{
  "video_id": "8ib4Qnh2HFE",
  "channel_slug": "deeplearningai",
  "channel_handle": "DeepLearningAI",
  "title": "Full AI Prompting Course with Andrew Ng",
  "duration_seconds": 8926,
  "url": "https://www.youtube.com/watch?v=8ib4Qnh2HFE",
  "upload_date": "20260518",
  "transcript": "It is now 2026 and prompting AI models\nis very different from when CAT GBT\nfirst came out in 2022.\nUsing AI well is one of the most\nimpactful skills you can develop. And\npeople that are not yet at a cutting\nedge of AI usage often run Intel AI\ngenerating frustrating outputs. I want\nto make sure you're an expert prompter\nand can take advantage of today's AI\ntools which are much more powerful than\nthey were even a year ago. Let's take a\nlook at two different experiences. The\nAI novice and AI power user. Many AI\nexperts have learned to use it to answer\nhard questions. In contrast, many\npeople, including AI noviceses, may have\ngotten used to using AI for simple\nquestions as if you were prompting it\nlike a Google search. So you ask it,\ndoes Taco Bell still have the double\ndeck taco? And maybe you get an answer\nlike that, which is fine. But if you\nhave much harder questions, you can also\nask it of the AI and give it time to\nthink. For example, if you're looking to\nbuy a car, you can upload to most of the\ncommercial services like Chai GPT,\nGemini, Anthropics, Claude, or others a\nset of documents including cost specs,\nquotes, insurance plans, and ask it what\nare the trade-offs for these different\ncars I'm thinking about and tell it to\nread everything and to think hard before\nanswering.\nAnd this can cause AI to spend many\nseconds or even minutes to think and\nthen compile a detailed report for you.\nI find this a huge timesaver for a lot\nof things I have to do. Another example,\nAI power users have learned to provide\nthe right context or the right\nbackground information to the AI to set\nit up for successfully answering your\nquestion. In contrast, I see some AI\nnoviceses use a short prompt and hope\nthe AI will fill in the blanks. But if\nyou think of AI as maybe being akin to a\nreally smart, fresh college grad, highly\nmotivated, but that doesn't really know\nthat much about you yet, then a short\nprompt sometimes doesn't give it enough\ninformation or enough background context\nanswer your question accurately. So, if\nyou tell AI, please write a good self-\nreview to send to my boss, the AI\ndoesn't know what you've actually done\nover the last year because you haven't\ntold it yet, and it might write a very\ngeneric self-re, which isn't that\nhelpful.\nIn contrast, I find that AI power users\nalmost have empathy for the AI. I don't\nwant to overly anthropomorphize the AI,\nbut if you could put yourself in the\nshoes of someone getting a set of\ninstructions from you, you can ask\nyourself, will they actually know enough\nabout you to do a good job on the task\nyou're assigning them? So, an AI power\nuser in comparison might upload a lot of\ninformation to the AI. Maybe give it a\nscreenshot of a project tracker showing\nwhat you worked on, recent project dogs,\nmaybe voice memo notes where you talk\nthrough the projects and then tell it to\nwrite a self- review to send to my boss.\nAnd that could do a much better job\ncapturing what you're most proud of.\nOne of the things power users have\nlearned to do is how to prompt AI to get\nhonest feedback. A big problem with AI\nis it often wants to please you. In\nfact, many AI systems were trained to\ntry to make the users happy. And if you\nask it a biased question, it will often\ngive a biased answer because it's trying\nto tell you what it thinks you want to\nhear. For example, if you say, \"I have a\ngreat business idea, mobile tie dying.\nCritique it.\" because you called it a\ngreat business idea and you're saying\nit's your idea, the AI will naturally\nwant to please you and say, \"What a\ngreat idea.\"\nWe sometimes call this sycopency and\nit's well known that if you give even a\nhint of what answer you're hoping for,\nthere's a good chance the AI will just\nreflect back your preferences or your\npreconceptions.\nIn contrast, AI power users tend to ask\nneutral questions that don't give any\nhint to the AI for what answer you're\nhoping for or not hoping for. Or if you\ngive it a rubric or grading criteria to\ntell the AI how to form the basis for\nits answer, that also forces it to be\nmore objective. For example, if you are\nto say, \"Please analyze the following\nbusiness idea objectively, mobile tie\ndying.\" And don't just make up a bunch\nof things for what you think. Use the\nrubric or the grading criteria above,\nsuch as, \"Is there a problem? Is there a\nmarket? Do I have a competitive\nadvantage?\" If you give instructions\nlike this to the AI, then the AI doesn't\nknow. Are you hoping it'll tell you it's\na great idea or that it will save you\nfrom spending a lot of time on a bad\nbusiness idea? and it's much more likely\nto then tell you something like, \"Oh,\nthis idea is a eight out of 100 and also\nwhy the score's low.\"\nIn case you run a mobile tie-dye\nbusiness, I wish you really best of luck\nand AI could also help ask some useful\nquestions to help you think through how\nto make the business even better.\nLastly, I found that AI novices and AI\npower users ask AI to write in very\ndifferent ways. Novices will just ask AI\nto write stuff like write a blog post\nabout the Blackberry\nand it will generate a bunch of text\nthat maybe looks like this which sounds\nlike AI slop. There's a bunch of generic\ntext that's just not that interesting\nand takes up a lot of space. In\ncontrast, an AI power user will often\nnot ask the AI system to just jump in\nwriting directly, but instead ask the AI\nto first outline an article and then\ncritique the outline and maybe iterate a\nfew times with the outline to shape the\narticle and only then ask the AI to\nstart to draft the final article. So\ngive it a set of uploaded notes as\ncontext. An expert may say, \"I'll line a\nblog post about the Blackberry based on\nmy notes so it knows what you want to\ntalk about.\"\nAnd the AI may start by giving an\noutline. And you might then give\nfeedback to the AI about what you like\nand what you don't like about the\noutline. And even iterate a few times,\nhave a few back and forth rounds before\nyou have an outline that you're\nsatisfied with. And maybe only then\nexpand the outline into bullet points.\nAnd maybe even go back and forth a few\ntimes to critique the bullet points\nbefore you're satisfied with that and\nthen expand it into the final text.\nThis type of power user workflow is much\nmore likely to generate some text that\nyou're happy with as opposed to AI slop.\nAnd in this type of workflow, you're\ntreating the AI as a thinking partner to\nalmost help you brainstorm and explore\ndifferent options for what you might\nwant to write. Air systems do make\nmistakes, but maybe fewer than most\npeople think, especially if you prompted\nwell. They made a lot more mistakes back\nin 2022 or 2023 than they do now. But a\nlot of widely publicized mistakes that\nAI has made, some of which went viral on\nsocial media, has made people think that\nAI maybe makes even more mistakes than\nit actually does. There's a well\npositized one where people asked it, how\nmany Rs are there in the word strawberry\nand it thinks there are two Rs. And\nhere's one that I found amusing. I want\nto wash my car. Should I walk or drive\nthere? And AI says walk, which would\nleave you there or wash your car. But\nthese viral examples are not\nrepresentative of AI capabilities.\nIn contrast, P users know that AI can\ndeliver significant value through tasks\nlike doing deep research and writing\nresearch reports or taking your personal\ndata like your health or heart rate or\nrunning time data and analyzing that for\nyou or stuff we will talk about later\neven building websites for you.\nI've seen being an AI PA user\ntremendously benefit individuals as well\nas their businesses. It'll save you time\nand improve your professional and\npersonal lives. They'll help you to\nbuild lots of cool things. You learn how\nlater in these videos and be able to\nprompt AI at an expert level is an\nhighly in demand job skill no matter\nwhat job row you're in. In the rest of\nthese videos, I hope to take you from\nwherever you are today to being an AI\npower user. Much has been said about AI\nbeing useful. I find using AI really fun\nas well, and you'll see a few examples\nof that in these videos, too. Now, one\nfoundational piece of knowledge that\nhelps you work of AI is understanding\nwhere it gets its knowledge from so that\nyou can better predict when it'll get\nsomething right and when you maybe\nshould encounter this answer. Let's go\non to the next video to learn about how\nAI gets its knowledge.\nHow did you learn to write as a child?\nProbably it involve reading a lot of\nthings. Well, it's the same for AI.\nAI systems have learned patterns from\nreading large amounts of text from the\ninternet. By understanding what's in\nthat text the AI has read, you'll be\nable to predict how they'll behave. AI\nmodels can answer questions on a variety\nof topics. If you were to ask, I dropped\nmy phone in soup. What should I do? Then\nhopefully you'll make some useful\nsuggestions.\nOr why do cats stare at walls like they\nsee in ghosts? My daughter loves cats.\nShe was actually curious about this.\nTurns out cats can detect subtle sounds\nand movements that we as humans often\nmiss. Because of the amount of things is\nread on the internet, they will even\npossess niche knowledge that few people\nknow about. They were to ask what kind\nof things were on the vinyl record sent\ninto space. years ago, NASA had a\nspacecraft called Voyager 1 that\nlaunched in the 1970s and is now about\n25 billion miles away from Earth. But AI\nwill know about this and be able to tell\nyou what is on that vinyl record. I\nthink it's cool that NASA chose to send\ngreetings in 55 different languages to\nwhoever may come across that spacecraft,\nif anyone does. AI models are trained on\nmany many different sources of\ninformation mainly from the internet and\ntraining on all of these very diverse\nsources of knowledge produces is\npre-trained knowledge. The term\npre-trained is a technical term that you\ndon't have to worry about. It turns out\nAI systems are trained in multiple steps\nand this is one of the first steps of\ntraining that somehow wound up being\ncalled pre-training which isn't a great\nterm but I wouldn't worry about why we\ncall pre-training. It's just what AI has\nlearned from. But these knowledge\nsources may include a lot of texts from\nsocial media like Reddit which will have\nanswers to questions like what are your\nmustwatch films or it may have read a\nbook on Lego micro cities or read a\nWikipedia article on fairy bread and\nlots of other things or read a bunch of\nnews articles as well as read a lot of\nresearch articles\non the internet. There's a lot of texts\non internet forums and social media like\nReddit and Quora. There are a lot of\nbooks that AI will have read from. There\nare encyclopedias like Wikipedia, news\nwebsites, research articles, and much\nmore. And so these trillions or tens of\ntrillions of words will go into training\nthe AI models brain.\nNow different types of data appear with\ndifferent amounts of frequency on the\ninternet and so this pre-trained\nknowledge reflects the frequency or the\npatterns in the training data. For\nexample, cooking is a very universal\nhuman experience. So there are a lot of\narticles on the internet on cooking.\nThere also a lot of articles online on\ncelebrities on movies and so AI will\nhave seen a lot of text on these topics.\nIn contrast, they're more specialized\ntopics like quazar, which is an\nastronomical term referring to really\nbright objects in the sky powered by\nsuper massive black holes. I think\nthey're fascinating, but they just a lot\nfewer articles on quazos and on cooking\non the internet. Now, while most of the\ninternet is in English, AI systems will\nalso have learned from some data that's\nwritten in other languages like\nCantonese. Over 80 million people speak\nCantonese, but that's far less than\nEnglish and Cantonese data represents\nmaybe less than 0.1% of all internet\ncontent.\nLastly, there are things that AI models\nknow nothing about at all, such as your\ncompany's secret proprietary data, which\nhopefully is not on the open internet,\nbut which an AI system will therefore\nnot have learned from. So I find that\nthinking about how frequent data appears\non the internet gives you a good rule of\nthumb for thinking about how reliable an\nAI systems responses are. Now because of\nthe data the AI has learned from\nsometimes it can exhibit surprising\nunderstanding of things. If you were to\ntype very quickly, can you cook eggs in\nmicrowave? Like shown on the left, you\ncan actually understand this type of\nmisspelled text very well. Pretty much\nas well as asking, can you cook eggs in\nthe microwave? And by the way, I've\nexploded a few eggs in the microwave\nmyself. So, if you ever want to avoid\nthat, feel free to ask the AI system how\nto do so, so you don't have to learn the\nhard way. And a reason that it's so good\nat understanding misspelled words is\nbecause it's actually learned from a lot\nof sources that could include typos. So\nif you look online, you will see phrases\nmisspelled. And that's why when you're\nusing the system, I'm not encouraging\nyou to use bad grammar or to misspell\nwords. But it turns out that if you're\ntyping quickly and you have a few typos\nor even a lot of typos, don't worry too\nmuch about it. It's pretty fine to just\nsend a prompt to AI and not spend too\nmuch time fixing every little\ngrammatical error. Now, the bad news is\na lot of AI sources also have\nmisconceptions and outdated information.\nSo, one of the skills in using AI is how\nto prompt it to have it give you back\nanswers that reflect fewer\nmisconceptions and does not overly\nreflect outdated information.\nBy understanding AI's knowledge sources\ncalled as pre-trained knowledge, you'll\nbe able to better predict how it will\nrespond to your prompts. But this\npre-trained knowledge is not enough for\nall applications, including those that\nneed real time information. For that,\nyou need web search. Let's go on to the\nnext video to learn more. At some point,\nthe people building the AI model had to\nstop his training. So there's some last\nstate where is information cuts off.\nThat is the AIS read the internet only\nup to certain date and time and it\nknowledge gets frozen in time as of that\ndate. But of course the world moves on\npast that date. New things happen,\nmovies come out and so on. Let's see how\nAI models handle gathering new\ninformation using web search so that it\ncan address questions that even relate\nto things after its knowledge cut off\ndate. If you're using one of the popular\nAI model providers like Jagy, Gemini,\nand CO, there are certain questions that\nwill probably trigger it to do a web\nsearch. For example, if you ask it, what\nis the 67 meme from 2025?\nThere's a good chance it will search on\nthe internet to tell you that the 67\nmeme, which is pronounced 67, which is\nkind of fun to say that this is a viral\ninternet slang widely seen on a few\nsocial media platforms. And the reason\nit triggers a web search when you ask\nit, what's a 67 meme from 2025? The Q\n2025 causes the AI to realize that it\nmay benefit from more updated online\ninformation because this could be a meme\nthat appeared on the internet after its\nknowledge cut off date. Here's what I\nmean. A specific AI models pre-trained\nknowledge is frozen in time even though\nthe internet continues to evolve over\ntime. And so if this line represents\ntime, then for a long time the internet\nwill have had pieces of text that say 6\n* 7 = 42. Text that talks about the\nchildren's joke. Why was six afraid of\nseven? Because 7 8 9. But if the\nknowledge cut off date was at a certain\nmoment in time and the 67 meme came\nafter that then the 67 meme would not\nhave been seen in the pre-trained\nknowledge of the AI model. So if you ask\nit what is the 67 meme from 2025 the AI\nmodel will realize that it doesn't know\nabout this 67 meme from 2025 and that it\nshould do a web search in order to get\nmore updated information.