Y Combinator Summaries & Notes

Early Days & Problem Identification Harshil was a coder with no initial interest in finance or startups. He encountered the difficulty of accepting online payments in India while building a side project. Noticed that it was easier to accept cash than digital payments, which contradicted the purpose of technology. Realized this was a significant, unsolved problem affecting many startups. Pivoting Strategy & GTM Initially planned to target educational institutes for fee payments but found low customer interest. Pivoted to serving startups, who were eager for digital payment solutions. This pivot was a crucial decision that led to early traction. The Regulatory Moat Razorpay faced a year-long wait for approvals and licenses before its first live transaction, a significant "gestation period" for a tech business. The complexity of regulations created a moat, deterring competitors due to the high barrier to entry. Regulations, though challenging, are fair and apply equally to all, fostering long-term trust and reliability. Conviction & Customer Focus Despite monthly doubts, the conviction came from direct customer feedback: founders confirmed the problem and the lack of solutions. "Make something people want" was the guiding principle; customer validation provided energy. Customer love after onboarding reinforced the belief in solving a critical problem. Navigating Crises: The Bank Pull-Out A bank partner suddenly stopped supporting Razorpay just before Demo Day, shutting down payments for 50+ live merchants. The team's fundamental principle was to maintain trust through transparent, human communication. They personally called every affected customer, explaining the situation and their actions, even enduring abuse. This crisis solidified the importance of human touchpoints and trust in B2B relationships, especially in finance. Long-Term Vision & Acquisition Offers Received early acquisition offers from global payment companies. Believed these companies underestimated India's complexity and growth potential. Razorpay's vision required a long-term perspective and significant investment that global players struggled to grasp. Chose to remain independent to achieve their vision for India. Capital Efficiency & B2B Logic Grew 40x between 2017-2020 with remarkable capital efficiency. Investor concern: Razorpay's interest income from fixed deposits exceeded its burn, making it profitable. B2B businesses are logical: value is exchanged for payment; excessive burning is unnecessary. Focused on adding value consistently, knowing customers would leave if value decreased. Early UPI Bet Became the first payment gateway to go live on UPI in September 2016, before major banks integrated. This early bet, made when UPI was doubted, positioned them to capture market share during demonetization. Leveraged being small to take calculated risks that larger, slower competitors couldn't. Embracing AI & Reinvention AI is a fundamental shift requiring recalibration and leadership focus. AI tools allow founders to return to "building mode," away from pure management. Razorpay reinvented its entire platform based on how they would build it today with AI. The strategy is to act like a startup, proactively adopting AI changes rather than reacting. AI will reduce build time, making execution speed the primary differentiator. Founder Evolution & Advice Learned the critical difference between "manager mode" and "founder mode." Founders must remain deeply involved in core product vision and direction. No one will care about the company as much as the founder; this never changes. Advice for aspiring founders: Find a problem you can commit 10 years to solving; AI makes building easier but doesn't change the core requirement of deep problem connection and sustained effort.

Introduction to Recursion in AI Recursion is proposed as the next scaling law in AI, improving reasoning performance at inference time rather than just increasing model size. Hierarchical Reasoning Models (HRM) and Tiny Recursive Models (TRM) are highlighted as key papers demonstrating this approach. Limitations of Traditional RNNs and LLMs RNNs suffered from "back prop through time," leading to vanishing/exploding gradients and error accumulation, especially with long contexts. LLMs, while efficient at training time with parallel processing, lack inherent reasoning and compression in the time direction, requiring the entire context for each step. LLMs struggle with tasks requiring sequential reasoning steps (like sorting) that exceed their fixed number of layers. Hierarchical Reasoning Models (HRM) HRM uses multiple levels of recursion inspired by the brain's different operating frequencies. It employs a "deep equilibrium learning" (DEQ) method, specifically truncated backpropagation through time (T-BPTT), to avoid backpropagating through all recursion steps. The key innovation is the "outer refinement loop," which scales and improves performance significantly. HRM achieved state-of-the-art results on ARC Prize with a small parameter count (27 million) and no pre-training. Tiny Recursive Models (TRM) TRM simplifies HRM by collapsing the hierarchical levels into a single network with shared weights. It further refines the backpropagation by backpropagating through only one full latent recursion step. TRM achieves even better performance (87% on ARC Prize 1) with a smaller model (7 million parameters) compared to HRM. TRM demonstrates that recursion can provide compute depth without parameter depth. Broader Implications and Future Directions Recursion is crucial and not going away; adding it to models shows significant benefits. Truncated backpropagation through time (T-BPTT) and outer refinement loops are powerful ideas to explore further. Combining the efficiency of recursive models with the scale of giant LLMs could lead to breakthroughs. Future research could focus on making recursive models more general-purpose and integrating them into latent spaces for enhanced reasoning.

