{
  "video_id": "PSWUr5E_OKY",
  "channel_slug": "dataindependent",
  "channel_handle": "dataindependent",
  "title": "Become An AI Engineer in 2025 | The 6 Step Roadmap",
  "duration_seconds": 961.0,
  "url": "https://www.youtube.com/watch?v=PSWUr5E_OKY",
  "upload_date": "",
  "transcript": "if you haven't noticed Ai and language\nmodels are changing the skills needed to\nbe a successful engineer the people who\nare learning these new skills are\ngetting jobs that pay up to\n$435,000 a year they're creating apps\nusing language models that generate\nmillions of dollars at 90% margins with\nno employees they're building features\nthat took months in minutes and they're\nusing AI engineering but there's a\nproblem these new AI models are\nnon-deterministic and English is now a\nprogramming language what and frankly\nthe industry is still figuring out the\nbest practices of using them however\nwe're now entering a golden era for AI\nengineers and the ones that add to their\nskill sets will be ahead of the curve\nthere are six new AI engineering skills\nthat are actually solving these problems\nand in my opinion every engineer is\ngoing to need to learn them to be\nsuccessful for those that are new to the\nchannel my name is Greg I led a team of\ndata scientists at Salesforce driving\nproduct analytics for billion dooll\nproducts I've taught over 10,000 velers\nhow to build their first AI applications\nand now I'm in the absolutely privileged\nposition to be making money doing what I\nlove building and teaching with others\nthis video is all about telling myself\nwhat I wish I would have known about\nworking with llms and some of these\nareas I completely overlooked I'm going\nto outline the road map of skills that\nyou need to be a successful AI engineer\nalong the way I'm going to highlight\nspecific jobs that these skills are\nrequired for I'll show you which experts\nare demonstrating these skills at a\nprofessional level and I'll give you a\nlist of resources to help you go further\nand you invest in these skills you're\ngoing to go off and build way more than\nyou ever thought that you could if\nyou're remotely serious about getting\ninto AI engineering then you're going to\nwant to watch the full thing and by the\nway every link and mention in this video\nis going to be Consolidated into a\none-pager if you want to get access to\nit links in the description the first\nstop on the road map of AI engineering\nis working with models now there's four\npopular models this is going to be open\naai anthropic meta and Google and each\none has their own personality check out\nhow arjin Canan CEO of resides puts it\nwe found that the open AI models tend to\nbe the best analysts the uh anthropic\nmodels tend to be the best writers and\nthe uh Gemini models broadly tend to be\nthe best detectives if you want to think\nabout it that way right like they're\ngreat at finding needles and hay stacks\nwhen looking at large contexts the\nimportant part to understand is that\nmost usage is going to be from text to\ntext however there is text to speech\ntext to image video to image really any\nmodality that you may want and this is\nthe multimodal part of working with\nthese models the next piece you're going\nto want to understand is how to work\nwith these models apis now you're going\nto get very used to seeing open. chat\ncompletions doc create but then there's\nalso a lot of different model types GPT\n40 01 whisper Sona highq Opus there's\nreally a ton to learn you're also going\nto want to understand streaming batch\nprocessing prompt caching and assistance\nnow for each one of these on the\ndocument that I talked about beforehand\nwe're going to have definitions and\nresources to go learn more about these\nnext up is going to be local versus open\nsource now if you don't trust open AI\nanthropic Google or meta well then you\ncan go run your own model on your own\ninfrastructure and popular examples for\nthese are going to be with open router\nor o llama now this is up to you now in\nmy opinion what you're not going to have\nto worry about yet is with regards to\nmodel training fine-tuning or model\nrouting now all these are very important\nbut I don't think that they're Mission\ncritical for an AI engineer right when\nthey're starting that's more of an\noptimization skill that they can learn\nlater on down the road for resources if\nyou want to learn more check out my pre\nprevious AI show andt tell videos and we\nhave different guests that talk about\ntheir experiences with working with\nmodels now another one I would highly\nsuggest is the Lex Freedman podcast with\nthe cursor team this is a bit more\nadvanced look into Model Management but\nI thought it was really really good and\nfor people who to follow I think Justine\nTony is absolutely fabulous for this now\nheads up again once more this video is\ngoing to have a ton of links and\nresources it's too many to put in the\ndescription of this video so I have