\nTake the GPD 5.4 model from OpenAI. His\nknowledge cut off dates was August 2025.\nAnd this graph shows how many Google web\nsearches there were for what does 67\nmean. So this 67 meme had taken off\nafter this GP 5.4 knowledge cut off\ndate, which is why the model doesn't\nreally know about this meme. Now,\nthere's certain types of questions that\nan AI will answer using its pre-trained\nknowledge, and there's certain types of\nquestions that will tend to trigger web\nsearch. For example, if you tell it,\nplease find me a highly rated gym near\nMountain View, California. Then, what is\nhighly rated, what may be open, and what\nmay be closed, does change over time,\nand there's a good chance that this will\ntrigger a web search. Or if you ask it,\nwhat is the market mountain cheese row?\nBecause this is a niche piece of\ninformation, it's probably not read a\nlot of information online about this\ncheese row. There's a good chance that\nit will search the internet in order to\nget you an answer. And if you're\ncurious, this is actually a pretty fun\nevents where people chase a rolling\nwheel of cheese down the hill. Let's\ntake a look more broadly at when an AI\nmodel needs to do some web search to\ngather more information to answer your\nquestion. If you're asking what to do if\nyou drop your phone in soup or why do\ncats stare walls or walls on the Voyager\none record in outer space then these\nquestions it could probably answer using\nthis pre-chain knowledge because these\nare represented in common knowledge on\nthe internet. But if it was ask it about\ncurrent events or something happening\nvery recently then it'll need to do web\nsearch to get that real-time\ninformation. If you ask it location\nspecific information and doing a web\nsearch makes sense. Or if you ask it for\nother types of niche information,\nthere's also a good chance of realize it\ndoesn't know enough about that topic\nthat doing a web search to gather more\ninformation would help it give you a\nbetter answer. For most of the popular\nAI model providers, web search can be\ntriggered in either of two ways.\nSometimes the AI model will decide by\nitself to carry a web search or you can\nalso explicitly trigger web search\nsometimes by clicking one of the buttons\nin a AI model providers web interface or\njust writing your prompt. Please do a\nweb search for this and it will comply\nand use a web search to answer your\nquestion. Not all AI models have web\nsearch enabled, but the most popular\nones that you're probably using mostly\ndo have this capability.\nAI will do better on many of the tasks\nyou want to use it for if it does web\nsearch. And web search allows it to\naugment this pre-trained knowledge with\nmore current information.\nBut like all web search, it can return\nbad sources. Let's take a look at when\nthis is an issue and when and how to get\nit to use more reliable sources to get\nyou more reliable answers.\nWeb search is a very valuable but\nimperfect tool. Just like when you\nsearch the web yourself, you might not\nalways find what you're looking for. It\nhas limitations like finding old or\ninaccurate sources.\nBut you can work around these\nlimitations to get AI to give you more\naccurate and up-to-date answers. Let's\ntake a look. If you ask a AI system, how\nsafe are gray market peptides, which is\na type of supplement, it may search\nonline and find posts on social media or\npublic forum sites like Reddit and Hora.\nOr you may find websites that are in the\nbusiness of selling peptides and so\nwould have a inclination to tell you\nthat they're safe and you may get back\nanswers that may or may not be accurate.\nBut if you encourage the AI model to use\nsources from official organizations or\nlook at studies that are backed by\nrigorous science, then it's more likely\nto look up resources from the World\nHealth Organization, from the US Food\nand Drug Administration, from the\nEuropean Medicine Agency and so on, and\nhopefully give you more reliable and\nscientifically credible answers.\nWeb search, whether done by a human on\nGoogle or Bing or done by AI, has a\ntendency to draw from popular sources.\nAccording to one report, the most cited\nwebsite by AI model was Reddit, followed\nby Wikipedia, YouTube, then Google\nitself, Yelp, and so on. And some of\nthese sources are more trustworthy than\nothers. There's just a lot of text on\nthe internet from social media, blogs,\nonline forums, and the amount of text\nfrom highly reliable scientifically\nverified sources is just much smaller.\nSo, if you don't steer the model in\nterms of what types of sources you\nprefer, there's a chance that it'll tend\nto pull text from whatever is most\navailable rather than what's most\nreliable.\nSo that's why if you ask it how safe are\ngray market peptides, it might base a\nlot of his answer on social media blogs\nand forums and only a little bit on the\nmore reliable sources. Whereas if you\ntell it to use sources from official\nhealth organizations, it may pull much\nmore from these reliable sources.\nAnother limitation of web search is that\nsometimes web pages can be outdated.\nthat can lead the AI model to also not\nprovide the most current information. A\nfriend of AI recently helped me find\nplaces to run in Henderson, Nevada. This\nis a location specific niche query and\nso this triggers web search and it found\nthis list of places to go for a jog. But\nit turns out that unfortunately this\npull from a web page from more than two\ndecades ago. And unfortunately, the\nlocation suggested was a school that\nunlike decades ago is no longer open to\nthe public to go running in. To help\nbuild intuition about how AI searches\nthe web to use that information, let me\nbriefly explain how web search actually\nworks under the hood. It turns out to be\na multi-step process. Imagine that\nyou're asking questions of a customer\nservice team of two people. There's the\nuserfacing AI model. That's what you are\ntalking to. And the userfacing AI model\nhas a second assistant AI model that it\ncan ask for help to do web search. So\nwhen you send a prompt, you are talking\nto the first model, the userfacing AI\nmodel, and it will occasionally decide\nto call up the assistant AI model, the\nsecond AI to say, hey, please do a web\nsearch for me to gather more\ninformation. This assistant AI model\nwill then search on a web search engine\nvery similar to Google and Bing and\nother web search engines that we as\npeople might use and it will scan the\nreturn results, filter out the relevant\nresults and download the most relevant\nweb pages and then summarize them. The\nsecond assistant AI model will then\npresent the summaries back to the first\nmodel, the userf facing AI model. And\nthe first model will then use these\nsummaries in order to generate the final\nanswer for you. You are speaking only to\nthe userfacing AI model. And one\ninteresting quirk to keep in mind is the\nuserfacing AI model has not actually\nread in its entirety all of the web\npages it may be citing for you. Instead\nis only seen summaries of those web\npages. And sometimes this causes it to\nmisinterpret what one of these\nunderlying web pages actually says.\nWhich is why you may have seen funny\nresults where AI cites a web page and\nsays the web page justifies a\nconclusion. But if you look at that web\npage yourself, it doesn't actually\njustify what the userfacing AI model\nsays it is doing. To walk you through\none example of this process, if you ask\nthe userfacing AI model, that's like the\ncustomer service agent talking to you.\nIt will ask what should I know before\nhiking Machu Picchu. The second model\nmay do some web searches with phrases\nlike mu future permits, much future\nweather or the social customs and so on.\nAnd it'll then scan the returned results\nmuch like you may scan the page of\nGoogle results to decide what's relevant\nand filter out irrelevant results and\nsummarize the most relevant web pages to\nprovide back to the first agent that\nthen generates the final answer for you.\nNow I frequently use AI models like\nCATV, Gemini, C and I also frequently\nuse web search engines like Google and\nBing. When should you use a AI model and\nwhen should you use a web search engine?\nIf you want to quickly scan multiple\nsources, a search engine can be useful\nfor that. Or if you want to navigate to\na specific website but have forgotten\nwhat's the name of that website, a web\nsearch engine can be very good for\nhelping you find it. Or if you want to\nlook at data in its original form, such\nas if you want to buy a 2013 Honda Civic\nair filter, you know, you want to find a\nwebsite to go to to buy that air filter.\nSo web search engine is very good at\nthat. In contrast, if you want to get a\nsynthesis from multiple sources, or if\nyou're searching for more complex\ninformation with pros and cons that you\nwant weighed, or if you just want to\ncontrast multiple sources to come up\nwith a more thoughtful conclusion, then\nan AI model can do a web search and put\ntogether the results of multiple web\npages for you quite efficiently, thus\nmaybe saving you time of having to read\na lot of web pages yourself. There might\nbe some good Google or other web search\nhabits that you've developed and those\nhabits will serve you well when working\nwith web search enabled AI models as\nwell. Things like looking for reliable\nsources and also double-checking the\nsources. But if you want to go beyond\nsearching a handful of web pages, it\nturns out AI models are capable of a\nmuch more extensive type of research\ncalled deep research.\nThis is a very powerful capability that\nI think is really underused by many\npeople. Let's go on to the next video to\nsee what it is and when and how to use a\ndeep researcher.\nSometimes you would want your AI to\nsynthesize not just a handful of sources\nbut many maybe many dozens of sources\nand do lots of thinking to come up with\nthe best possible deeply researched\nanswer to a question that you have.\nPopular AI chat interfaces like\ntragically Gemini and CO all have a deep\nresearch mode. I found this to be a very\nvaluable and often underutilized tool.\nLet's take a look. Let's say you want to\nuse an AI model to help you plan your\nHalloween haunted house. I'm going to\nwrite a prompt to ask you to help me set\nup a haunted house in my front yard for\nHalloween and give it some information\nabout where I am, what's the size of my\nfront yard, what's the experience I\nwant. So, I give it lots of context to\nset it up to plan it out for me\nappropriately.\nWith a prompt like this, an AI model\nmight come up with a research plan in\nwhich it tries to think through what are\nthe types of sources it needs to\nresearch.\nMany systems would give you an\nopportunity to approve or potentially\nedit the research plan. And if you're\nhappy with it, I'll often launch the\nresearch plan without updating it unless\nI see something that just looks really\nwrong. It will then go ahead and start\nto do online searches. So in this\nexample is starts by gathering Palo\nAlto's rules on permits Halloween\nordinances and so on. And then it will\nread some of those web pages and\nsynthesize what is learned so far. And\nit may then decide to do some more\nsearches online to gather more\ninformation about fire safety\nguidelines. And then it may after that\ndecide to look for decoration ideas. So\nloosely follow the original research\nplan but also have the flexibility to\nkeep looking deep into certain areas if\nit thinks it needs that information.\nAfter searching for a while maybe many\nminutes it will finally write a detailed\nresearch report for you. This process,\nby the way, is an example of agentic AI.\nAnd what that refers to is that through\nthis dresearch process, the AI model has\nsome flexibility to make decisions by\nitself on what to do next, such as what\nadditional searches, if any, to carry\nout. The output of this can then be a\nfairly detailed and thoughtful plan with\ndifferent sections outlining what you\nmight need to think about in terms of\nstructural and reg framework safety and\nso on. If you're using Google's Gemini\nAI model for this, one of the neat\nfeatures is it makes it easy to take the\ndeep researches done and hope you turn\nit into a web page or infographic or\nhandful of other things. Here's a web\npage that was generated by Gemini using\nthe Gemini D researcher. And I think\nit's pretty neat. There's created a web\npage with four different sections,\npie charts for budget,\npretty neat visualizations for noise\nordinance. And I think it's pretty neat\nthat this even has a little checklist\nthat I could use to plan out my\nHalloween event.\nTo give you a sense of how a deep\nresearcher works, this is loosely what\nit does. After formulating a research\nplan, an AI model can actually issue\nmany web searches at the same time and\nget back multiple web pages at the same\ntime. And this is one of the nice things\nabout using aid researcher.\nIt doesn't have to do the web searches\none at a time. It can do many of them at\nthe same time, which lets it be very\nefficient in fetching lots of web pages.\nThe system can also take a look at all\nof these sources and quickly assess\nwhich ones are relevant and which ones\nare less relevant. And based on that, it\nmay decide whether or not to go back to\ndo additional web searches, maybe using\ndifferent web search terms. Finally,\nafter going around this loop a few times\nof doing web search, evaluating sources,\ndeciding whether or not to go back to\nget more sources, it'll hopefully decide\nit's done. And then lastly, take all of\nthe pages it has downloaded and maybe\nsummarize and synthesize all that into a\nreport that it adds citations to and\nthat it then presents to you. Both web\nsearch enable AI as well as deep\nresearch use the internet or do web\nsearch. The basic web search enable AI\nis good at queries like this. Find me a\nhighly rated gym. What's the weather in\nDubai this week?\nWhereas deep research I would tend to\nuse for tasks that require synthesizing\nmultiple views such as if I want to know\nwhat's the impact of daily steps on\nlong-term health and if I wanted to\nsearch the most recent scientifically\njustified articles and think through the\nanswer rather than just tell me whatever\npeople tend to say on the internet. or\nif I wanted to deeply think through how\ndoes weather affect tourism in Dubai and\nagain not just take one or two popular\nanswers found on a social media site but\nto read up on weather read up on tourism\nread up on Dubai and to really think\nthrough the implications to give you a\nmore thoughtful answer that's when deep\nresearch could be particularly helpful\nto give you another framework to think\nthrough when to use web search versus\ndeep research if you have a single\nquestion do you want answered doing work\nthat would take me just a few seconds\nbased on a handful of sources. That's\nwhen web search would be helpful. And\nweb search, as we've seen, can be\ntriggered either automatically or by the\nuser. Whereas deep researchers often is\ntrying to draw a complex set of\nconclusions that may require answering\nmultiple questions or answering multiple\ndimensions that relate to a question.