AGI Timeline and Components AGI is estimated by Demis Hassabis to be around 2030. Key components for AGI likely include large-scale pre-training, RLHF, and chain-of-thought, which are believed to be part of the final architecture. Unsolved areas crucial for AGI are continual learning, long-term reasoning, and aspects of memory. It's possible current techniques can scale to solve these, or one or two major breakthroughs are still needed. Continual Learning and Memory The brain handles continual learning and memory consolidation gracefully, a process studied by Hassabis during his PhD. DeepMind's early Atari program DQN used experience replay, inspired by neuroscience. Current context window usage feels like "duct tape," and even large windows have a cost for relevant information retrieval. Processing live video requires significantly more than a million tokens for extended context. Agents and Reinforcement Learning Agents are seen as the path to AGI, requiring active systems to solve problems. DeepMind's early work on AlphaGo and other games involved agent systems designed for goal accomplishment and active decision-making. Many current foundation model techniques, like chain-of-thought, have roots in AlphaGo's pioneering work. Reinforcement learning and search techniques from AlphaGo and AlphaZero are considered highly relevant today. Agents are still in their early stages; full adoption into workflows is just beginning. Model Efficiency and Open Source A core strength of Google DeepMind is distilling frontier model capabilities into smaller, efficient "flash" models. This is crucial for serving billions of users across various Google products with speed and low latency. There's no current known informational limit to how smart smaller models can become through distillation. Google DeepMind is a proponent of open source, exemplified by the release of Gemma models, which have seen significant downloads. Open models are strategically important for edge devices like Android and robotics due to their inherent vulnerability. Multimodality and Scientific Breakthroughs Gemini was built to be multimodal from the start, offering advantages for world modeling and robotics. Multimodal capabilities are essential for AI in the real world, understanding physical context and intuitive physics. While inference costs are dropping, they are unlikely to become free due to Jevons paradox and ongoing innovation. Isomorphic Labs is working on drug discovery, aiming for a virtual cell simulation within approximately 10 years. Key scientific domains ripe for breakthroughs include materials science, drug discovery, climate modeling, and mathematics. AI as a Tool for Science AI is viewed as the ultimate tool for advancing scientific understanding and discovery. AlphaFold is a prime example, impacting millions of researchers and drug discovery processes. Future AI applications could include modeling full cellular systems and complex biochemical processes. The "AlphaFold moment" is expected across various scientific domains, requiring a similar breakthrough. Building Frontier Companies Startups advancing AI should combine AI advancements with deep technology areas like materials or medicine. Interdisciplinary teams, especially those involving "the world of atoms," are seen as defensible. Deep tech ventures are not easy but can lead to lasting impact. AI's rapid advancement means deep tech journeys must account for AGI potentially emerging mid-journey. Scientific Reasoning and Discovery AI systems are getting closer to genuine scientific reasoning, moving beyond pattern matching. True scientific discovery may require creativity and the ability to go beyond known patterns, potentially through analogical reasoning. A test for AGI could be whether it can independently produce groundbreaking theories like Einstein's special relativity. The ideal conditions for AI-driven breakthroughs involve massive combinatorial search spaces, clear objective functions, and sufficient data or simulators.

Replit's Vision and Evolution Replit aims to enable anyone to build and deploy scalable apps from an idea without technical expertise. The platform has evolved from solving the development environment to abstracting code entirely through AI agents. Replit's mission shifted from "making programming accessible" to "creating a billion new developers." Target Audience and Market Shift Replit targets "tech adjacent" users like product managers and designers, not traditional software engineers who may prefer complexity. The focus is on creators and a new generation of "AI native developers." The company sees a market gap in specialized software for various industries (e.g., physical therapy, pool maintenance, sports clubs). Product Capabilities and Agent Development Replit allows building personal, enterprise, and entrepreneurial software, including complex health tech apps. Enterprise use cases include faster product development and internal tools/automations (e.g., quote configurators). Agent 4 introduced parallel agents, asynchronous design capabilities (canvas), and teamwork features. Replit can now generate a mobile app from a web app and deploy across platforms. Future of Work and Skills The future company will be comprised of "builders" and "salespeople," with sales evolving into transformation consultants. Builders' roles will become higher-level, focusing on automation and abstract tasks. Key skills for the future include understanding possibilities, continuous learning, idea generation, and an entrepreneurial mindset. The concept of "post-prompting" is emerging, with agents responding to high-level goals.

AI as the Company Operating System AI should be the operating system of a company, not just a tool. Every workflow, decision, and process should flow through an intelligent layer. Implement "closed loop" systems where information is captured, fed back into AI, and improves processes over time. Making Your Company Queryable The entire organization must be legible to AI. Record meetings, minimize DMs/emails, and embed agents across communication channels. Build custom dashboards for all company metrics (revenue, sales, engineering, etc.). Example: An agent with access to tickets, Slack, customer feedback, and sales calls can propose more accurate sprint plans. Teams using this have cut engineering sprint time in half. AI Software Factories This is the evolution of test-driven development. Humans write specs and tests; AI agents generate and iterate on code until tests pass. Example: StrongDM's AI team built a system with no handwritten code, only specs and tests. This enables the "thousandx engineer." New Management Hierarchy and Roles Classic management hierarchies are obsolete; the intelligence layer handles information routing. Aim for minimal human middleware. Three employee archetypes: Individual Contributor (IC): Builder/operator, everyone contributes to building/operations. Directly Responsible Individual (DRI): Focuses on strategy and customer outcomes, owns specific results. AI Founder: Leads by example, demonstrates massive capability gains. Focus on maximizing token usage over headcount. Startup Advantage Startups have a major advantage as they can build AI-native from day one without legacy systems. This allows for operating a thousand times faster than incumbents.

