Full episode · 30 min
Andrej Karpathy: From Vibe Coding to Agentic Engineering w/ Stephanie Zhan
Andrej Karpathy says agentic engineering can speed up professional software work without lowering its quality bar.
TDTraining Data · Stephanie Zhan
- Andrej Karpathy is co-founder of OpenAI, former head of AI at Tesla, and founder of Eureka Labs.
The best minutes
- 2:24Software 3.0 makes some apps unnecessary5 min
- 9:29Verifiability helps explain where AI excels4 min
- 19:22Engineers still need to own judgment and system design4 min
Also worth watching
- 15:42Agentic engineering raises output without lowering the quality bar4 min
- 25:15Agent-first services should let agents handle setup and deployment2 min
- 27:40You can outsource thinking but not understanding2 min
Skip the opening setup, the founder discussion about automatable work, and the animals-versus-ghosts framing.
Summary · 3 min read
Andrej Karpathy co-founded OpenAI, served as head of AI at Tesla, and founded Eureka Labs. At AI Ascent 2026, he discussed how agentic tools are changing programming and software development.
Andrej Karpathy says agentic tools began producing code he could trust around December, and that agentic engineering now aims to preserve professional software quality while increasing speed. He describes Software 3.0 as programming through prompts and context, with neural networks enabling new forms of information processing rather than simply speeding up existing software. He argues that AI capability is jagged: models excel where tasks are verifiable and labs train for them, but still need human oversight. Karpathy says people remain responsible for taste, design, and understanding, while software and infrastructure should become more agent-native.
Minute by minute
- 0:37Agentic tools changed coding
Andrej Karpathy says newer models began producing code that worked without his corrections around December. As he trusted them more, he started vibe coding and filling his side-projects folder, a shift he says went beyond the ChatGPT-style use many people had seen the year before.
- 2:24Prompts become the programming interface
Karpathy describes Software 1.0 as writing code, Software 2.0 as training neural networks with data, and Software 3.0 as programming through prompts and context. He points to OpenClaw’s agent-based installation: instead of maintaining a complex shell script for different computers, users give instructions to an agent that can inspect and configure the environment.
- 4:46Some apps can disappear
Karpathy built Menugen to turn restaurant-menu photos into illustrated menus, using code, OCR, and image generation. He says he was surprised when a prompt to Gemini using NanoBanana produced the result directly, making much of his app unnecessary. He sees this as a shift toward neural networks processing information without conventional software in between.
- 7:33Neural networks may run the show
Karpathy imagines systems that take raw audio or video and generate a suitable interface, with neural networks doing most of the work. He suggests the neural network could become the host process while the CPU acts more like a co-processor, but says the path there remains uncertain.
- 9:41Verifiable tasks create jagged skills
Karpathy says models often excel at verifiable tasks such as math and code because labs can train them with reinforcement-learning rewards, while capability remains uneven elsewhere. He contrasts models that can refactor large codebases with ones that advise walking 50 meters to a car wash. He also says chess improved from GPT-3.5 to GPT-4 partly because chess data was included in training.
- 13:36Verifiable domains invite fine-tuning
Karpathy says founders may be able to build valuable reinforcement-learning environments in domains the major labs have not prioritized, then fine-tune models on them. He argues that almost everything can be made verifiable to some extent, including writing judged by a council of LLMs, and says everything is ultimately automatable.
- 15:42Agentic engineering protects quality
Karpathy distinguishes vibe coding, which raises the floor of what everyone can build, from agentic engineering, which aims to preserve professional software quality while moving faster. He describes agents as powerful but fallible and says coordinating them without sacrificing quality is an engineering discipline. In his view, skilled practitioners can gain far more than 10x.
- 17:17Hiring should test real projects
Karpathy says strong agentic engineers invest in their tools and make use of their features. He argues that hiring should move beyond puzzles toward large projects, such as building a secure Twitter clone and then having agents try to break it.
