{
  "video_id": "g8um2AEf5ZA",
  "channel_slug": "deeplearningai",
  "channel_handle": "DeepLearningAI",
  "title": "AI Dev 26 x SF: Andrew Ng: The Future of Software Engineering",
  "duration_seconds": 1161,
  "url": "https://www.youtube.com/watch?v=g8um2AEf5ZA",
  "upload_date": "20260520",
  "transcript": "It's really nice to see all of you here.\nIn addition to the AI prompting for\neveryone course announcement, there's\none other significant announcement I\nwant to make in this presentation today,\nand you'll be the first people in the\nworld to hear about this. And I'll do\nthat toward the end of this\npresentation.\nWhat I want to do today is share with\nyou what I'm seeing with the future of\nsoftware engineering as it relates to AI\ncoding, and touch on context up, and a\nnew thing that I'll announce a little\nbit to share with you more I think how\nwe can hopefully help developers\nnavigate to this future.\nA year ago at the first AI dev,\nI presented a slide like this, where I\ntalked about thinking of software as\nassembling many building blocks to\ncreate these wonderful things.\nSo, it turns out that I often think of\nsoftware, tools, frameworks, APIs,\nas akin to LEGO bricks. And if all I\nhave is one type of white colored LEGO\nbricks, I can build some stuff, but not\nthat interesting.\nIf you mix in some black LEGO bricks,\nyou can build more interesting things.\nMix in some blue ones, mix in some red\nand yellow ones, and so on. And as the\nnumber of disparate building blocks you\nhave grows, the way you can combine them\nto create interesting software grows\nexponentially or grows combinatorially.\nOver the last year, the rise of AI\ncoding agents has made it far more\neffective for all of us to quickly put\nthese building blocks together to create\namazingly cool stuff. And this includes\nboth AI building blocks, like large\nlanguage models you can prompt, rag\nagentic workflows, and many more, as\nwell as non-AI building blocks like UI\ncomponents, persistence of databases,\nidentity or off managers, and so on. And\nit turns out because of AI coding,\nthese building blocks are proliferating\nat a speed like we've never seen. There\nare more and more building blocks every\nsingle day, both AI and non-AI, for all\nof us to use.\nAnd this is wonderful, especially when\nAI coding lets us\nget AI agents to assemble these building\nblocks for us.\nNow, when I speak to different teams in\nstartups, academia, also very large\ncompanies, one question that's often\ndiscussed and debated is how much should\nwe use AI coding agents? And there's\nactually one popular poll question,\nwhich is I'm not going to ask this\nquestion, but I see many people ask how\nmuch of your coding is AI versus human?\nAnd, you know, there'll be some teams\nthat'll say, \"Oh, yes, my team does 80%\nof our coding by AI.\" And, you know,\nthat's pretty good. Um, my own coding is\nnot 80% AI, it's actually pretty much\n100% AI. And I want to share with you\nwhy I think this even last gap makes a\nbig difference. Oh, by the way, in case\nthere's anyone here that's 0% AI, it's\nokay too, but I encourage you to\nincrease that percentage, right?\nUm, but what I see among teams is if 80%\nof your code is AI, then in terms of\ntime spent, it actually looks like this,\nright? AI coding is time, human to\nreview that 2% of code just takes\nforever. Whereas a team that's 100% AI\ncoding just gets a lot more done in a\nshort time.\nSo,\nI find that this is true not just for\nwriting code, but also reviewing code.\nAnd I know that a lot of teams that\nstill have AI write code, but have a\nhuman review every single line of code.\nBut for myself and a lot of my teams, we\nfind that if I got to review the code,\nthen I become the bottleneck, and it\nalso doesn't work. And I'm not saying,\nyou know, if you're if you're working\nfor NASA building and you need to write\n50 lines of code for a spaceship,\ngo ahead, write it by hand. It's totally\nfine. So, this is not a religion where\nwe should never be allowed to write code\nby hand. But, what I'm seeing is um many\nof the frontier teams are trending\ntoward 100% almost 100% of code written\nby AI. And this creates a real\nacceleration that getting only 80% of\nthe way there feels very different from\num both for writing and reviewing code.\nNow,\nas coding is done by AI, there are\nfascinating second-order effects.\nI think um last July in the batch um\nnewsletter, I wrote about the product\nmanagement bottleneck. And was this\nobservation that um when I'm trying to\nbuild stuff,\nyou know, we'll often have an idea,\nbuild the prototype, that's software\nengineering, and then go to users, get\nfeedback, that's PM or product\nmanagement work, use those ideas to come\nand refine the software.