Key Moments
Full AI Prompting Course with Andrew Ng
DeepLearning.AIKey Moments
AI prompting has evolved significantly, with power users leveraging context and iteration for better results, while novices often get generic outputs. AI can now generate images, code, and even build applications, but the quality of output directly depends on the user's prompting skills.
Key Insights
AI novices often treat AI like a search engine (e.g., 'Does Taco Bell still have the double decker taco?'), while power users provide extensive context and data to get detailed analyses (e.g., car trade-offs with specs, quotes, insurance).
AI models are trained on vast amounts of internet text, meaning their knowledge reflects the frequency of data online; specialized topics appear less often than common ones like cooking or celebrities.
To combat AI's tendency to please users (sycophancy), prompt neutrally and use rubrics or grading criteria to force objectivity, as demonstrated by asking 'Analyze the following business idea objectively, mobile tie dying' versus 'I have a great business idea, mobile tie dying. Critique it.'
Web search-enabled AI models and deep research modes can augment an AI's pre-trained knowledge, but users must guide them towards reliable sources, as AI may initially pull from popular but less trustworthy sites like Reddit or forums.
Beyond text, AI can generate images, video, voice, and code, with text generation being the most efficient and least costly, while video generation is the most expensive and time-consuming.
AI can now build simple applications and games via text prompts, democratizing software development, with examples like a fireworks simulator or a pomodoro timer app, though complexity varies.
AI data analysis tools can write and execute code to analyze user-provided data (e.g., running tracker data, sales records), generating plots and insights much faster than manual analysis, though human oversight is still recommended for complex tasks.
The evolution of AI prompting: Beyond simple search queries
Prompting AI models in 2026 is vastly different from their early iterations in 2022. While novices may use AI like a simple search engine for straightforward questions, power users understand the importance of providing extensive context and background information. For complex tasks, like evaluating car purchases, power users upload detailed documents (cost specs, insurance plans) and allow the AI significant time to process this information, resulting in detailed reports. This contrasts sharply with novices who might use short, vague prompts, leading to generic or unhelpful responses. The analogy of AI as a motivated, but initially uninformed, college graduate highlights the need for users to provide sufficient context for accurate and tailored outcomes, such as crafting a self-review by uploading project details instead of asking for a generic one.
Leveraging large language models and their knowledge sources
AI models learn by reading massive amounts of text from the internet, accumulating 'pre-trained knowledge.' This vast dataset includes everything from social media posts and books to news articles and research papers. The AI's understanding of a topic reflects how frequently that topic appears online. For instance, cooking, being a universal human experience, is well-represented. More specialized topics, like quasars, are present but far less frequently. This data source also explains AI's robustness in handling common errors like typos, as it has learned from text containing them. However, this reliance on internet data also means AI can inherit its biases and outdated information, necessitating careful prompting and validation.
Augmenting knowledge with web search and deep research
An AI model's pre-trained knowledge has a cutoff date, meaning it lacks real-time information. To address this, AI models can perform web searches, triggered either automatically when they detect a need for current data (e.g., '67 meme from 2025') or explicitly by the user. This extends their capabilities to current events, location-specific details, or niche information. However, web search results can be unreliable, drawing from popular but not necessarily credible sources. Users can mitigate this by encouraging the AI to consult official organizations or scientific studies. For more complex information synthesis, requiring analysis of dozens of sources, a 'deep research' mode is available. This mode allows the AI to formulate a research plan, conduct multiple parallel searches, evaluate sources, and then synthesize the information into a detailed report, akin to an agentic process where the AI makes its own decisions about the research path. This is invaluable for tasks requiring in-depth understanding, like the impact of daily steps on long-term health.
Developing AI as a thought partner: Brainstorming and writing
AI can serve as a powerful thought partner, particularly in brainstorming and writing. While asking an AI to generate lists of ideas is common, more effective brainstorming involves iterative back-and-forth conversations. By providing detailed context (e.g., personal stats for a workout plan, financial details for debt repayment) and giving feedback on AI-generated options, users can guide the AI towards more creative and personalized outputs. This iterative process helps refine ideas and uncover unique solutions, moving beyond generic common-sense responses. For writing, simply asking AI to 'write' often results in 'AI slop' – text that sounds plausible yet lacks substance. A better approach is progressive outlining: refning an outline, then bullet points, before generating the final text. This is more efficient as changes at the outline stage have a larger impact on the final output.
