{
  "video_id": "6qs_ScUelHY",
  "channel_slug": "dataindependent",
  "channel_handle": "dataindependent",
  "title": "4 Reasons Why AI Won’t Work",
  "duration_seconds": 566.0,
  "url": "https://www.youtube.com/watch?v=6qs_ScUelHY",
  "upload_date": "",
  "transcript": "AI has a ton of hype and the supporting\nfans are getting borderline ridiculous\nit's more profound than fire makes money\nusing only AI I certainly think that\nthis is like a revolutionary moment in\ntechnology but what about the reasons\nwhy AI might fail to live up to the hype\nlarge language models have no concept of\ntruth I'm on a mission to figure out how\nbusinesses will create more value with\nAI but I don't want to fall into the\nFanboy trap one-sided thinking or lose\nmy first perspective I decided to go on\na hunt to find the top reasons why AI\nwon't work I challenge myself to see if\nmy opinions would change so let's go\nfind out now I'm not talking about the\nAI doomers or trying to guess why AGI\nisn't possible what happens if we build\nsomething smarter than us that we\nunderstand that poorly what I'm talking\nabout is the argument for why AI won't\nprovide more value in the workplace but\nwhat tipped me over the edge to do this\nin a recent article Gary Marcus goes as\nfar as to say what if AI turned out to\nbe a dud\nI sent out tweets asking others why they\nthought AI may not work I even asked\nchat gbt to give me some ideas I\nConsolidated them and picked the ones\nthat I thought had the strongest\narguments let's talk about the four\nreasons why AI could fail reason one\nhallucinations stop progress AI\nhallucinations in his article Gary\nMarcus points out there's no reason to\nthink that the hallucination problem\nwill be solved soon Emily Bender\nLinguistics professor at the University\nof Washington goes as far as to say this\nisn't fixable the hallucination problem\nrefers to when language models\nconfidently make up information which is\npretty hard for the user to know that\nit's happening now sometimes this is a\nfeature like when you're writing a poem\nabout a made-up topic that the language\nmodel doesn't know about however\nhigh-stakes business decisions rely on\ntrusted information it's not great when\nthe language model makes up information\nabout a sales call you just had or what\nabout when that lawyer submitted a bogus\ncase law and he actually got caught if\nhallucinations continue to be a problem\nthen businesses will opt out of gen AI\ntools all together I agree this is a big\nproblem and a blocker and a wider\nadoption if this doesn't get fixed then\nyes of course generative AI will have an\nissue however the answer is a mix of\nresponsible AI use and tooling you\nshould always verify that any Source\nyou're using is true the theme of AI in\n2023 is to have it augment your work not\nreplace it as for tooling a lot of\nincredible people are working on the\nhallucination problem things like more\nadvanced retrieval methods like Lang\nChain's parent document retriever or\nllama index's retrieval agents or better\nyet guardrails for chat Bots like\nnvidia's Nemo reason number two models\nare too complex to be reliable as AI\nsystems become more complex their\nbehaviors can become more unpredictable\ndo we understand everything about why\nthe model does one thing and not one\nother thing certainly not a member of\nopen ai's technical staff also said they\naren't sure why chat GPT doesn't produce\nthe same output each time or another way\nwhy it's not deterministic you see\ndeterministic code is nice because you\ncan in theory debug and Trace why an\noutcome happened non-deterministic code\nleads to unintended consequences or\nerrors and a load-bearing process for a\nbusiness can't afford to be undebuggable\nif these issues aren't addressed\nbusinesses may see AI as a liability as\nopposed to an asset determinism is one\nof the utmost important features of\ntraditional programming if you shake\nthat Foundation you shake everything\nelse that comes Downstream in addition\nto this openai recently released a post\non how they're able to explain and score\nevery neuron within gpt2 this is huge\nfor as Sam puts it we are pushing back\nlike\nthe fog of War more and more it's only a\nmatter of time before they have this\ncapability for later stage gbt models\nlastly in darkest interview with Dario\nthe CEO of anthropic Dario says the\nwords mechanistic interpretability at\nleast 15 times I don't know what's going\non inside mechanistically and I think\nthat's the whole point of mechanistic\ninterpretability this is the ability to\npeek into the model and understand\nwhat's happening after a quick break\nwe're going to review two more reasons\nwhy AI could fail and I'll share whether\nor not my views on AI have actually\nchanged if you've been following me for\na while you know how much I love easy to\nset up and out of the box functionality\nin just a few lines of code which is why\nI want to tell you about today's sponsor\nsingle store single store Powers fast\nreal-time analytics powered AI\napplications you can easily store\nvectors for a semantics search get\nstarted in your language of choice and\neven expand your data outside of vectors\nwith support for structured\nsemi-structured time series full text\nspatial and key value data seriously if\nyou have a range of data they can store\nit the best part they even have\nnotebooks directly in the browser that\nyou can connect your database with no\nneed to hassle with connections when you\nget logged onto single store you can get\nan easy web browser interface real-time\nmonitoring and familiar SQL tooling\nsingle store makes it easy to get set up\nwith up to six hundred dollars in