Stanford CS153 Frontier Systems | The AI Native Company: How One Founder Becomes a 1000x Engineer

Introduction to CS 153 and Y Combinator's Role

• The class CS 153 is a composite of several different classes taught at Stanford by Silicon Valley leaders, including Peter Thiel's class on how to start a startup, which became the book "Zero to One", and Terry Winograd's class on computers and the open society 10s.
• Garry Tan, a founder who learned from these classes, is now back to talk about his work and update the YC philosophy with Diana Hu, making it a "closed the loop moment" 2m6s.
• The class CS 153 is a systems class that covers topics from the chip layer to the application layer, and there is a full rewrite of systems going on to unblock bottlenecks on frontier progress 4m42s.
• One of the things needed to unblock bottlenecks is capital, and YC has a system to try and scale the deployment of capital in Silicon Valley, similar to the system developed by Mark and Ben Horowitz 6m15s.

The SAFE and Standardization in Venture Capital

• The compute bottleneck is a major issue today because we are in the pre-standardization of compute era, similar to the pre-standardization of electricity during the Industrial Revolution, and standards and institutions are needed to make compute a stable resource 8m10s.
• In the capital world, the introduction of the SAFE by Paul Graham and Jessica Livingston was a new standard for how capital should be allocated, and it has had a profound impact on the venture capital industry 12m20s.
• The Simple Agreement for Future Equity, also known as the SAFE, was introduced by Y Combinator as a two-page legal document that standardized seed stage funding for early-stage startups, and this move had a significant impact on the history of Silicon Valley 10s.
• The introduction of the SAFE was a response to the venture capital bottleneck that existed at the time, where it was difficult for innovators to get time with venture capitalists and secure good deals, despite the reduced marginal cost of innovation due to the rise of the cloud and SaaS era 2m6s.
• The SAFE became a standard for funding early-stage startups, and Y Combinator's enforcement of it helped to standardize seed stage funding, with its impact still being felt today, and companies like Amp are looking to Y Combinator as a spiritual ancestor for their work in standardizing agreements for future compute 4m6s.

Systems Design and Unblocking Bottlenecks

• The concept of systems design is not limited to engineering, but can be applied to any domain to accelerate progress and unblock bottlenecks, and this idea is being explored in the context of the lecture and the work being done by Y Combinator and other companies 6m42s.
• Garry Tan, a Stanford class of '03 alum, introduced himself and discussed the significance of the current time, where new standards are being established, and he expressed his hope that the students in the room will be the ones to establish these new standards and build the cognitive layer for society 10m10s.
• The lecture is focused on the idea that the current generation of students will be building the foundation for the rest of society, and that this will involve creating a new cognitive layer, with code and other technologies playing a key role in this process 12m20s.

The Impact of AI on Innovation and Growth

• The current time is considered exciting due to the capabilities being unlocked by AI, with companies in portfolios experiencing unprecedented growth, going from zero to tens of millions of dollars in revenue within a year, which would have taken four or five years before, and having access to hundreds of millions of dollars in capital 10s.
• AI is expected to change the unit of production, and building a company now means being AI native, which involves humans working in concert with agents, memory, and evaluation, as well as a customer loop, and this concept will be explained in detail to help implement these ideas and remake society 2m6s.
• The personal story of a founder is shared, who in 2008 got into YC, raised $4 million, hired 10 people, created Posterous, and sold it to Twitter for $20 million three years later, and now with a $200 a month Cloud Code Max plan, anyone can create similar software in just 5 days 4m30s.
• A six-person team can now hit $10 million in revenue using the concepts being discussed, and the story of creating Gary's List and Gstack is mentioned as an example of the speed and productivity that can be achieved with AI 6m40s.

