On Karpathy's "Software is changing (again)"

Posted on Sun 29 June 2025 in ai, it

I watched Andrey Karpathy's talk Software is Changing (Again). I might have found it inspiring a while ago; now I felt just disconnection.

  • He shows this room filled with 9-to-13-years-olds, hunched over laptops, humming OCD-like rythms, in a windowless, spotless white prison where the only consolation is fake plants, on Aeron chairs and standing desks, with their shoulders already higher than their ears when their bones are still halfway through the development, and imparts this futuristic hope that just makes me wanna puke:

    I find this such a wholesome video, I love this video. Like, how can you look at this video and feel bad about the future. The future is great. My god, you really need to go and touch some grass. AI is unleashing a generation of wildly creative builders.

    Which is akin to saying "We're shaping the next generation of police officers who'll ensure our society stays safe", and 5 years down the line they're checking tickets at the train station while somebody gets robbed round the corner. Will you come down on this planet, Andrej?

    The future is bleak

  • He gives this wonderful definition of LLMs as "stochastic simulations of people", which is quasi-accurate, as we've never trained LLMs on people. We've trained them on text written by people.

  • Why are we taking advice on the future by somebody who worked on Tesla Autopilot for 5 years, and deemed self-driving cars imminent in 2012? Please explain it to me. You only need to visit the wiki page for Tesla Autopilot to figure out that this is now a 12-year-old fake news:

    Since 2013, Tesla CEO Elon Musk has repeatedly predicted that the company would achieve fully autonomous driving (SAE Level 5) within one to three years, but these goals have not been met.

  • He envisions a world where we are but verifiers of the AIs work. Please kill me before. I write software so that I don't have to do work, and I rely on algorithms, not on crystal balls. I stopped doing math homework in middle school because I had implemented factorization and division and primality algorithms and then used them to do my homework. Me not having to verify the answers was exactly the point of implementing those. Why would I build something I need to constantly vet? It doesn't free me; it only dumbs me down, it disempowers me, and puts me at the mercy of a digital bullshit.

  • He shows this AI-built app, MenuGen, that generates pictures from a menu. What am I looking it. He says "it took 1 day of work". What are you talking about: it takes 1 days to code it yourself if you know what you're doing anyway; and if you don't know what you're doing, you should go back to being a high-stakes E-staff member. Just don't mess with programming. Then he goes on to ramble on how it would take 1-2 weeks to build the devops/integrations side - which is exactly the time it takes now to an engineering team, and what 80% of efforts are already directed to in any professional setting. You sucker, have you really worked in this field decades?

  • "We should make the web more accessible to LLMs". Swap web with streets and LLMs with cars and you'll get a large reason why our current world sucks. But please go ahead. You'lld find me with the head in the microwave.


ai it