Introduction to the Agent Era and GStack The video introduces the "agent era" of software development, emphasizing teamwork, process, and review as key to utilizing AI agents effectively. GStack is presented as an open-source tool that transforms Claude Code into an AI engineering team. The creator coded more in the past two months than in all of 2013, highlighting the productivity gains with AI. GStack's "Office Hours" Skill GStack's "Office Hours" skill is modeled after Y Combinator's partner sessions, designed to reframe product ideas through forcing questions. It helps explore business models, pain points, and the core value proposition of a product. The tool can assist in identifying potential issues, like a lack of strong evidence for user demand. It can also help in developing a "wedge strategy," where an initial simple solution leads to a larger, more lucrative business model. Browser Automation and AI Capabilities GStack incorporates browser automation, allowing AI to log in, navigate, and download documents (like 1099s) directly. This can happen in a visible browser, not just the cloud, enhancing transparency. The tool can leverage different AI models, suggesting Claude Opus for broad ideas and Codex for detailed implementation. GStack includes adversarial review to identify and auto-fix issues in design documents, improving their quality. Design and Development Workflow "Design Shotgun" is a visual brainstorming tool within GStack that generates multiple AI-driven design options. Users can select their preferred design, which is then locked in. The process includes various review stages like "Plan CEO review" and "Auto Plan." Post-coding, GStack offers "Review" for bug catching and "Ship" to ensure PR readiness. A key feature is the integration of Playwright and Chromium for browser automation, enabling actions like taking screenshots, filling forms, and running regression tests. Scalability and Future of Development GStack aims to achieve a "level seven" software factory, allowing for parallel development on multiple projects and features. Automated QA is crucial to handle the increased output from AI agents. The "Ship" tool acts as a final check before merging code. The creator manages numerous parallel Claude Code sessions and open-source PRs daily. GStack is available on GitHub, promoting a new era of accelerated software development.

Stripe's Website Redesign Drivers The previous Stripe website, launched in 2020, was still functional but no longer reflected the company's evolved business and broader user base. Key areas for improvement included articulating the expanded product suite beyond just payments, updating visuals to match brand sophistication, and clarifying the narrative. The business had grown to serve a wider range of clients, including AI companies needing usage-based billing and large enterprises for various financial infrastructure needs. New Website's Core Principles and Features The website's purpose was redefined as a manifesto, showcasing Stripe's identity, mission, and values through design choices and messaging. A significant new element is the GDP counter prominently displayed, emphasizing Stripe's scale and trustworthiness. The "bento" section visually represents the diverse product offerings (payments, billing, issuing, etc.) using imagery and minimal text. Interactive modals provide more product details without leaving the homepage, offering a progressive disclosure approach to information. Design Process, Animation, and AI Integration Careful attention was paid to animations within the bento cards to make the site feel alive and express the company's meticulous approach to its services. Animations were fine-tuned to be engaging without being distracting, with interactive elements providing feedback to the user. Animated metrics add visual interest and loosely communicate concepts like global scale and uptime, designed to be beautiful and informative. AI was used to accelerate prototyping and experimentation, particularly for image generation and exploring interaction paradigms, but it doesn't replace craft or taste. Iterative Design and Decision-Making The design process involved extensive exploration and iteration, including weeks of experimenting with visual elements like the website's wave/gradient. A custom tool allowed for detailed tweaking of wave properties (blur, grain, rotation, texture, color) to find the perfect fit. Decision-making for key elements involved a process of down-selecting options and discussing them with leadership to ensure alignment with the brand and user experience. The team prioritized getting design elements "right," even if it meant delaying launch, to ensure a polished and joyful user experience. Various bento layout variations were explored, including compressed, section-based, and accordion styles, with the visual bento ultimately chosen for its user-friendliness in a "browse mode." User Experience and Company Culture The emphasis was on creating a visual and "kinder" browsing experience, allowing users to explore at their own pace. AI-assisted image generation required meticulous attention to detail to ensure realism and brand consistency, highlighting that AI speeds up exploration but doesn't replace craft. Designers are focused on creating "exceedingly easy to use," powerful, and joyful experiences that push the status quo forward. Design systems are evolving with AI to manage components and scale design efforts, encouraging designers to go beyond "good enough" and explore new interaction paradigms. The "walking the store" practice, where all employees use Stripe's products as customers, is crucial for maintaining user empathy and identifying cross-product user experience issues. Multiple perspectives (engineers, product leaders) during "walking the store" sessions help uncover different user pain points and improve the overall composite experience. The pursuit of "Minimum Viable Quality Product" (MVQP) balances progress with maintaining user trust, especially when incorporating new technologies like AI.

