Adopting ChatGPT
Enabling people to use ChatGPT
Augmenting workflows
Workflow: Standardized series of tasks to achieve an end goal
工作流:实现最终目标的标准化系列任务
Aims: 目标
-
Highest-quality output 最高质量的输出
-
Shortest timeframe 最短时间
A standard workflow 标准工作流
-
Example: Summarizing a project scoping document
-
30 pages -> summary of key findings
-
-
We extract the key findings and compile the summary
提取关键发现并编写摘要
-
Using a spelling and grammar checker to proofread
使用拼写和语法检查器进行校对
A ChatGPT-powered workflow
-
Example: Summarizing a project scoping document
-
30 pages -> summary of key findings
-
-
ChatGPT allows us to reverse the roles 角色互换
-
Human now becomes the proofreader 人工校对
-
Huge time-savings! 节省大量时间!
-
Allow us to focus on more creative tasks
Leaders
Use cases:
-
Compose emails
-
Draft presentations
-
Brainstorm strategic ideas
-
Summarize meeting notes
Technical roles
Use cases:
-
Recall code syntax
-
Generate examples
-
Explain code
-
Troubleshoot errors
-
Write documentation
HR and people teams
Use cases:
-
Brainstorm employee engagement and wellbeing initiatives
-
Communicate more effectively and efficiently
Marketing
Use cases:
-
Write social media posts
-
Copyediting content
-
Generating marketing copy
-
Search Engine Optimization(SEO)
Sales
Use cases:
-
Generate outreach templates
-
Personalize outreach content
-
Brainstorm strategies
-
Summarize information
Identifying use cases for ChatGPT
GPT
-
Can be inaccurate
-
No predictability in response
-
Subject matter expertise is still very important
-
Need consent to process the data
-
Must adhere to data governance laws, such as GDPR
-
Legal counsel may be able to enable the use case
-
Can claim ownership over ChatGPT output
-
Other considerations such as copyright infringement may prevent ownership.
-
-
Legal and ethics -> Coming up!
Rule-of-thumb 经验规则: Don't ask ChatGPT to do something that we
couldn't do ourselves
Example 1: Brainstorming ideas in HR
-
Improve employee wellbeing
-
Is this a suitable use?
-
Situation doesn't require definitive answers
-
HR Manager can verify response
-
Sensitive data isn't required
-
Response won't be used
Example 2: Healthccare recommendations
-
Customers input their symptoms and receive recommended action
-
Is this a suitable use?
-
Use case requires certainty due to implications of poor recommendations
-
ChatGPT cannot provide this level of certainty
-
Ownership and privacy
Ownership and privacy
-
Ownership and privacy are key considerations when validating the suitability of ChatGPT
-
Neglecting them can risk financial penalties, lawsuits, and brand damage
Who owns the response?
-
Assuming compliance, users can claim ownership over the response
Ownership and copyright
-
Copyright: the rights of the owner of the intellectual property(IP) to use or distribute the material
-
If generated content resembles copyrighted property, infrigement claims can be made
Who owns the prompt?
As between the parties and to the extent permitted by applicable law, you own all input
Prompt privact
-
ChatGPT is being continuously developed and improved
-
OpenAI may use prompts and responses for performance improvements
-
May need to opt-out of usage agreement
-
Risk of breaching data governance laws
Data governance
-
Govern how data can be collected, stored, and used
-
Example: GDPR governs data usage impacting EU citizens and residents
-
Use cases must adhere to data governance laws
AI ethics
-
Ensure data is used with people and society's best interest in-mind
-
Ask whether the use will negatively or positively impact people
Advancements in generative AI
Performance improvements
-
More human-like content
-
Handle more complexity
-
Greater reliability
What's driving the improvements?
Large Language Models(LLMs)
-
Learns from a huge text dataset
-
Algorithms detect patterns in text
-
Fine-tune the model by rating responses
What's driving the improvements?
-
Amount of training data will increase
-
Usage data will help in fine-tuning
Building balanced datasets
Challenge: Ensuring data is high quality and balanced
-
Quantity of data makes detecting bias prior to training difficult
Goal: Develop more robust bias mitigation procedures
Opportunities for misuse
-
Misrepresenting AI-generated content
-
Creating malicious content(e.g., spam)
Intervention by lawmakers:
-
Regulations could help or hinder AI advancement
From generalized to specialized
-
ChatGPT is a generalizable model
-
Generative AI models will become more specialized
-
Example: a model specifically designed to write long and complex code