\nSo, AI coding has made that 10 or 100\ntimes faster.\nAnd so, this means that deciding what to\nbuild or the product management work\nbecomes the new bottleneck rather the\nactual building. So, what I started to\nsee, like starting a year and a half\nago, was this very strange trend where\nwe used to have product manager to\nengineering ratios of, you know, one PM\nwould keep eight engineers busy, right?\nThese ratios typically vary in a small\nbut one to eight PM to engineer, one to\nseven common in many companies. But, the\nratio started to trend toward, you know,\none to two was really weird, and then\none to one, which is even more weird.\nAnd what I'm seeing is rather than one\nengineer and one PM, the only thing that\ncan move even faster is you take those\ntwo people and collapse them into a\nsingle human. And I find that engineers\nwho shape products or product managers\nthat can code can move really fast. And\nthat speed is a huge advantage.\nBut not everything can be done by a\nsingle person team. And so I find that\nfor more and more AI natives, the trend\nis that very small teams, they're all\nkind of generalists where the engineers\ncan do some PM work, the PMs can do some\nengineering work, and everyone kind of\nknows almost everything, and the small\nteam can move really fast.\nI know that in some Silicon Valley\ncircles, it's trendy to talk about the\njob apocalypse. This idea that AI would\ntake all jobs almost all jobs and the\nrewriting in this phase.\nUm, frankly, I'm not really seeing that.\nAnd in fact, with this profile of\nengineer, the future AI engineer has a\nvery bright future.\nWhen my teams are hiring for AI\nengineers, these are these are what we\nlook for, right? First, have to be able\nto use coding agents uh effectively, be\nit Cloud Code, Gemini's AI, or Codex,\nopen code, or something else. Um, and\nI find that having a robust knowledge of\nthe amazing building blocks out there,\nwhich is a challenge because there's so\nmany of them. So many and they change so\nquickly. That is a challenge, but\nknowing these building blocks lets you\nassemble things very quickly. I'll come\nback to the building blocks a few times\nin this presentation.\nAnd then lastly, I find that people with\ngeneralist skills, be it basic product\nmanagement or other things as well, can\nbuild much faster. So there's a\ncombination of skills on the ability to\nhow to build things as well as the\nability to know what to build. And um,\nI'm seeing massive\nunmet demand for a lot more engineers\nwith anywhere near this mix of skills.\nFrankly, even my teams, we can't find\nenough of these people.\nUm,\nand this is why\num, and and maybe going beyond the\nproduct management bottleneck, I want to\ntake that one step further, which is\nwhen software engineering speeds up 10\nor 100 times X, everything else seems\nslow in comparison. So in addition to\nthe product management bottleneck, I'm\nseeing that um\nthere's often a design bottleneck cuz\nyou design something and she implements\nit or even better, maybe the designer\njust implements it in code without going\nthrough a design tool like, you know,\nFigma, and that even speeds it up. And\none thing that used to be painful, but\nis even more painful now, is the legal\ncompliance bottleneck, right? If you\nspend 3 months writing code and legal\ntakes a month to sign off, it's okay,\nbut if you take a day writing code,\nyou're going to wait a month for legal,\nit's like, boy.\nUm and even marketing bottleneck. For\nmany of our teams, we write code and\nship products so fast that marketing has\na hard time scrambling to keep up with\nwhat these engineers are doing to figure\nout how to tell people about it. So, I'm\nseeing design bottlenecks, legal\ncompliance bottlenecks, marketing\nbottlenecks, sales bottlenecks, a whole\nwhole bunch of other things.\nAnd this is why I find it increasingly\nsmall teams of people that includes\nengineers, but also generalists, can\nmove really fast. And a piece of good\nnews is, it turns out that an engineer\nthat's not a marketer,\nbut that knows how to prompt AI well,\nmaybe they're not an amazing marketer,\nAI actually makes really weird marketing\ndecisions, but, you know, they could do\nsome basic marketing work. And in fact,\nif a team needs software, product,\ndesign, legal, and marketing, and it's a\nteam of two,\nby definition, these two people have to\nhave some skills in in all of these\nfunctional areas, and I find that smart\nAI native teams, they can use AI to help\nwith some of these other functions can\nmove really fast. Um not advising anyone\nto be a lawyer, but but frankly, I I I\noften have AI, you know, deal with legal\nstuff as a first draft and then take it\nto a real lawyer to sign off before I,\nright, before I uh launch something, but\nI find this AI native team moves really\nfast. Something I don't have time to\ntalk about is how to structure multiple\nAI native teams to work together because\nagain, not everything can be done by a\n10-person team, but if you have a\n100-person team of many AI native teams\nwith limited communication and clear API\nboundaries, that also helps to scale\nbeyond a single team.