The GPT Moment for Robotics The equation for starting a robotics business has changed due to decreasing upfront costs. The mission is to build a model that can control any robot for any physically capable task at a high performance level. The approach is to build a strong base model with common sense knowledge, then use a mixed autonomy system that improves over time with real-world exposure. Why Robotics is Difficult and Recent Breakthroughs Robotics has three pillars: semantics (language models help here), planning, and real-time control in dynamic environments. Seikhan demonstrated using language model common sense knowledge in robotics, reducing the need for robot-specific data. POMoE and RT2 (Robotic Transformer 2) showed that adapting powerful vision-language models with robotic data allows knowledge transfer to low-level actions, enabling tasks with unseen objects or concepts (e.g., "move the coke can to Taylor Swift"). These initial models were single-embodiment (worked for specific robots). Scaling and Cross-Embodiment The insight is that data from multiple robots can teach a more abstract notion of robot control, leading to better generalization. Open Cross Embodiment and Robotic Transformer X showed scaling laws for robotics by training across multiple hardware platforms, a first in the field. A generalist model trained on data from 10 different robot platforms performed 50% better than specialist models optimized for individual platforms. OpenX was a large collaboration within the robotics community. The Data Problem and Its Scale The data problem in robotics has two parts: data generation and data capture. There's a lack of an "internet of robotic data," requiring operational effort for collection. Solving robotics could contribute significantly to GDP, justifying investment in data collection. The approach focuses on cross-embodiment to easily consume data from diverse robot sources. No two robot platforms are the same, and platforms drift over time, making multi-robot data ingestion more robust than optimizing for a single, changing platform. Emergent Properties and Current State Emergent properties are being seen in large robotic models, allowing for zero-shot task performance that previously required extensive data collection. Tasks requiring precision, reasoning with multiple objects, and deformable objects (like laundry folding) are being tackled. Current state allows for scaling robot deployment if tasks can tolerate mistakes and a mixed-autonomy system (human oversight) is feasible. Examples include folding laundry in a real laundromat (collaboration with Weave) and picking/placing items in pouches for shipping (collaboration with Ultra) in real e-commerce warehouses. These demonstrations show autonomy at scale, ready for deployment. Technical Insights: Cloud-Based Inference A surprising insight is that most robot evaluations are cloud-hosted, even for complex demos. This is enabled by tightly coupling system, hardware, and model development. Real-time inference is achieved by burying it within the robot's control loop, querying the cloud API as needed, rather than waiting for actions to complete. Algorithmic improvements like "real-time chunking" and pre-computation ensure smooth transitions between predicted action chunks. This approach simplifies robot hardware, reducing the need for powerful on-board compute and expensive hardware that quickly becomes obsolete. The cloud-based approach allows for decoupling hardware control from semantics and planning. The pace of progress has been faster than expected, with real deployment and scaling being serious considerations two years into the company's life. Starting a Robotics Company Today Robotics was historically vertically integrated with high barriers to entry. Physical Intelligence aims to provide an AI foundation for others to build upon, enabling faster autonomy onboarding. The new playbook for a vertical robotics company: understand existing workflows, identify opportunities for maximum impact, be scrappy with hardware and data collection (models compensate for inaccuracies), and implement mixed-autonomy systems to reach economic break-even for scaling. Upfront costs are significantly lower, focusing on use case identification and data collection rather than proprietary autonomy stacks. This allows companies to differentiate themselves on the components that matter most. The Future: A Cambrian Explosion There is a strong belief in a "Cambrian explosion" of robotics companies across many verticals due to lower costs and accessibility. It no longer requires decades of robotics experience to start; scrappiness and speed are key. The focus shifts from building vertically integrated systems to identifying use cases and leveraging foundational models. The promise is to enable a massive increase in robotic applications beyond traditional "dirty and dangerous" industrial tasks. Physical Intelligence aims to enable this explosion by providing foundational models, publishing research, and open-sourcing models (PI Zero, PIO5 are the same as internal models). Enabling the Future and Company Building The infrastructure for large-scale, general-purpose robotics (data collection, management, annotation, evaluation) was largely missing, prompting Physical Intelligence to build much of its own software. There's an opportunity for services that support robotics companies (e.g., remote teleoperation, data collection, annotation). A tight, collaborative loop in model development (data collection -> training -> evaluation -> improvement) is crucial. An "automated robotic research scientist" is a desired future development to bottleneck progress. Current language models can assist with simple failure modes by providing recommendations based on textual descriptions of failures. A fundamental understanding of the physical world, which is missing in current foundation models, is needed for more advanced automation. Using cloud-based inference with large models has shown significant improvements in compute utilization (e.g., a prototype "pre-training on call" improved utilization by 50%). The company's mission is to create this Cambrian explosion by focusing on the model as the bottleneck and enabling others to build on their work. Success is defined not just by their own models on their robots, but by their models performing useful tasks on *other people's robots*. The founding team, many from Google's robotics team, enjoy working together and believe their combined strengths increase the chances of success.

