Ads skipped

The Dangerous Illusion of AI Coding? - Jeremy Howard

166K views · Mar 3, 2026 · Science & Technology

Comments · 511

  • @MachineLearningStreetTalk · 6 months ago · pinned

    RESCRIPT: <a href="https://app.rescript.info/public/share/BhX5zP3b0m63srLOQDKBTFTooSzEMh_ARwmDG_h_izk">https://app.rescript.info/public/share/BhX5zP3b0m63srLOQDKBTFTooSzEMh_ARwmDG_h_izk</a><br>PDF: <a href="https://app.rescript.info/api/public/sessions/62d06c0336c567d6/pdf">https://app.rescript.info/api/public/sessions/62d06c0336c567d6/pdf</a>

    14

  • @JZ50702 · 6 months ago

    The first half of the interview shows Howard has correct understanding of LLM and the second half shows his heart is in the right place. He has my respect.

    118

  • @Gornius · 6 months ago (edited)

    <a href="https://www.youtube.com/watch?v=dHBEQ-Ryo24&amp;t=1685">28:05</a> That&apos;s the most important part for me. When I write code organically I often stumble upon parts of the codebase that I don&apos;t need to implement a feature. Management or somebody else who is not software engineer might think it means I lose time navigating a codebase, but every time I am surprised how often I can solve some issues immediately, because I instantly know what I need to do.<br><br>It stuck to me how first developer in our team who automated most of his work with AI told us that he&apos;s doing an AI detox, because he realized he lost this ability. Things that months ago he would have solved immediately now required involving AI AND him understanding the months of codebase changed by AI.<br><br>I use AI in coding mainly to navigate through documentation and I think it actually made me more productive. If you think about it, it&apos;s processing natural language - exactly the use case that LLMs were designed for. Of course from time to time it hallucinates some methods or classes that don&apos;t exist, but gives enough context so I can find it in documentation far sooner that if I had googled it or decided to find it manually.

    13

  • @thedreadedgman · 6 months ago

    the section comparing vibe coding to gambling slot machines is absolutely genius! I love this, wonderful video

    41

  • @AlternativeTakes · 6 months ago

    this is by far the most intelligent and non bullshit take on genai coding.

    163

  • @ernestooropeza6150 · 5 months ago (edited)

    I feel like I just discovered one of the best YouTube channels to accompany me on my computer science journey.<br><br>This is coming from an applied math major who is currently binging the CS curriculum in my last semester.

    2

  • @andreafeelsfantastic · 6 months ago

    Watching this video has DRAMATICALLY improved the quality of YouTube’s recommendations for me on what else to watch. So I would owe you my thanks even if this weren’t one of the best conversations I have ever heard about AI, which it absolutely is.

    3

  • @alanc497 · 5 months ago

    my favorite tech/code podcast, with the best measured takes who don&apos;t go into hype or doom or crazy extrapolation of the present, just thoughtful observations, thank u

  • @ericmichel3857 · 6 months ago

    I am not a software engineer but I see the same thing happening everywhere. I am an MRI service engineer and have been doing this in various capacities for almost 40 years. Back in the day the systems were far more analog. Calibrating and testing systems took days with elaborate setup and adjustments, and we traced circuits to component level. It gave me a much deeper understanding of how the system functioned and how things can go wrong and interact with each other. Now, what was a row of cabinets is a single board or module and is 90% digital with built in sensors and diagnostics. Calibrations that took days and now you drop in a phantom and the system calibrates itself in about an hour, which is great when it works. And when something fails the diagnostics and log data practically tell you what part to replace.<br>The problem is all the newer engineers really have very little idea what is going on and become glorified part changers which is all that is needed nine times out of ten. However, on those times it doesn&apos;t work they have no way of understanding how or why and they just keep changing parts till it works, or they call in someone like me. I try to explain what is happening and why, the problem is the automated systems work so well they don&apos;t really need to know these things in order to do the job effectively and so they never really learn it well. The thing is as more and more people with this expertise and understanding retire, there is virtually no way to replace it. I have seen cases where they replaced an entire $80K cabinet that only needed a $15 part, but until I got there they had no idea. <br>The problem is you don&apos;t need an army of people with deeper understanding and experience to keep these systems running most of the time. And on those rare occasions where that is required , they bring in someone like me, well for now anyway. The thing is it is not practical or efficient for everyone to take the time and effort. And they see it as a huge waste of time, and for the most part they are not wrong. You can try to force them to learn it as suggested in this talk about forcing software engineers to write their own code for a few years. Good luck with that, I heard an eight year complaining about why she had to learn to read and write because the computer will just tell her what it says and write it for her too. I am sure there are many skills and knowledge we have already lost in our modern society, it may be that this is inevitable.

