

  • Can and cannot do
  • Common use cases: writing, reading, chatting

Generative AI Projects

  • Lifecycle of the project
  • Technology options: Prompting, RAG, Fine-tuning

Implications on Business and Society

  • Potential for automation and augmentation
  • Responsible AI, concerns

注意:加快2x没问题🉑️(andrew ng的语速挺慢)

Generative AI For Everyone (non-tech, 手把手零基础小白ok, free if audit) 要求:零基础ok,入门级non-tech内容

  • Text image audio video generation for own work+business
  • E.g. ChatGPT, Google Bard, Microsoft Bing, Midjourney
  • AI world – supervised learning (labelling things), unsupervised learning, reinforcement learning
  • Repeatedly predict the next word

Generative LLM Applications as thought partner

  • Write – brainstorm, write press release, translate
  • Read – proofread, summarize an article, summarize call center conversations, reputation monitoring, sentiment analysis
  • Chat – specialized chatbot,

What can they do and cannot do?

  • Can a fresh college graduate to complete it with correct instructions?
  • Knowledge cutoffs
  • Making things up – hallucinations
  • Text count limit
  • Not really good with tables data
  • Bias in society

Tips for prompting

  • Be detailed and specific, provide context
  • Guide the model to think through the answer
  • Experiment and iterate
  • Don’t overthink the initial prompt

Generative AI Project (E.g. Food Ordering Chatbot)

  • Repeatedly find and fix mistakes to improve performance, lifecycle:
    • Scope project
    • Build/improve system
    • Internal evaluation
    • Deploy and monitor

Cheap Option 1 – Retrieval Augmented Generation (RAG) – Reasoning Engine, NOT source of information

  • Retrieve relevant text (e.g. upload PDF, based on website articles) from documents to answer questions

Cheap Option 2 – Fine-tuning

  • Provide datasets (e.g. in certain style or structure, speaking style)
  • Gain specific knowledge (e.g. medical, financial)
  • To get a small model is OK (cheaper, can run on phone)

Very Expensive – Pretraining an LLM  

Part 3: Business & Society

  • Writing assistant
  • Marketer brainstorm
  • Recruiter summarizing comments
  • Programmer initial Python draft

Replace jobs?  Automation or Augmentation (human to double check)

  • Doesn’t replace jobs, but automate/speed up tasks
  • Technical feasibility and business value (costly! Can save money?)
  • Not just the most iconic role, can help save time and cost in some tasks

Common roles

  • Software engineer
  • ML
  • Product Manager
  • Prompt engineer?
  • Impact on knowledge workers

Concerns about AI

  • Concern 1: Humanity worst biased impulses
  • Reinforcement learning from human feedback (RLHF)
    • Human can score the generated answer to reduce bias
  • Concern 2: Job Less
  • Concern 3: Harm – self-driving cars, stock crash, unjust cass in court

Artificial General Intelligence (AGI)

  • Can do anything that a human can

Responsible AI

Fycompa 18日

本来我都唔信新藥會帶嚟personality/behaviorial change,但原來真係會lol

最近嘅我異常喜悅、好純、唔燥底、d怒火細咗好多,對人loving、grateful同體諒咗好多。身邊嘅人都好似好受了,話我overall開心咗,多野講咗。。

orz呢d 係咪euphoric mood嘅感覺?姐係食k仔個種high 咗!!(呢個係真係寫咗出嚟嘅side effect)

omg,我係全日都high咗。。係好開心呀,但好累 =_=

做自媒體的心得

接近3年了,followers 7萬1千人,粉絲群人數3,712人。其實心得真係好多,不能盡錄。試下再次寫寫

  • rule number zero係要堅持✊唔好成為數據(有幾多like、幾多view)嘅奴隸。不停trial and error直到搵到最適合自己嘅發佈頻率、發佈時間、style等等
  • 個人IP人設好重要,沒錯,雖然「內容為王」。但更重要係,要比人知道你係邊個,先會有追看性,先有fan retention!
  • 個人IP要鮮明獨特
    你不單是一個內容農場content generator/另一個普通的內容平台,更需要讓followers喜歡你的這個人本身,發掘個人魅力(冇,都要吹到有!)。內容每一篇都要時不時提醒粉絲“你是誰”。這樣才會有fans retention保留住粉絲,讓粉絲對妳不離不棄(就算係廣告post都比個like)。
  • 有輸出,就需要輸入!增值自己
    人的經驗是有限的,例如我大學4年的經驗和感悟係得咁多,就係咁多。所以必須不斷update自己。現在流行什麼?大家喜歡什麼?要根據自己的fans profile來設計內容curate content,自己要不停學習。這是一個終身學習的過程。
  • 你的business model是什麼?
    Productize yourself,把自己變成產品。Earn with your mind, not with your time 靠腦賺錢,而不是時間賺錢。雖然好似好功利,但由day 1 開始,就需要想好長遠來說,你要如何變現$$$。在未來,什麼類型的品牌可以跟妳合作?你嘅audience base 要target有高spending power高消費嘅精英人群。
  • 成功密碼要繼續(多參考其他爆文),錯了ok但不要一錯再錯
    成功過的內容,要復盤為何為受歡迎?是否發布時間選對了?然後就要keep住個勢,例如個format,內容方向,標題和圖片等等。(內容垂直了一段時間後,就可以開始diversfy了)如果曾經fail過,就不要再踏入呢個禁區了。例如我一寫大學經驗就會掉粉,那就不要重複犯錯了。
  • 做influencer自媒體,要厚面皮。
    相信做startup都係,要厚面皮。由零到一係非常難,所以找到機會就要推銷自己。成為你identity身分的一部分。身分identity推動行為behavior,行為behavior推動結果end result。
  • 所以第一步,你嘅帳號名稱、定位都要諗清楚。要夠特別,一睇就講到你嘅identity!而且容易記得。(我覺得我個名係超級好嘅 XD)
  • ultimate aim除咗「為愛發電」地分享,都係要變現賺錢。好多途徑,例如直播帶貨、賣課程,賣產品,幫品牌出post。我就係專供出post。一篇稿費大概係10%粉絲量。
  • 每一個post都要預留一個「廣告位」(視頻或者首圖裡面),品牌brand先會覺得你有commercial value,可以打廣告落你度。
  • 但為咗粉絲要習慣你d 廣告post,聽到每一個post都做下一d推薦,咁樣先唔會反感unfollow。
  • 標題and 首圖 太重要!要有個consistent style,觀眾先會有印象。多d參考最近嘅爆文流行d乜,大陸trend真係轉得太快了。最好係有阿拉伯數目(量化)、同埋英文字係title,會更吸睛catchy。
  • 內容方面,唔好太淺(覺得你冇料到,唔需要follow),又唔好太深(打擊咗讀者自信,趕客),要6/10嘅深度,加少少英文字,就ok架啦。
  • credibility,除咗妳自己,都可以quote下d名人(Bill Gates被我無限使用)。
  • 內容長度,因為xhs讀者嘅閱讀時間太快,盡量要concise簡短。cut得嘅無謂字就cut,例如「可以」「需要」,就變「可」「需」。
  • Fans retention 都可以係微信/xhs開粉絲群,線下聚會,線上chat等等。但好累,懶人應該維持唔到。