\nBut if we come back to the work that\nneeds to be done to create software,\nwriting software code is just a small\npiece of it that's the third thing\nrapidly, but I just don't see the AI job\napocalypse happening anytime soon. Um, I\nfind that the business media often has\nto be more accurate than the popular\npress because the popular press can tell\nexciting stories, but the business media\nis paid to get it right, and I've been\nencouraged that over the last month or\ntwo a lot of the business media has been\ngetting the story right that this AI job\napocalypse now just missing Tesla\nvehicle's job apocalypse, job apocalypse\nbeing delayed, a Federal Reserve Bank\nsmall highlight, sorry if I can't read\nit, um, highlighted few CEOs need to\nreduce their workforce or study by the\nFederal Reserve Bank of Philadelphia.\nSo, I'm actually um, and and to be\nclear, I know a lot of people are\nfeeling job insecurity at all levels of\ninsincerity. Um, I think we need to do\nsomething to address that, and I know\nmany people have been laid off because\nof over hiring for the pandemic or\nhiring from zero interest rates period.\nSo, I'm not saying the job situation is\nis perfect, but there's a lot of it\nthat's not that pretty, but the job\napocalypse is possible to talk in some\ncircles. I just don't see it happening,\num, which is why I'm eager to keep on\ninvesting in all of us becoming better\nat building these things.\nNow,\none of the hardest things to do,\nexciting most important hardest is as\nwe're becoming better and better\nbuilders each of us, how to master these\nbuilding blocks. I want to share with\nyou a framework of parallel skill\ndevelopment.\nHere's what I mean.\nWe talk a lot about how our coding\nagents are becoming more capable, and\nthis includes agent skills like the\nAnthropic agent skills and so forth, but\nmore generally our coding agents are\njust getting better and better.\nAnd then\nthere's the education or training world\nI've been a part of for a long time that\nfocuses on building human skills.\nBut looking at the future, I think we\nincreasingly need parallel skill\ndevelopment where as our coding agents\nbecome more capable,\nour people with people need the\ncomplimentary skills to help us to drive\nthe coding agents in the appropriate way\nso that we can partner together to build\nnew things.\nSo, I want to briefly touch on two\nthings in the rest of this presentation.\nOne is Context Hub, which I talked about\nbefore, but I'll spend a couple of\nminutes on. And the second, which is the\nannouncement I'll make toward the end,\nis Code Dream.\nSo, um Context Hub. It turns out\nbuilding blocks are proliferating so\nfast that most of our coding agents,\nwhich have a knowledge cutoff date,\nyou're far in the distant past,\nsometimes a year old, sometimes a month\nold, will often hallucinate or use\ndeprecated APIs or not know about the\nlatest tools you want to use. So, it\nturns out even today, if you ask Auto\nCode um to call the OpenAI API, it will\nuse the chat completions API, which is\nolder and deprecated, and call older\nmodels because most of the training data\non the internet uses this older API,\neven though the newer APIs have been out\nfor a long time.\nUm\nContext Hub is a tool that some\ncolleagues, uh Vivek Prasad, uh Sanyam\nHota, and I built um in order to give\nyour AI agent most up-to-date\ndocumentation. And it turns out that um\nif you run Context Hub\nwith uh Auto Code, it will generate code\nwith the newer responses API. Um maybe\njust to show you really briefly, I don't\nknow, chop search OpenAI.\nUm so, this returns to this of packages.\nAnd then chop get\nthis actually right OpenAI/chat.\nUm\nI did some more.\nAnd so, this will fetch the latest\ndocumentation on open AI that knows\nabout\num\nright knows about GPT 5.5 has the knows\nabout the responses API and so on. And\nthis is a tool that you are welcome to\nuse of but it's actually you know built\nfor your coding agent to use and so this\ngenerates\num like all right a 600 line markdown\nfile right kind of long for a human to\nread or you can you're welcome to read\nit to give you all the context it needs\nto\nmake these and other API calls\ncorrectly. So I pray many of you too\nright? I've had my coding agents\nhallucinate API calls\ndefinitely you know open AI Gemini calls\nmake bad calls to various database\nservices there's just so many tools our\ncoding agents don't know about but\nfeeding in this latest context I think\ncan be very helpful.