All Segments
Andrej Karpathy: Software Is Changing (Again)
32 segments available
Andrej Karpathy's keynote on June 17, 2025 at AI Startup School in San Francisco. Slides provided by Andrej: https://drive.google.com/file/d/1a0h1mkwfmV2PlekxDN8isMrDA5evc4wW/view?usp=sharing Chapters: 00:00 - Intro 01:25 - Software evolution: From 1.0 to 3.0 04:40 - Programming in English: Rise of Software 3.0 06:10 - LLMs as utilities, fabs, and operating systems 11:04 - The new LLM OS and historical computing analogies 14:39 - Psychology of LLMs: People spirits and cognitive quirks 18:22 - Designing LLM apps with partial autonomy 23:40 - The importance of human-AI collaboration loops 26:00 - Lessons from Tesla Autopilot & autonomy sliders 27:52 - The Iron Man analogy: Augmentation vs. agents 29:06 - Vibe Coding: Everyone is now a programmer 33:39 - Building for agents: Future-ready digital infrastructure 38:14 - Summary: We’re in the 1960s of LLMs — time to build Drawing on his work at Stanford, OpenAI, and Tesla, Andrej sees a shift underway. Software is changing, again. We’ve entered the era of “Software 3.0,” where natural language becomes the new programming interface and models do the rest. He explores what this shift means for developers, users, and the design of software itself— that we're not just using new tools, but building a new kind of computer. More content from Andrej: https://www.youtube.com/@AndrejKarpathy Thoughts (From Andrej Karpathy!) 0:49 - Imo fair to say that software is changing quite fundamentally again. LLMs are a new kind of computer, and you program them *in English*. Hence I think they are well deserving of a major version upgrade in terms of software. 6:06 - LLMs have properties of utilities, of fabs, and of operating systems → New LLM OS, fabbed by labs, and distributed like utilities (for now). Many historical analogies apply - imo we are computing circa ~1960s. 14:39 - LLM psychology: LLMs = "people spirits", stochastic simulations of people, where the simulator is an autoregressive Transformer. Since they are trained on human data, they have a kind of emergent psychology, and are simultaneously superhuman in some ways, but also fallible in many others. Given this, how do we productively work with them hand in hand? Switching gears to opportunities... 18:16 - LLMs are "people spirits" → can build partially autonomous products. 29:05 - LLMs are programmed in English → make software highly accessible! (yes, vibe coding) 33:36 - LLMs are new primary consumer/manipulator of digital information (adding to GUIs/humans and APIs/programs) → Build for agents! Some of the links: - Software 2.0 blog post from 2017 https://karpathy.medium.com/software-2-0-a64152b37c35 - How LLMs flip the script on technology diffusion https://karpathy.bearblog.dev/power-to-the-people/ - Vibe coding MenuGen (retrospective) https://karpathy.bearblog.dev/vibe-coding-menugen/ Apply to Y Combinator: https://ycombinator.com/apply Work at a startup: https://workatastartup.com
Segments Timeline
Welcome to the Future of Software
Andrej Karpathy introduces his keynote on the evolution of software in the AI era, emphasizing the unique timing for students entering the industry. He highlights the fundamental changes in software over the past 70 years and the rapid evolution seen in recent years, setting the stage for a discussion on Software 3.0.
"Please welcome former director of AI Tesla Andre Carpathy. Hello. Wow, a lot of people here. Hello. Um, okay. Yeah. So I'm excited to be here today to talk to you about software in the era of AI. And ..."
From Software 1.0 to 3.0
Karpathy explains the transition from Software 1.0, which involves traditional coding, to Software 2.0, where neural networks are trained rather than explicitly programmed. He introduces the concept of Software 3.0, where large language models (LLMs) are programmed using natural language, marking a significant shift in how software is developed.
"for 70 years. And then it's changed I think about twice quite rapidly in the last few years. And so there's just a huge amount of work to do a huge amount of software to write and rewrite. So let's ta..."
The Rise of Programmable Neural Networks
In this segment, Karpathy discusses how neural networks have evolved from fixed-function computers to programmable entities through LLMs. He emphasizes the importance of prompts as a new form of programming, showcasing how this shift allows for more accessible software development using English.
"I think the hugging face is basically equivalent of GitHub in software 2.0. And there's also model atlas and you can visualize all the code written there. In case you're curious, by the way, the giant..."