Company Genesis and Core Technology BillionToOne is a next-generation molecular diagnostics company that detects DNA in blood samples. Their core technology addresses the "needle in a haystack" problem of finding rare DNA fragments (e.g., fetal DNA, tumor DNA) among billions of other DNA molecules in blood. The key innovation is adding synthetic DNA to the sample before amplification. This allows them to quantify amplification noise and errors, enabling accurate detection of extremely dilute DNA signals. This process transforms a difficult biology problem into a manageable mathematical one. Early Commercialization and Growth BillionToOne's initial focus was prenatal genetic testing, addressing conditions like sickle cell disease and cystic fibrosis. This prenatal test quickly became widely used, achieving nearly 20% market share. The company's founders, two PhD students, started with limited resources but rapidly developed and commercialized their test within two years of founding. Early growth was challenging, with initial slow adoption, requiring a strategic shift to market directly to patients to influence physician adoption. They have scaled operations to process over 600,000 tests annually and built a state-of-the-art lab. Expansion into Oncology The same core technology used for prenatal testing is applicable to detecting cancer DNA in blood (liquid biopsy). BillionToOne launched an early version of their cancer test in 2023, demonstrating their ability to execute in multiple markets simultaneously. Their strategic plan involved a phased approach: prenatal genetics first, then late-stage cancers, and finally early-stage cancer detection. Future Vision and "Holy Grail" of Cancer Detection The ultimate goal is to achieve ultra-sensitive Minimal Residual Disease (MRD) testing for stage 1 cancer patients, with the potential for even earlier detection before stage 1. This capability is described as the "holy grail" of cancer detection. They aim to detect microscopic levels of remnant tumor DNA after surgery in stage 1 and 2 cancer patients. The long-term vision includes developing a universal cancer screening test for early detection in the general population. Company Culture and Strategy BillionToOne seeks "interdisciplinary people" rather than just an interdisciplinary team, fostering rapid iteration and innovation within small research teams. They operate with a lean, decentralized structure, with small R&D teams reporting directly to the founders, akin to "many startups within the larger company." Their strategy emphasizes making tests accessible and affordable, differentiating them from a similar phased approach by other companies. The company culture embraces challenges, with a saying that "pressure is a privilege," attracting individuals motivated by difficult, impactful work.

Company Announcement & Funding Variance is emerging from stealth mode, announcing a $21 million Series A funding round. The company has been building in stealth for the past 3 years. What Variance Does Variance builds purpose-built AI agents for risk and compliance. They automate content review, fraud reviews, and identity reviews at scale. Their AI agents are used by large companies, including Fortune 500s and marketplaces. Secrecy and Sensitive Data Variance operates with a degree of secrecy because they handle sensitive data and issues. The company's tools are used to combat "bad guys" but are built "for the good guys." Marketing their use cases could inadvertently create more fraud or abuse. They see themselves as a "secret weapon" for their customers. Customer Use Cases GoFundMe: Variance AI agents review fundraisers at scale to verify legitimacy, ensure money goes to the correct recipients, and prevent funds from being sent to sanctioned countries. Example of GoFundMe Fraud: During crises, fraudulent fundraisers mimicking real ones (e.g., for a deceased public figure's family) emerge, and Variance AI agents identify the legitimate ones. Marketplaces/Gig Economy: Verifying seller identities and complex ultimate beneficial owner (UBO) verifications for platforms like marketplaces and gig economy apps (e.g., delivery drivers requiring ID verification). KYB (Know Your Business) Verifications: Verifying that individuals signing up for business are truly linked to the businesses they claim to own, even when multiple shell companies and complex ownership structures are involved. Technology & Data Variance's AI agents use three main building blocks: compliance documents, standard operating procedures, and data (internal/external). They integrate with various data sources, pooling unstructured data and accessing hundreds of business registries and the open web. Access to the open web allows their AI agents to perform similar investigative tasks as human analysts (e.g., Googling names) to trace complex fraud rings. A significant technical challenge was handling and aggregating **petabytes of unstructured data** scattered across multiple systems, sometimes requiring scraping data directly from human-facing UIs. Evolution of Fraud Detection Previously, fraud systems relied on a patchwork of deterministic systems (rules engines, classifiers) and slow, inconsistent human analysts. Variance's AI agents provide a **fully self-healing system** that can materialize features, reason over unstructured data, and eliminate the need for specialized classifiers or human reasoning, allowing for faster iteration and new product lines. They have detected complex fraud rings, including state-sponsored actors pushing narratives during elections, and even potential threats of physical harm. Team & Culture Variance maintains a **lean team of 12**, with 5 software engineers. They are heavily reliant on AI coding agents, significantly amplifying their output. The company fosters a strong **ownership culture**, where engineers are empowered to manage their own work and take initiative. They are hiring for backend and frontend roles, recognizing the importance of strong investigative tools for complex cases that require human review. Origin Story & Early Challenges The co-founders, Karine and Michael, met while working on the fraud engineering team at Apple. They were driven by a desire to solve the inefficiencies and limitations they observed in fraud detection at scale. Landing their first enterprise customer, IA (specifically Ask Media Group), took eight months and involved convincing them to use LLMs for marketing content review, a task previously done manually by a large team. The company started building with LLMs just before ChatGPT's widespread adoption. Resilience & Vision Karine experienced a severe accident where she was hit by a truck, breaking her spine and leg, highlighting the founder's critical role and the need to scale processes beyond a single individual. Despite the setback, the founders remained committed to their mission, driven by a sense of duty to improve the industry with their expertise. Their strong initial hypothesis about the problem and solution has guided their development, rather than pivoting frequently. They believe their deep understanding of the problem and commitment to solving it resonate with customers.

