{
  "video_id": "FMpgh2QfX_g",
  "channel_slug": "deeplearningai",
  "channel_handle": "DeepLearningAI",
  "title": "AI Dev 26 x SF | Anush Elangovan: Impact of AI on Software",
  "duration_seconds": 864,
  "url": "https://www.youtube.com/watch?v=FMpgh2QfX_g",
  "upload_date": "20260519",
  "transcript": "Good morning, San Francisco. I was told\nto yell in here, so I hope you can hear\nme. Okay, good. At the back. Okay. Um,\nthis is great to be here in the city uh\nwhere every wave of software innovation\nhas either started or has come to reckon\nitself here.\nStarting from mainframes,\nclient server. I know these all like you\nknow some of you know these words but\nmainframe client server the web mobile\ncloud and now AI. But I will tell you\nthis AI is coming at you so much faster\nthan each of those other transitions.\nAnd I've been able to live through those\ntransitions. this is happening in the\norder of months and weeks compared to\nthings that took um decades or or years.\nSo before I get started on talking about\nthe impact of AI on software, I do want\nto spend a minute uh acknowledging\nAndrew uh thanking I want to thank him\nfor something he did about like 12 years\nago. uh we were a small startup working\non gesture recognition and and u you\nknow uh this was pre-AII right before\nthe the term AI was popular and he\nawarded us the uh best AI startup award\nin 2014 and that team now runs most of\nAMD's software uh AI software team so\njust a legendary like visionary that can\nsee things 10 12 years ahead so it's\ngreat to be here with partnering with\ndeep learning AI and you know um looking\nforward to learning from him on what is\nlike 10 12 years from now. So\nbefore I get started, I do want to um\nyou know preface this with you know this\nthere there's a there's a slide that I\ncannot unsee or something that came up\non social uh and it's attributed there.\nYou can go check it out. It was called\nthe K-shaped future of software\nengineering and I completely agree with\nwhat this assessment looks like and it\nis a uh a way to frame your thought on\nAI and how AI can be consumed and\nprocessed. So what we will look for in\nthis and this I think is backed by a a a\npaper that came out of Harvard that did\nthis analysis. Um what rises in the top\narm of the K-shaped future is systems\nlevel thinking uh judgment and taste,\nintuition, uh problem framing, how how's\nyour harness set up? So all of these are\nlike higher level constructs and that is\nwhat is going to be accelerating. So\nyou'll have a a section of your team\nthat is just supercharged with that,\nright? and they'll be generating\nmonstrous amounts of code but they have\nthe framing of the problem. They know\nthey know how how to solve the problem\nand they're thinking at it still the\nfirst principles thinking um is still a\na big part of it. What falls is whether\nyou know how to code in a specific\nlanguage, right? Whether you know how to\nuh format your uh Python or syntax. And\nyes, that's still important when you if\nif you need to go look at it, but\nincreasingly um you do not need to know\nany of that, right? And that is just\nintermediate language for your AI agents\nto consume. Um so it it is very uh you\nknow um important to acknowledge this\ntransition and we do see this in the\nindustry. We are seeing, you know, teams\nthat just suddenly become 100x more\nproductive. And there are people that\nthe wingspan of what they're working on\nmoves from like higher levels of the\nstack, lower levels of the stack. And so\nhow you measure the outcome is not\nnecessarily uh necessarily lines of\ncode. It is outcomes for your business.\nRight? So can we can we unlock a\nspecific um use case? Can we do\nsomething that we just thought is not\npossible? And we have been um at AMD\ntrying to do this quite a bit. And I'll\nI'll hit a few samples of what we\nthought was not possible like 6 months\nago and now it's just like an overnight\nproject that someone sits down, you\nknow, with a cup of coffee and and they\nhave it done, right? Okay.