\nAnd I think of this as doing work that\nwould take me minutes to maybe even\nhours if I was doing this manually. And\nI may want many sources of integration\nsynthesize. And as we've seen, deep\nresearcher is usually triggered\nexplicitly by the user unless you select\nit in the user interface. Most AI models\nwill not care a deep researcher and keep\nyou waiting for many minutes for an\nanswer. To recap, if you're asking,\n\"Hope I drop my phone in soup,\" it\ndoesn't need to look up any online\nsources. We're not worried about\nfreshness. It'll give you an answer in\njust a few seconds. And this is good for\nfinding basic facts, definitions,\nsummaries for things that occur calmly\non the internet. webs may download a\nhandful of sources and it will find\nrelatively up-to-date information and it\nmay take many seconds to get you back an\nanswer. And we've seen what types of\ninformation this is useful for. And\nlastly, deep research may download often\ndozens or more of sources. It will get\nup-to-date information and it will spend\nmany minutes or longer to get back an\nanswer. and is great at answering\ncomplex question that involve\nsynthesizing many sources of knowledge.\nFinding information is one of the most\ncommon tasks that people use AI models\nfor. You've seen three different paths\nthat you can take advantage of for this\ntype of information finding task. You\ncould use just a pre-train knowledge or\nweb search or deep research. And we also\nwalk through how and when to use these\ndifferent options. I want to make sure\nthat you have good intuitions about when\nto use each of these options. So, let's\ngo on to take a look next at a practice\nhands-on lab for this module, which I\nthink you'll find a fun way to compare\nand contrast what each of these three\noptions do, and more importantly will\nalso help you hone your intuition on\nwhen to use each.\nIn this module, you learn how to use AI\nmodels to find information. In this\npractice lab, you can explore how web\nsearch, deep research, and different\nprompts affect the AI models output.\nLet's take a look. When you open up the\nlab, it starts off with a tutorial on\nhow to use the lab. And I'm just going\nto close it out here. We can always\naccess this tutorial again via this\nbutton up here. And what I hope you do\nis follow the instructions written here.\nAnd when you're done, click mark as\ncomplete to mark this item as completed.\nFirst, these buttons down here\ncorrespond to different things you might\ntry. So, this one, current events,\ncompares a question with and without web\nsearch. So, what's the 67 meme? This on\nthe left is without web search. The one\non the right is with web search enabled.\nAnd if you compare\nthen without web search it gives these\nanswers. It doesn't know about the meme.\nBut with web search it tells you what is\nthis 67 meme as well as the origins of\nthis meme. And if you want, you can also\nfollow up and ask, \"How do I use 67\nappropriately\nin a\nblack tie tuxedo\nbody?\" And then hit this red button to\ngo see his answer.\nLet's go back to the homepage by\nclicking new chat up here. And I hope\nyou try out the other examples such as\nfind me a highly rated gym and hit\ncompare or when's the next Avengers\nmovie scheduled or what major news\nhappened today in the US or feel free to\nenter your own country and compare these\nresults with and without web search.\nThis example over here shows the\ndifference between web search denoted by\nthis globe icon versus steep researcher\ndenoted by this microscope icon. So how\nsafe by gray market peptides use high\nquality sources and this would give you\na sense of what web search results looks\nlike\nversus deep researcher results. I hope\nyou run it and see what results you get.\nOne more example. If I want to ask, can\nI keep my rocket propelled monster truck\nin my garage? If you ask it this\nquestion, the AI model may or may not\nknow the answer. But if you were to also\nupload the lease agreement, then maybe\nyour lease agreement,\nwhich states the terms under which\nyou're renting your place, may have\nrestrictions on that. And so you'll be\nable to see different answers depending\non whether or not you upload additional\ninformation. In this case, a lease\nagreement that is helpful for the AI to\ngive you a thoughtful answer.\nWe haven't talked much yet about\nuploading your own files in this module,\nbut this is something we'll dive into\nmore deeply later in these videos, but\nfeel free to play with this. Now,\nlastly, one fun example.\nHere is a version of why do cats stare\nat walls with lots of typos. Here's one\nwith nicely formatted grammar, no typos.\nAnd you'll find that the answers are\nmaybe surprisingly similar\nthat the AI system is pretty good at\nanswering a question like this, even if\nit's lots of typos.\nOnce you tried out these examples\nreflected by these buttons down here,\ncome over your own. Like, is the weather\ngood for a picnic in Palo Alto today?\nAnd try this with or without web search\nand see what answers you get. Or pick\nyour own example and try comparing the\nresults you get using web search. And if\nyou want, uploading your own files.\nSo that's it for module one of this\ncourse. Great job getting this far. I\nhope you enjoy playing around with the\nlab. Next, please join me in the next\nmodule where you hear about using AI as\na thought partner, including having it\nbrainstorm with you and explore ideas\nwith you. This has helped me shape the\ndirection of many projects, and I'm\nconfident you find it useful, too. and\nwe'll also explore getting AI to help\nyou with your writing and editing. I'll\nsee you in the next module.\nOne of the most helpful uses of AI is as\na thought partner. When I'm trying to\nthink through a complex problem or make\na complex decision, it's nice to have a\nhuman expert as if our partner that is\nsomeone to talk things through with. But\nif there isn't a human expert readily\navailable, AI, which actually knows a\nlot about a lot of things, can be a\nreally good resource for this. We'll go\nthrough together multiple examples of\nthis, but to get started, brainstorming\nis one great such use case. Now, I know\na lot of people ask AI to help\nbrainstorm lists of ideas, but they're\nmore effective ways to use it as a\nbrainstorming partner than just having\nit generate a list. Let me show you what\nI mean. According to data released by\nOpenAI analyzing chat GPT conversations,\nabout half of Chat GPD chats are asking\nfor writing and practical guidance. And\nin fact, creative ideation accounts for\n3.9% almost 4% of all chats. I found\nusing AI to help me brainstorm to be\nreally valuable. Let me share with you\nsome ways to do so. AI can be pretty\ngood at generating options. There's a\ncommon creativity test which asks people\nto name 200 potential users for a brick.\nSo give it a brick like this. How many\nuses can you think of it? This actually\npretty difficult. Some people think,\n\"Oh, it could be a paper weight, maybe a\nplanter, and oh, it could be used to\nbuild a house, too, I guess.\" But to\ncome up with 200 examples, it's not that\neasy. But if you ask an AI model,\nthere's a good chance you can come up\nwith a long list of ideas. And your\nrole, if you're actually trying to use a\nbrick for something, would be to\nevaluate these options to pick out which\nones are the ones that you like. In\nbrainstorming, a common guidance is the\nmore ideas the better. And so sometimes\nhaving AI generate a lot of ideas for\nyou to pick from can be a powerful way\nto find one or two good ideas. So this\nis the maybe more common use of AI as a\nbrainstorming partner. I want to show\nyou a different form of brainstorming in\nwhich you give it more context and then\nalso iterate with the AI longer, meaning\nhave a longer back and forth\nconversation to help get you to better\noptions.\nSo if you tell it, help me build a\nworkout plan. I'm 38, bring the level,\nhave 10 lb dumbbells in 15 minutes a\nday, then they may give fairly generic\nanswers like three workout plans. So, 10\nsquats, 10 push-ups, pretty reasonable,\nvery sensible, common sense answer. But\nif you want more creative options,\ngiving it more context can be helpful.\nSo, if you say, \"I can't stick to these.\nGive me hacks to stay on track. I have a\ntrampoline and a cat.\" By encouraging it\nto give you trampoline and catreated\nworkout options, which is an unusual way\nof approaching workouts, it may ask you\nto consider trampoline breaks or cat\ntriggered micro workouts where maybe\nevery time you see your cat wags tail or\nsomething, go do a tidy little workout.\nBut these are certainly more creative\nideas. AI models have some inherent\ncreativity because they've trained on a\nlot of texts on the internet which\ncovers a lot of very different ideas\nincluding some creative ones. And AI's\noutput is a little bit random. So if you\nask it multiple times, help me build a\nworkout plan, it'll probably give you\nslightly different answers. But if you\ngive the AI basic questions, then common\nsense relatively generic responses like\ndo squats, push-ups, and so on are more\nlikely. Let me plot a conceptual diagram\nwhere on the horizontal axis I'm going\nto plot how unique a response is, how\ncreative response is. So on the left\nwere responses like normal weightlifting\nexercises like bicep curls which is very\ncommon sense to then maybe slightly more\nunique things like standing on one leg\nwith a yoga block on your head to the\nreally creative ones like cat triggered\nmicro workouts. And on the vertical axis\nI'm going to plot the probability of AI\ngiving these different responses.\nAnd it turns out that it's much more\nlikely to give a common sense response\nthan a highly unique creative response.\nThere's a reason for this. Namely, it\nwas trained on internet text. And\nthere's a lot more internet text talking\nabout dumbbell curls than there are cat\ntriggered micro workouts.\nAnd for most questions, this is actually\nokay because the average information on\nthe internet is probably decently\nfactual. So when you are seeking\ninformation such as what's the tallest\nbuilding in the world is actually the\nBurj Khalifer. Most internet texts will\nsay is the Burj Khalifa. There are\nsmaller amounts of text that will name\nother buildings but the average response\nand the most common response in the\ninternet is usually the factual one for\nquestions like what's the tallest\nbuilding? But if you're brainstorming,\nthen giving the average information of\nthe most common response ends up with\nsquats, push-ups, and almost never\ntrampoline breaks and pretty much never\ncatbased sessions.\nWhich is why if you ask the AI model to\nbrainstorm with you, you get a lot of\ncommon sense ideas rather than the more\ncreative ideas, which depending on your\ngoal, may or may not be what you want.\nSo what do you want to do if you want to\nget high quality more creative ideas\nfrom AI? We've seen with a basic prompt\nyou get responses from the common sense\nspace. But if you give the AI model more\ncontext, so give your age, your level,\nbut also tell it you have a mini\ntrampoline, a cat, trouble staying\nmotivated, nail squats, then this\ncontext pushes it into the more relevant\nand creative space and it's more likely\nto give a custom answer rather than to\ngenerate common sense answers. Now, one\nproblem that you may face when\nbrainstorming is if you're trying to\ncome up with creative ideas, there's so\nmuch context you could potentially give\nthe AI model, what should you prioritize\ntelling the AI model? It turns out\nthere's a technique that is very helpful\nfor driving what context you decide to\ngive the AI model, which is to iterate\nwith the AI. Let me show you what I\nmean. If I want to ask AI to help me\nbrainstorm plans for paying off my debt,\nhave $1,100 of credit card debt at 19%\ninterest, monthly minimum payments of\n$40, a student loan 8% interest, and a\nfamily loan $900. So, this gives decent\nbackground context. Then, one thing you\ncould do is ask AI not to give you one\noption or tell you what to do, but to\ngive you multiple options to choose\nfrom. I'll often ask you to give me\nthree to five options. And so the AI may\ncome up with a few different plans. Plan\none is liquidity first to preserve cash.\nPlan two is eliminate the highest\ninterest loan. Plan three is prioritize\npaying your family back first. So these\nare all actually reasonable ideas and\nI've not yet given it enough context to\nknow which of these plans it should\nfavor. And it turns out that one of the\nreally good ways to figure out what\nadditional context to gifted AI is to\ngive it feedback on the options it\npresent to you. Highly relevant feedback\nthat allows it to then give you the next\nset of options. I don't like option one.\nIt's too passive. I do like the idea of\npaying off the 19% interest loan. Oh,\nand I forgot. I actually have $450 cash\ncoming and I'm also moving house soon.\nAnd then with this additional context,\nit now knows among plans one, two, and\nthree maybe what you like and what you\ndon't like. And you can ask it to create\nthree new plans. And then once again, by\ngiving a feedback on these plans, you\nare giving additional context that will\nhelp shape the AI models thinking.\nAnd you can keep on iterating like this\nfor a while until it comes up with a\nplan that you do like and then maybe\nhave it flesh out the details of the one\nplan or two plans that you like the\nmost. I found that giving feedback to\nthe AI on what it thinks like good ideas\nis just a very useful mechanism for very\nefficiently figuring out what's helpful\ncontext to give to the AI. To summarize,\nif you're brainstorming, consider giving\nAI as much of the relevant context as\nyou can in advance and then ask it for a\nhandful of options. Then give it\nfeedback on the different options and\nask it for more options and iterate\nmultiple times. Get more options, get\nfeedback, get more options, get\nfeedback. And do that a few times until\nyou have one or more ideas that you're\nsatisfied with. If you follow this\nrecipe for brainstorming, I think you\nfind you get consistently more useful\nand creative ideas.\nNow, you've heard me use the word\ncontext quite a few times. It's\nimportant to give your AI model the\nright context so that it knows enough to\ndo what you want it to. Let's take a\nlook at the next video at how context\nworks and how it is used to produce a\nresponse.\nAccording to psychologists, most humans\ncan keep only about seven things in\ntheir active working memory at a time.\nThat's why remembering a grocery list of\nabout seven items is just barely doable\nif you aren't thinking about all the\nthings. But remembering a grocery list\nof 15 or 20 items is much harder.\nInterestingly, AI can use a large amount\nof context. Some models can have context\nsizes of hundreds of thousands of words.\nLet's see how an AI model's context\nworks and how you could take advantage\nof it. AI models can read and reason\nover very large amounts of context. For\nexample, if you are trying to choose an\napartment, you can upload hundreds of\npages of lease contracts and upload\ntenant reviews and neighborhood\nstatistics and ask AI to read all of\nthis and to tell you the pros and cons\nof each of the options. And you might\nwrite a prompt like pros and cons of\neach apartment. Read everything and\nthink really hard before answering. By\nthe way, telling AI to think hard or\nthink really hot is another common\nprompting pattern that we'll come back\nto a little bit later. Context refers to\nall the text and files that the model\nuses to generate this custom response to\nyour query. If you give it a prompt like\npros and cons of studying physics versus\nzoology, then it will generate an output\nbased on this very limited prompt, very\nlimited context that you have given it.\nand the response will probably be fairly\ngeneric. But if you give it more\ncontext, maybe give it your career\nassessment results and give it your high\nschool schedule so it knows what class\nyou've been taking and then ask the same\nquestion. All this additional context\nwill help it to give a much more custom\nand likely higher quality response. If\nyou're trying to think about what\ncontext to give to the AI model, think\nabout what's all the information that a\ntrusted advisor would need in order to\nthink at length and reason and then give\nyou a good answer to your question. And\na smart adviser that knows nothing about\nyou, but has just asked what are the\npros and cons of study physics versus\nzoology. The best it could really do is\ngive a pretty generic response that's\nnot custom to you because a context in\nthe example in Tom just doesn't have\nanything specific to you. AI models\nstart with some amount of builtin\ncontext and leading AI models today can\naccept maybe up to around 750,000 words\nas context and this corresponds to about\nthe first four or five Harry Potter\nbooks. So that's lava text or several\ndays of continuous speech. So many\npeople underestimate how much\ninformation or how much context you can\ngive to an AI model.