Generative AI Fundamentals: Capabilities, Applications, and First Experiment

Reading objectives and introduction

This page provides a English translation. After reading, you will be able to explain the differences between generative AI and common analytical and predictive AI, identify text, image, audio, and multimodal applications, recognize the tools listed in the original text, determine whether a given task is suitable for use, and begin small-scale trials.

A key shift in generative AI is its ability to create content, rather than merely analyzing existing data. Starting from this distinction, we will then explore its applications, limitations, selection methods, and prompt engineering techniques. Following the main text, four corresponding Chinese video script segments are retained.

What is generative AI?

From assessing existing content to generating new content

Generative systems can write articles, reports, emails, and marketing copy; generate images and code based on descriptions; synthesize speech and music; engage in conversations; and during translation, attempt to preserve context and tone.

The original material uses a market team as an example: writing introductions for 500 new products, which originally might take several weeks, can now be quickly generated with generative tools producing a large number of initial drafts. This example illustrates the potential of initial draft generation, but it does not mean that the information for each product is accurate, nor that the review process has been eliminated.

Ability does not equal understanding or reliability.

The original material outlines four limitations: the system does not understand content in the way humans do; it may generate "hallucinations" that sound plausible but are incorrect; it may reproduce biases present in its training data; complex reasoning and ethical judgments still require human review.

The technical comparison here is between "discriminative" and "generative": the former is typically used for classification or prediction, while the latter learns the distribution or patterns in data and generates new samples. The attention mechanism in Transformers helps the system handle relationships and context within sequences.

Note: Artificial intelligence is not equivalent to discriminative models, and generative methods are not newly introduced. The table is used to differentiate common tasks and should not be considered a complete technical history. Transformer is an architecture; generative models also include other architectures; the ability to "produce text resembling human writing" does not prove the system possesses human consciousness, nor can the name of an architecture alone determine performance across all reasoning tasks.

II. What kinds of content is it suitable for creating?

text

The original material cites a 70%–80% reduction in writing time as a case study and provides an anonymous example of financial reports shrinking from days to seconds. These cases do not offer verifiable research or measurement methodologies and therefore cannot be considered general claims of effectiveness. Actual work still requires time for data preparation, verification, revision, and approval.

Image

Common approaches include text-to-image generation, image editing based on instructions, style transfer, and rapid generation of multiple visual concepts. Retail teams can experiment by placing products in different scenarios, reducing the need for some product shoots and concept development costs; however, product appearance, text, branding, and usage permissions still need to be verified.

Sound and multimodal

Text-to-speech converts articles into audible content; music generation creates music in various styles; voice cloning attempts to replicate specific vocal characteristics. Multimodal applications combine capabilities such as text, images, and sound, for example, producing illustrations, narration, and voiceovers simultaneously.

The original material takes educational content production as an example: multimodal tools may speed up production and enhance presentation methods. Whether they can better support learning needs further validation; judgment based solely on production speed is insufficient.

Three. Tools listed in the original materials

The following is a tool map retaining the original content, designed to help understand product categories. It does not reflect current product rankings, and does not represent the features, fees, or availability of these services as permanent or unchanged.

Regarding voice, the original text lists Amazon Polly, Google WaveNet, ElevenLabs, as well as AIVA, Amper Music, and Soundraw among music services, reflecting the ecosystem at the time of content compilation. Do not infer that any of these products still operate in the same manner today.

Many platforms offer graphical interfaces, free trials, or some free tier, as well as APIs for system integration. However, "being able to try" is different from "long-term free access" or "uploaded data not being used for other purposes." Always check the current terms and account settings before actually getting started.

Four. How to determine if a task is appropriate

Generative AI is better suited for repetitive but varied or creative tasks, drafts that will still be edited by humans, large-scale personalization, and a high volume of content with similar structures.

For tasks requiring absolute factual accuracy without verification, sensitive communication demanding fine attention and empathy, deep ethical judgment, and important decisions made without human oversight, special caution should be exercised.

Evaluation can begin with four questions:

• Time value : How much time is currently spent on manual work? What is the value of this time?
• Error risk : What consequences would arise if the content is incorrect?
• Creative request Does the task benefit from change, personalization, or creativity?
• Number of entries How many similar pieces of content are needed to be produced?