\nUm when when that K divergence happens,\nwhat is your mode? And I keep repeating\nthis. Um I I think speed is the mode and\nyour I'll call it intent velocity is\nwhat you want to measure. So you you\nknow how fast can you take an idea that\nyou're thinking about to actual\nproduction and everything else in\nbetween is just it's just a means to get\nto the end right um so intent velocity\njust internalize that it's like okay I\nneed to go do this and how fast can I do\nthat right and if you uplevel that it\nall comes back to how fast you move\nbecause what was cool and new last week\nisn't that today, right? And so what you\nwant to do is try to adapt and and be\nready for change. Um, so you know,\nwinners operate in parallel. This is\nanother one that I I think you want to\ninternalize. It is not necessarily about\nsingular throughput. It's about how much\nyour um your agent span\nu you know uh can cover, right? And so,\nfor example, at night, I have like four\nor six agents that are crunching away\nand they're doing something in parallel.\nSo, you want these things to just be\nrunning autonomously at night or when\nwhen when I'm doing this keynote, you\nknow, my agents are crunching away. This\nis something that you couldn't do in the\npast. And you want to set the intent and\nthen unlock them to start running,\nright? Um, and this is going to be a\nflywheel. the faster you move, the\nfaster you're going to go up that K\ncurve and the divergence is going to be\nstark, right? And and as leaders who\nhave large uh organizations, I think the\nresponsibility will lie on us to see how\nyou reduce the divergence but not\nnecessarily slow down the upper arm of\nthe um you know um uh output of the uh\nupper arm.\nSo um let's talk about rockom uh for a\nsecond. This is just uh an AMD\nportfolio. AMD has traditionally had a\nvery very strong uh hardware portfolio.\nWe we have a pervasive hardware strategy\nstarting from laptops, desktops,\nworkstations, um edge AI, cloud and data\ncenter. But what we wanted to do next\nwas try to make sure that we brought in\nthe software layer and AMD has always\nhad a open ethos and open software\nmindset. Um and so that was our founding\nlike you know pillar in terms of how we\napproach software and then the second\none was how we brought up abstractions\nright like you didn't have to program\ndeep in the details of GPUs but the cool\nthing is what's happening now with AI we\nlayer all of this with AI and we're now\nable to unlock software acceleration\nthat has not been possible in the past\num in the order of days and I'm going to\ngo through a few of the examples that I\nhave here um that we'll we'll talk\nabout. But the acceleration is just\nhuge. And what I would take away for\neach of your efforts is how you apply at\napply AI at every layer of the stack. Uh\nand luckily for AMD, it's fully open\nsource. So all of the frontier models\nunderstand our stack end to end. And so\nit's really really easy to program on\nAMD with AI.\nSo I'm going to uh touch on three\ndifferent um you know project three or\nfour projects. Uh I I'm not going to\nkeep it too long and I can chat about\nthis after the keynote. Uh so we did\nsomething called geek which is one of\nthe the big pain points we had which was\nperformance on you know uh the software\nand how we achieve it. So now we have an\nagent loop that just understands what\nthe customer is running and it\nautonomously just optimizes the software\nnon-stop. Um and this has shown us like\nimmense amount of uh outcome right like\nand customers actually see the value\nthey are able to serve tokens much\nfaster and and this is uh this is what I\nmean by you want to aim for your like\noutcome we wanted faster time to\nperformance and geek was one of those\nprojects that provided us that\num we did another one again this is uh\nuh it's like Rosetta um which is like we\ncan we can translate machine ISA from\none GPU to another on the fly. Years\nago, this would this was literally like\nI know of implementations that took four\nfive years and\nI know like 200 300 people to implement\nbecause each ISO you have to like each\ninstruction you have to make sure it's\nmapped correctly. It's it tests well. It\nwas a real software engineering project.\nThe first prototype of this took about\nlike 48 hours and I think it's about\nlike a few billion tokens of uh um I I\nforget it was cloud cloud code or Opus\n4.6 and now we are shipping it in\nproduction. You can actually take an old\num older generation hardware, run it on\nnewer generation hardware or vice versa.