a\ndock for one of those and if you want\nthat dock link is in the description for\nthose next up on the road map becoming\nan AI engineer is understanding the art\nof prompting now I know a lot of people\nwill argue with me and say prompting\nisn't serious or that it is a fad but\nthe way that I see it is with prompting\nwhat you're really doing is just trying\nto elicit a behavior from a model now\ntoday that's English tomorrow it might\nbe something else so let's not argue\nabout the future but let's agree today\nthat effective prompting or the art of\neliciting the behavior from a model that\nyou want is a skill that's needed now\nthe art of prompting goes extremely deep\nas a practice so we won't go into all\nthe details in this video but I'm going\nto share four techniques that have\nworked for me in the past now the first\none that I love is Chain of Thought or\nthink out loud this is going to be when\nyou have the model explain its thought\nprocess first rather than have the model\nderive its answer early in its output\nnext it's hard to get performance from a\nprompt if you don't include examples so\nmake sure you include those and then of\ncourse there's a slew of other tricks\nlike using XML tags or using pre-filled\nprompt messages if you want to find out\nwhat those are those are on the dock as\nwell the next technique you're going to\nwant to master is with structured\noutputs now it's interesting because\ncomputers do much better with structured\ndata like Json or tables however\nlanguage models they deal with plain\ntext literally the output is a text that\nyou have the normal output now when the\nmodel responds back to you with valid\nJson well then you can go and give that\nto other programs in your computers or\nyou can start to use other tools outside\nof your language models and if you want\nyour language models to use things like\nsearch queries or tools or integrate\nwith other systems while structured\noutputs are going to be the way that\nyou're going to do this now let's listen\nto what TBO says and he sold his\nbusiness for $8 million now he's heavily\ninvolved with AI I think I think the the\nsentence that I tpe the most is output\nin jasm do not output anything else\nabout it so it's basically structured\noutputs only that come out from it yeah\nand it's super reliable I'll put a link\nto how to deal structured outputs on the\ndock next up as an AI engineer you're\ngoing to come up to the question around\nhow to do prompt management and one of\nmy favorite people to follow in this\ncase is Jared the CEO of prompt player\nhe has a super cool perspective on this\nit's interesting to see that we kind of\nall go down the same path first we're\ngoing to put our prompts right in our\ncode then we're going to have txt files\nthen we're probably going to use an\noutside manager prompt manager like a\nprompt lier for resources we have a ton\nof these first I want you to check out\nthis anthropic job and it's not current\nanymore but it was there that paid up to\n$375,000 a year for a prompt engineer or\nlibrarian h a prompt librarian okay next\nup we're also going to have a\nwell-rounded introduction to optimize\nprompting and I highly suggest reading\nEugene's post and for a huge Deep dive\nwell Elvis is the master in this one he\nis the prompt engineering guide that I\nhighly suggest going and checking out\nand not only that but we also have\nresearch directions as well so if you\nwant to check out this paper that Google\njust came out with around prompting\nlinks in the dock as well the third\nskill on the AI engineering road map is\naround context or retrieval so far we've\nonly been working with the model as is\nthat is we've relied on the data that\nthey model new at its time of training\nand this is okay for G reasoning but\nwhat about when you want to combine it\nwith the context it knows about you or\nyour users or maybe your role the\npopular term for this is retrieval as in\nhey we're going to go retrieve data and\nbring it back to our language model so\nit can generate a better response this\nis where the term retrieval augmented\ngeneration or rag comes in because\nyou're doing generation with retrieval\nbut don't be intimidated because this is\nliterally just putting data in your\nprompt like you would for any other\nanalysis I do this all the time when I\ncop copy and paste information about\nmyself into chat GPT that's a really\nmanual form of of retrieval but I'm\nstill retrieving information about\nmyself the common way to do this is with\nmatching relevant documents based off of\na user's query and what's amazing is\nthat the most popular way to do this has\nbeen with embeddings and semantic search\nlet's break both these down very quickly\nbecause they're super important so\nembeddings are a vector representation\nof a body of text if that sounds\ncomplicated it's actually pretty simple\nthis is literally just a list of numbers\nof vector and we do this because\ncomputers