free\ncredit when you get started they have\nIntegrations right into langchand and\none-on-one support from one of their\nexperts if you have questions single\nstores for anyone who's looking to get\nstarted with a no-brainer database\nsolution in order to power their AI\napplications it's perfect for Indie devs\ndata engineers and companies looking to\nexplore new data options it's being used\nby analytics companies Fortune 500s and\nsome of today's biggest consumer apps\nthat includes DBT for their use of\nreal-time analytics and Heap who uses\nsingle store as the query layer on top\nof their customer facing Data Systems\nsingle store is the platform for low\nlatency access to massive data sets if\nyou're interested in learning more links\nin the description reason number three\nmodels won't get better in the same\ninterview with Dario darkesh asks hey if\nscaling laws end up plateauing before\nreaching human level intelligence what's\nyour explanation for why this would\nhappen and Dario says well we could run\nout of data we could run out of compute\nor maybe it's that we don't have the\nright architecture maybe this is true\nand maybe we're reaching an asymptote of\nintelligence with GPT models and they\nwon't actually get better it's well\nknown that scaling laws are fairly\npredictable predictable to like\nsometimes even to several significant\nfigures which you don't see outside of\nphysics companies could simply choose to\ninput more money and get a smarter model\nmy hypothesis is that they'll continue\nto do this as long as they see the\nreturn on investment Additionally the\nactual language model itself is only one\npiece in a larger orchestration of tools\nit's been amazing to see the\nadvancements in various areas like the\nStanford simulation study or retrieval\ntools like metals and betting platform\nthat reacts to user feedback or\nsuperstructures around the language\nmodel like Lang change finally I'll\nclose this counter argument with a\nquestion what do you think is the\nlimiting factor of getting more value\nfrom AI the actual model itself or our\nunderstanding about how to work with the\nmodels that we already have I'd argue\nit's more of the latter than we think\nreason number four Advanced prompting is\ntoo hard most generative AI relies on\nprompts from a user to perform an action\nyou can have the best tool in the world\nbut if you don't give it proper\ninstructions then it'll just spin in\ncircles in everyday life an employee\nmight be able to perfectly articulate\nthemselves to another colleague however\nwhen it comes to speaking to a language\nmodel it requires a different set of\nskills it could be that the barrier to\nentry for writing clear instructions to\na language model is simply too high for\nthe common team member to learn my take\nis that we're already seeing a rise in\nthe tide on a workforce's ability to\ncreate prompts in 2023 many people were\nlearning prompting for the first time\nchat gbt is free so literally anyone\ncould try it out and as they learned\nthey shared their thoughts along the way\nall those learnings build on top of each\nother and help the next person ramp up\nmore quickly I'm also expecting the\namount of detail a person will need for\ntheir desired output will also go down\nsimply people will get better at\narticulating themselves and you'll also\nneed to do less articulation in order to\nget what you want okay so we reviewed\nfour counter arguments to Ai and\ncounters to those counters I challenge\nmyself to understand the other sides of\nthe argument to see if my mind would\nchange about Ai and did it no there's\nway too much value being derived today\nand I'm having way too much fun look\nthere's no doubt that AI has its\nchallenges right now but I'd press you\nto find a new technology that didn't\nmost technologies have years of tooling\nand best practices as their Foundation\nbefore they get Mass adoption however\nMass adoption with AI took two months\nand we're still building the tools and\nbest practices we're trying to\nretroactively fill that Gap and there's\ngoing to be growing pains along the way\nhowever I see too many examples of\nlanguage models already providing value\nto not get excited like creating Dynamic\ninterfaces on the fly or automating\nsummaries and meeting notes of sales\ncalls what about chat Bots that allow\nyou to have conversations over massive\namounts of data oh yeah and the entire\nvisual generation space for pictures and\nvideo and that's just what's come out in\n2023 and while I can't tell what the\nfuture is going to be I'm excited enough\nto keep pulling on this thread with you\nby the way if you like this type of\ncontent and want to go on a journey with\nme to see how AI will play out in\nbusiness I send out my learnings to a\nsmall group of email subscribers links\nin the description\nthank you",
  "transcript_chars": 10084,
  "ingested_at": "2026-05-15T04:41:35.376579+00:00",
  "source": "channel",
  "yt_meta": {
    "view_count": 3136,
    "like_count": 130,
    "channel_id": "UCyR2Ct3pDOeZSRyZH5hPO-Q",
    "categories": [
      "Science & Technology"
    ],
    "tags": [
      "ai hype",
      "ai expectations",
      "ai complexity",
      "ai instructions",
      "ai challenges",
      "ai overviews",
      "LLM",
      "language models",
      "chatgpt",
      "the ai problem",
      "problem with ai",
      "how ai will affect jobs",
      "how artificial intelligence will affect jobs",
      "impact of artificial intelligence on jobs",
      "impact of ai on business",
      "why ai won’t work",
      "the argument against ai",
      "don’t use ai for your business",
      "ai",
      "artificial intelligence"
    ]
  }
}