{
  "video_id": "dzPsXaSl6PQ",
  "channel_slug": "dataindependent",
  "channel_handle": "dataindependent",
  "title": "How I Fine-Tuned An AI Clone - Can You Tell The Difference?",
  "duration_seconds": 518.0,
  "url": "https://www.youtube.com/watch?v=dzPsXaSl6PQ",
  "upload_date": "",
  "transcript": "so I'll show you the video and then you\ntell me which one you think it is real\nor fake no fake real real fake fake the\nhardest part of cloning someone with AI\nis coming up with what they'd actually\nsay I've devoted entire tutorials trying\nto mimic the tone of someone it wasn't\nbad but the audio voice was off I didn't\nhave video and their words didn't quite\nfeel right but recently I got my hands\non two new tools that I think that can\nactually help me create a convincing\nclone of myself at the end I want to\nwalk up to 10 people and see if they can\ntell the difference hello oh good\npeople when you try to digitally clone\nyourself you need to hit three key\npieces content look and sound let's\nstart off by trying part one matching my\ncontent and tone let's try the easiest\nmethod first simply talk like Greg camad\nso my first job out of college was with\na tech startup called innovate Tech H it\nhas no idea who I am this method's not\ngoing to work let's try something else\nyou see the words that come out of\nsomeone's mouth are derived from years\nof EXP experience memories and\npersonalities it's what makes them\nuniquely them method two tone\ninstructions okay so I want a language\nmodel to talk like me but where do I\nbegin it doesn't know who I am so that\nprobably means I need to give it\ninstructions on how to talk like me but\nyou see I'm not great at describing what\nI sound like I think a language model\nwould be a little bit better I'm going\nto give it samples of me speaking from a\nrecent podcast that was on and if you're\ndoing this at home just use any\ntranscript of you talking I'm going to\nask it what's the best way to describe\nhow Greg talks let's see what it says\nGreg's pace is steady and measured\nGreg's mood is predominantly\nenthusiastic and passionate Greg's tone\nis authoritative yet approachable H here\nare the instructions to talk like Greg\nbut wait it also doesn't have\ninformation about my background or\ndetails that I've shared I'm going to do\na retrieval process and give the\nlanguage model stories and memories that\nI've shared on the podcasts so it can\nhave more background context about how\nto answer questions like me all right\nlet's see if it can talk like me yet\nsomething still isn't quite right by the\nway the full code is available to email\nsubscribers if you'd like to join\ntotally free links in the description\nmethod three examples you know a picture\nspeaks a thousand words maybe an example\nwould speak a thousand words as well I'm\ngoing to put in raw examples from the\npodcast of my speaking Style with the\nums and the O's and everything in\nbetween let's ask the language model\nonce more hm now we're getting somewhere\nthat's looking a little bit better\nthere's a problem though GPT 4 tokens\nare pricey and it doesn't really fully\nsound like me we need something with a\nbit more Firepower you know I want to\nsee if fine-tuning can help fine tuning\nis the process of giving model validated\ndata so it learns the output you're\ntrying to give the model give it enough\ndata and you hope that its performance\nis going to improve I've avoided fine\ntuning up until this point because the\niteration speed on development is really\nmuch slower it's way easier to try\nprompt only methods first you should try\nbetter prompting and better retrieval\nbefore fine-tuning one more time you\nshould try prompting and better\nretrieval before fine-tuning I also want\nto try an open source model to see if\nit's easier to manipulate I thought this\nwas going to be a big pain in the butt\nbut then I was introd uced to today's\nvideo sponsor gradient. a they make\nhosting open source models super easy\nand it's way cheaper than gp4 all you\nneed to do is just pass it samples in a\nfew lines of code and your model's ready\nto go they have direct Integrations to\nlangon and the output was also way\nquicker than chat GPT let's give fine\ntuning ago method four fine-tuning with\nsynthetic prompts when you want to do\nfine tuning you need to have high\nquality input and output pairs which are\nvalidated you're basically telling the\nmodel hey given this input this is what\nyour output should say the more you do\nthis the more you'll reinforce the\nbehavior of your training set my first\nthought was to synthetically generate\ninput prompts with the GPT 4 based off\nof my speaking sentences for example if\nI said last night I went to dinner with\nmy family I would hope that GPT 4 would\ngenerate a synthetic prompt along the\nlines of hm what did you do last night\nI'm going to use a bunch of the lines\nthat I set on the podcast to generate a\nbunch of data points and then use\ngradient for fine tuning all right let's\ncheck the output what was your first job\nout of college and did you like it hm\nyeah that is totally off method five\nconversational prompt pairs maybe I need\nto mimic the real life natural\nconversation a bit more maybe I should\nhave the podcast questions be the input\nand then my response is the output let's\ntry that out where you first got\ninterested in data kind of two pivotal\nmoments you know I think I'm thinking\nabout this the wrong way when you find\ntune you're reinforcing the emphasis\nthat a certain output should go along\nwith a certain input but I want the tone\nto be reinforced not