The New Programming Paradigm
Karpathy elaborates on the implications of programming in English, noting that this new paradigm not only changes how we write code but also integrates natural language into the coding process. He reflects on the growing presence of English in code repositories and the significance of this evolution for future developers.
"difference if you're doing sentiment classification for example you can imagine writing some uh amount of Python to to basically do sentiment classification or you can train a neural net or you can pr..."
Lessons from Tesla Autopilot
Drawing from his experience at Tesla, Karpathy shares insights on how the autopilot software transitioned from traditional C++ code to neural networks. He illustrates how Software 2.0 has effectively replaced much of the original code, showcasing the transformative power of neural networks in real-world applications.
"blew my mind a few uh I guess years ago now I tweeted this and um I think it captured the attention of a lot of people and this is my currently pinned tweet uh is that remarkably we're now programming..."
Navigating Multiple Programming Paradigms
Karpathy emphasizes the importance of being fluent in multiple programming paradigms as the industry evolves. He discusses the need for developers to understand when to use Software 1.0, 2.0, or 3.0, and the implications of these choices for software functionality and design.
"kind of software and it's eating through the stack. We have three completely different programming paradigms and I think if you're entering the industry it's a very good idea to be fluent in all of th..."
LLMs as Utilities and Fabs
In this segment, Karpathy explores the analogy of LLMs as utilities, comparing their operational model to electricity. He discusses the capital expenditure involved in training LLMs and how this investment parallels the infrastructure needed for utility services, highlighting the growing demand for reliable AI services.
"between these paradigms. So what I wanted to get into now is first I want to in the first part talk about LLMs and how to kind of like think of this new paradigm and the ecosystem and what that looks ..."
The Intelligence Brownout
Karpathy discusses the implications of LLM outages, likening them to an intelligence brownout where reliance on these models leads to a decrease in overall cognitive capabilities. He reflects on the societal impact of such dependencies and the need for robust AI systems.
"very utility- like demands out of this API we demand low latency high uptime consistent quality etc. In electricity, you would have a transfer switch. So you can transfer your electricity source from ..."
LLMs as Operating Systems
Karpathy draws parallels between LLMs and operating systems, noting that they represent complex software ecosystems rather than simple utilities. He discusses the competitive landscape of LLM providers and the potential for open-source alternatives to emerge, similar to the evolution of operating systems.
"have properties of utilities. I think it's also fair to say that they have some properties of fabs. And the reason for this is that the capex required for building LLM is actually quite large. Uh it's..."
The Future of LLM Ecosystems
In this concluding segment, Karpathy speculates on the future of LLM ecosystems, suggesting that as these models become more complex, they will require sophisticated tool use and multimodal capabilities. He emphasizes the importance of understanding these developments for anyone entering the tech industry.
"strong kind of analogies to operating systems. Uh in that this is not just electricity or water. It's not something that comes out of the tap as a commodity. uh this is these are now increasingly comp..."
LLMs as Operating Systems
In this segment, Karpathy elaborates on the analogy of LLMs as new operating systems. He compares the structure of LLMs to traditional computing systems, explaining how they orchestrate memory and compute for problem-solving. He discusses the implications of this analogy, suggesting that LLMs are becoming increasingly complex and are central to the future of computing.
"because these are just simple LLMs, but we're starting to see that these are going to get a lot more complicated. It's not just about the LLM itself. It's about all the tool use and the multiodalities..."
The 1960s of LLM Computing
Karpathy reflects on the current state of LLM computing, likening it to the 1960s era of computing where resources were centralized and expensive. He discusses the limitations of personal computing with LLMs and the potential for future developments in this space. He encourages the audience to think about how they might contribute to this evolving landscape.
"uh more analogies that I think strike me is that we're kind of like in this 1960sish era where LLM compute is still very expensive for this new kind of a computer and that forces the LLMs to be centra..."
LLMs: A New Kind of Computer
In this segment, Karpathy describes LLMs as a new kind of computer, emphasizing their unique properties and capabilities. He discusses how LLMs differ from traditional operating systems and the implications of their widespread availability. He highlights the democratization of technology, where LLMs are accessible to everyone, not just corporations or governments.
"Maybe some of you get to invent what what this is or how it works or uh what this should what this should be. Maybe one more analogy that I'll mention is whenever I talk to Chach or some LLM directly ..."