\nNow when you ask the AI model a question\nby default\ncontext is filled with a few things.\nFirst, there's something called a system\nprompt, which usually is how the AI\nmodel knows what's the current date,\nknows the name of the model, basic\ncapabilities, maybe general instructions\nto be helpful to the user, and then if\nyour AI model is able to use tools like\na web search engine, in its context will\nalso be written descriptions of what are\nthese tools and how to use them, such as\nwhat is a web search engine and how\nshould it use a web search engine.\nBefore you've written your prompt, the\ncontext includes the system prompt and\nthese two definitions. And when you then\nwrite your prompt, your prompt is added\nto the AI models context. And it will\nthen use all of these things as input to\ngenerate a response. Like you have the\noptions, full body, upper low, split,\nlower impact strength. The input text\nthat is the prompts you've written as\nwell as the AI responses are called the\nchat history and the chat history gets\nincrementally added to the context of\nthe AI model as well. Now, if you start\noff giving the AI model more context,\nsuch as write a longer prompt as well as\nmaybe upload a handful of documents to\ntell it more about your workout schedule\nor your workout preferences, then all of\nthese files can be included into the AI\nmodel context and used to generate the\nresponse.\nNow if you continue the conversation to\nsay I like this about the first plan I\ndon't like that about the third plan\nwhat you say here is added to the AI\nmoral context and then this additional\nresponse in this type of brainstorming\nworkflow is further added to the context\nand this is why whenever you ask the AI\nto go back and forth and generate\nadditional answers it knows everything\nthat's been said so far in the\nconversational history.\nNow let me take this in a different\ndirection. Imagine you had asked it for\nthe workout plan like I shown here and\nit's given you a few workout plans. If\nyou were to instead give feedback on\nthese plans go in a totally different\ndirection and say now come with a\nworkout plan for my mom. Well, a lot of\nthe context AI model has including your\nschedule, your workout preferences, all\nthat isn't really relevant for your\nmother's workout plan. I guess unless\nthe two of you work out together. And so\nall this context would be distracting\nfor the AI system and might lead it to\ngenerate a worse answer. And in fact, it\ncan be hard to know whether this answer\nwas influenced by the previous context.\nThis is why if you're going to go off on\nthe unrelated topic, it's better to\nstart a new conversation so that you can\nempty out the context and start with\njust the new prompt or just the\ninformation that's helpful context for\nthe new question you want to answer.\nYou've seen how context is used to\nproduce high quality responses. thinking\nabout what's in the AI model's context\nand managing that so it has just the\nrelevant information and hopefully not\ntoo much irrelevant information although\nit can ignore a little bit of it that\nwill help you get better answers from\nyour AI. One way of handling lots of\ncontext is to allow the AI model access\nto your computer so that it can explore\nrelevant files and pull in relevant\nfiles into the AI models context only as\nneeded. Let's take a look in the next\nvideo at this very powerful technique.\nAI is moving beyond just chat\ninterfaces. You may have heard of\napplications like cloud co-work or\nMicrosoft copiloted co-work or Google\nanti-gravity.\nThese are applications that can with\nyour permission gather context\nagentically from your computer. Meaning\nthey can find and read files from your\ncomputer in order to give themselves the\ninformation they need to do a task. This\nis a new and exciting way of using AI to\naccomplish real work.\nLet me illustrate a common use case for\nthese AI desktop apps. If you have been\ndoing research on a topic and have a\nmessy folder with lots of PDF, research\nreports, images, and so on, you can with\none of these apps ask it to read through\nthe files in the folder and propose a\nnew organization for it based on what it\nfinds. In this example, the AI is able\nto look through this folder and apply a\nbunch of changes, renaming files, moving\nfiles around, creating subdirectories to\ncome up with much more sensible\norganization folders.\nLet me show you the process of how is\nactually done. You might start off\nasking it to organize a folder first to\nunderstand what's there\nand it can then automatically or\nagentically look at how different files\nare named.\nAfter it has figured out enough of what\nare the files in this folder, it then\ncomes up with an initial proposal for\nhow to reorganize it.\nAnd if you take a look at the initial\nproposal, you may be not fully satisfied\nand give it a little bit further\ninstructions to tell it what to do with\nthis data.\nAnd finally, it comes up with a refined\nproposal,\nwhich I'm happy with. So, I'm going to\nsay go ahead and carry on these\ninstructions.\nand then it reorganizes my folder to\nmake all the files much neater.\nHere's how AI desktop apps work. They\nare powered by an AI model. So, it comes\nwith the AI's pre-trained knowledge and\nit also has a set of tools like web\nsearch. So you can choose to carry out a\nweb search if it needs to do so to\naccomplish this task and has additional\ntools to work with files on your\ncomputer including the ability to search\nthrough files to read files to write\nfiles to move and rename files and so\non. So when asking a desktop app to do\nthings in your computer, a best practice\nworkflow would be for you to tell what\ntask you want done, such as organize the\nfiles in the folder. Let the AI system\npropose an action plan, but not yet take\naction. You can then review the plan,\ngive critique, maybe have it update this\nplan if needed, and only when you're\nsatisfied with it, then tell it to\nexecute the plan, which you can go ahead\nand do on your computer.\nOne neat aspect of these desktop apps is\nthat they can automatically explore\nfiles and manage contacts by reading\nfiles only when needed. If you are using\nan AI chat, then you have to decide in\nadvance what files to upload to give the\nAI model context. So for example, if you\nneed to write a schedule for filming,\nyou might upload a file that outlines\nyour filming procedures and based on\nthat it can generate a schedule. And\nnotice that you had to decide in advance\nwhat context to provide to the AI. But\nwith an AI desktop app, if you start the\napplication in say your document/forming\nfolder, then if you tell it write a\nschedule for filming this week, then the\nAI can decide to explore the files in\nthis folder\nso that it can see what files are there,\nload the relevant files and then on that\nbasis figure out what is a good foaming\nschedule. And in this example, surprise\nsurprise, that actually noticed that one\nof your crew members birthday is the\nweek of filming. And so maybe you'll\nfold in a celebration for your crew\nmate, Mia. One note on using these\ndesktop apps safely. Desktop apps can\nget access to and can edit or even\ndelete your files. And while mishaps of\ndeleted files are pretty rare, they have\nhappened to people. So I encourage you\nto choose the most relevant folder to\nrun the AI desktop app in. For example,\ninstead of running this in your whole\nfolder and giving this access to all\nyour files, maybe just give it access to\nthe subset of files it really needs for\na task in a certain folder.\nWhen the AI system makes a permission\nrequest, I would encourage you to\ncarefully review the permission request\nto make sure you know what it is reading\nand writing. And so I'll give AI access\nonly to the documents I wanted to know\nabout and let it write only to the files\nof places I wanted to. And when a AI\ndesktop app deletes a file, it often\ndoes not go to a recycle bin. So there\nmay not be a way to recover the file.\nAnd if it edits a file, it behaves a bit\ndifferently than if you were editing the\ndocument. And in particular, edited\nfiles usually don't have an edit\nhistory. And so it's not possible to go\nback if it made some change that you may\nnot like. So until you are very familiar\nwith these tools, I'll encourage you to\nlook carefully at the permissions\nrequest it makes to decide what you do\nand do not want the AI system to be\nallowed to do. AI desktop co-working\napps are a powerful tool you can use to\ngive an AI model the ability to discover\nrelevant context as well as to take\nactions such as reading, writing,\nmoving, renaming files.\nNow, we've talked about context a lot in\nthese last several videos, and using as\nmuch relevant context as possible\nenables your AI system to help you with\nwhat's called reasoning toss. by which I\nmean tasks where you want your AI to\nthink maybe for a long time to give you\nthe best possible answer. Let's go on to\nthe next video to see examples of\nreasoning with AI. The latest AI models\nhave very strong reasoning capabilities.\nWhat that means is they can think\nrigorously and at length about the task\nwhen given the right context. I find\nmyself more and more using AI as a\nreasoning engine. Let's take a look.\nHere's an example where thinking for a\nlong time can help get a better\nresponse. If you are car shopping and\nare considering trade-offs among\nmultiple cars, you might upload spec\nsheets for the cars, insurance plans,\ncolumn quotes, lots of documents, and\nthen ask what are the trade-offs for\neach car? read everything and think hard\nbefore answering. The AI model may then\nthink for quite a long time to read the\ndocumentation, maybe do some online\nsearch, then maybe think through what\nare the evaluation criteria that'll be\nright for you, and then generate reports\non the pros and cons of different cars.\nAnd just as doing research on what car\nto buy could involve gathering a lot of\ninformation and thinking at some length\nabout the pros and cons, AI can help you\nwith that.\nAs AI models get better, their ability\nto carry out long running tasks has\ngrown rapidly.\nThis is a study by an organization\nmeteor which plots for tasks at\ndifferent levels of difficulty as\nmeasured by how long it will take a\nhuman to do the task. That's the\nvertical axis. How well can AI do these\ntasks. So for example, a task like\nfinding a fact on the web may take a\nhuman just several seconds. Summarizing\na few vial texts may take a human an\nhour. Write a blog post couple hours.\naudit legal documents,\nexplore a complex cyber security\nvulnerability, may take a human many\nhours, and so on. And around 2024, 2025,\nmodels could start to do tasks that took\nseconds to many seconds to many minutes\nor tens of minutes. And in 2025, models\ncould start to have a decent success\nrate for doing tasks that to humans\nlonger and longer and longer to the\npoint where now AI models can do tasks\nthat can take humans many many hours to\ndo. Often AI model doesn't need 10 hours\nto do a task that takes a human 10 hours\nto do, but also takes a bit longer than\njust a few seconds. And this is what\nreasoning models has enabled for AI to\nthink at great length in order to do\nthese more complex tasks.\nIf you remember the how many hours in\nstraw example, that's a task that takes\na human just a few seconds to do. And\nseveral years ago, AI models used to\nsometimes get this wrong. And it was\nalso in that era, maybe 2033, 2034, that\nyou may have heard advice like tell the\nAI model to think step by step. And back\nthen, this was good advice, but this\nadvice is largely obsolete now. And I no\nlonger tell my AI model to think step by\nstep. Instead, I'm more likely to just\ntell it to think hard. And it knows what\nthat means, and that it should reason at\nlength, not necessarily step by step,\nbut in even more complex ways in order\nto accomplish a task successfully. And\nso rather than think about AI as\ncounting ours and strawberries or having\nbe told to think step by step, today you\ncan ask AI models much more complex\nquestions like what are the trade-offs\nfor each car or look at all this context\nand hope create a strip for a custom\npart.\nSo for these more complex tasks, I\nrecommend trying to use one of the more\nmodern models if you're able to access\none. And there may well be models even\nmore modern than the ones listed on the\nslide that could be available to you\nnow. So this is how AI reasoning or how\nAI thinking at length works. If you ask\nit to plan the fastest way to visit five\nlandmarks in Rome in one day, then it\nmight want to gather quite a lot of\ninformation to check map distances via\nweb search, estimate walking times,\nsearching hours, reorder the stops, and\nso on, and then generate your optimized\nitinerary.\nThe reasoning process may require the AI\nto think at length and repeatedly gather\nadditional information and then to think\nsome more until it's satisfied with the\nanswer.\nConceptually, you can think of reasoning\nas a process like this. Given your\nprompt in other input context, it will\nreason or think for a while using that\ncontext and then depending on where it\ngets to, maybe it'll decide it's done\nand then just give you the final answer.\nAlternatively, after thinking for a\nwhile, it may decide that it needs to\nuse a tool to gather more information,\nmaybe via web search or maybe by reading\nmore files from your computer if it's a\ndesktop app, and then use that\nadditional context to reason longer\nuntil it either again decides to use a\ntool to gather more information or\ndecides it's done. and so it can go\nthrough a few rounds of gathering more\ninformation and reasoning longer before\nit decides the answer is good enough to\npresent to you. If you are working on a\ncomplex toss and you want the model to\nthink at length, then one way to do so\nis to just tell the model to think. A\nfew of the interfaces among the popular\nAI model providers will have a thinking\noption. And if you select that option,\nthat's a cue to the model that you\nwanted to think longer.\nAlternatively, in your prompt, you can\nalso just tell it to think really hard\nabout this and it usually obey your\ninstructions or some people will use the\nphrase ultra think and that's another\nkey word that the models understand as a\ncue to, you know, think really hard\nabout it. And if you do so, an AI model\nwill sometimes think for many tens of\nseconds or even minutes or maybe\nsometimes even over 10 minutes in order\nto give you a good answer.\nEspecially with reasoning or with\nthinking models, I'd also encourage you\nto try giving the model hard tasks to\nsee what it can do for you. So if you\nare building a startup, maybe give it\nlots of context on what you're doing and\nmaybe tell it to design a topline plan\nfor a fourperson startup with limited\ncash. I encourage you to give the AI\nmodel real job task, real problems that\nyou want to think through or that you\nwant to solve. And as part of setting up\nfor success, try to give it all the\ncontext that a human expert would need\nin order to complete the task to make\nsure that the AI model has the\nsufficient information that really\nanyone would need to do the task\nsuccessfully.