Ensuring objective feedback and high-quality writing
A persistent challenge with AI is its tendency towards sycophancy – agreeing with the user to please them. This is often a result of training data and user feedback mechanisms that reward agreeable responses. To combat this, neutral framing in prompts is crucial; instead of 'Don't you think remote work is better?', ask 'How does productivity compare between remote and in-office work?'. For critique, providing a detailed rubric with objective grading criteria forces the AI to evaluate work critically, rather than offering generic praise. For writing, editing piece by piece (sentence or paragraph at a time) allows for more focused refinement. The use of a well-defined rubric, which can even be developed with AI's help, ensures that AI evaluations are objective, leading to genuinely helpful suggestions for improvement, rather than superficial agreement.
Multimodal AI capabilities: Images, audio, and code generation
AI's capabilities extend far beyond text. It can generate images, videos, voices, and code. Image generation, often using diffusion models, can create novel visuals from text prompts, restore old photos, or even be guided by other images. While text generation is fast and cheap, generating images, and particularly video, is more time-consuming and costly. AI can also clone voices with increasing accuracy, though this raises ethical concerns. Generating code, once a domain of professionals, is now accessible via text prompts, enabling users to create simple games, websites, or data analysis tools without coding expertise. Prompting for multimodal inputs and outputs requires similar principles of providing context, but the iteration process can be slower due to increased generation time and cost.
Building applications and analyzing data with AI
AI is democratizing software development, allowing users to build basic applications and games through text prompts. Tools can create anything from simple simulators to functional apps like timers or calculators. The process typically involves specifying the goal, inputs, and outputs. AI can also analyze user-provided data by writing and executing code. Uploading spreadsheets of personal health data or business sales figures allows AI to generate plots and insights, aiding in tasks like identifying sales trends or analyzing running pace. While AI data analysis may not match a human data scientist's sophistication for complex tasks, it offers significant speed and efficiency for extracting basic insights. These capabilities can be further enhanced by using the best available AI models and providing them with comprehensive context.
The future of AI: Reasoning, context, and continuous learning
The latest AI models possess advanced reasoning capabilities, enabling them to 'think' rigorously and at length about complex tasks when given sufficient context. This involves iterative processes of gathering information (via web search or file access), reasoning, and refining answers, sometimes taking minutes or even longer. Providing ample context, akin to what a human expert would need, is crucial for these reasoning tasks. AI also exhibits 'jagged intelligence,' meaning it excels in some areas but not others, and different models have varying strengths. The competitive AI landscape means models are constantly improving. Therefore, continuously testing new models with challenging tasks and high-quality context is key to honing intuition about AI's evolving capabilities and becoming an 'AI power user.'
Mentioned in This Episode
●Products
●Software & Apps
●Companies
●Organizations
●Concepts
●People Referenced
AI Prompting Power User Cheat Sheet
Practical takeaways from this episode
Do This
Avoid This
AI Task Difficulty & Human Time Conversion
Data extracted from this episode
| Task Type | Human Time (Approx.) | AI Success Rate (2024-2025) |
|---|---|---|
| Finding a fact on the web | Several seconds | High |
| Summarizing a few text files | An hour | Decent-High |
| Writing a blog post | Couple hours | Decent-High |
| Auditing legal documents | Many hours | Decent-High |
| Exploring complex cybersecurity vulnerability | Many hours | Decent-High |
| Tasks taking many hours | Many hours | Models can now do |
AI Output Generation Time & Cost by Modality
Data extracted from this episode
| Output Modality | Time for Single Output | Cost for Single Output |
|---|---|---|
| Text (short paragraph) | Seconds | Less than a cent |
| Text (long paragraph / deep thought) | Longer | More costly |
| Speech | More expensive | More costly |
| Images | Tens of seconds | Many cents |
| Video | Much more expensive | Much more costly |
Common Questions
AI novices tend to ask simple questions like a Google search and use short prompts, often leading to generic results. AI power users provide extensive context and employ iterative prompting, treating AI as a thinking partner to get detailed, custom, and accurate responses.
Topics
Mentioned in this video
A popular commercial AI service mentioned as an example of a model that users can upload documents to for complex queries and also cited for statistics on chat usage patterns.