The original material presents a case of creating numerous quarterly investment summaries for a large number of clients with similar structures: analysts require significant time investment, content needs personalization, and risks can be controlled through review, making it suitable to first generate a draft. The figure "saving 80% time" in the material has no independent source and should be treated as an anonymous teaching example, not as a directly reusable prediction.

Five. How to Start the First Time

Put the requirements into the prompt

Specific tasks, target audience, tone, context, and output format can help the system align the generated result with the intended goal. They cannot guarantee factual accuracy, but they can reduce deviations caused by ambiguous requirements.

Start with low-risk tasks, save effective prompts, verify facts, tone, and appropriateness before use, and establish simple usage boundaries for the team.

a workflow

First, confirm the key issues the quarterly report needs to address, then use the tool to draft the structure and content; have individuals with relevant expertise verify facts, correct errors, and supplement analysis; if needed, refine individual paragraphs with the tool; finally, have a human approve and issue the report.

This approach combines generative capabilities with human expertise. After becoming familiar with the tools, gradually expanding to more complex tasks is typically easier for identifying issues than starting with full automation.

Video transcript one: What is generative AI?

Let's first establish a fundamental distinction: what's the difference between traditional AI systems and generative AI? This concept helps explain why generative capabilities expand the use of artificial intelligence in workplace applications.

Many common traditional systems employ a discriminative approach, analyzing existing data to perform classification or prediction. For example, examining an email to determine whether it is spam or normal communication; analyzing historical patterns to predict customer behavior; or identifying objects in images based on learned features.

Generative systems create new content. They don't just classify emails; they can write a complete email based on given conditions; they don't just predict customer churn; they can draft retention strategies; they don't just identify images; they can generate new images. Moving from analysis to creation, it transforms how businesses and creative tasks can use AI.

Today's generative systems can produce coherent, stylistically consistent articles, reports, and marketing copy; generate realistic images based on descriptions, aiding visual prototyping and content creation; and draft code based on natural language specifications, helping to shorten certain development processes.

In terms of sound, applications range from speech synthesis to music composition. Translation attempts to preserve meaning, context, tone, and cultural nuances. With capabilities spanning multiple media, it is not merely an isolated tool but can serve as a platform for creating diverse types of content.

However, these limitations must be understood together. Such systems perform complex pattern recognition and statistical computations based on training data and do not possess the same understanding or consciousness as humans. They may generate content that appears authoritative and logically sound, yet contains factual errors, commonly referred to as "hallucinations."

They may also reproduce biases present in the training data and fail on complex logical or ethical judgments. Therefore, the most effective implementation requires an awareness of both capabilities and limitations, combining generative capabilities with human oversight and domain expertise, rather than simply relying on its ability to provide quick answers.

Note: The use of "traditional" versus "generative" here to distinguish common tasks does not constitute a complete technical classification; discriminative and generative methods have coexisted for a long time. Specific reasoning capabilities must be validated individually according to the model and task, and cannot be universally inferred from the above generalizations.

Video transcript two: Which applications are more effective?

First, let's look at text. It is one of the early types of generative AI applications that have widely entered business operations, originally divided into four categories.

First is content creation, including blogs, reports, emails, and marketing copy. The original explanation states a 70%–80% reduction in writing time, indicating potential, but emphasizes speeding up initial drafts rather than fully replacing writers; this figure does not provide a verifiable source.

The second is document summarization. When dealing with long reports, tools can quickly extract key points, aiding research and information processing. However, whether the summary omits important content still needs to be verified by referring to the original document.

Third is customer service chat. Modern systems can handle certain multi-turn and common queries, allowing service personnel to focus on more complex cases. The ability to respond does not mean the issue has been correctly resolved; escalation and verification mechanisms are still required.

Fourth is code generation. By describing software requirements in natural language, program drafts can be obtained to assist in development and lower the barrier for non-programmers to create prototypes.

Image generation has also seen practical applications. Text-to-image generation transforms detailed descriptions into visual content; image editing allows users to modify existing images using instructions; style transfer helps maintain a visual direction; and concept development rapidly generates multiple variations.

