Introduction to ChatGPT

Provide examples if necessary...

Another way to superpower our prompts is to provide examples. For example, let's use ChatGPT to create a list of example customer names, ages, and occupations that we can use to test our internal systems. We want the results in a very specific format: the full name and age separated by a comma, followed by the occupation in brackets.

Opportunities for misuse

As AI-generated content becomes more human-like, there will be more opportunities for misuse, such as representing AI-generated content as human-generated or using AI to create malicious content, such as spam emails.

Limitation 3 - Context tracking

ChatGPT has the ability to build on information and context from earlier in the conversation, so follow-up corrections can be made. However, if the topic of the conversation shifts multiple times, ChatGPT can struggle to keep track of the context and generate inaccurate or irrelevant responses. A good rule of thumb is to keep a conversation to one topic and create new conversations for different topics.

Generating and debugging code

ChatGPT is able to generate template code, explain why code isn't working, and even make suggestions for improvements!

Why utilize ChatGPT?

ChatGPT is able to perform many time-consuming tasks with greater efficiency. This workflow of AI doing the legwork and a human providing the finishing touches saves a substantial amount of time and money. Implementing ChatGPT into products will also allow for greater personalization of content, delivering more value to customers.

Generative AI

ChatGPT is an example of a generative AI model. Generative AI is a subset of artificial intelligence and machine learning, where a model creates new content based on patterns in information it has already seen. In ChatGPT's case, this generated content is text, but other models exist for image, audio, and even video generation.

Summarizing text

ChatGPT is great at summarizing text or concepts for a particular audience, which is useful when summarizing reports for different stakeholders or interpreting complex information.

Limitation 2 - Training data bias

ChatGPT was trained on a massive text dataset from a variety of sources, including books, articles, and websites, but this data could contain biases. The model may learn these biases and produce biased responses.

Limitation 1 - Knowledge cutoff

ChatGPT was trained on data up to a certain date, and the model isn't connected to the internet or other external sources, so it isn't aware of events beyond this date.

What is ChatGPT?

ChatGPT, by OpenAI, is a conversational language model, which means it can answer questions or perform tasks that yield a text-based response.

Ownership and copyright

Copyright prevents people from using intellectual property without prior consent from the owner. If AI-generated content bears similarity to content protected by copyright, we're at risk of receiving an infringement claim from the owner of the IP.

Data ethics

Data ethics is a field dedicated to ensuring that data is used with people and society's best interests in mind. we should still ensure that enabling our use case won't adversely impact society and, desirably, bring an overall positive impact.

Data governance

Data governance laws govern how data and information can be collected, stored, and used to protect people's personal data. One of the biggest and most established data governance laws is GDPR, which governs data usage in EU countries.

Building balanced datasets

Due to the sheer quantity and unstructured nature of the training data, reducing bias before training the model is difficult. As generative AI begins to become an integral part of day-to-day life, more robust procedures to mitigate bias will need to be developed.

Validating a use case part 3

Enabling a use case involving sensitive data is difficult. We need to acquire the necessary consent to process the data, as well as ensuring that the applicable data governance laws, such as GDPR or CCPA, are being adhered to. For these use cases, it's best to seek legal counsel that specializes in data governance.

Validating a use case part 4

Finally, ask whether ownership over the response is required. Providing that users comply with OpenAI's terms of use, they can claim ownership over the output in many cases, but other considerations, such as copyright infringement 侵犯版权, may prevent ownership.

From prompt to response part 2

Finally, the generated text, or response, is returned to the user.

How does ChatGPT interpret a prompt?

First, ChatGPT identifies the broad topic of the prompt. In this case, the phrases job description, data scientist, and New York help the model identify that the prompt is about a job description for a specific role and location.

three key considerations for owning the response

First, we cannot claim ownership over responses that are considered non-unique, as they can also be generated for other users. Second, we can't claim that responses from ChatGPT were human-generated. Finally, ChatGPT can't be used to infringe on a person's rights, which includes copyright infringement.

Limitation 5 - Legal and ethical considerations

It's easy to fall into one of these legal gray areas if the use cases for ChatGPT aren't properly scoped so that ownership and legal implications are well-understood and accepted.