AI Coding Agents and Productivity

• The concept of AI coding agents is introduced, which can make engineers 10x to 100x more productive, and at Anthropic, they are about a thousand times as productive as Googlers were in 2005, with the example of writing a million lines of code using Claude code 8m30s.
• The common misconceptions about AI coding, such as it being "AI slop" or only producing boilerplate code, are addressed, and the importance of creating a software factory and achieving high test coverage to get to production is highlighted 10m50s.
• The plan-eng-review skill is used about 20 times a day to achieve 80-90% test coverage, ensuring that the shipped product is usable and reliable, and this approach has been controversial, with some people taking issue with the focus on lines of code (LOC) as a metric 10s.
• LOC can be a flawed metric, but having tests and measuring whether the product works for users and customers is a more important indicator of success, and the goal is to write dense and concise code that serves its purpose 42s.
• The experience of achieving 87,000 stars on one project and 13,000 stars on another, with over 100,000 GitHub stars and 15,000 daily users, has led to the realization that a software factory approach can be effective, and tools like G stack are enabling this 2m6s.

Skills and Personas in AI Development

• The discovery of the usefulness of pulling out specific personas from the latent space, such as the Office Hour skill, which distills the essence of Y Combinator's office hours with partners, has been a key insight, and this skill has been distilled into a potent and open-source tool 4m30s.
• Other skills, such as plan CEO review, have been developed to help with product development, asking questions like "what is the 10x version of this product?" and "what is the platonic ideal of this?", and these skills are designed to help create a roadmap for building a product that is a straight line from the current state to the desired outcome 6m15s.
• The approach of using coding agents and pulling out the latent space for a particular vibe or thing has made it easier and better to use the product, and has led to the development of a range of skills that can be used to improve product development and achieve success 8m0s.

Challenges and Realities of AI Development

• The concept of "boiling the ocean" refers to taking on a task that seems too large or overwhelming, but with the help of coding agents and models like Claude, it's possible for one person to do the work of 500 to 1,000 people, challenging traditional expectations of what a founder or small team can accomplish 10s.
• The models themselves have not yet caught up to this new reality, often providing inaccurate estimates of the time required to complete a task, as seen when Claude estimates a task will take 3 weeks but is actually completed in about an hour 2m6s.
• The use of G stack and open-source tools like OpenClaw and Hermes Agent are teaching new primitives on how to think about code, markdown, and their interaction to achieve real work, and are being explored by companies like YC 4m42s.
• Building agentic systems can be challenging, and they often break due to incorrect usage of deterministic work, such as code, or latent work, such as tasks that should be handled by a Large Language Model (LLM), highlighting the need to combine these approaches effectively 6m15s.
• A concrete example of this challenge is curating the experience of people at events, which can be easily done with tools like Claude or ChatGPT for small groups, but becomes difficult for larger groups, requiring a combination of latent and deterministic approaches 8m30s.

Tools and Frameworks for AI Automation

• The concept of a "skill" is being redefined with the help of LLMs, allowing for real work to be done with tools like skill files, which are essentially runbooks or markdown files that can be used to automate tasks and investigate complex topics 12m0s.
• The concept of creating a system where any human being or agent can look at a set of steps and understand what to do, and even make it call code, is a simple yet powerful idea that is being utilized in tools like Open Claw and Hermes 10s.
• The cloud code revolution has changed the way people write code, and similarly, Open Claw and Hermes are revolutionizing the way non-technical and process-oriented tasks are handled, allowing users to automate tasks such as making calls or booking appointments 2m6s.
• The use of resolvers is an important aspect of creating a great agent, as it allows the agent to load instructions only when needed, making it more efficient and effective, and enabling users to create custom skills and code paths, such as a skill pack for checking signatures 4m30s.
• The ability to create custom code and automate tasks is empowering, and tools like Open Claw and Hermes are making it possible for users to create their own custom solutions, such as fixing issues with Claude.md, without having to wait for others to develop and ship them 6m20s.
• The concept of a "skill pack" is being used to create custom code paths for specific tasks, such as checking signatures, and is an example of how users can create their own custom solutions using tools like Open Claw and Hermes 8m40s.