    40

  • @markusvanalmsick537 · 6 months ago

    Let&apos;s also give credit to Theo Gray for the Mathematica notebook interface. He wrote the interface 40 years ago.

    13

  • @mrjoeing450 · 4 months ago

    What a great post! &nbsp;I love this and also how you showed the actual articles.....Brilliant!🎉. &nbsp;I truly enjoyed this! &nbsp;Thank you! &nbsp;Cheers!

  • @dimakarpov1956 · 6 months ago

    It is so satisfying to find some where an idea you creating yourself: coding != programming - that is wrote multiple time across my diaries.The programming is decomposition the problem into components, and defining contracts between them. Because the real demand is to adopt the application to the constantly changing requirements of the market. That&apos;s so neat, to find the same in the title of a video.

    3

Up next

LIVE

The FTC Is Investigating OpenAI — Here’s Why

Prof G Markets and Platypus Economics with Justin Wolfers · 194K views

LIVE

Why More Data Cannot Replace Prior Structure - Alexander Mattick

Machine Learning Street Talk · 24K views

LIVE

Build to Last — Chris Lattner talks with Jeremy Howard

Jeremy Howard · 21K views

LIVE

This is how the AI bubble pops | Cory Doctorow

The Tech Report · 691K views

LIVE

AI Chatbots for Small Business: How to Improve Customer Support Without Losing Trust

AI Business Flywheel · 67 views

LIVE

AI Skills with Matt Pocock

The Pragmatic Engineer and Matt Pocock · 204K views

LIVE

Strange Geometric Shapes Found Inside AIs — Tom McGrath

Machine Learning Street Talk · 40K views

LIVE

Keynote: After the AI Hype – What’s Real, and What’s Next - Richard Campbell - 2026

NDC Conferences · 427K views

LIVE

The real AI threat isn't what you think - Yanis Varoufakis & Wolfgang Munchau | The Econoclasts

UnHerd and Econoclasts · 612K views

LIVE

The Uncomfortable Truth About AI “Reasoning” | World Science Festival

World Science Festival · 640K views

LIVE

Skill Issue: Andrej Karpathy on Code Agents, AutoResearch, and the Loopy Era of AI

No Priors: AI, Machine Learning, Tech, & Startups · 1M views

LIVE

Are the hyperscalers coming for your services revenue? | The AI Take, Ep 11

1KE · 32 views

LIVE

Model Collapse Ends AI Hype

Theos Theory — AI & Science Counter-Perspectives and 2 more · 337K views

LIVE

Software engineering at the tipping point

Google for Developers · 455K views

LIVE

Is there an AI mental health crisis?

Syntax · 229K views

LIVE

Modern Architecture 101 for New Engineers & Forgetful Experts - Jerry Nixon - NDC Copenhagen 2025

NDC Conferences · 177K views

LIVE

What Building an AI Scientist Actually Requires Beyond Intelligence — Edward Hughes

Machine Learning Street Talk · 22K views

LIVE

Yann LeCun's $1B Bet Against LLMs [Part 1]

Welch Labs · 1.2M views

LIVE

Full Walkthrough: Workflow for AI Coding — Matt Pocock

AI Engineer and Matt Pocock · 1.6M views

LIVE

Andrej Karpathy: From Vibe Coding to Agentic Engineering w/ Stephanie Zhan

Sequoia Capital · 1.5M views

LIVE

AI Expert WARNS: "You're Not Ready For 2027"

The Diary Of A CEO Clips · 1.8M views

LIVE

How AI will change software engineering – with Martin Fowler

The Pragmatic Engineer · 323K views

LIVE

How Many Narrow AIs Could Behave Like One Superintelligence - Daniel Kokotajlo and Thomas Larsen

Machine Learning Street Talk · 42K views

LIVE

I'm done coding with AI

Brett Codes · 1M views

LIVE

Benedict Evans: OpenAI’s Moat Problem & the Future of Software

The MAD Podcast with Matt Turck · 29K views

LIVE

The Man Who Calls BS On AI: They’re LYING About AI, 2027 Is When It All Breaks! | Ed Zitron

The Diary Of A CEO and Better Offline · 8.9M views

YouTube, with the door locked.

Aegis plays a clean stream instead of YouTube's player, so pre-roll ads, trackers, and fingerprinting never ride along. Drop Shields any time if you want the official player back.