The Psychology of LLMs
Karpathy introduces the concept of LLMs as 'people spirits,' explaining their emergent psychology and cognitive quirks. He draws parallels between LLMs and human-like memory, referencing the movie 'Rainman' to illustrate their capabilities and limitations. He discusses the dual nature of LLMs, possessing superhuman abilities in some areas while also exhibiting cognitive deficits.
"and I think it informs maybe some of the uses of how we want to use this technology or like where are some of the first apps and so on. So, in summary so far, LLM labs LLMs. I think it's accurate lang..."
Cognitive Limitations of LLMs
In this segment, Karpathy delves deeper into the cognitive limitations of LLMs, discussing their propensity for hallucination and memory issues. He compares LLMs to characters from movies like 'Memento' and '51st Dates,' highlighting their fixed weights and context window challenges. He also addresses security concerns related to LLMs, such as susceptibility to prompt injection and data leakage.
"chunk. Um and um this simulator of course is is just is basically there's some weights involved and we fit it to all of text that we have on the internet and so on. And you end up with this kind of a ..."
Cognitive Limitations of LLMs
Karpathy elaborates on the cognitive limitations of LLMs, comparing them to characters in films like 'Memento' and '50 First Dates' who struggle with memory retention. He points out that LLMs lack the ability to consolidate knowledge over time, which affects their performance and reliability in tasks requiring contextual understanding.
"expertise over time. LLMs don't natively do this and this is not something that has really been solved in the R&D of LLM. I think um and so context windows are really kind of like working memory and y..."
Partial Autonomy in LLM Applications
In this segment, Karpathy introduces the concept of partial autonomy in LLM applications, using coding as an example. He advocates for dedicated applications that leverage LLM capabilities, such as Cursor, which enhances user experience by integrating LLMs for context management and task orchestration.
"So what I want to switch to now is talk about the opportunities of how do we use these models and what are some of the biggest opportunities. This is not a comprehensive list just some of the things t..."
The Importance of User Interfaces
Karpathy emphasizes the significance of user interfaces in LLM applications, arguing that graphical user interfaces (GUIs) are essential for effective interaction. He explains how GUIs facilitate easier auditing of LLM outputs and improve user experience by making complex tasks more manageable.
"properties that I think are um useful across all the LLM apps. So in particular, you will notice that we have a traditional interface that allows a human to go in and do all the work manually just as ..."
The Autonomy Slider Concept
In this segment, Karpathy discusses the 'autonomy slider' feature in LLM applications, which allows users to control the level of autonomy granted to the AI. He illustrates this concept with examples from Cursor and Perplexity, highlighting how users can adjust the AI's involvement based on task complexity.
"audit the work of these fallible systems and to go faster. I'm going to come back to this point a little bit uh later as well. And the last kind of feature I want to point out is that there's what I c..."
Human-AI Collaboration Dynamics
Karpathy explores the dynamics of human-AI collaboration, stressing the importance of keeping humans in the loop. He discusses the need for effective verification processes to ensure that AI outputs are reliable and secure, advocating for a balanced approach to leveraging AI capabilities.
"So, I guess my question is I feel like a lot of software will become partially autonomous. I'm trying to think through like what does that look like? And for many of you who maintain products and serv..."
Best Practices for AI-Assisted Coding
In this segment, Karpathy shares insights on best practices for AI-assisted coding. He emphasizes the importance of clear prompts to improve verification success and discusses the need for incremental changes in coding workflows to maintain control over AI outputs.
"accessible to LLMs. So, one thing I want to stress with a lot of these LLM apps that I'm not sure gets as much attention as it should is um we we're now kind of like cooperating with AIS and usually t..."
Education in the Age of AI
Karpathy reflects on the implications of AI and LLMs for education. He proposes the need for distinct applications for course creation and delivery, emphasizing the importance of structured interactions with AI to enhance learning outcomes and maintain effective educational practices.
"assisted coding. If I'm just bite coding everything is nice and great but if I'm actually trying to get work done it's not so great to have an overreactive uh agent doing all this kind of stuff. So th..."
Keeping AI on a Leash
Karpathy explores the concept of keeping AI 'on a leash' in educational contexts. He suggests that effective AI applications should include structured courses and syllabi to prevent the AI from getting lost in vague prompts. This approach enhances the reliability of AI in educational settings.