Skill Development and Validation

• The concept of "skillify" is introduced, which involves taking a task that has been completed once and turning it into a reusable skill that can be applied in various situations, allowing for a higher level of abstraction 10s.
• To create a skill, one must write the code, test it, and then refine it through various steps, including writing unit tests, LLM evals, integration tests, and resolver triggers, to ensure the system works as intended 2m6s.
• The process of skillifying a task is complex and involves multiple steps, similar to how compliance teams work in finance organizations, where a significant amount of effort is dedicated to ensuring the system functions correctly 2m6s.
• The importance of testing and validation is emphasized, including the use of unit tests, LLM evals, and integration tests, as well as the need for a resolver trigger and an LLM as judge eval to ensure the skill is triggered correctly 4m30s.
• The concept of a "check resolvable" is mentioned, which is important for avoiding duplicate skills and ensuring that the system is efficient 6m20s.

G Brain and Knowledge Systems

• A new project called G Brain is introduced, which is a three-layer memory system built on top of a knowledge wiki, and is designed to capture and track knowledge and beliefs over time 8m10s.
• The G Brain system includes features such as vector search, ARR fusion, backlinks, and a graph database, and is intended to track the development of ideas and hunches over time, allowing for the identification of patterns and connections 10m0s.
• The goal of G Brain is to create a knowledge system that can capture and track the evolution of ideas and beliefs, and to provide a platform for founders and individuals to develop and refine their thoughts and concepts 12m0s.

AI Native Companies and Dynamic Ontologies

• The concept of building an AI native company involves creating a specific schema for a use case, and the goal is to make it fully dynamic to support various users, including researchers, journalists, and politicians, with different schemas 10s.
• The idea of a dynamic ontology is being explored, which is a concept learned from Palantir, and it is necessary to support all types of users, whether they are building for themselves or for others 2m6s.
• The primitives being learned include skills, resolvers, and check resolvables, which map to company concepts, such as employees with capabilities, org charts, and internal processes, as well as audit and compliance 2m6s.
• A skill is equivalent to an employee with a capability, a resolver is similar to an org chart, and a check resolvable ensures that the resolver works for a set of tasks, which is comparable to audit and compliance in human organizations 4m30s.
• The concept of a trigger eval is also being explored, which is a latent space squishy operation that needs to be checked, and it is comparable to performance reviews in an organization 6m40s.

Open Loop vs. Closed Loop Systems

• Building an AI native company requires changing how companies are run, from an open loop to a closed loop system, with a tight feedback loop, to avoid error accumulation and systems becoming more erroneous 10m0s.
• The difference between open loop and closed loop systems is important, with open loop systems being lossy and prone to error accumulation, and closed loop systems providing a concrete and tight feedback loop 12m0s.
• The concept of closed-loop systems, such as PID controllers, can be applied to companies using AI, allowing them to take lossy information and turn it into a closed-loop system, which can improve decision-making and efficiency 10s.
• In traditional companies, information is often stored in people's heads, side conversations, and meeting notes, which can be lossy, but with AI, this information can be embedded into the decision-making process using agents like Hermes or open claw, which can have read access to every artifact the company produces 1m5s.

AI in Company Operations and Organization

• These agents can connect to various tools, such as GitHub, Discord, and meeting recordings, to suggest the best next items to work on or bug fixes, and can help build a self-healing system 2m6s.
• Companies that have implemented this approach have seen significant improvements, such as one employee generating $1-2 million in revenue, and cutting sprint times in half while producing 10x more work 4m30s.
• This approach can also lead to a flatter organization with less need for middle management, and three main roles: individual contributors who build and ship, direct responsible individuals (DRIs) who orchestrate and own outcomes, and AI founders who embody the use of AI tools to drive the company forward 6m40s.
• The concept of a DRI, or direct responsible individual, is familiar to some companies, such as Apple, where every outcome can be traced back to a particular owner, and the DRI works with individual contributors to make sure goals are met 8m10s.
• The role of an AI founder is to live at the edge of the future, trying all the latest tools and driving the company to run fast and adapt to changing circumstances 10m30s.