Introduction to NDIA and Program Synthesis NDIA is a new AGI research lab aiming to build a new branch of machine learning, an alternative to deep learning, focusing on program synthesis. The goal is to create a new learning substrate that is much closer to optimal than current parametric deep learning models. This approach involves building extremely concise symbolic models of data, contrasting with deep learning's reliance on fitting parameters of complex curves. NDIA proposes "symbolic descent" as the equivalent of gradient descent in this symbolic space. Limitations of Current AI and the Case for Alternatives The current industry focus on scaling LLMs, while productive, is seen as potentially counterproductive due to everyone working on the same thing. François Chollet believes the current LLM stack may not be the foundation for AI in 50 years and aims to leapfrog towards optimal AI. He acknowledges that such ambitious projects have a low chance of success but are worth pursuing if successful. The success of coding agents is attributed to code providing a verifiable reward signal, enabling automation in domains with formal verification. Domains like English language composition or law are harder to automate due to a lack of natural formal verifiability, leading to slower progress with current LLM stacks. Defining and Measuring Intelligence: The ARC AGI Benchmark Chollet's definition of AGI is a system that can approach any new problem or domain, model it, and become competent with human-level efficiency (data and compute). He believes current technology can already automate any domain with verifiable rewards, but achieving human-level learning efficiency across arbitrary tasks requires a different approach. The ARC AGI benchmark was created to capture the idea of intelligence as skill acquisition efficiency, inspired by ImageNet for reasoning. ARC V1 and V2 focused on producing causal models from given data, while V3 measures "agentic intelligence." ARC V3 requires an agent to explore an interactive environment (like a mini video game) without instructions, measuring exploration efficiency, goal setting, planning, and execution. V3 aims to measure fluid intelligence, the ability to efficiently explore and model new environments, matching human efficiency. V3 is designed to be more resistant to "harness" strategies used to saturate V2, using a private set of environments with novel concepts. NDIA's Approach and Future Vision NDIA's approach involves deep learning-guided program search, using deep learning to guide exploration in a symbolic search space. Chollet predicts that AGI, when achieved, will be a small codebase (megabytes), distinct from a large knowledge base. He believes that retrospectively, AGI might be found to be a codebase under 10,000 lines of code, achievable with past computing resources. NDIA aims to remove humans from the improvement loop for self-improving, compounding systems. The NDIA approach is described as "science incarnate," recreating the scientific method algorithmically through symbolic compression. Chollet believes human intelligence is messy but its underlying principles can be reimplemented more efficiently. The development of NDIA started with a symbolic learning vision, focusing on replacing parameter curves with shortest symbolic models. The Evolution of ARC AGI and Future Prospects ARC V1 was initially difficult for LLMs, showing that scaling alone wasn't enough for fluid intelligence. The emergence of reasoning models led to significant performance improvements on ARC V1, signaling new capabilities. ARC V2 was saturated by large-scale targeting using reinforcement learning loops and "harnesses" to generate and solve tasks, demonstrating a new post-training paradigm. This saturation indicates models becoming more useful in specific domains through better training and verifiable reward signals, not necessarily "smarter." ARC V3 moves beyond modeling static data to testing agentic intelligence in interactive environments. Future ARC benchmarks (V4, V5) will focus on continual learning, curriculum learning, and invention. Chollet predicts AGI might be achieved around the early 2030s. Advice for Aspiring AI Researchers and Developers There is room for startups exploring alternative AI approaches beyond LLMs, such as scaled-up genetic algorithms or alternative architectures. Researching older, less-invested-in approaches from the 70s and 80s is recommended. Promising approaches should demonstrate scalability and allow for improvement without constant human intervention (recursive self-improvement or decoupling from human effort). For open-source projects, focus on a simple, intuitive API, usability, informative documentation, and community building. Hiring enthusiastic users ("fans") can be beneficial for building successful projects. AI progress should be viewed as empowerment, with expertise in a domain allowing individuals to leverage AI tools for their benefit.