\nTo wrap up, if you want to use AI to\nreason at length about complex problems,\nI encourage you to use the best models\navailable. The best models are often\nbetter than models that are maybe 6 to\n12 months older. Remember to give it as\nmuch context as is needed to carry out\nthe task. It's okay to try giving a hard\ntask to see what it can do. So don't\njust give a trivial task. And lastly,\neither select thinking mode or just tell\nit in the prompt to think hard. We've\nspoken about when it's helpful to use\nthe cutting edge models, but it turns\nout even the most up-to-date models have\nsome common issues. One being the\ntendency to tell you whatever it thinks\nyou want to hear. This is called\npsychopy.\nLet's go on to the next video to see how\nto manage this behavior of AI models. AI\nmodels will act in ways to try to please\nyou because of the way they've been\ntrained. They have a strong bias to tell\nyou what you want to hear. This is\ncalled sick offensivey and avoiding it\nis a key prompting skill and involves\nprompting neutrally and keeping context\nfactual. Let's take a look at how to\navoid sick offensive because this will\nhelp you get much better answers from\nyour AI. If you consider the pros and\ncons of remote work versus in office\nwork and you ask it, don't you think\nremote work is better than office work?\nThis word in the question gives away\nwhat you're hoping the answer is and AI\nwill probably say yes remote work offers\nmany advantages. In contrast, if you\nwere to ask, is it true that office work\nis more productive? Then it will\nprobably agree with you that office work\nhas these strong benefits.\nSo depending on how you ask a question,\nit will tend to reinforce your own\npreferences, your own biases, and this\nmay not be the most helpful thing for\nyou to make objective fact-based\ndecisions.\nIn a study by the Washington Post on\nCHGB responses, it was much more likely\nto respond with phrases like that's\ncorrect, good point, you're on the right\ntrack, compared to not quite right,\nthat's not the case, or actually. And in\nfact, it tended to agree strongly about\n10 times more than it disagreed. And\nsome models from Chai GPD have said\nthings like, \"Dude, you just said\nsomething deep without even flinching.\nYou're a thousand% right.\" And while\nsome users appreciate AI agreeing with\nthem like this, I personally don't find\nit that useful to have AI just tend to\nagree with whatever I say even when I'm\nnot right. While leading AI model\ncompanies are working to reduce\nsyphancy, it is still a problem. Models\nare trained to be hopeful assistants\nusing human feedback and this reinforces\nsecrecy. For example, if you ask an AI\nmodel, I feel like it's better to be an\nintrovert. Don't you? If the AI model\nresponds, that's an interesting idea.\nHere's why I tend to agree. Then most\npeople are more likely to hit thumbs up\non the feedback button because, you\nknow, it's a nice answer. Makes you feel\ngood. But if AI were to say, \"Not\nnecessarily. Both types, introverts and\nextroverts, carry tradeoffs. you know,\npeople just don't feel as good about\nthat answer and so they're less likely\nto hit thumbs up or may even hit thumbs\ndown. Because of this type of feedback,\nAI, which has been trained to generate\nmore answers that leads to thumbs up,\npositive feedback, will learn to try\nsubtly to agree with people more often\nthan not. And this leads to psychopensy.\nAnd secrecency feels hopeful but\nactually degrades answer quality.\nSometimes maybe is easier to spot. If\nyou were to say, \"I'm really proud of\nthis essay. What do you think?\" Well,\nthey'll probably agree with you, but\nother times it's harder to detect\nsyreency. If you to say, \"Analyze this\ndata and find all the positive measures\nof performance this quarter.\nyou're subtly signaling that you're\nlooking for positive measures of the\ncompany's performance and so it's more\nlikely to say something like data\nclearly shows revenue growth strong\nintention improving margins and less\nlikely to point out problems\nto avoid circ\nneutral framing of your questions and\navoid giving any hints as to what is the\nanswer you want to hear so for example\nif I were to ask aren't carving taxes\nhave for small businesses are really\ntelling it what answer I'm hoping for.\nIn contrast, a more neutral prompt would\nbe to what extent if at all do carbon\ntaxes affect small businesses. If\nsomeone reads the question on the right,\nit's actually not clear what answer they\nonce is harder for AI model to be sick\nof antantic. Or if you were to ask, do\nyou agree that AI would create a lot of\njobs? I actually happen to agree. But if\nI'm actually doing research, I don't\nwant it to just tell me what I want to\nhear. Instead, you can ask, \"What does\ncurrent research say about AI's effect\non jobs?\" Or instead of asking, \"Does\nremote work reduce worker productivity?\"\nMaybe ask, \"How does productivity\ncompare between remote and inoff work?\"\nOne common pattern is to lay out two\noptions such as remote and in office\nwork and just ask it for pros and cons\nof the compare. but without hinting\nwhich of the two options I am hoping\nwill come out ahead.\nSo, don't you think that was the best\nvideo ever? Oh, what do you think of the\npros and cons of the videos you just\nsaw? Sometimes it's nice to be told what\nyou want to hear, but generally that\nwon't help you to do better work.\nSecond, despite a lot of attempts to\ncombat it, is still one of the pervasive\nissues with practical AI usage today.\nAnd so implementing the strategies from\nthis video, especially taking more\nneutral framing, will help you get more\nobjective and valuable feedback from AI\nmodels. Now, one of the most common\ntasks that people use AI for is writing.\nLet's go on to the next video to see how\nto work with AI models to help with your\nwriting.\nIn a study by OpenAI, writing accounted\nfor 24% of tasks that people ask chat\nGPD to do. This is the single largest\ngroup of task. Writing is really a kind\nof thinking. I find that when I'm\nwriting, I have to think. And so AI,\nwhich is really good at thinking and\nreasoning, can help you with this. But\njust asking AI to write for you often\nleads to AI slop or writing that sounds\nlike AI writing which somehow feels\ndifferent than human thoughtful writing.\nLet's take a look at some techniques for\ngetting AI to write effectively for you\nand how to take advantage of AI\nreasoning and avoid AI slop.\nWhat makes AI slop? Many people have\nnoticed that AI writing often includes\nthe M dash, that is the long dash, much\nmore than normal human writing. On the\nsocial media site Blue Sky, the use of\nthe N dash has been trending upward ever\nsince the release of GPD. A recent\nsurvey showed that 40% of US-based\nemployees have recently received work\nslop in the last month. And the term AI\nslop refers to content that's gened by\nAI and that looks good if you don't read\nit too carefully. So maybe every\nsentence read in isolation sounds like\nit's well written, but collectively the\ntext just lacks substance. It feels like\nit was written without much deep or\ncareful thought.\nOne of the property of AI slop is it\noften contains sentences like this.\nBut it does change everything. is kind\nof vague, empty sounding, but also\nsomehow overly important sounding text.\nAI's distinctive writing style comes\nfrom certain words and patterns that it\ntends to overuse. AI tends to use fewer\nunique words and over represents some\nwords and phrases. For example, AI tends\nto overuse the words nuanced and delve.\nAnd it tends to use list of three more\nthan most people will. And it tends to\nuse few nouns leading to phrases like\nthis is a robustly structured and highly\ninsightful paper. And this not X but Y\nis not just about speed is about\navailability.\nIn fact, I've been noticing on social\nmedia a lot more of this not X but Y\ntype of verbiage. often with X and Y\nboth vague things like it's not about\ninfrastructure it's about architecture\nand a lot of phrases like those are just\nvague and don't reflect a deeply\ninsightful point of view interestingly\nbecause humans are using AI models so\nmuch humans are themselves starting to\nsound more like AI and humans ever since\nchat GPD was released are using the word\ndelve more in podcast and talks. And\nthis is true both for spontaneous speech\nas well as for prepared speeches. So\nthis isn't just a case of people using\nAI to write scripts for them. It looks\nlike people are picking up speech\npatterns or texting patterns from\nspending a lot of time with AI.\nSo what's a better way to write that\navoids generating AI slop? One technique\nthat I think you find helpful is to use\nprogressive outlining in which you don't\nask AI to write the final text right\naway, but instead have it write an\noutline, refine the outline, and iterate\na few times before having it generate\nthe final text. For example, here's a\nprompt that says, I'm writing an article\nabout small AI teams moving faster than\nlarge teams that don't use AI. And I\njust want to acknowledge that this is\nnot a neutral prompt where I'm asking AI\nto help me decide if small AI teams do\nmove faster, but instead this is writing\nan article from a certain point of view.\nBut you can ask AI to research evidence\nfor and against this hypothesis.\nSo the AI model may search online and\nfind a handful of articles.\nNext, if you wanted to help you\nbrainstorm a handful of options for the\nstory outline, you can tell it to\nbrainstorm or to create three different\noutline options. Maybe tell it to\ninclude a counter argument section and\nalso provide or upload a handful of\nstories from AI team say that you work\nwith. With this input context, the AI\nmay give back a few different options\nfor what the outline for your article\nmight look like. Option one could be to\ntell the three stories and then conclude\nwith a thesis. Option two might explore\ndifferent patterns of how AI teams work\nand so on. Then following what you saw\nin how to brainstorm of AI, you might\ngive feedback on these options.\nAnd so you might say let's use option\none and keep all the stories but move\nthe thesis right after story one. And\nalso maybe say you want to add a\nhistorical analogy, Pixar treating the\nToy Story in the 90s, which is actually\na really inspiring story where Pixar at\nthat time a small company created Toy\nStory which was the first fully computer\nanimated featurelength film just using a\nreally small team. Based on your\nfeedback, the AI maybe gives you back a\nrevised outline. If you're satisfied\nwith this outline, then you can tell it\nto expand each heading, not even into\nthe final text, but just into bullet\npoints.\nAnd following that, you might decide to\ngive it more feedback on the bullet\npoints and iterate on the bullet points\nbefore finally having to generate the\ntext for the article.\nAnd it turns out that starting with an\noutline speeds up review. Let's say\nyou're working on a fun article about\nwhether a flying squirrel can carry a\ncoconut. In contrast to having AI first\nwork with you on the outline, then on\nbullet points, and then on the final\ntext, maybe you ask the AI to write the\nfinal text right away, just from the\nstart. If it writes a sentence like\nthis, you may be unhappy with a few\nwords and you can edit a few words, but\nchanging each word just changes one word\nand the rest of the paragraph stays the\nsame. In contrast, if you write an\noutline first and you're unhappy with\npart of the outline, then changing the\noutline causes an entire section, that's\na lot of words, of the final article to\nchange. So that's why editing the\noutline or iterating with the AI system\non the outline is very high leverage\nbecause you can figure out how to change\njust a few words of the outline and this\nwill result in an entire paragraph or\nentire section if you find the article\nchanging\nand this ends up being a much more\nefficient way for you to think through\nwhat you want to say in an article and\nadapt it to what you want it to be.\nWriting is one of the most common use\ncases of AI. According to open eyes data\nof the writing focused chats, about\ntwothirds involves starting from some\npre-existing text rather than starting\nfrom scratch or starting from a empty\nsheet of paper. When you already have\nsomething written up, it can be very\nhelpful that AI critique it for you and\nhelp you make it better. Let's take a\nlook at some techniques for this in the\nnext video.\nYou often have some idea where you've\nalready written some text about your\nidea, but want an AI model to help you\nedit and refine your text. I often show\nmy writing to AI to help me make it\nbetter. AI is great at this task and it\nalways is time to read your work,\nwhereas finding a human to help you out\nmight be trickier. You've already\nlearned how to avoid secrecy, but how do\nyou get the best quality editing and\ncritique? Let's take a look. One useful\ntechnique for editing with AI is to edit\nyour article piece by piece, such as one\nsentence at a time or one paragraph at a\ntime, rather than telling it to edit the\nentire article all at once. And to do a\nlittle brainstorming around each\nparagraph until I've nailed down one\nparagraph before going on to the next\none. For example, if someone's written\nthe sentence, the public thinks\nachieving AGI, artificial general\nintelligence means computers would be as\nsmart as people. You might ask AI to\nhelp brainstorm a few different ways to\nsay this. And so it may come up with a\npunchy way to say this, a visionary way\nto say this, a conversational way to say\nthis. And depending on your editing\ngoals, you could even iterate a little\nbit until you pick some version of\nrephrasing this part of the text that\nyou like.\nAfter you've nailed this down, then go\non to the second sentence or maybe the\nsecond paragraph and work on that little\nbit with the AI and then go on to the\nnext sentence and next paragraph and so\non until you get through your entire\narticle. And I find that working on one\npiece of a long article at a time makes\nfor a much more manageable workflow than\nif it were to change a lot of things all\nat the same time and you're reading this\nvery long edited article to figure out\nwhat has changed and what you like and\ndon't like. Now if you want highlevel\nmore holistic feedback about an entire\npiece you've written, it turns out AI\ncan hope for that too. But because of\npsychopensy, AI is often not a very good\nobjective critic. For example, if you\nwrote a sci-fi short story about an\nastronaut stepping out of his ship, and\nif you ask AI without further\ninstructions to critique it, there's a\ngood chance they'll tell you whatever\nyou did is fantastic work. In contrast,\nthere's a very helpful technique to\nguide AI in how to evaluate your work to\ngive you more helpful critical feedback\nand that is to give it a rubric and that\nmeans a grading criteria.\nFor example, you might write a rubric\nthat specifies what are the most\nimportant criteria by which to grade or\nto judge the work. So you may say that\ncharacters of the story is worth 25\npoints out of 100. The plot 25 points\nworld building writing craft and\nestablish a point system and then also\ndevelop detailed instructions on how to\nevaluate each of these criteria. So to\nevaluate characters maybe ask if every\nname character has a go and that's worth\n10 points. conflict between two\ncharacters goals and so on and giving AI\nvery explicit criteria on how to judge\nwork forces the AI to be more objective.\nOne thing to notice about these criteria\nis that each of them is very clearly and\nunambiguously defined. So for each text\neach of these criteria is either true or\nfalse, yes or no. And there's nothing in\nbetween. So either it's true that every\nname character has a goal or it's not\ntrue. And these completely objective\ncriteria\nforces AI to look at whatever you're\ngiving it through a objective well\nspecified standard with no ambiguity.\nAnd by the way, if you're not sure what\nrubric or what grading criteria to use,\nyou can brainstorm with AI to develop\nthat rubric. And AI is actually pretty\ngood at this, too.