"work as well I'm currently interested in uh what education looks like in um together with kind of like now that we have AI uh and LLMs what does education look like? And I think a a large amount of th..."
Lessons from Tesla's Autonomy
Drawing from his experience at Tesla, Karpathy discusses partial autonomy in AI products. He reflects on the evolution of self-driving technology and the importance of human oversight, cautioning against overestimating the capabilities of autonomous agents while highlighting the ongoing challenges in achieving full autonomy.
"autonomy product and shares a lot of the features like for example right there in the instrument panel is the GUI of the autopilot so it's showing me what the what the neural network sees and so on an..."
The Iron Man Analogy
Karpathy uses the Iron Man suit as a metaphor for the dual nature of AI technology—both as an augmentation and an autonomous agent. He emphasizes the need for a balance between building augmentations and agents, advocating for products that incorporate partial autonomy while maintaining human involvement.
"so I think like like this is software is really tricky I think in the same way that driving is tricky and so when I see things like oh 2025 is the year of agents I get very concerned and I kind of fee..."
The Future of Programming: Vibe Coding
In this segment, Karpathy introduces the concept of 'vibe coding,' where programming becomes accessible to everyone through natural language. He discusses the implications of this shift, highlighting how it democratizes software development and allows individuals without extensive training to create applications.
"autonomy products. And these products have custom gueies and UIUX. And we're trying to um and this is done so that the generation verification loop of the human is very very fast. But we are not losin..."
Building Menu Gen: A Vibe Coding Experience
Karpathy shares his personal experience of creating the app 'Menu Gen' through vibe coding. He describes the ease of building the initial prototype but contrasts it with the complexities of deployment and integration, illustrating the challenges developers face in making their applications operational.
"and which tweet like fizzles and no one cares. And I thought that this tweet was going to be the latter. I don't know. It was just like a shower of thoughts. But this became like a total meme and I re..."
Building for Agents: A New Paradigm
In this segment, Karpathy discusses the emergence of AI agents as new consumers and manipulators of digital information. He proposes the idea of creating software that directly interacts with these agents, using examples like robots.txt for web crawlers. This shift represents a significant change in how developers approach software design, focusing on making systems more accessible to AI.
"the reason for this is this was just really annoying. Um, so for example, if you try to add Google login to your web page, I know this is very small, but just a huge amount of instructions of this cle..."
Making Documentation LLM-Friendly
Karpathy emphasizes the importance of adapting documentation for LLMs, advocating for markdown formats that are easier for AI to understand. He shares his experience using LLMs to generate code from documentation, highlighting the potential for improved accessibility and usability in software development. This segment illustrates the need for a paradigm shift in how technical documentation is created.
"domain and you can instruct uh or like advise I suppose um uh web crawlers on how to behave on your website in the same way you can have maybe lm.txt txt file which is just a simple markdown that's te..."
The Future of LLMs and Documentation
In this segment, Karpathy discusses the necessity of revising documentation to eliminate instructions that LLMs cannot execute. He cites examples of companies like Vercel that are transitioning their documentation to be more LLM-friendly. This highlights the evolving relationship between developers and AI, as well as the importance of creating tools that facilitate this interaction.
"to understand. This is great. Um maybe one simple example from from uh my experience as well. Maybe some of you know three blue one brown. He makes beautiful animation videos on YouTube. Yeah, I love ..."
Unlocking LLM Potential with New Tools
Karpathy introduces various tools that enhance LLM accessibility, such as GitHub repositories that can be ingested by LLMs for easier interaction. He discusses the potential of these tools to streamline the development process and improve the usability of AI in software applications. This segment showcases the innovative approaches being taken to integrate LLMs into the development workflow.
"then, of course, there's a model context protocol from Enthropic. And this is also another way, it's a protocol of speaking directly to agents as this new consumer and manipulator of digital informati..."
The Iron Man Suit Analogy: Future of LLMs
In the concluding segment, Karpathy reflects on the future of LLMs, likening their development to the early days of operating systems. He emphasizes the need for infrastructure adjustments to accommodate AI agents and the exciting possibilities that lie ahead. This segment encapsulates the overarching theme of the talk: the transformative potential of LLMs in software development and the collaborative future between humans and AI.
"it. One more note I wanted to make is that it is absolutely possible that in the future LLMs will be able to this is not even future this is today they'll be able to go around and they'll be able to c..."