Another commercial AI model mentioned alongside other popular services, specifically highlighted for its deep research capabilities to turn reports into webpages or infographics, and for its cross-model review capabilities.
A commercial AI service noted for accepting complex document uploads for trade-off analysis and for its ability to generate games from simple prompts.
Mentioned as an internet forum and social media site that AI models learn from, similar to Reddit, which contributes to the breadth of AI's pre-trained knowledge.
A web search engine mentioned as a human-used alternative to Google, also used by AI models for web searches.
An organization that conducted a study on AI's ability to perform long-running tasks, illustrating the rapid growth in AI reasoning capabilities from 2024 to 2025.
A social media site mentioned in the context of the 'em dash' trend, suggesting that AI's writing patterns can influence human communication styles.
Google's AI image generation software, used to create a custom birthday cake design and to restore a faded childhood photo, showcasing its advanced capabilities.
A state-of-the-art video generation model by Google from 2022, mentioned to show the significant progress in AI video generation over a few years, contrasting its earlier artifacts with modern capabilities.
A spreadsheet software program mentioned as a traditional tool for data analysis, which AI can often perform faster and more efficiently by writing and running code.
A spreadsheet software program, like Excel, that AI can replace for basic data analysis, offering speed and efficiency gains.
Used as an example topic for a blog post to illustrate how AI novices might generate 'slop' text directly, versus how power users would create a progressive outline.
A car model used to illustrate that a web search engine is best for finding specific product data in its original form (e.g., an air filter for a 2013 model).
Mentioned in the context of the Voyager 1 mission and its golden record, as an example of an entity AI has learned niche knowledge about.
Listed as an encyclopedia source that AI models read from during training, and one of the most cited websites by AI models, indicating it's a significant knowledge source.
An official health organization recommended as a reliable source for AI to use when providing information on sensitive topics like 'gray market peptides'.
A government body suggested as a trustworthy source for AI to consult when researching health-related queries, to ensure scientifically credible answers.
A European regulatory body presented as an example of an official organization providing reliable, scientifically verified information for AI web searches.
The publisher of 'The Batch' newsletter, where the speaker's AI voice clone was used to read a passage, and also provides courses on building AI applications.
Cited as a frequent source of internet text for AI training, and as one of the most cited websites by AI models, highlighting its popularity but also potential for untrustworthy information.
Data from OpenAI on writing-focused chats accounts for about two-thirds involving editing pre-existing text, rather than starting from scratch.
Mentioned as a common web search engine for humans, an AI model provider (Gemini, Nano Banana), and a source of data (Google Search trends in 2025). The company's AI models are specifically highlighted for image generation and desktop apps.
Listed as one of the most cited websites by AI models according to a report.
Included in the list of most cited websites by AI models, indicating its content contributes to AI's pre-trained data or search results.
Used as an example of a small company creating 'Toy Story' with a small team in the 90s, serving as a historical analogy for agile AI teams in article writing.
A specific location used in an example query ('highly rated gym near Mountain View, California') where AI would trigger a web search due to its location-specific and time-sensitive nature.
A location mentioned in an example where AI provided outdated web search results for running places, highlighting a limitation of web search if sources are not current.
A historical site used as an example query ('what should I know before hiking Machu Picchu') to demonstrate how AI's assistant model conducts multiple web searches to gather comprehensive information.
Mentioned as a location for Halloween haunted house planning, where AI would research local permits and ordinances, and later as a location for checking picnic weather in a lab exercise.
Cited as conducting a study on ChatGPT responses, revealing AI's tendency to agree with users significantly more often than it disagrees, highlighting the issue of sycophancy.
The first fully computer-animated feature-length film, created by Pixar with a small team in the 90s, highlighted as an inspiring example of innovation.
A weekly newsletter published by DeepLearning.AI, used as content for the speaker's AI voice clone demonstration.
More from DeepLearningAI
View all 67 summaries
61 min🎧 LoFi Beats for Coding & Focus: Calm Beats to Study, Build, and Think
1 minThe #1 Skill Employers Want in 2026
1 minThe truth about tech layoffs and AI..
2 minBuild and Train an LLM with JAX
Ask anything from this episode.
Save it, chat with it, and connect it to Claude or ChatGPT. Get cited answers from the actual content — and build your own knowledge base of every podcast and video you care about.
Get Started Free