\nBut beyond that I want to share with you\na project I've been working on to help\nwith human skill development. Um for a\nlong time you know my team's been\nworking on a variety of online courses\nand please keep on taking online courses\nI think it's an important part of how we\ncan all stay up to date with the\nbuilding blocks and other tools and\nwe've been exploring one other\nalternative experience that um I'm\nannouncing here for the first time today\nand that's the project we've been\nworking on called Code Dream. Um\nthe Wi-Fi here is not great so I'm going\nto see if the live demo works. It well\nmay not. But let's see if this\nOkay.\nWelcome to Code Dream. This is not a\ncourse it's a conversation where I want\nto show you how to use a coding agent\ntogether with tool called context hub to\nwrite code using up-to-date API\ndocumentation.\nI say this is a conversation because I'd\nlike you to imagine we're on a video\nconference call\nand you can interrupt me at any time\nusing that little microphone button down\nbelow.\nThe Code GG has created a voice agent\nthat has information you need to try to\nanswer your questions as I will.\nSo, feel free to stop me if there's\nanything that you want to chat about. Or\nif you like, you can just lean back and\nwatch the presentation.\nHere's the idea, all right.\nWhat are we going to learn about in this\nconversation?\nWe're going to learn how to use Context\nHub to keep coding agents up to date\nwith the latest information. Then, we'll\nbuild a joke generator app together.\nSo, this is not a slide, right? I just\nclicked into this and let's let's tell a\ndad joke about\nbad Wi-Fi.\nWhy did the computer show later where\nwas bad connection? Okay, not bad, could\nbe worse.\nAnd the other thing about this being\nJavaScript is um you can also copy paste\nfrom these slides. So, for example, this\nis a slide with instructions on how to\ninstall Context Hub and how to prompt\nopen code and so on. So, I should copy\non hitting command C\nand then go into the Codoji build\nenvironment. And um this is a a terminal\nin my browser\nwhere I can\nyou know, paste that command. I just\ncopy pasted from my slide to run npm\ninstall chob. And then um I can do\nthings like\nstart up open code, which is open source\nversion of a coding agent kind of like\nyou know, alternatives to cloud code or\nI zero I open I codex. And then I don't\nknow, write some\nPython uh\nwhatever.\nRight. And so, this is a um learning\npractice environment that uh\nlets you practice using modern coding\nagents with tools like Context Hub.\nUm\nand so\nwe have been working\non this Codoji build environment for\nquite some time.\nAnd um I'm excited about exploring this\nconcept where in s- in lieu of taking a\ncourse, you can come on to a video\nconversation and have a conversation\nwith me or an AI version of me to try to\ngain new building blocks and code in\nthis more modern AI coding way.\nAnd I am\nhappy to announce that Code Realm is\navailable as of today in preview.\n[applause]\nSo, please check it out. Um, and the\nfirst experience will show you how to\nuse Code Assist app and give us\nfeedback. We're eager to keep on working\nto make this better.\nSo, just to wrap up, um,\ncoding agents are allowing us to\nassemble building blocks quickly to\ncreate software, and that acceleration\nis real. But, many of the downstream\nimplications about whether new\nbottlenecks are are still being worked\nout, and I'm eager to help our entire\ncommunity keep moving forward because\nthe job openings are there. The job\napocalypse is not being nearly as bad as\npeople are hyping it up to be, and I\nthink it's important that we keep on\nlearning these building blocks and\nlearning these skills.\nAnd for parallel skill development,\ncoding agents are\nworking well, but give them access to\nlatest building blocks, given the right\ncontext, um, context help. will help.\nAnd for human skill development, I'm\nexcited also to share with you the Code\nRealm conversations, and all of you in\nthis room are the very first people in\nthe world that I've shared this with uh\noutside outside our team. Um, and I hope\nyou will check this out and that'll help\nyou to master some new building blocks\nand master some new skills.\nSo, with that, um,\nreally excited to be here. Thank you for\nall coming, and thank you all very much.\n[applause]",
  "transcript_chars": 18127,
  "ingested_at": "2026-05-21T19:15:59.114789+00:00",
  "source": "retry-no-transcript",
  "yt_meta": {
    "view_count": 1390,
    "like_count": 49,
    "channel_id": "UCcIXc5mJsHVYTZR1maL5l9w"
  }
}