Jun 19, 2025 · 39m · y-combinator
Andrej Karpathy: Software Is Changing (Again) · Y Combinator
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In his keynote at Y Combinator's AI Startup School, Andrej Karpathy explores the historic shift toward Software 3.0 powered by Large Language Models. He analyzes LLM system architectures, cognitive limitations, human-in-the-loop application design, and the need for agent-first digital infrastructure.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
The partners as informed peer 0.0 Guest teaching 0.0 Guest disagreement 0.5 The partners pushing back 0.0
speaking balance: gold is the partners, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 24:38 Pushing Back on 'Year of the Agent' Hype
Karpathy challenges industry claims that full agentic autonomy is imminent, comparing it to the 12-year timeline of autonomous driving.
Hardest push from the partners ▶ 0:01 Introductory HousekeepingNo host pushback occurs in this solo lecture; the host only introduces Karpathy at the beginning.
Biggest teaching moment ▶ 9:30 Operating System Architecture of LLMsKarpathy systematically breaks down why LLMs should be viewed as 1960s-style cloud operating systems rather than simple commodities.
The partners hold their own ▶ 0:01 Host IntroductionThere is no host debate or substantive host interjection in this monologue format.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The partners as informed peer | Guest teaching | Guest disagreement | The partners pushing back | Why |
|---|---|---|---|---|---|---|
| The Evolution of Software: 1.0, 2.0, and 3.0 | 0 | 0 | 0 | 0 | This is a solo keynote lecture by Andrej Karpathy tracing the transition from Software 1.0 (explicit code) to 2.0 (weights) and 3.0 (prompted LLMs in English). As there is no host interaction, host-side metrics are zero. | |
| Part 1: Framing LLMs as Utilities, Fabs, and Operating Systems | 0 | 0 | 0 | 0 | Karpathy compares LLMs to utilities, semiconductor fabs, and 1960s-era centralized operating systems with time-sharing. The presentation proceeds without interruption or pushback. | |
| Part 2: Understanding LLM Psychology and Cognitive Deficits | 0 | 0 | 0 | 0 | Karpathy outlines LLM psychology, framing models as people spirits with vast memory but severe cognitive deficits like anterograde amnesia and jagged reasoning. | |
| Part 3: Partial Autonomy Apps and the Human-in-the-Loop | 0 | 0 | 2 | 0 | Karpathy advocates for partial autonomy over fully autonomous hype, noting that self-driving took over a decade and agents will require human-in-the-loop verification. | |
| Vibe Coding, Accessibility, and DevOps Bottlenecks | 0 | 0 | 1 | 0 | Karpathy discusses coining vibe coding and reflects on building Menugen, highlighting how UI/DevOps deployment remained a manual slog compared to prompt-driven coding. | |
| Building Infrastructure for Agents and Final Summary | 0 | 0 | 0 | 0 | Karpathy concludes by presenting methods to optimize web infrastructure and documentation for LLM agents (llms.txt, Markdown docs, MCP) before giving his final summary. |
Statements from this episode (16)
Hugging Face is the GitHub of the Software 2.0 era
“And now actually what we have is kind of like an equivalent of GitHub in the realm of software two point oh, and I think the hugging face is basically equivalent of GitHub in software two point oh, and there's also model Atlas, and you can visualize all the co…”
LLMs establish Software 3.0 where natural language prompts are programs
“And I think what's changed, and I think is a quite fundamental change, is that neural networks became programmable. With large language models. And so I see this as quite new, unique. It's a new kind of a computer. And so in my mind, it's worth giving it a new…”
Tesla Autopilot neural networks systematically replaced explicit C++ code
“As we made the autopilot better, basically the neural network grew in capability and size, and in addition to that, all the C++ code was being deleted and kind of like was And a lot of the kind of capabilities and functionality that was originally written in o…”
LLMs function as CPUs and operating systems orchestrating context memory
“LLMs are kind of like a new operating system, right? So, the LLM is a new kind of a computer. It's setting, it's kind of like the CPU equivalent. The context windows are kind of like the memory, and then the LLM is orchestrating memory and compute for problem …”
Modern LLM cloud infrastructure mirrors 1960s mainframe time-sharing
“We're kind of like in this 19 sixties-ish era where LLM compute is still very expensive for this new kind of a computer, and that forces the LLMs to be centralized in the cloud, and we're all just sort of thin clients that interact with it over the network, an…”
Mac Minis are uniquely suited for batch-one local LLM inference
“Mac minis, for example, are a very good fit for some of the LLMs, because it's all, if you're doing batch one inference, this is all super memory bound, so this actually works.”
LLMs invert historical tech diffusion by reaching consumers before enterprise
“LLMs, like, flip, they flip the direction of technology diffusion, ah, that is usually, ah, present in technology. So, for example, with electricity, cryptography, computing, flight, internet, GPS, lots of new transformative technologies that have not been aro…”
LLMs display jagged intelligence with superhuman skill alongside basic errors
“They display jagged intelligence, so they're going to be superhuman in some problem-solving domains, and then they're going to make mistakes that basically no human will make, like, you know, they will insist that 9.11 is greater than 9.9, or that there are tw…”
Continuous knowledge consolidation remains unsolved in LLM research and development
“LLMs don't natively do this, and this is not something that has really been solved in the R&D of LLMs, I think. And so context windows are really kind of like working memory, and you have to sort of program the working memory quite directly because they don't …”
AI handles generation while human productivity bottlenecks on verification speed
“We're now kind of like cooperating with AIs, and usually they are doing the generation, and we as humans are doing the verification. It is in our interest to make this loop go as fast as possible, so we're getting a lot of work done.”
Waymo autonomous vehicles still rely heavily on human teleoperation
“Like, you may see Waymo's going around, and they look driverless, but, you know, there's still a lot of teleoperation and a lot of human in the loop of a lot of this driving.”
2025 is not the year of AI agents, it is the decade
“When I see things like, oh, twenty-twenty-five is the year of agents, I get very concerned, and I kind of feel like, you know, this is the decade of agents, and this is going to be quite some time. We need humans in the loop.”
Developers should build Iron Man suits, not fully autonomous robots
“Working with valuable LLMs and so on, I would say, you know, it's less Iron Man robots and more Iron Man suits that you want to build. It's less like building flashy demos of autonomous agents and more building partial autonomy products”
Vibe coding will become a gateway drug to software development
“I think this will end up being like a gateway drug to software development.”
Generating code is easy; DevOps and deployment are the true bottleneck
“The code of the Vibe coding part, the code was actually the easy part of Vibe coding Menugen, and most of it actually was when I tried to make it real so that you can actually have authentication and payments and the domain name and the virtual deployment. Thi…”