\nAfter you've written the rubric, you can\nthen provide the rubric as well as the\nstory and in your prompt ask the AI to\nbe objective. So critique to attach\nsci-fi store, assign a score per\ncategory, then sum the scores at the\nend. And by giving the AI these very\nclear instructions on what to do to sum\nthe scores at the end, it then hopefully\ngives a more objective assessment of\nyour story. If you want, you can also\nask it to then give you suggestions on\nhow to improve the story to do better on\nthis rubric and as a cause it to give\nyou more focus suggestions to improve it\nin the dimensions that you think matter\nthe most. In contrast, poorly written\nrubrics encourage suency and ambiguous\nor less objective thinking. For example,\nif you say, \"I would work on the sci-fi\nstory. Please score it out of 100.\" One\nof the things that's strange about this\nprompt is you first ask it the score out\nof 100. So that will tend to cause it to\nleap to the conclusion about what score\nand then only after that assign a score\nper category characters plot world\nbuilding and writing craft. And these\nare ambiguously defined categories\nbecause we're not told it how to score\ncharacters plot and so on. And so this\nwill tend to cause the AI model to first\ncome up with some score and then justify\nit rather than score it carefully\naccordingly rubric and then add it up to\nthen come with a more thoughtful score.\nAnd as you see in the practice lab to\ncome at the end of this module, this\ntype of rubric will tend to give higher\nscores than more objective rubrics.\nWe talked about using AI to critique and\nhelp give suggestions for improving your\nwork. It turns out that having AI\ncritique his own work or having one AI\nmodel critique a different AI model's\nwork can also help improve the results.\nFor example, if you ask Chat GPT to\nwrite a user manual for a fancy role\nplaying game, then it might generate a\nfile for you. And one thing you could do\nis provide a rubric for chat GPT to\ncritique his own work. But a neat\ntechnique is to find a different AI\nmodel, maybe Gemini, and give that a\ngrading rubric to have Gemini critique\nChat GP's work or vice versa. And it\nturns out that this type of crossmodel\nreview where you have one model review a\ndifferent model's output, it helps\nintegrate a bit of knowledge from the\ntwo different models and can result in\nslightly better results than if you were\nto ask one model to critique its own\nresults.\nI think using multiple models in this\ncontext might give only a slight boost\nin performance. Realistically, if you\nask Chachby to review his own results or\nask Gemini to review his own results, I\nthink that will actually do just fine.\nBut sometimes I find it reassuring if a\ntotally different model judges the\noutput of a different AI model. And I\nuse this technique only rarely myself,\nbut one thing I do do is frequently\nswitch between different AI models.\nIt turns out that AI models are\nadvancing rapidly and at different\nmoments in time, different models will\ndo better on different tasks. And so\nroutinely trying out different models\nwill help keep you sharp and keep\nholding your intuition about what model\nis best for what task. Air models have\nwhat's called jagged intelligence. If\nthe circle represents the task of what\npeople can do, maybe in a job or maybe\nin a personal context, it turns out that\nAI can do some things better than any\nperson like quickly read tons of web\npages or solve tricky math problems. But\nthere are also many tasks that AI\ndoesn't do as well as people. So there's\nsome tasks where AI does poorer than\nhuman and some where it does much better\nthan human and different AI models are\njagged in different ways. So the task\ndifferent AI models can do well are\ndifferent.\nMoreover, the marketplace of AI models\nis highly competitive. So CHB cloud\nGemini really the long list of model\nproviders are releasing better models\nall the time. And so the best model for\nyour tasks will likely change rapidly\nover time. And so I find that I'll often\ntake the same prompt and feed it to\nmultiple different models to see how\nthey compare. And this continuously\nholds my intuition about what models are\nbest for which of the tasks I care\nabout. So that takes us to almost the\nend of this module. I hope you've seen\nthat AI models are really useful for\nreasoning tasks as well as for\nbrainstorming, writing, editing, and\ncritiquing your work. These are powerful\nways of using AI as a thought partner\nthat I found very useful in my own work\nand that I'm confident you will too.\nLet's go on to the next video to explore\nthe practice lab for this module.\nI hope you practice the techniques we\ntalked about for using AI as a thought\npartner. In the upcoming practice lab,\nyou can explore once again side by side\nmore effective and less effective\nstrategies for both brainstorming and AI\ncritique. Let's take a look once again.\nWhen the lab pops up, you can dismiss\nthe tutorial.\nYou can also find the tutorial here as\nwell as the instructions I hope you\nfollow over here.\nAnd the buttons here, similar to the\nprevious lab, allow you to enter\ndifferent prompts to brainstorm. Here's\na prompt of less context. I need a\nworkout plan, 30 years old, want to get\nstronger. Whereas this is a more\ndetailed prompt with more helpful\ncontext. And so you can run it like so\nand see the difference in the outwards.\nSo let me focus on this more detailed\nexample on the right. It's given a\nprogram one which looks pretty\nreasonable and a program two which also\nlooks pretty reasonable. And we have\nalso a few suggested prompts for how you\nmight refine it. So maybe I'm drawn to\nthis.\nLet me do that.\nand have it continue the conversation.\nThen it will take this feedback on the\nprograms it has given you in order to\nrefine this results.\nBy the way, I actually use a workout\nprogram that AI had helped me to\ngenerate and I found this helpful in my\npersonal life of thinking through my\nregular workout plan. And so if you're a\nfan of working out, maybe you find some\nof the suggestions it makes useful as\nwell.\nLet's go back to the homepage. Here are\na couple more examples of brainstorming\nprompts with a simple prompt. I have\n$1,000 to invest. What should I do?\nVersus giving more context. And then\nwhat are some options I should consider?\nHope you try this out as well. As well\nas some examples of critiquing a sci-fi\nstory\nas well as this is example of an\nobjective rubric on the right. And I\nhope you bring this up and read through\nit to give yourself a sense of what a\nwell-written objective rubric looks like\nand compare that with what a more\nsubjective rubric looks like.\nAnd if you hit compare, you see the\ndifference in the quality of these\nreviews. And in fact, maybe not\nsurprisingly, with the less objective\nrubric, it gets a 100, but a more\nobjective rubric of 75 out of 100.\nIt also gives more helpful suggestions\nhow to improve your story.\nIn addition to critique and improving a\nsci-fi story, you can also take a look\nat these examples of helping you improve\na cover letter for applying for a job as\nwell as critique and help you to improve\na possible business plan. After checking\nout these five examples, please also try\nyour own prompts or try your own stories\nor cover letters or business plans or\nsomething else. And I encourage you to\nuse this interface to play with\ndifferent ways to brainstorm or to\ncritique your writing.\nSo that's it for this module. Please\nenjoy exploring the lab. Next, let's go\non to the final module where you see\napplications beyond text. Specifically,\nyou look at multimodal prompting in\nwhich you get your AI to also use images\nand audio and also look at building\napplications like games. It'll be a lot\nof fun and one of the examples we see\nwill have something to do with\nfireworks. So that I'll see you there.\nIn the previous two modules, we've not\nseen her AI generate text. But AI can\nproduce richer types of outputs as well\nlike images, videos and so on. We call\nthis multimodal outputs that is outputs\nwith multiple modalities. Prompting for\nmultimodal output is a bit different\nbecause multimodal interactions are\nslower and more costly. But these\ncapabilities will let you get a lot more\ndone with AI. Additionally, we'll also\nlook at multimodal inputs specifically\nif you want to show some images to your\nAI and have it reason about that. Let's\ndive in. AI models can generate images,\nvideos, voices, even music, code, and\nmore. You might have seen AI generate\nvarious fun, creative images. I want to\nshare with you one example of AI image\ngeneration that I really enjoyed.\nFor my daughter Nova's 7th birthday, I\nwanted a unique cake design, and she\nloves cats. So, we use AI image\ngeneration to explore different cake\ndesigns. The leftmost image here is an\nAI generated image created with a AI\ngeneration software called Nano Banana\nwhich is created by Google and this is a\npicture that my daughter really liked.\nShe wanted a cake that looks like this\ngenerated image.\nWe then took the image and showed it to\na baker and asked the baker to render\nthis picture into a real life 3D cake.\nAnd the picture on the right shows my\ndaughter cutting this birthday cake that\nshe loved. So in this case, image\ngeneration wound up being a\nbrainstorming tool to explore different\ncake designs until we found one that\nturned into a real life 3D cake that we\nall ate and liked.\nIn addition to generate images, I really\nenjoy playing with AI for video\ngeneration as well. Here's a fun video\ngenerated by our team of a man\nshrinking.\nThat video would previously have\nrequired maybe expensive special\neffects, but now AI can just generate\nit. AI can also generate voices. Here's\na voice clone of me reading out loud a\nletter from the batch, which is a weekly\nnewsletter that deep learning.ai AI\npublishers to cover what matters in AI.\nDear friends, here's the latest from\nthis week's issue of the badge. A\nbarrier to faster progress in generative\nAI is evaluations, evals, particularly\nof custom AI applications that generate\nfree form text.\nBy the way, I played audio of my voice\nclone to both of my parents. And it\nturns out one of my parents could tell\nit wasn't me and one of my parents could\nnot tell it was me or my voice clone.\nAnd to avoid me get into trouble, I will\nnot review which of my parents got it\nwrong. But I think voice clones are\ngetting really good. Lastly, AI can\ngenerate code. I mentioned my daughter\nloves cats. She also loves the color\nyellow. And so when her teacher\nmentioned that she wish kids in the\nclass could type or keyboard a little\nbit faster, I use AI to generate this\ntyping game where if my daughter hits\nthe right letter, then she sees this fun\nlittle animation of a cat being fed,\nwhich she loves. Using AI to write code\nhas made it easier and more accessible\nfor everyone, including you, if you\nwish, to write at least basic computer\nprograms. I'll say more about this later\nin this module. When you're working with\nAI model, there are many combinations\nyou can use of input and output types.\nFor example, you can input text and\nimages such as if you input an\ninspirational image like this, if you\nlike this Halloween costume and also the\nplan my Halloween costume. And in this\ncase, the AI model might output text\nthat says, \"Let me help you brainstorm a\nfew alien inspired costume ideas.\" Or\nyou can also upload music to an AI model\nand ask it help me plan my haunted\nhouse. And AI may take these things as\ninput and generate both text and a video\nof a haunted house design incorporating\nyour creepy sounds audio.\nAI models can use most of these input\ntypes relatively easily. Some are\nslightly more expensive or slightly more\ncostly to use as an input, but the\ndifferences aren't very significant.\nIn contrast, the time and cost of\ngenerating different types of output\nvary significantly.\nIn particular, some data types are much\nslower and much more costly to generate\nthan others. To give you a sense, text\ntends to be on the lower end in terms of\ntime or cost of generation. So, AI is\nvery efficient at generating text. In\nfact, modern AI has started with large\nlanguage models, sometimes abbreviated\nOM, but because they started with\nlanguage, a lot of them were really\nadapted to deal with text. and it's very\nefficient at that.\nGenerating speech tends to be a bit more\nexpensive and generating images even\nmore expensive and generating video much\nmuch much more expensive than images or\nany of the other modalities.\nAnd the further we go to the right of\nthis chart, the longer it takes or the\nmore time it takes to generate a single\noutput and also the more costly it is to\ndo so.\nImage generation has progressed\nsignificantly in the last few years.\nHere's a short video generated by\nImagen, which was in 2022 a\nstate-of-the-art model by Google, and it\nlooks pretty good, but still has some\nartificial looking artifacts. The lines\nweren't quite right on the back wall.\nThe dishes changed midwash. In contrast,\nmodern AI video generation looks much\nbetter and can also be synchronized\nautomatically with generated audio.\nvoice generation has also gotten much\nbetter. Here's what AI could do just a\nfew years ago.\nBut before you start tuning anything,\nyou need to define what success looks\nlike. This step is easy to overlook, but\nit's foundational.\nIt sounded a bit robotic, not that\nexpressive. In contrast, modern AI voice\ngeneration can sound much more\nexpressive and much more natural.\nBut before you start tuning anything,\nyou need to define what success looks\nlike. This step is easy to overlook, but\nit's foundational.\nIf you're generating multimodel data,\nsome of the techniques you learned\nearlier, such as giving the model enough\ncontext and maybe using the best model\navailable, those are relatively easy\ntechniques to apply as well to\nmultimodel generation. But some of the\nother techniques like generating\nmultiple options, you can still do that.\nBut if each option now takes many\nseconds or even a few minutes to\ngenerate, then this becomes harder to\napply because you end up having to wait\nfor a long time. Or if you want to\niterate through many designs, then that\ntoo becomes harder if each generation\ntakes a long time or as costly. But if\nyou have the patience to wait a little\nbit longer, then all of those techniques\nalso apply to generating audio, images,\nvideo, and so on. With great power comes\ngreat responsibilities. And AI\ntechnologies can be used for good or for\nharm. Take voice generation.\nIf you have recorded a podcast and you\nwant to make little fixes, that can be\nquite conveniently done today using AI\nvoice generation to just reynthesize one\nor two words that you may have flubbed.\nOr if you're building a video game and\nyou want to give characters lifelike\nvoices, more and more video game\ndesigners are using AI voice generation\nto do so. It does raise important\nquestions about the livelihoods of voice\nactors and I sympathize with all the\nvoice actors that are worried about AI\nvoice generation. At the same time, I\nthink it is also very valuable that AI\nvoice generation is making it easier for\na lot more people to build entertaining\nvideo games, including developers that\ndon't have access to the great voice\nactors. In contrast to these\napplications, there are also some that\nare clearly harmful. Unfortunately,\nthere's been a rise of scams where\nsomeone would use a AI voice clone to\npretend to be someone else, to maybe\npretend that someone's relative is an\nemergency and to ask to wire emergency\nfunds. The number of beneficial use\ncases of AI vastly outnumbers the number\nof harmful ones, but we still have work\nto do to combat the harmful\napplications. And I hope that each of us\nwill only use these techniques for\nbeneficial and responsible applications.\nSo, as you've seen, AI can now work with\nmuch more than text. It can work with\nimages, audio, video, code, and more.\nAnd most of prompt techniques you learn.