Touch the AI: integration of AI to the business environment?

Original name

Šáhni si na AI: integrace umělé inteligence do firemního prostředí?

Author(s)

Martin Krček

Length

51:56

Date

13-11-2023

Language

Czech 🇨🇿

Rating

⭐⭐⭐☆☆

  • ✅ Introduced a new perspective of AI capabilities and use-cases, especially from the business point of view.

  • ⛔ I dont want to buy ChatGPT app fo 25 USD.

  • ⛔ Buy Teams for 6 USD a month for automatizing MS Teams (summary of meetings, etc..).

  • ⛔ The woldwide economical and ecological impact was not mentioned.

"The goal of OpenAI is a friendly AI to make the word better and support creativity: "Give me a sad story" - the story will be sad but ends nicely."

""But I think in the future, perhaps the real pictures will signed by blockchain or anything else proving that it was created by human.""

Conversation tree

Technologies for conversations:

  • IBM Watson

  • MS Composer

  • Google Dialogflow Can be conected to Slack or Messenger

Benefits

  • Speed

  • Clarity

  • Secure answers

Challenges

  • Flow diagram

  • Unknown intents

  • Huge scope (conversation design) + NLU + time to develop

  • Localisation (including broken language)

What business wants?

  • More dialogs

  • More smalltalk

  • More intents, trainings, entities

Generative AI

Mantra: understand, read, write, count, draw, now can see

Brief history

  • 1966 Eliza

  • Late 1980/1990s Statistical Language Models

  • 2000s Neural Probabilistic Language Model

  • 2017 Transformer Models and Attention Mechanisms

  • 2018 BERT

  • 2019 GPT-2 and T5

  • 2020 GPT-3

  • 2021-2022 LaMBDA, xlarge, Chinchilla, CodeGen, InCoder, mGPT, PaLM, OPT-IML, Minerva

  • Nov 2022 ChatGPT

  • Dec 2022 GPT 3.5

  • Feb 2023 Google Bard and LLaMa

  • Mar 2023 GPT-4

  • Apr 2023 BloombergGPT, StableLM, Dolly 2.0, Titan, BingChat

  • May 2023 PaLM2

Cosine similarity - things are converted to vectors and compared (for similarity)

OpenAI

  • The goal of OpenAI is a friendly AI to make the word better and support creativity.

  • "Give me a sad story" - the story will be sad but ends nicely.

  • Training data: Huge language models created by people: Common Crawl, Wikipedia, WebText2, Book1 and Book2.

  • GPT = Generative Pre-trained transformer.

  • Limited knowledge of internet to June 2021.

  • No longer open (no sources available anymore).

Gen AI tools

Top 5 activities for developers

  • Write

    • blog article about AI generation

    • self yearly evaluation

  • summary about the conference

  • Conversations

  • Summary

  • Brainstorming

    • I want to increase my salary ideas

  • Write emails

  • Programming

    • Translate from JS to Java or C#

    • Describe what the code does

    • Code interpreter: Generates Python code to read attached Excel and analyses it (describes and draws data)

  • See: Describe a photo

  • Integration: Find my insurance contract, find my scheduled meeting from Gmail

How to implement into business

  • Can no longer give information to ChatGPT

    • Personal ID (email, phone, photo…​)

    • IP address, Browser, OS, Device

    • Prompt and content

    • All conversations including history

    • Trainings on my content

    • GDPR

    • Data storage and processing in the EU

Consumers and enterprises have different needs:

  • Consumers (Bard + MakerSuite):

    • Plan a 30day trip to Patagonia

    • Create a valentine poem

    • How to make GF pancakes?

    • Create a jazz song for a bday card

  • Enterprises (Vertex AI + Duet AI):

    • How do we control OUR data

    • How do we deal with fraud and security

    • How will we control costs

    • We need to be accurate and explainable

    • How do we integrate our existing data and applications

Gen AI tools + API:

  • Google : cloud.google.com/vertex-ai

  • Microsoft: oai.azure.com

  • OpenAI: platform.openai.com/playground (private training playground)

API:

  • Measuring temperature (exactness vs creativity)

  • The completionTokens, promptTokens and totalToken are returned to estimate the calculation costs.

Use cases:

  • Marketing

    • Content generator, product description (name, short, long)

    • Sentiment Analysis: From ChatBot, comments, social media appstores (traffic light + suggestion)

  • Search

    • Search on INTERNAL documents, sheets, PDF, images

    • Developer documentation on Confluence pages

    • Internal system - whisper in customer center

    • QnA

    • Preparation:

      • Prepare documents

      • What to do when we have 2+ pages

      • Embedding documents

      • Save embedding to vector DB

      • Local, Postgres, AirTable..

    • Answering:

      • Get question from user

      • Embedding the question

      • Search in Vector DB (using cosine similarity)

      • Get document

      • Create AI query to LLM (get answer from the document)

      • Send answer and link to document to the user

    • Opportunity:

      Manager: "No need embedded, Microsoft can select a directory/storage for cognitive search (Azure Cognitive Search) "

      Me: "Ok, it is a good way, it will cost 30k USD a month, but I can do cheaper"

  • ChatBot and Voice

    • HR ChatBot (how many vacation days do I still have)

    • Internal Support Bot

    • Public Support

    • Integrations:

      • Phone: AudioCodes, Avaya, Twilio, Voximplant

      • Dialogflow: Dialog CX, Phone Gateway, Dialogflow, Messenger, Messenger from Facebook, LINE

      • Open source: Google Chat, Azure Bot Service, Microsoft Teams, Discord, ServiceNow, Slack, Spark, Telegram, Twilio, Twitter, Viber

      • Independent integrations: Diagflow API

Gen AI

  • Benefits: Fast and easy to learn, local knowledge, always answers, localisation and data sources

  • Challenges: Testing (always different answer), speed, content security, can I trust?

Automation

  • jaRobot.cz: articles written and published by AI

    • sometimes too productive

    • context-sensitive (speaking in Czech triggers describing Prague for a trip ideas)

How to start

  • ASK people what they want (customer support, marketing, HR)

  • SHARE knowledge (AI coffee, virtual regular calls either local or global)

  • start WORKING (start to code, MVP, PoC, Hackaton and workshops)

  • Do not give up

QaA

  • Author rights: EU wants to accompany AI image with a text that the image is fake. But I think in the future, perhaps the real pictures will signed by blockchain or anything else proving that it was created by human.