- 19:33People still own judgment
Karpathy says people remain responsible for aesthetics, judgment, specifications, and oversight. He gives Menugen’s account system as an example: an agent matched Google and Stripe accounts by email, even though users could use different addresses. He says people must define the right design, such as tying funds to unique user IDs, while agents fill in implementation details.
- 23:32Treat models as ghosts, not animals
Karpathy says the ghost framing helps him think about models as statistical simulations shaped by pretraining and reinforcement learning, rather than as animal intelligences. He says yelling at them will not change how they work; the framing mainly encourages suspicion and careful evaluation rather than offering a fixed recipe for improvement.
- 25:15Infrastructure needs to serve agents
Karpathy says much of today’s software and documentation is built for humans, and argues for agent-native tools and infrastructure. He wants users to describe a task to an agent and have it handle steps such as deploying Menugen, without manual service configuration. He also anticipates agents representing people and organizations and handling some coordination with one another.
- 27:40Understanding remains essential
Karpathy says people may outsource thinking but cannot outsource understanding: someone still has to decide what to build, why it matters, and how to direct agents. He uses LLM-generated knowledge bases to reorganize information and prompt new insights, but says humans remain responsible for understanding what the models do not.
Published Apr 29, 2026.
Transcript · 25 min read
Training Data, hosted by Stephanie Zhan, with Andrej Karpathy.
Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.
we're so excited for our very first special guest he has helped build modern AI then explain modern AI and then occasionally rename modern AI he actually helped co-found OpenAI right inside of this office was the one who actually got autopilot working at Tesla back in the day and he has a rare gift of making the most complex technical shifts feel both accessible and inevitable. You all know him for having coined the term vibe coding last year, but just in the last few months, he said something even more startling, that he's never felt more behind as a programmer. That's where we're starting today.
Thank you, Andre, for joining us. Yeah, hello. Excited to be here and to kick us off. Okay, so just a couple months ago, you said that you've never felt more behind as a programmer. that's startling to hear from you of all people can you help us unpack that was that feeling exhilarating or unsettling yeah a mixture of both for sure well first of all I guess like as many of you I've been using agentic tools like lot code adjacent things for a while maybe over the last year as it came out and it was very good at you know chunks of code and sometimes it would mess up and you have to edit them and it was kind of helpful and then I would say December was this clear point where for me, I was on a break so I had a bit more time.
I think many other people are similar. And I just started to notice that with the latest models, the chunks just came out fine. And then I kept asking for more and just came out fine. And then I can't remember the last time I corrected it. And then I just trusted the system more and more. And then I was vibe coding. And so it was kind of a, I do think that it was a very stark transition. I think that a lot of people actually, I tried to stress this on Twitter and or X because I think a lot of people experienced AI last year as chat GPT adjacent thing but you really had to look again and you had to look as of December because things have changed fundamentally and especially on this like agentic coherent workflow that really started to actually work and so I would say that yeah it was just that realization that really had me go down their whole rabbit hole of just, you know, infinity side project.
Read the whole transcript
My side projects folder is like extremely full with lots of random things and just by coding all the time. So yeah, that kind of happened in December, I would say. And I was looking at the repercussions of that since.
“your programming now turns to prompting. And what's in the context window is your lever over the interpreter”
I love the menu gen progression and dichotomy that you laid out. And I think even I'm sure many folks here followed your own progression of programming from last October to early January, February this year. If you extrapolate that further, what is the 2026 equivalent for building websites in the 90s, building mobile apps in the 2010s, building SaaS in the last cloud era, what will look completely obvious in hindsight that is still mostly unbuilt today? Well, going with the example of Menugen, I guess, so a lot of this code shouldn't exist and it's just neural networks doing most of the work.
I do think that the extrapolation looks very weird because you could basically imagine, I don't think, yeah, so you could imagine completely neural computers in a certain sense. You feed raw videos, like imagine a device He takes raw videos or audio into basically what's a neural net and uses diffusion to render a UI that is kind of like, you know, unique for that moment in a certain sense. And I kind of feel like in the early days of computing, actually, people were a little bit confused as to whether computers would look like calculators or computers would look like neural nets. And in the 50s and 60s, it was not really obvious which way it would go.