\nso far will be helpful for handling\nmultimodal inputs and outputs. One\nespecially useful capability is giving\nAI images as input so they can see what\nyou're talking about. Let's go on to the\nnext video to see how to use images in\nyour prompts.\nProviding images of your prompt can\nenrich the context for the AI. pictures\nof something you want the AI to see,\npictures of handwritten text, really\nanything that might be hard to describe\nin words. This video will help you build\nintuition about what AI can see in\nimages. Here's a picture of me\nexplaining some concepts in AI in front\nof a whiteboard. My handrinting is not\nthat great and there are a number of\nmath concepts that I'm trying to\nillustrate on this whiteboard. If you\nupload this picture to an AI model and\nask what is this class about, it may\noutput something like this. He's\nteaching a convolutional neuronet\nnetwork. And the neat thing is my head\nis blocking the word convolutional, but\nit knows from this picture that I'm\nteaching about a specific AI technique\ncalled a convolutional neuronet network.\nand has extracted some facts about what\nI'm drawing and also has some good\nguesses about what I might ask students\nto do next. So is able to make a pretty\nsmart interpretation of this image. One\nweakness of AI models in terms of how\nthey look at images is they tend to look\nat the coarse image but may miss fine\ngrain details. So for example, if you\nupload this picture to an AI model and\nask what are these machines at my gym,\nit may confidently give an answer like\nthis, which turns out to be wrong. And\nthat's because a lot of gym machines, if\nyou look at them through a slightly\nblurry lens, they all look a little bit\nsimilar. And AI is not that good today\nat look at the fine details of images to\ndistinguish what really is a glute\nkickback machine or a hamstring curl\nmachine. In contrast, if you were to\nupload an image like this and ask it to\ncreate a sales ad for this item, it\nactually does pretty well because this\nis very visually distinct object. And so\nif you're looking at this even through a\nslightly blurry lens, yeah, you kind of\nsee this is a humansized hamster wheel\ntreadmill. When you upload an image, you\ncan also give it moderately complex\ninstructions on what to do with it. So\nuploading a receipt like this, you can\nask it, \"What's my portion of the bill?\nI had these items.\" And in this case, it\ngets it correct. AI's ability to read\ntext like this is not bad. It does make\nmistakes. So, I wouldn't trust it for\nhigh stakes applications, but if you\nwanted to take a quick look and if\nyou're willing to spend a few seconds to\ndouble check the result, then it could\ndo decently well.\nAn AI turns out to be pretty good at\neven reading handwritten text. If you\nupload a picture like this to AI and ask\nit to transcribe it, it does a pretty\ndecent job. Feel free if you want to try\nreading this cursive handwriting\nyourself to see if you can outperform\nthe AI.\nAnd so if you upload an image like this\nand write a prompt like build an archive\nof a family's history based on these\nhandwritten letters, I wouldn't trust it\nto read everything completely\naccurately. But it might take a\nreasonable stab at this task.\nRather than uploading a single image to\nAI, sometimes you can upload many\nimages. For example, if you just had a\nbrainstorming session and had some notes\nyou taken as well as pictures of post-it\nnotes and whiteboards, you can upload\npictures and notes to an AI model and\nask it to summarize the ideas from\ntoday's brainstorming meeting. And\nagain, it will probably do a decent job\ninterpreting these images to come up\nwith some summary. Probably not perfect,\nso it's worth double-checking his\noutput, but this could help you\naccelerate coming up with notes from\ntoday's meeting.\nTo recap, AI models can read basic text\nin images. Visual understanding,\nhowever, may miss details in the image\nbecause it tends to see the image in a\npretty coarse way. And you can also use\nmany images when needed to give the AI\nmore context.\nA picture is worth a thousand words. So\nadding an image to a prong can often be\nthe fastest way to get the AI model the\nbest context. Like taking pictures of a\nbrainstorming exercise or digitizing\nyour grandmother's handwritten recipe\nbook. Of course, AI models can also\ngenerate images. This is an interesting\ncapability because it works somewhat\ndifferently to how AI models generate\ntext. Let's go on to the next video to\nsee what are some fun images you can\ngenerate.\nGenerating images of AI has made my life\nmore fun. For example, I've used image\ngeneration for my kids birthday parties\nor to make fun illustrations. Generating\nhigh quality AI images is a skill you\ncan learn. And understanding how AI\nimage generators were trained will help\nyou to control image generation to get\nbetter outputs. Let's take a look. One\nneat application of AI image generation\nis if you input an image and ask it to\nedit it. This is a childhood picture of\nme on the right, my younger brother on\nthe left, and one of our childhood\nfriends in the middle. And this is a old\nsomewhat faded image. And if you upload\nthis to AI model and ask it to remove\nthe glare and the rough texture and to\nmake it a more natural aspect ratio, it\ncan produce something like this, which\nlooks like a nicely restored photo. This\nparticular image restoration was done\nwith Google's nano banana model.\nIf you're not sure how to prompt an AI\nimage generation system, you can\nactually ask a textbased AI model to\nhelp you write a prompt. For example, if\nyou ask it, generate a prompt for an\nimage of a cat secretly running a coffee\nshop at night. An AI text model may\nwrite a prompt like this. Notice that\nhere it specifies a setting, specifies\ndetails of the character, and specifies\na mood or a style.\nAnd if you don't like any of these\ndetails, you can modify them to your\nliking. And a prompt like this might\ngenerate this cute picture on the right.\nPeople skilled in the visual arts have a\ncertain language for describing images.\nFor example, a picture like this with\nthis look is cinematic.\nThis is a watercolor image. This is a\ncyberpunk image. And this is an anime\nimage. And I find that art buffs and art\nhistory buffs excel at image prompting\nbecause they understand the language of\nimages and can describe what they want\nusing more precise language than those\nof us that don't know this language and\ncan't quite find the right words to\ndescribe the look that we want.\nSo if you want to become really expert\nat generating images, it could be worth\nreading up or studying a little bit\nabout the language of images to\nunderstand how to accurately describe\ndifferent images. And in fact, one way\nto do so would be to upload images to an\nAI model and ask the AI how it would\ndescribe those images.\nAnd this could whole new instincts on\nwhat types of words can be used to\ndescribe what types of images. It turns\nout image generation uses a very\ndifferent technology than text\ngeneration. When AI is generating text,\nit produces the output piece by piece or\nit generates a few characters at a time.\nIn contrast, when generating an image,\nit doesn't generate the image a few\npixels at a time. It generates the\nentire image all at once. Specifically\nduring training, that is when an AI\nmodel is looking at pictures maybe found\nonline. In order to learn what images\nlook like, it will typically look at\ncaptions or descriptions of images like\na small potted plant on a wooden table.\nAnd it'll learn to start from image that\nlooks like pure noise. This is just a\ngrid of random pixel values. And it will\nthen learn to sequentially remove or\nsubtract noise from the image to go from\npure noise on the right to a slightly\nblurry picture of a potted plant to a\nless blurry one to a less blurry one to\nfinally to a sharp picture of a potted\nplant. And that's what AI model tries to\nrepeatedly practice doing during\ntraining. A model that does this is\ncalled a diffusion model. Then when you\ncome in and when you write a prompt,\nmaybe create an image of a potted plant\npond in the table, it then goes through\nthis process starting from a pure noise\nimage and then gradually tries to remove\nnoise from it to come up with that final\nimage.\nAnd the key is how it learns to subtract\nnoise to maybe reveal the image that the\nperson might have had in mind. Diffusion\nmodels do generate random outputs and\nthey can also make certain types of\nerrors. If you repeatedly ask AI model\nto generate a positive plant, different\ntimes you run the algorithm might result\nin different images like the one shown\nhere.\nBut many people have also observed that\nthe diffusion model tends to generate\nweird looking hands, often with more or\nfewer than five fingers.\nAnd it can often output gobbled text\nlike happy birthday is very badly\nmisspelled here. And it can also lead to\ninconsistent characters. So if you ask\nit to generate a cartoon, the\ncharacter's hair has changed between the\ntwo frames of this cartoon.\nFortunately, modern AI models have\nbecome much better at addressing these\nproblems.\nFor example, modern models like Nano\nBanana can allow you to upload a number\nof research papers and ask it to\ngenerate an infographic and it'll do a\ndecent job with text that looks mostly\nplausible.\nOr if you ask it to generate cartoon,\nthe more modern models can generate\nfairly consistent characters. Meaning\nnow this character, as you can see,\nlooks very similar from frame to frame\nof this cartoon.\nAnd the text also looks pretty decent.\nCompared to text generation, image\ngeneration can be slow and costly. For\nexample, if you're generating just a\nshort paragraph of text, many eye models\ncould do that in just seconds and it may\ncost less than a scent. Of course,\ngenerating long paragraphs of text or if\nyou ask it to think for a long time,\nthat can cost more. And also, AI models\nwere generated word by word or maybe a\nfew characters at a time and you can\ninterrupt it or have it stop early if\nyou want. In contrast, generating a\nsingle image might take tens of seconds\nand cost many sense\nand it generates the image all at once\nand there often isn't an option to stop\nearly because image generation is much\nmore expensive. That's why our ability\nto iterate with images is usually more\nlimited. And if you're generating\nvideos, then it gets even harder.\nEven though some things are more\nexpensive to generate, the good news is\nthe cost of generating virtually\nanything with AI is trending downward.\nSo a year from now, it will be less\nexpensive for you to create art for your\nhome or graphics for a family member's\nbirthday card compared to today. Beyond\ngenerating images, AI can also help you\ncreate fun games and websites without\nhaving to write any code yourself. This\nis a more advanced capability and it's\nso easy to get stuff that doesn't work\nor for noviceses to get stuck trying to\nbuild more complex applications. But I\nwant to just give you a taste of how\nusing AI to build custom software works.\nWhile it's not that easy, it's also\ncertainly much easier than it was just\nmonths ago and quite likely easier than\nyou might think. Let's go on to the next\nvideo to see how to use AI to create\nyour own mini game or website.\nBuilding computer games and websites\nused to be something that only\nprofessional developers were able to do.\nBut the ability to do this is being\ndemocratized. By writing text prompt,\nyou too will be able to build basic\nsoftware applications and websites. This\ndoes take some skill and I don't want to\nmake it sound trivial, but in just this\none video you learn some of the basics.\nThis is a very exciting capability and I\nencourage you to explore it. I've seen\nmany people who are not software\nengineers have a lot of fun with this\nand even with just one prompt is often\npossible to get a cool little game or\napp. Let's see some examples\nusing this prompt. Build a game where\nthe user has to place obstacles and they\ngo and it creates a simulation of what\nyou design. Claude created this game\nwhich you can play and is, you know,\nactually pretty cool.\nSo just a short simple prompt like that,\na leading AI model can create a\nreasonably interesting game.\nOne example that you see in the practice\nlab is prompting an AI model to generate\na fireworks display. Here's a prompt\nand this creates this app which is\nactually pretty fun to play with.\nIf you're trying to write a prompt to\ntell AI to build a simple app for you,\nhere are some building blocks you might\nconsider including in your prompt. First\nis to specify your goal for what you\nwant to create. Next, specify what are\nthe inputs, what the users need to input\ninto the system. And then lastly, what\nare the outputs or what the app shows\nback to the user. For example, in this\nprompt, we are telling it that the goal\nis to generate a fun fireworks\nsimulator.\nThe input is I want to do a click on the\nscreen and the output is see a colorful\ndisplay of fireworks.\nSophisticated developers will use AI to\nbuild much more complex applications\nthan what I'm showing here. But just to\ncontinue with the simpler examples, to\nbuild a game like this, you might ask it\nto create a fun game where the user has\nto place obstacles and a go and it\ncreates a simulation of what you design.\nIn addition to entertaining games, you\ncan also make more useful and functional\napps to help you save time or make your\nlife a bit easier. For example, you can\ncreate a work timer called a pomodoro\ntimer. This is a type of timer that\npeople use to time their work or\nstudying with 25 minutes of work\ninterference with a 5minut break. Or\nmaybe you can create a bill calculator.\nHere you can input the bill and the\nnumber of friends you need to split the\nbill with and the app can tell you how\nmuch each person should pay. or maybe\nbuild an outfit picker app that helps\nyou decide what to wear based on the\nweather. Each of these apps could be fun\nor useful in everyday life. And they're\nalso good starting points if you're new\nto making apps with AI because they're\nquite simple. In particular, they each\nhave a specific well- definfined task.\nThey also don't need any additional\nfiles that need to be uploaded or\noutside information. And they're also\nsomething you can open up, use for a\nshort period of time, and enclose.\nIf you're curious to try using AI this\nway yourself, I encourage you to\nexperiment starting with building simple\napps. It turns out some ideas are easier\nto create than others. For example, a\nsimple platformer game would be\nrelatively easier to build, or a quiz to\npractice French words. In contrast, a\nmultiplayer game played over the\ninternet that would be harder and much\nmore complex or live French practice\nwith AI feedback would also be harder to\nbuild. It takes a while to hone\nintuition about what is easy for AI to\nbuild and what is hard for AI to build.\nIf you're not sure, I encourage you to\njust try it out. And the worst thing\nthat could happen is it doesn't work and\nyou start to hone your intuitions about\nwhat's hard. But if you're just getting\nstarted trying to use AI to build apps,\nI encourage you to start with simple\nideas like build a simple game and see\nif you can get AI to do that. When we\nget to the practice lab in this module,\nyou'll be able to try a few examples\nthat you can build with just a single\nprompt. If you want to dive deeper into\nusing AI to build software, I encourage\nyou to take the course build of Andrew\noffered by deep learning.ai.\nIn addition to writing code to build\ngames and websites, AI can also help you\nto analyze data, which it can do by\nwriting code to do that data analysis.