necessarily my\nactual content of what I say I don't\nwant to use fine tuning as a way to turn\nthe language model into a database and\nhope that it can recite facts I can do\nother retrieval and prompting techniques\nto tell the language model what to say\nbut I need the language model to learn\nhow to say it I think I need to make\nthis a lot more simple method six you\nare Greg camad let's use those same\nlines that I said on the podcast but now\ninstead of a synthetic prompt or a\nconversational prompt simply The Prompt\nwill be you are Greg camer now in my\nprompt I'm going to tell the model that\nit's Greg camad so hopefully it'll lean\ninto that fact with fine tuning let's\ntry try asking it again okay now we're\ndefinitely getting somewhere but I think\nthere's one last thing to try method\nseven talk in the style and tone of Greg\ncamad I don't want to try and convert\nthe entire model into Greg camarad\nthat's a lot of model to push around if\nI wanted to do that the chance was\nlikely earlier in the model making\nprocess what I really want to do is just\naffect the cherry on top I want to get\nit good at mimicking My Style I want to\nfind tune on its sounding like the style\nand tone of Greg camad not you are Greg\ncamra so it'll notice sound like me\nrather than embody me it's a subtle\ndifference but I think it's going to\nwork out let's try it what's your first\njob out of school and did you like it my\nfirst job out of college was in\ncorporate finance and budgeting I\nenjoyed the analytical side of the work\nbut I wanted to do something more\nimpactful that looks good let's roll\nwith that one wow what a journey but I'm\nglad we were able to fine-tune our\nfine-tuning approach to recap our final\nmodel and prompt will be a fine-tune\nmodel with tone instructions with\nexamples and relevant background content\ncontext so it can answer my question\nokay so that's a third of our battle\ndone hopefully the other two parts are\ngoing to be a lot easier part two mimic\nhow someone looks and how they sound a\nwhile back I saw an awesome video\ngeneration demo going around from a\ncompany called heyen I reached out to\nhen's founder through Twitter DMS an\nemail and I got a nice response but uh\nno access quite yet but to my surprise\njust last week I got another email from\nhey Jen saying that it was my time I\ncouldn't wait to try it out I logged on\nto ha Jen and then recorded a 2-minute\nvideo sample I had to wait till the next\nmorning until it was done fine-tuning a\nvideo but when I woke up I got a\nglorious notification that said my model\nwas ready well let's check it out\nwelcome to the new era of video creation\nwith hey Jen simply type your script to\nget started wow that is really solid I'm\nsuper happy with those results great now\nwe finished all three steps content\naudio and video so this is awesome but\nwhat do other people think could they\ntell it is a clone I'm going to show 10\nrandom people and then I want to see\nwhat you think I asked my clone what was\nyour first job out of college and did\nyou like it I got my iPad I packed in my\nbag and I headed to a tech Meetup in San\nFrancisco so Greg how many people do you\nthink will actually know it's a clone I\nthink like half are going to get it\nright let's go figure out if people\nthink it's an AI or it's real\nGreg I want you to tell me whether or\nnot you think it's real or fake Is that\nreal me or is that clone me tell me real\nor fake no fake F you think it's fake\nyeah fake why do you think it's fake\nyeah because low resolution or the face\nfake okay I go with fake fake fake real\nreal fake real fake fake all right now I\nwant to see what you the audience thinks\nwhich one do you think is real Greg and\nwhich one do you think is clone Greg let\nme know what you think in the comments\nmy first job out of college was in\ncorporate Financial Planning and\nAnalysis my first job out of school was\ndoing corporate Financial Planning and\nAnalysis for big\ncorporations the fact that this mimics a\nperson so well brings up a lot of\nquestions ethical too what should\ncreators disclose to Their audience\nshould you tell someone they that you're\nusing an autogenerated clone for me\npersonally I value human relationships\nand authentic connections I'm not going\nto be using this deceptively I will\nabsolutely disclose if I'm ever doing\nthis and it'll likely just be for\ntesting out purposes if I were to do\nthis though the verticals that interest\nme the most are going to be short form\ncontent news recap and personalized the\nsales videos if you want to see more on\nmy tone Journey you can check out the\nwebinar that I did with Harrison Chase\nCEO of Lang chain and by the way email\nsubscribers got Early Access to this\ncontent so if you want to join totally\nfree links in the description and don't\nforget to check out gradient. a so you\ncan find tune your own\nmodel",
  "transcript_chars": 10026,
  "ingested_at": "2026-05-15T04:41:16.651569+00:00",
  "source": "channel",
  "yt_meta": {
    "view_count": 5238,
    "like_count": 170,
    "channel_id": "UCyR2Ct3pDOeZSRyZH5hPO-Q",
    "categories": [
      "Science & Technology"
    ],
    "tags": [
      "How to create an ai clone of yourself",
      "step by step guide to creating a digital clone",
      "how to make a clone with ai",
      "how to mimic another person using ai",
      "how to replicate myself using ai",
      "ai clone",
      "fine-tune using ai",
      "fine-tuning an ai clone",
      "fine tune using gradient",
      "greg kamradt",
      "real or fake clone",
      "can you tell my ai clone is a fake"
    ]
  }
}