And, of course, we went down the calculator path and ended up building classical computing. And then neural nets are currently running virtualized on existing computers. But you could imagine I think that a lot of this will flip and that the neural net becomes kind of like the host process and The CPU has become kind of like the co-processor So we saw the diagram of you know intelligence compute is going to of neural networks is going to take over and become the dominant Spend of flops so you could imagine something really weird and foreign when where neural nets are doing most of the heavy lifting They're using tool use as just like, you know Historical appendage for some kinds of like deterministic tasks, but what's really running the show is these neural nets that are networked in a certain way.
So you can imagine something extremely foreign as the extrapolation,
If you were a founder today and thinking about building a company you are trying to solve a problem that you think is tractable, something that is a domain that is verifiable but you look around you think oh my gosh well the labs have really really started getting to escape velocity in the ones that seem most obvious math coding and others what would your advice be to the founders in the audience so I think maybe that comes to the previous question of I do think that verifiability because it let me think so verifiability makes something tractable in the current paradigm because you can throw a huge amount of RL at it. So maybe one way to see it is that that remains true even if the labs are not focusing on it directly. So if you are in a verifiable setting where you could create these RL environments or examples then that actually sets you up to potentially do your own fine-tuning and you might benefit from that. But that is fundamentally technology that just works.
You can pull a lever if you have huge amount of diverse data sets of RL environments, et cetera, you can use your favorite fine-tuning framework and pull the lever and get something that actually works pretty well. So I don't know what the examples of this might be, but I do think there are some very valuable reinforcement learning environments that people could think of that I think are not part of the... Yeah, I don't want to give away the answer, but there is one domain that I think is very, oh, okay, sorry. I don't mean to vague post on the stage, but there are some examples of this. On the flip side, what do you think still feels automatable, only from a distance?
I do think that ultimately almost everything can be made verifiable to some extent, some things easier than others. Because even for things like writing or so on, you can imagine having a council of LLM judges and probably get something reasonable from this kind of an approach. So it's more about what's easy or hard. So I do think that ultimately, yeah, I think... Everything. Everything is automatable.
“agentic engineering is about preserving the quality bar of what existed before in professional software.”
“You basically still have to be in charge of the aesthetics, the judgment, the taste, and a little bit of oversight.”
You wrote a little bit about this with a very thought-provoking piece around animals versus ghosts. And the idea is that we're not building animals. We are summoning ghosts. And these are jagged forms of intelligence that are shaped by data and reward functions, but not by intrinsic motivation, or fun, or curiosity, or empowerment, things that kind of came about via evolution. Why does that framing matter, and what does it actually change about how you build and deploy and evaluate or even trust them? Yeah, so, yeah, I think the reason I wrote about this is because I'm trying to wrap my head around what these things are, right, because if you have a good model of what they are or are not, then you're going to be more competent at using them, and I do think that I don't know if it has, I'm not sure if it actually has like real power. I think it's a little bit of philosophizing, but I do think that, I think it's just coming to terms with the fact that these things are not, you know, animal intelligences, like if you yell at them, they're not going to work better or worse, or it doesn't have any impact.
and it's all just kind of like these statistical simulation circuits where the the substrate is pre-training, so like statistics, and then but then there's RL bolting on top so it kind of like increases the appendages and... Maybe it's just kind of like a mindset of what I'm coming into or what's likely to work or not likely to work or how to modify it. But I don't actually, I don't know that I have like, here are the five obvious outcomes of how to make your system better. It's more just being suspicious of it and figuring it out over time. That's where it starts. Okay, so you are so deep in working
“What is the thing I should copy paste to my agent?”
but yeah, I think everyone here is excited about that.
I really like the visual analogy of sensors and actuators. I actually hadn't thought about that. It's super interesting. Right.
“you can outsource your thinking but you can't outsource your understanding”