\nLike we're doing code design of hardware\nthat's coming out in 2 3 years running\non current generation GPUs seamlessly at\nnative speed. And this would have been\njust like something that we we we can't\nwe couldn't even plan right like\nsoftware engineering our thing is like\nwe were limited by oh why would you do\nsomething like that? You're you're doing\nsomething too hard, right? Now there is\nno such thing as too hard. You just have\nto have an intent saying I'm going to\nattempt that too hard thing. Don't tell\nme not what not to do and just go\nattempt it. Right? You you have to frame\nthe problem. You have to guide it. You\nhave to understand where are the\npitfalls so you don't get lost. But the\nyou you are now limited by your ability\nto like think forward and not by what\nyou're like churning out in terms of\nlines of code.\nUm, another one that we we we've been\nlike, you know, beat up on a little bit\nwas like Llama CPP and and how it\nintegrates with AMD GPUs. There was a\nlittle hard problem to make sure that\nyou could seamlessly move tensors\nbetween CPU, GPU, NPU. And again, we did\na new runtime that has like zero cost\noverhead to move tensors seamlessly and\npartition data and and compute across\nCPU, GPU, NPU. the first integration is\ngoing to come out on Llama CPP so that\nyou can now use your laptops and use all\nof the silicon that's available. I do\nbelieve that um the ability to consume\nAI is going to shift to wherever there\nis compute AI will consume it and and we\nwant to enable that on your laptops too.\nSo that that was another uh uh good um\noutcome that we had. Um we also built a\ntokenizer. One of the uh you know\nbottlenecks of uh LLM inferencing is the\nfirst part of it which is token which is\nthe tokenizer. And there had been a area\nthat not a lot of people had focused. Uh\nit was one guy and 200,000 lines of code\ngenerated and we have like the world's\nfastest tokenizer now. Again it's all\nopen source. It's it's more tokens for\nthe future tokens to learn from. And\nthat's why it's a flywheel effect\nbecause all what we created now you have\na reference your your models will learn\nfrom it and know that hey there's a\ntokenizer that works fast, right? So the\nnext time you ask for building\nsomething, it's going to have this in\nits pre-training data and know how to\nexecute and move faster. And that's why\nit's a compounding flywheel effect of um\nof impact.\nI will um I I will you know uh finish\nwith a few thoughts. Uh one is you know\num obviously the K shape is real. Uh\nit's moving really fast. In December I\nthought um agents were um prompts in a\ncron job. I I don't know how many got\nthat but but it literally is like you\nknow it's just a yeah sure it's a it's\nan agent. Um but now there is no way I I\nthink I can do my job um without having\nthe agents running continuously and they\nare like off the bugs that we have on\nthose models and those uh those projects\nthere are agents that are monitoring\nthem for issues that are filed PRs that\nare filed and they're going to respond\nto that automatically and I'm not even\neven involved. All I've told the agent\nis that make sure every bug is\nrecreated. Every bug is fixed with\nanother is fixed and then there's\nanother test that validates all of it\nand then commit it if the CI is green.\nAnd those projects that I showed you are\nactively being, you know, um monitored\nand and worked on. So the change is\nreal, the change is fast. And what I\nwould recommend is or what what based on\nwhat I've seen is one it it is changing\nour trade and and the skills that we\nhave to um you know to operate. And so\nit's uh the analogy I'll use is like\noxygen masks. First put one on yourself\nand then help the person next to you. Uh\nmake sure you bring them along too. But\nyou do need to put that oxygen mask on\nand start like you know just really\nleaning into agentic AI because uh the\nfuture is coming very fast. Thank you.",
  "transcript_chars": 12053,
  "ingested_at": "2026-05-21T19:18:22.071835+00:00",
  "source": "retry-no-transcript",
  "yt_meta": {
    "view_count": 364,
    "like_count": 7,
    "channel_id": "UCcIXc5mJsHVYTZR1maL5l9w"
  }
}