can compare numbers a lot more\neasily than they can compare words so if\nwe have two vectors together which\nrepresents two different pieces of text\nwe can actually compare those and do\nsimilarity metrics between the two and\nthen all of a sudden we can tell hey how\nsemantically close are these two\ndifferent pieces of text now the other\nkey word here is semantic or semantic\nsearch and this means that you're going\nto be searching by the meaning of a word\nor phrase rather the than the exact\nkeyword so all of a sudden ocean is\nsemantically similar to water which is\nvery cool right and you'll see that\nthere are immediate problems with doing\nvanilla retrieval you'll find that users\nqueries aren't really detailed enough or\nyour app isn't returning the exact right\ncontext that you need to answer a\nquestion or perhaps maybe there's too\nmuch fluff in your prompts and this is\nwhere the world of advanced retrieval\ncomes in and there's techniques like\nenhancing users query before it goes\ninto your search or improving the raw\ndata chunking strategies or improving\nhow you split your long texts into\nsmaller pieces which is really your\nindex strategies for resources I have an\nentire series on retrieval called\nfullstack retrieval.com if you're\ninterested to learn more you can go\ncheck that out there's also the Lang\nchain documentation which is a great\nplace to get ramped up to Concepts also\nfor those that are ready I've made a\ntool called chunk viz.com that actually\nhelps you visualize different chunking\ntechniques the next AI engineering skill\nthat you're going to want to master is\norchestration this is when you go just\nbeyond a single API call to a model and\nyou start creating systems that work\ntogether in concert now the base levels\nof this look like working with an\norchestration framework something like\nLang chain that's going to help you\nadopt the foundation orchestration\npatterns now at its simplest form this\ncould just be chains right chains are\njust when you sequentially put different\nmodel calls together however the\nadvanced version of this is when you\nstart to dip into the world of Agents\nnow agents may seem complicated and\nwe're still arguing about what a the\ndefinition of an exact agent is however\nmy definition is going to be similar to\nwhat Harrison Chase says the the core\nidea of Agents is using the language\nmodel as as a reasoning engine um so\nusing it to determine kind of like what\nto do and how to interact with the\noutside world and and and this means\nthat there is a non-deterministic kind\nof like sequence of actions that'll be\ntaken depending on the user input in\nEnglish agents are simply language\nmodels that have access to tools and can\ndecide when a job is specifically done\nbut I will say that the line is blurry\nfor what is an agent what is not let's\nnot argue about it there are jobs that\nwill pay up to\n$435,000 a year year for people who\nspecialize in agents now in the world of\nAgents the most popular Frameworks are\ngoing to be Lang graph crew Ai and\nHaystack or in the noode world you have\nthings like Lindy doai or langlow or\nhoe. another key important part with\nworking with agents is going to be\naround long-term memory now Lang chain\nhas a really cool blog post on this\nentire thing however I will say we're\nall still trying to figure it out and I\njust posted a tweet that asked everybody\nwhat they thought about long-term memory\nand uh there's no there's no solidified\nanswers for other resources I suggest\nchecking out Lang change has put out a\nstate of Agents report now this is\naround agent adoption within companies\nso if you're looking to build a company\nit would be pretty helpful to know what\nthe adoption patterns currently are this\none's very very cool the people I\nsuggest to follow in this world are\ngoing to be Harrison and Alex and by the\nway if you want to learn more about\nbuilding simple agents on top of\nunstructured data this is where my\ncohort comes in next up on the AI\nengineering road map and this is\narguably most important of all is\nevaluations and observability one of my\nfavorite lines about evaluations comes\nfrom Jason if you don't have evals then\nyou don't have a serious app evals are\nthe unit tests of your language model\napplications now in regular code it's\neasy to see if things break but what\nabout when your output is\nnon-deterministic and really Vibe based\nlike language models are well this is\nwhere evals comes in people will stay\naway from evals well because look how\npeople stay away from regular unit\ntesting in their code anyway but really\nalso because it's hard it's hard when\nyou want to thoroughly evaluate how good\na summary is or how good a natural\nlanguage output is these things are\ndifficult but there's evolving best\npractices one of the best resources I\nknow about this is going to be Hamill\nhe's one of the leading voices on evals\nand has a great resource