Company Origin and Mission Solugen reinvents America's trillion-dollar chemical industry using biology and chemistry. Founded with a scrappy prototype built from Home Depot PVC pipes, now a billion-dollar company. Mission: Use biology to create chemicals, enabling smaller, cleaner, safer, and more environmentally friendly chemical plants. Core Technology: Chematic Processing Combines the specificity of biology (enzymes) with traditional metal catalysts. Achieves significantly higher reaction yields (96% compared to traditional 60%). First company to fuse biology and chemistry in this manner. Uses enzymes from living cells (initially discovered in pancreatic cancer cells) and pairs them with novel metal catalysts. Process: Receives corn syrup, adjusts parameters for enzyme and metal catalyst reactions to oxidize corn syrup, then evaporates water for final product. Discovery and Breakthroughs Eureka moment: Found an obscure enzyme in pancreatic cancer cells that produces hydrogen peroxide, key to a new hydrogen peroxide production process. Key Insight 1: Organic enzymes can operate at industrial scale and efficiency. Key Insight 2 (Commercial): Instead of massive upfront funding, built a $10,000 reactor and started selling immediately, gradually scaling up. Innovation and Operations Solugen operates its own biology and metals labs to produce enzymes and catalysts in-house. Biology Lab: Grows microbes, extracts enzymes, and tests their capabilities at scale. Metals Lab: Pairs enzymes with appropriate metal catalysts, allowing for mix-and-match combinations. Traditional plants use fossil fuel feedstock; Solugen starts with sugar (corn syrup), leading to cleaner processes and fewer toxic byproducts. Early Stage and Growth First reactor built for $10,000 using PVC pipes from Home Depot. Started by manually operating the reactor and selling small volumes of peroxide to float spa hot tub owners. Identified supply chain inefficiencies by selling directly to end-users, bypassing multiple distributors. Accepted into Y Combinator (YC), deferred to secure initial customers. YC experience was like "grad school for customers," emphasizing deep customer understanding. Used $4 million seed round to build the first large pilot reactor (1500-gallon) in Houston, Texas. Secured first oil and gas customer through a targeted billboard campaign and direct outreach to a key decision-maker. Scaling Up: Bioforge 1 Built state-of-the-art plant, Bioforge 1, with components manufactured in five locations and assembled on-site like Legos. Plant uses corn syrup as feedstock, stored in large tanks (holding 4 rail cars). The 60-ft tall bubble column reactor is a scaled-up version of the initial PVC reactor. A small amount of enzyme can produce large volumes of product (one Coke bottle of enzyme yields 2-4 tanker trucks of product). Distributes products via trucks directly from the plant, filling them at 300 gallons per minute. Key strategy: Building factories near customers to reduce shipping costs and undercut competitors. Future Outlook Plans to build multiple different manufacturing assets using their core technology. Aims to solve complex customer problems that may not even exist yet. Focus on cultivating a culture willing to be wrong and solve future problems.

Company Origins and Pivot Emergent, founded by twin brothers Makund and Madav, went through Y Combinator in Summer 2024. The company has seen explosive growth, with 7 million apps built in 8 months since launch. Initially, the founders focused on automating software testing, but pivoted to general coding agents after realizing that solving verification could automate all of software engineering. They initially targeted enterprises but found it too slow, then shifted to a product for non-technical users after seeing the growth of platforms like Lovable and Bolt. Today, 80% of Emergent's users are non-technical, building apps for real businesses globally. Product and Technology Emergent allows users to build and ship production-ready software using AI agents. The platform was built with a focus on replicating the entire software development lifecycle: code reviews, automated testing, debugging, deployment, security, and hosting. They built their own infrastructure, including cloud sandboxes and a Kubernetes stack, to ensure consistency between build and deploy environments and provide rapid feedback to agents. Emergent utilizes a multi-agent architecture, delegating tasks to sub-agents for efficiency and managing context frugally. They developed a long-term memory for agents by aggregating trajectories and running them through a CI/CD process, enabling continual learning across sessions. The platform abstracts away complexities like API key management for non-technical users, allowing them to use an "Emergent LLM key." Emergent's internal QA engineer built an Asana clone using the platform, demonstrating its capability for complex applications and offering significant cost savings compared to traditional SaaS. User Base and Impact Primary users are small to medium business owners who previously relied on spreadsheets and manual processes or faced high costs for custom software development. Emergent has democratized software creation, enabling individuals with domain expertise but no coding background to build and launch their own applications. Examples include a clinical psychologist/equestrian coach who built an app combining psychology and horse riding insights, and a CRM for lawyers. The platform empowers "solopreneurs" by eliminating the "translation loss" often experienced when communicating ideas to developers. Emergent is not just about cost savings; it's about empowering individuals to pursue their ideas and gain autonomy over their lives and businesses. Future Outlook and Industry Trends The founders believe the coding aspect is only 20% of the job; taking an app to production and understanding user needs are crucial. They anticipate SAS workflows will be increasingly consumed by agents, requiring SAS companies to become "agent-first." The nature of software is changing, with a rise in "agentic" applications and longer agent task horizons. Emergent is experimenting with agent swarms and anticipates agents running 24/7 and collaborating on complex tasks. They are focusing on building better verifiers and custom fine-tuned verification layers to augment models. The founders believe that understanding customer needs deeply and building closer to them will be key to success in the evolving AI landscape. They see a trend towards personalized software and an "explosion of being able to start businesses that aren't venture funded," driven by individual passions and autonomy.