The Role of AI Founders and DRIs

• The revolution in agentic coding, which occurred with the release of Claude 4.5, has enabled innovations that can be brought into companies, but only if the founders are at the edge and building, allowing them to leverage these advancements 10s.
• To build agentic systems and avoid "AI slop," the concept of "taste" is crucial, as it cannot be delegated and is what will be durable, referring to the ability to discern what is good or bad, which is essential for evaluating systems and building good products 2m6s.

The Importance of Taste and Evaluation

• The cost of coding, or shipping code, is going to zero, but the cost of having a good "taste" to build something good is not, and this is what will set companies apart, with the actual judge of whether something is good being whether users really want it 2m6s.
• Generic benchmarks are not sufficient to determine whether a product or agent is working, as they do not take into account the specific needs and goals of the product, and instead, human evaluation and feedback are necessary to determine whether an agent is working correctly 4m30s.
• To evaluate an agent, it is necessary to go into the details and assess whether it followed instructions, provided correct answers, preserved customer trust, and met business goals, among other factors, which requires human judgment and cannot be fully automated 5m30s.
• The use of cross-modal evaluation, which involves having multiple models evaluate inputs and outputs and provide feedback, can help improve the performance of agents and enable them to learn and improve over time 8m0s.
• The process of capturing traces, or data, to evaluate and improve agents is contest-dependent and varies depending on the product being built, such as video applications, speech applications, or B2B SaaS, each requiring different approaches 10m0s.

Improving Systems Through Feedback and Traces

• To improve a system, it is necessary to convert failure cases into evaluations, detect when they fail, and replay them constantly into the system to self-heal and improve the system and prompts automatically, which is what Garry is describing 10s.
• The principles of building these systems can be applied by anyone, and it is encouraged to experiment with tools like Hermes, Claude, and G brain, which offer 40 skills to test and try out, and even create and open-source your own skills 1m42s.

Meta-Prompting and Human-Machine Collaboration

• The concept of meta-prompting not only applies to machines but also to humans, where individuals need to meta-prompt each other to improve and fuse with machines in a more profound way every day 2m6s.
• This is considered one of the best times in history to start a company, with many opportunities to grow and innovate, particularly by identifying painful workflows and becoming a forward deploy engineer to automate repetitive labor and handle complex domains 4m10s.

Examples of AI-Driven Startups and Growth

• Examples of companies that have achieved significant growth by applying these principles include Salient, Happy Robot, and Reduct, which have built agents to automate tasks in industries such as loan services, freight forwarding, and document processing 5m40s.
• To start a company in this fashion, it is often necessary to go undercover and learn the depths of an industry by shadowing or taking a job, and then automating repetitive labor and handling complex domains, as seen in the examples of Scale AI and Happy Robot 7m30s.
• The deployment of AI solutions in different industries is becoming increasingly prevalent, with companies like Anthropic posting graphs showing the growth of AI adoption across various sectors 9m40s.

Opportunities for AI Innovation and Entrepreneurship

• Many computer science graduates have a real fear of the job market after graduation, but there is a huge white space in various domains such as back office, finance, data, academics, cybersecurity, and customer service that has room for hundreds of AI unicorns to be started, potentially by these graduates 10s.
• The current state of AI penetration is at 50% according to a chart by Anthropic, indicating a significant opportunity for growth and innovation in other areas 42s.
• Historically, only the top 1% of companies at Y Combinator (YC) grew 10% week over week, but now companies are achieving an average growth of 3x within 3 months, which has never happened before in YC's history 2m6s.
• This new era of rapid growth and innovation presents an opportunity for individuals to create something with real impact, where customers will appreciate and pay for their products or services, and the customer base will grow by 10% every week 2m6s.
• The concept of a one-person frontier lab can evolve into a one-person company, and the necessary information to achieve this has been shared, making it possible for individuals to create their own successful companies 2m6s.