Imagine a retail team needing product visuals. Each product, version, and scenario must be reshooted, resulting in high costs and long turnaround times; AI can first generate preliminary concepts, reducing production bottlenecks and enabling the marketing team to quickly test ideas, compare visuals, and iterate on designs. Actual product appearance, branding, and rights still need to be verified.

Sound is another direction of expansion, enabling text to be converted into more natural speech, music creation, or the reproduction of specific sound characteristics. Further multimodal workflows combine text, images, and sound. For example, an educational content company could first produce explanations, then add illustrations and voiceovers. The original concept suggests this approach might significantly shorten production cycles and enhance expressiveness, but no verified data on time or learning effectiveness has been provided.

These examples share a common pattern: repetitive work that requires creativity or personalization; a large volume of similar content; outputs that are well-suited for human editing, rather than requiring absolute perfection on the first attempt. When assessing suitability, return to the characteristics of the task.

Video transcript three: When it's not advisable to rely on it directly

First, it's important to distinguish that confidence does not equal correctness. Generative AI might deliver answers in a calm, assertive tone, stating that the Eiffel Tower was built in 1999 or that sharks are mammals; when wrong, it may not explicitly indicate uncertainty.

This seemingly plausible yet factually incorrect output is called hallucination. Do not simply interpret it as the system intentionally deceiving like a human; it may arise from data gaps, training errors, and reliance on statistical patterns when generating content beyond reliable bounds.

When is it inappropriate to use? First, in tasks requiring high factual accuracy without human verification, such as diagnosis, legal judgment, and significant financial calculations, relevant professional expertise is required. Second, in highly sensitive situations that demand nuanced expression and empathy, such as layoffs or crisis communication, human judgment is essential.

Third, deep ethical judgments and high-consequence decisions should not be entrusted to systems without oversight. Systems can propose options, but the final decision on significant impacts must be clearly assigned to humans.

Before use, ask yourself: what if the output is wrong? If the output could cause harm, financial loss, damage to relationships, or legal liability, manual verification and oversight must be arranged.

A more cautious approach is to have the system work faster without replacing judgment. Let it draft, generate ideas, and handle routine tasks; always verify key content, especially material intended for external release or impacting significant decisions.

Video transcript four: Start experimenting with specific prompts

Why does generative AI sometimes disappoint and at other times prove helpful? A key factor is how the request is formulated, i.e., the prompt. Clearly specifying the goal and context leads to more useful results, but it is not the sole condition for success.

Suggest five marketing campaign ideas for a new organic skincare line targeted at consumers aged 25–40 who value environmental sustainability, emphasizing sustainability and natural ingredients.

Now look at the delayed email. Instead of simply saying "write a project delay email," say: "write a professional email explaining that the website redesign has been delayed by two weeks due to technical issues with the new payment system, including an updated timeline and a plan to make up for the lost progress." Clearer context and requirements lead to output that more accurately meets the actual need.

How to get started? Begin with low-risk tasks that save time, such as drafting emails, brainstorming, or creating outlines. Have someone with relevant expertise review the results, verify facts, and provide insights.

Second, document effective practices. Create a prompt library for frequently performed tasks, save useful prompts, and keep them for future reuse and comparison.

Third, before external use, check accuracy, tone, and appropriateness. Tools are more valuable when enhancing professional competence; they should not be treated as a substitute for professional judgment.

Start with small tasks, verify them, and then gradually try more complex applications. Now, select a task you frequently do, write a specific prompt, and see how the results differ from your expectations.

Related Reading

After understanding the capability, continue the comparison Model and deployment methods of generative AI , then use Applicability Practice Verify your task.

Reference and Extended Reading

The following is supplementary information provided by the editor for querying principles and practical usage methods, and does not serve as proof of anonymous case return values.

• Google: Overview of Generative AI : Chinese, query generation, application, and model categories.
• Microsoft: Getting Started with Generative AI Chinese, learning generative tasks, models, and responsible usage.
• NIST AI Risk Management Framework : English, for further reliability and risk management checks.