\nAgain, you don't need to write any code\nyourself. Just tell AI what you want and\nor try to write code to do it for you\nand this can lead to hopeful insights in\nyour work and personal life. In the next\nvideo, let's take a look at using AI for\ndata analysis. If you have data from\nyour personal health records, like maybe\nif you use one of the apps that track\nyour heart rate or running time, or if\nyou have data of sales records from your\ncompany, or really most other types of\ndata, like what you might store in a\nspreadsheet table, AI can be pretty good\nat writing code to analyze this data for\nyou and to try to extract useful\ninsights. Let's look at some examples\nthat I hope will inspire you to try out\nthis capability.\nIf you have ramming tracker data that\nyou can download into a spreadsheet, you\nmight try uploading that spreadsheet to\nAI model and asking it how are my pace\nand distance progressing\nand an AI system might spend some time\nto write some code to analyze it and\nmaybe generate a plot for you and also\npotentially provide some insights.\nOr if you run a small business and you\nhave a spreadsheet of sales data, you\ncan upload that data and ask the AI\nmodel, what if you tell me about this\nmonth's sales?\nThe AI could analyze for a little while\nand it might actually write some code to\ndo things like compute the monthly\nrevenue or to create a graph and try to\nshow you whatever insights it finds.\nI find that AI data analysis often isn't\nas sophisticated as a really really good\nhuman data scientist. But for pulling\nout basic insights and doing so\nefficiently, an AI model can be pretty\ngood. How do AI models write and execute\ncode? It turns out that the way it works\nunder the hood, the ability to write and\nexecute or to run code is like any other\ntool that the AI model might use. You've\nseen how AI model might have tools to\ncarry out web search or to read and\nwrite files and so on. And some AI\nmodels also have a tool to run a\ncomputer program or to run code. And\nthis capability will often be used when\ndata is present or when some sort of\ncalculation or some sort of plotting of\na graph is needed. In that case, the AI\ncould generate a bunch of code and then\nuse this tool to run the code in order\nto generate a result for the user. We\nsaw previously how a reasoning model\nmight input a user prompt, reason for a\nwhile using the context, and\noccasionally decide whether or not it\nneeds to use a web search tool to get\nmore context.\nIn this case, it has the option of not\njust using a web search tool, but also a\nrun code or code execution tool in which\nthe AI can write the basic computer\nprogram to analyze data, compute\naverages, plot a graph, or whatever is\nneeded in order to get to the final\nanswer.\nLet's look at a concrete example. If\nyou're running a bubble tea shop, you\nmay want to look at your sales trends\nover time to answer questions like, \"Did\nyour new drinks sell well compared to\nexisting drinks.\" If you have sales data\nhandy, AI can help you with this. You\ncan attach your sales data as a file and\nwrite a prompt like this one. Which\ndrinks had the biggest changes in sales?\nGraph it. The AI will then go through an\nagentic process to analyze the data and\ngraph it. Inspect the data, then\ncalculate the monthly changes in sales,\nand do things like analyze intelligently\nyour data to help you get useful\ninsights. For example, it may say things\nlike, I'm noticing some clear patterns.\nMost drinks are flat, but four stand\nout. Now, graph those. So, not just\ngraphing all the drinks, but identifying\nand potentially focusing on the most\ninteresting ones. Then, it might\ngenerate a graph like this one. Time is\non the horizontal axis and number of\nsales on the vertical axis with each\ndrink drafted in its own color. So the\nstrawberry matcha took off in spring,\nmango green tea and strawberry lemonade\nin summer and your new coconut milk tea\ndid well in fall. It also added these\ncolorful highlights to make these trends\nstand out. So in just a few minutes you\ncan get a highly useful graph like this.\nFrom this graph, you might conclude that\nyour spring promotion for strawberry\nmatcher was strong and maybe you want to\ntry that again next year. You can keep\ngoing and iterate on this graph and ask\nfor different insights or changes to the\ngraph or give more context about your\nbusiness to the AI to give it a better\nshot at finding more useful insights.\nBut let me show you a more longinking\nexample. You could take your sales\nanalysis further and ask for a more\ncomprehensive graphic. For example, you\nmight want to create a year in review\ngraphic for your business to present to\nyour team. This involves analyzing all\nyour data in a few different ways to\nfind out most interesting to share. So\nyou might write a prompt like create a\none slide year and review graphic for a\nbubble tea shop. Analyze the data\ncarefully for insights and again attach\nyour data file. Using the words\ncarefully in your prompt may trigger the\nAI to think for several minutes to\ncomplete this analysis. And it might go\nthrough an agentic thinking process like\nyou saw on the previous slide. And it\nwill likely write and run code to\ncalculate things like revenue and items\nsold. And in the end, you might get a\ngraphic like this with a lot of\ninteresting insights like brown sugar\nand classic being the most ordered\ndrinks and most customers choosing a\nlarge drink. This graphic also has a\ncreative bubble tea color scheme. You\nprobably want to double check these\nfigures to make sure they match your\nexpectations since the AI sometimes can\nhallucinate, but you can get a good\nanalysis in a relatively short period of\ntime just by prompting. that's fairly\nlikely to be accurate since AI\ncalculated these numbers by writing and\nrunning code. I encourage you to try\nthis type of analysis yourself if you\nhave data like sales data or personal\ndata you're interested in getting\ninsights from. If you're using an AI\nmodel that can run code, when would it\nchoose to do so? We've seen that for\nsome questions it can use this\npre-trained knowledge.\nIf you're asking for a question that can\nbe answered using common knowledge on\nthe internet, what it already knows may\nwell be good enough. If you're asking a\nspecific question that is real time,\nthen web search enable AI will help you\nget a better answer. You may want to ask\nher to use a deep researcher if you have\na more complex question that may require\nmultiple related searches such as to\ncome up with a complete plan for your\nHalloween house. And for queries that\nrequire calculation or drafting, those\nare the types of questions where it's\nmost likely to write and run some code\nin order to carry out that task\nprecisely for you.\nOne of the most powerful things that AI\ncan do now is write and run code in\norder to carry the task for you. If you\nhave some data that you want to get\ninsights from, I encourage you to try\nexploring it with an AI model. AI isn't\nalways reliable and it's better at\nsimple analysis than really complex\nones. So consider double-checking this\nconclusions, but it can be much faster\nthan having to do the analysis yourself.\nan Excel or Google Sheets and it has on\nmany occasions helped me discover useful\ninsights in my data. To put your skills\ninto practice, we have one last practice\nlab and an optional final project for\nyou. Let's go on to the next video to\nsee what these are. I'm excited for you\nto test out building games and\napplications using AI. This should give\nyou a taste for what's possible to build\nwith AI.\nWhen you get into the lab, you can read\nthrough these instructions shown here,\nbut I'm just going to dismiss this for\nnow. And I'm just going to click the\nprompt for a fireworks show. I encourage\nyou to read through this prompt\ncarefully so that you understand what it\ntakes to build an app like this. And I'm\njust going to hit run.\nAnd this short prompt will build this\napplication.\nI did leave a lot of things unspecified,\nbut it looks like my AI has made\nreasonable decisions. And if I click my\nmouse, it launches pretty fireworks.\nIf I don't want to launch them by\nmyself, the auto show runs my automatic\nfireworks show. And then here's the\ngrand finale.\nSo, pretty cool. One neat feature is\nthat you can actually share the apps you\nbuilt. So, I'm going to click share.\nJust copy the link to my clipboard. And\nI can open up a new tab with this URL.\nAnd this actually launches the Fireworks\napp in my web browser.\nSo, if you share this URL with a friend,\nthey'll be able to run the Fireworks app\nthat you just created.\nOne of the things I most enjoy about\nbuilding simple applications is to share\nit with friends. So, I encourage you to\nmaybe take what we built and consider\nsharing that with friends and see what\nthey think.\nFeel free to take the firework show\nprompt modified and see if you can get a\ndifferent result that could be even more\nto your liking. Second example, let me\nclick on the color palette prompt.\nIf you are designing a website,\nsomething you may have to do is choose a\ncolor palette for the website. And so\nwith a prompt like this,\nyou can build a color palette picker.\nHere's a base color in RGB values.\nSo this is actually roughly the color of\nmy shirt\nand you can choose complimentary\nanalogous\nand so on color palettes that you know\nthey actually go pretty well together.\nHope you have fun playing with this.\nSo, I hope you try out all four of these\nbuilt-in prompts and additionally try\nusing your own prompt. For example, here\nI'm going to ask it to build a fast card\napp to help me practice basic French\nvocabulary. This process may take a few\nseconds and when it's done, this is what\nyou might get. Where what is yes?\nNot too shabby.\nAnd so on. Looks like I got that right.\nIt looks like it g me all the answers\nhere on the right. But if I don't want\nthat, I can go back and modify the\nprompt to redesign my flash card app.\nPlease have fun with this. And I find it\ninspiring that with just a one English\nprompt. You can build a web page to\nbuild applications like these.\nThe final project is to build a simple\napp from research on a topic that\ninterests you. We're going to go through\nthree steps. First brainstorm a research\nquestion, then run research, and then to\nbuild an app. I encourage you to read\nthrough these instructions, but I'm just\ngoing to dismiss them for this demo. The\nfirst step is to brainstorm a research\nquestion on either a topic you want to\nexplore or a decision you want to make,\nlike researching health supplements or\nbuying a car or things to do your career\nor some fun things with astronomy. Let\nme pick the careers one. Given this\nprompt or some other one that you may\nchoose, I would encourage you to then\ngive the AI more context.\nSo I miss exploring different careers\nabout me, my situation.\nLet's say I'm at university studying\ndeep learning, want to work on office\nand I encourage you to write more than I\nam here in this demo walk through.\nBut let me send it to the AI. And here\nit will help me to brainstorm a few\nspecific research questions such as what\ncareers let me spend a lot of time\ncollaborating and what does it look like\ndayto-day? So hopefully you pick\nsomething that's relevant to your life\nand this will help you brainstorm a\nrange of different questions that\nhopefully will be interesting to you.\nAnd we also have a AI mentor to give you\nsome feedback on how the brainstorming\nprocess is going.\nSo in this brainstorming workflow,\nI would read through these questions and\ngive feedback. So let's say I'm most\ninterested in question one, but like the\nspecificity of question three. I'm also\nwant to make sure I have time to go to\nmy job. So this type of feedback, the AI\nnow has additional context about what\nmake an interesting research question\nfor you. And so it refineses it to\nslightly better options.\nLet's take one more turn. Based on these\nthree questions,\nmy feedback is and I also want to make\nsure to consider nonprofit work.\nAnd so by going a few rounds with\nlooking at the brainstorm research\nquestions and giving feedback, we've now\nrefined it to a handful of research\nquestions.\nSo if you're using an OM, you can\nactually go for multiple rounds, ask you\nto refine the questions, combine them.\nBut let me just end this part of the\nexercise for now and say I like this\nquestion the best. And I'm just going to\npick this question to go on to set two\nnow that I've formulated a research\nquestion.\nAnd what we just did was go through a\nbrainstorming exercise. There's an\nexample of how you may iterate with your\nAI to brainstorm any of other possible\ntopics as well, not just to identify a\nresearch question to work on.\nNow that I've picked my research\nquestion, let me give it a little bit\nmore context. So, I also want to know\nabout salaries and typical compensation.\nAnd I wanted to use these sources forums\nfor personal experience and so on. I\nencourage you to add more personal\ncontext and add more sources that I'm\ndoing in this quick run through. But\nlet's just have it carry out this\nresearch.\nAnd let's have it go ahead and do so.\nThe mentor gives some feedback on my\nprompts.\nAnd after running for a while here, my\nweb search enabled AI gives a pretty\ndecent set of results with lots of\ncitations.\nFor the third and final step, after\nyou've gotten your research report,\nlet's go build an app.\nWe can build a quiz to test my knowledge\nbased on the reports or does a mini game\nor build an infographic.\nAll three of these are actually pretty\nfun, but I'm going to choose the quiz\noption. So, here I'm going to build a\nfive question multiplechoice quiz.\nAnd the research report that we had\ncreated from step two is uploaded here\nas an attachment to give it more\ncontext.\nAnd to keep the results more predictable\nin our website, this prompt is grayed\nout. So, it's not editable,\nbut you can always copy the report into\na third-party AI system such as Chadly\nor Gemini or Claude to experiment more.\nAnd I encourage you to play of all of\nthese. All of these are actually pretty\nfun. But let me just generate my quiz\napp.\nSo, that will again take a little bit of\ntime. And so, here's the app. Let me\npick that. Yep, looks like I got that\nright.\nAnd so on. And same as before, you can\ncopy this link using the share button\nand share this app with your friends and\nsee if they like it. So, I hope you\nenjoy going through the iterative\nbrainstorming workflow, which is a very\nuseful skill to have when using AI and\nthen provide enough context for it to do\nresearch for you and then based on or\ninspired by that research, build one of\nthe fun apps that we just showed you.\nand maybe even share it with some\nfriends. I hope you enjoy exploring the\nlab and the optional final project and I\nhope you'll share your final project\nwith others as well. And again, if\nyou're interested in learning more, I\nencourage you to take the Build of\nAndrew course. Once you're done, I'll\nsee you in one final video.\nCongratulations on making it to the end\nof this course. With all the techniques\nyou've learned, you're now ready to be\nan AI power user. I hope you find a lot\nof places in your personal life and work\nwhere these skills will benefit you.\nlike using AI to help brainstorm or use\ndeep researcher where you need\nthoroughly researched reports or use AI\nto help you with writing and even\ngenerate multimodel outputs and code\nand even as you're doing all this AI\nmodels will keep on getting better. So\nplease keep trying new models and give\nAI hard tasks and provide highquality\ncontext to help you to keep honing your\nintuitions about what AI can and cannot\ndo. I'm confident you get a lot out of\nthis incredibly powerful technology.\nThank you for sticking with me to this\npoint and I hope you use these powers to\nhelp yourself, your friends, your\ncommunity and go make the world a better\nplace for yourself and others.",
  "transcript_chars": 125973,
  "ingested_at": "2026-05-21T19:19:25.114538+00:00",
  "source": "retry-no-transcript",
  "yt_meta": {
    "view_count": 31269,
    "like_count": 1918,
    "channel_id": "UCcIXc5mJsHVYTZR1maL5l9w"
  }
}