to go learn\nmore now on the observability side this\nreally breaks down into two different\npieces the first one is tracing now\ntracing is when you literally just start\nto log the different language model\ncalls you have so you can easily debug\nfor them now it's extremely difficult to\ndig into why your application\nperformance has degraded if you can't\nfind or see which llm calls you're doing\nthe best way I've seen to do tracing is\njust hook into an additional framework\nsomething like Langs Smith or another\nobservability platform make sure to add\na bunch of metadata to your calls so you\ncan easily pinpoint and D duplicate\nwhich calls you're actually making on\nthe other side you're going to want to\ndo cost management now if you're not\ntracking your costs then you're going to\nbe in for a bad time observability tools\nhelp you track every single llm call and\nthe latency errors as well for resources\nmy go-to observability tool is lsmith\nthey have a free tier which is super\neasy to get set up with and it's a\nreally easy interface for other tools\ncheck out gen trace and arise for\nproduct teams I suggest checking out\nAuto blocks and free play the final\nskill that is needed to be an AI\nengineer is going to be the mindset now\nI bluffed a little bit because this is\nreally a meta skill but it's really\nimportant to have a new mindset as\nyou're starting to build with AI now hey\nllms give us a new capability to work\nwith and when you have new capabilities\nyou have new use cases and what are\nthese things even good for check out\nthis sample of use cases that Elvis\ntalked about SAR of Hampton he put out\nan AI report about how all his companies\nare using AI which is super interesting\nand then I also put together in early\nsignals the next mindset pillar I want\nyou to adopt is the build first and\nbuild quickly mindset so check out what\nTBO says about building quick I I would\ntry to like get the first ID that to\nhave and start working on it right away\nand try to ship something uh to the\npublic as soon as possible and I would I\nwould treat about that even if it's\nblack even if it's shitty and even if\nit's not not that useful I would just\ntweet about it and expect a few of my\nfriends to test it and give me feedback\nthe friction is so low that execution is\nmore important than ever the next skill\nis around understanding the emerging new\ntool stack now I'm sure you've heard of\na few of these and they could take up an\nentire video each but if you're building\nfront ends it's too cheap not to work\nwith something like v0 to get\ninspiration from then there's cursor and\nwind surf for if you're looking for a\nnew AI powered IDE and then not only\nthat anthropics projects is very popular\namongst the developers see what Nicole\nHeadley CEO of Head Start has to say\nabout anthropic projects we're just top\nusers of CLA and CLA projects uh we love\nit I think it's the greatest thing that\nhas ever been developed if he took it\naway from me I would really struggle to\nrun my business which is probably like\nthe highest NPS you can get for a\nproduct the next mindset pillar is\naround scaling llm apps now I got to\nattend open a eyes Dev day and I heard\nan amazing talk by Colin and Jeff of\nopen AI in it they highlight three key\nareas of scaling LL Maps improving\nperformance reducing cost and reducing\nlatency and by the way I packaged up all\nthese links and resources in a single\nblog post completely free link for that\nis in the description if you want it now\nif you've gotten this far in the video\nthere's a good chance that you're on\nyour way to becoming an AI engineer and\nI want to keep pointing you in the right\ndirection and this is only the beginning\nthe fact is is that the landscape is\nchanging so much which really that\nshould encourage you because there's\nnever been a better time to build or be\na builder who's actually executing now\nI'm going to be stepping a few students\nthrough these best practices Hands-On\nthrough a course called building with\naims AI engineering patterns for Scrappy\ndevelopers the whole point of this is I\nwant you to ship an app using these\nexact six skills you're going to learn\nnew tools along the way and you're going\nto be surrounded by a cohort of Builders\nwe're also going to have guest speakers\nfrom industry who are going to come in\nand chat with us about this not only\nthat but then we're also going to have\nai credits so you're going to get a\nspeed boost to building with some of\nthese tools in general the cohort will\nbe running ongoing if you want to get\nmore information about that link is in\nthe description and the doc as well",
  "transcript_chars": 18917,
  "ingested_at": "2026-05-15T04:40:18.671961+00:00",
  "source": "channel",
  "yt_meta": {
    "view_count": 203446,
    "like_count": 6593,
    "channel_id": "UCyR2Ct3pDOeZSRyZH5hPO-Q",
    "categories": [
      "Science & Technology"
    ],
    "tags": []
  }
}