It’s the holiday season so I would like to send a message of hope to everyone. GPT-4, Claude, and Bard were prompted to create a haiku about hope and world peace.
GPT-4 produced the best haiku because it’s more beautiful and poetic:
Dawn’s light breaks the night,
Seeds of hope in bloom, so bright,
Peace on earth takes flight.
Best wishes and good luck in 2024!
Reinforcement Learning from Human Feedback (RLHF) has been employed to train and fine-tune large language models to achieve better performance for artificial general intelligence (AGI). Here are the explanations of RLHF (1) one from ChatGPT at simple level, and (2) the other from Brian Roemmele (@BrianRoemmele) that is concise and descriptive.
1) From ChatGPT:
RLHF stands for "Reinforcement Learning from Human Feedback." It's a technique used in machine learning where a model, typically an AI, learns from feedback given by humans rather than solely relying on predefined datasets.
This method allows the AI to adapt to more complex, nuanced tasks that are difficult to encapsulate with traditional training data.
In RLHF AI initially learns from a standard dataset and then its performance is iteratively improved based on human feedbacks.
The feedback can come in various forms, such as corrections, rankings of different outputs, or direct instructions. The AI uses this feedback to adjust its algorithms and improve its responses or actions.
This approach is particularly useful in domains where defining explicit rules or providing exhaustive examples is challenging, such as natural language processing, complex decision-making tasks, or creative endeavors.
Environment: ChatGPT Plus, GPT-4.
Objective: Explain complex topics with summary and mind map.
ServiceNow has recently been named as a leader in Governance, Risk, and Compliance (GRC). GRC, one of 36 workflow automation products of ServiceNow, is backed by AI innovation throughout the risk management lifecycle. The product helps corporations to connect business, security, and IT, and to handle risk with confidence.
Using the combination of GPT-4 + VoxScript + AI Diagrams, I prompted GPT-4 to explain ServiceNow GRC from the perspective of the CEO Bill McDermott. The result described the product’s features with sophisticated business language.
Then GPT-4 was prompted to create a mind map visualizing ServiceNow GRC product. This mind map organizes the key aspects and provides excellent summary of the intelligent product.
On a different note, ServiceNow has a family of 36 AI-powered workflow automation products that can help your company on the digital transformation path. You can peruse the company website to determine which products are beneficial to your company's needs.
Environment: ChatGPT Plus, GPT-4.
Objective: Explain complex topics with summary and mind map.
Using the combination of GPT-4 + VoxScript + AI Diagrams, I've prompted GPT-4 to explain ServiceNow IT Service Management (ITSM). ITSM is a premier AI-integrated product of ServiceNow that transforms the impact, speed, and delivery of IT for corporations.
This time GPT-4 was prompted to explain from a sales executive's perspective. The result listed the key features and their benefits accordingly.
Environment: ChatGPT Plus, GPT-4.
Objective: Explain complex topics with summary and mind map.
GPT-4 has demonstrated it can analyze, visualize, and transform numerical data. Today’s post will look into how GPT-4 analyzes textual data.
First, we need to download 2 plugins: (1) VoxScript for web searching and data retrieval, and (2) AI Diagrams for creating mind maps. This combination of GPT-4 + VoxScript + AI Diagrams can explain complex topics into easy-to-understand summary and then create a visual mind map.
GPT-4 was prompted to explain ServiceNow’s IT Operations Management product (ITOM). ITOM, one of ServiceNow's premier AI-enabled products, has many capabilities that group key applications and capabilities into packages to keep digital services running 24/7.
Verdict: GPT-4 is remarkable as it makes complex topic easier to understand with its concise summary and mind map. However, keep in mind that GPT-4’s training data is currently up to April 2023 so the output might not be up-to-date. As a result, for topics that are time-sensitive, GPT-4 might not be the right tool to gather the information.
Environment: ChatGPT Plus, GPT-4
Data: Preppin Data, 2023, Week 1
Objective: Transform data without prompting GPT-4 step-by-step
This video from the Information Lab tested whether ChatGPT can transform data. However, the approach was not prompting GPT-4 to transform data step-by-step. Rather the approach was uploading two files (one input and one output) and asking GPT-4 to produce an output file similar to the uploaded output file.
Let’s test GPT-4 the same way and see how it performs.
1) Output 1: Upload 2 data files and prompt GPT-4 to produce output file without providing step-by-step instruction.
GPT-4 provided the steps how it transformed data and produced the output.
3) Output 3: Finally, upload another 2 data files and prompt GPT-4 to produce output file without providing step-by-step instruction.
In all 3 scenarios, GPT-4 transformed data, provided the steps, and produced outputs. Note that these data sets are considered simple. The results are posted here.
Verdict: For simple data sets, GPT-4 can transform data without step-by-step prompting. This is quite remarkable because GPT-4 is acting like human intelligence with cognitive ability to think through the process, transform the data, and provide the steps.
For complex data sets, GPT-4 cannot transform data without step-by-step prompting yet. However, Sam Altman the CEO of OpenAI has recently said that the company has discovered an emergent new cognitive capability. Perhaps with this new capability of critical thinking, the next version of ChatGPT will eventually be able to transform complex data sets.
Environment: ChatGPT Plus, GPT-4
Data: Preppin Data, 2023, Week 31
Objective: Use GPT-4 to transform data and fill in the missing IDs in file ee_dim_input.csv.
This data set was selected because it deals with HR data and HR data is almost alway complex. The requirement is to fill in the missing IDs. The data set has 2 csv files: (1) ee_dim_input.csv contains the list of employees and (2) ee_monthly_input.csv is a monthly snapshot of employees who worked during the month.
1) Upload Data: upload csv files.
The next 3 steps are for removing duplicated rows.
2) Remove Duplicates: remove duplicated rows with same employee_id in file ee_monthly_input.csv and use the output as a lookup table. The output file is named ee_monthly_input_cleaned.csv
3) Double-Check for Duplicates: double-check for duplicates for field ‘employee_id’ in file ee_monthly_input_cleaned.csv. Result confirmed that there’s no duplicates.
Double-check for duplicates for field ‘guid’ in file ee_monthly_input_cleaned.csv. Result confirmed there’s one duplicate.
4) Remove Duplicates: remove duplicated row. The output is named ee_monthly_input_cleaned_updated.csv file.
The next 2 steps are for filling in the missing values in fields ‘employee_id’ and ‘guid’.
5) Fill In Field ‘guid’: link 2 files together by field ‘employee_id’ and fill in missing values for field ‘guid’. The output is named ee_dim_input_updated_guid_linked_to_cleaned.csv file.
6) Fill in Field ‘employee_id’: link 2 files together by field ‘guid’ and fill in missing values for field ‘employee_id’. The output is named ee_dim_input_updated_employee_id_linked_to_cleaned.csv file. This file is the final result.
Verdict: It took me more than 1 hour to transform this data set. The challenge was knowing how to provide the proper prompts so GPT-4 would do what were needed. The final result met the requirement.
GPT-4 is remarkable as it can transform data just like Tableau Prep or Alteryx. The billion-dollar question is to figure out how to integrate GPT-4 within the corporate IT systems so that it can connect to databases, work with millions of records, and refresh data on schedule.