Creating marketing content

Let's ask ChatGPT to write a tweet encouraging people to acquire data literacy skills. ChatGPT is already being applied to streamline many different marketing tasks, including creating email templates, writing blog post titles and descriptions, and copyediting large bodies of text.

How does ChatGPT interpret a prompt? part 3

More context could be added to the prompt for greater personalization, such as key role-specific skills and company culture information.

How does ChatGPT interpret a prompt? part 2

Next, ChatGPT attempts to understand what the prompt is requesting. In this case, the verb, write, and job description indicate the task being given to ChatGPT. Data scientist and New York provide additional context to help populate the response with relevant information.

Validating a use case part 2

Next, we should evaluate whether someone is able to verify the quality of the response. A good rule of thumb is to not ask ChatGPT to do something that we couldn't do ourselves given enough time. If we can't verify the correctness or quality of the result, it would be irresponsible to begin using it to drive decisions or surfacing it to customers.

Ownership and privacy

Ownership and privacy are two key considerations when scoping ChatGPT's suitability for a particular use case. Neglecting to consider these factors, especially for companies, can bring with it financial penalties, lawsuits, and damage to customer trust and brand image.

Prompt engineering

Prompt engineering is the process of writing prompts to maximize the quality and relevance of the response.

Standard chatbot vs. ChatGPT

Standard chatbots are usually designed to return a predetermined response to a limited number of questions. ChatGPT is far more generalizable, as it uses its understanding of language to interpret the question or task and determine the most appropriate response.

What's driving the improvements?

The amount of available training data will continue to increase, so models should develop a deeper understanding of language, including complex language expressions, such as sarcasm and idioms. Additionally, many generative AI models collect usage data, including user ratings of the generated content. This data allows developers to continue to fine-tune the model while it is live.

Writing tips for prompt engineering

The first guideline is to be clear and specific. In the prompt, we should specify roughly how long we want the resulting summary to be: one page, one paragraph, or even one sentence. Second, keep prompts concise. This means removing any unnecessary information or linguistic frills that don't provide extra context for the task at hand - these extras will only dilute the important information and keywords. Finally, use correct grammar and spelling in prompts.

Validating a use case

The first question to ask ourselves is if we require a high degree of accuracy in the response? ChatGPT can be inaccurate, and there's no predictability in its response If the use case requires certainty in the quality of the response, such as government policy advisory, ChatGPT would be an unsafe option.

Performance improvements

We expect that generative AI models will produce content that much more closely resembles human-generated content. We also expect generative AI models to tackle more complex instructions and questions with greater reliability.

A ChatGPT-powered workflow

We need to summarize the key findings for other stakeholders, We can ask ChatGPT to summarize the document for us with a well-engineered prompt; then, we proofread the final review for accuracy before sending.

Limitation 4 - Hallucination

hallucination, which is when the model confidently tells us inaccurate information. This often occurs when attempting to go beyond ChatGPT's knowledge cutoff or abilities.

Summarizing text Example

t would be better if we could boil it down to two key sentences. Notice that we don't have to instruct it to summarize GDPR again; ChatGPT will actually remember the information and context from earlier in the conversation. The ability for ChatGPT to remember the conversation history and for the user to provide follow-up corrections to responses are two extremely powerful capabilities of the model.

Demystifying the LLM

the LLM was shown a huge amount of text data, which is like the large building block wall. From this data, it builds its understanding of the structure of language by looking at the frequency and order of words, which are like the differently-colored building blocks. the training data, and the sheer amount and variety of data used to train ChatGPT is a large part of its success. The model itself used complex algorithms to detect these language patterns in the training data, and it was fine-tuned through iterative processes that included rating the quality of the responses. So when we provide a prompt to ChatGPT, it is essentially trying to complete a building block wall, using its understanding of the training data to generate the words that are most likely to follow the prompt.

From prompt to response

the user writes a question or instruction, which is more generally called a prompt. This prompt is passed as an input to a large language model, or LLM. LLMs use complex algorithms to determine patterns and structure in language. These patterns are then used to interpret the prompt and generate new, relevant language in response to it.

Augmenting workflows 加强

we develop workflows to achieve some end goal through a standardized series of tasks. These workflows are often molded through years of experience, experimentation, and innovation to achieve a high-quality end goal in a timely manner.