Science's Retinal Implant (Prima) Science's BCI treatment involves a tiny silicon chip implanted under the retina. This chip acts as a retinal stimulator, bypassing damaged rods and cones. Patients wear glasses with a camera and laser projector that sends images to the implant. The implant absorbs laser light and excites cells, restoring some vision for those who have lost sight due to retinal degeneration. A large clinical trial in Europe showed significant positive effects, and approval is being sought. The Nature of Brain-Computer Interfaces (BCIs) BCIs are not a single product but a category of technologies for different applications. They can be used to restore lost functions (sight, hearing, movement) or for structural neural engineering (enhancing cognition, treating mental health issues). BCIs are moving beyond restoring functionality to potentially augmenting human capabilities. Different modalities, like ultrasound and implantable chips, will serve different purposes. Neuroplasticity and Learning While there are critical periods in early development, the brain remains significantly plastic throughout adulthood. The brain can learn to control neural activity through feedback, as seen in cortical motor decoders. The brain adapts to new inputs and can learn to interpret them, even if the initial wiring was for different functions. The brain's plasticity is often stable due to its adaptation to reality, forming "basins" in an "energy surface." The Qualia of Artificial Vision and Beyond The qualia of Science's Prima implant is described as normal, albeit black and white with a limited field of view. Blind patients' brains, deprived of visual input, may generate internal perceptions that need to be dissociated from real input during rehab. The potential qualia of ultra-high bandwidth bio-hybrid neural interfaces are difficult to imagine, with conjoined twins offering a glimpse into shared conscious experience. Future of BCIs and Healthcare Within 10 years, BCIs may approach native visual acuity, including color and a wider field of view. BCIs are seen as a neural engineering approach to medicine, potentially more effective than drug discovery for certain conditions. The goal extends beyond restoring function to fundamentally reframing medicine and human capabilities. BCIs are poised to impact vision, hearing, balance, motor control, and potentially longevity. Technical Aspects and "The API of the Brain" The brain's input/output is through cranial and spinal nerves, which can be considered its "API." Understanding this API allows for new ways to interact with the brain's information processing. Progress in AI has led to a unification with neuroscience, with AI models exhibiting representations similar to those in the brain. BCI development is limited by the ability to record and stimulate neural signals, with the retina's layered structure being a key area of study. Science stimulates bipolar cells in the retina, which is a critical processing step, allowing for image formation in the mind's eye. Science's Bio-hybrid Approach Science is developing bio-hybrid neural interfaces by culturing engineered neurons onto implants. These engineered neurons are hidden from the immune system, avoiding the need for immunosuppressants. This approach aims to create new biological connections without genetically modifying the patient's brain. The concept is compared to growing a new cranial nerve or the "ponytails" in the movie Avatar, forming a new biological interface. Neural Representations and Latent Spaces The brain contains "representations" of concepts, like hand activity or objects, which can be mapped. Deeper brain areas exhibit abstract representations, similar to latent spaces found in AI models. This convergence of AI and neuroscience suggests that AI models are on the right track in understanding brain function. The "Smartphone Dividend" and Motor Decoding The development of efficient, small, and low-power electronics, driven by the smartphone industry, has been crucial for implantable BCIs. Closing the skin over implants is important to prevent infection, requiring highly efficient electronics that don't generate excessive heat. Motor decoding, enabling cursor or keyboard control, has been a foundational BCI application since the late '90s. The Vessel Program and Profusion Technology Science is also working on profusion technology for life support, inspired by cases like a teenager kept alive on ECMO. The goal is to improve profusion systems to be more accessible and allow for higher quality of life, moving beyond "bridge to nowhere" scenarios. This involves refining the technology to make it portable and integrate seamlessly with the body, addressing issues like skin healing around implants. Early Days and Motivation Max Hodak's early interest in BCIs was fueled by science fiction, particularly "The Matrix," and a fascination with the brain as a computer. He co-founded Neuralink with the motivation to "upgrade humanity" in the face of advancing AI, preventing humans from being left behind. The initial Neuralink team was formed from a small community of researchers. Hodak emphasizes the importance of high agency and persistence in pursuing a clear vision, but also the value of learning from experienced individuals and companies. The Future Horizon and Exceptional Change Hodak believes we are in an "era of takeoff" for BCIs, marking a significant new phase for humanity. He predicts that the first people to live to a thousand may already be alive, driven by technological advancements. The next 15 years are expected to bring transformative changes comparable to the early impact of the Industrial Revolution. BCIs and AI are seen as parallel yet distinct forces that will reshape intelligence availability, human agency, and the human condition.

AI Design Trends and Pitfalls AI design tools enable easier creation, but accepting all AI suggestions can lead to common, unoriginal designs. Trends like purple gradients and fading-in sections are becoming ubiquitous due to LLMs being trained on popular examples. While AI offers superpowers, founders must remain in control, acting as editors to ensure originality and avoid "AI slop." Website Review: New.ai Features a very purple color scheme, a common AI-generated trend. A distracting line following the user down the page was implemented likely because it was easy with AI, but adds no value. Contrast issues make text hard to read. Tasteful hover animations on cards are a good use of AI, enhancing the brand and reinforcing messaging. Navigation hover effects that cause menu items to fade out are counterintuitive and distracting. Website Review: Rosebud AI Continues the purple gradient trend, leading to a lack of brand originality. Features an interactive 3D game demo, which is engaging but its connection to the product isn't clearly explained. Non-standard navigation and potentially confusing elements like a following top bar hinder usability. The combination of a red logo with purple accents and the use of emojis can appear lazy or unharmonious. Cursor-following light effects on game examples are visually appealing but may not be worth the development effort if not AI-assisted. Website Review: Get Crux Exhibits scroll jacking and automatic fade-in sections, which can be disorienting. A button that constantly chases the user is distracting and makes it difficult to focus on content. "Meteor" animations and blurry hero screenshots detract from the user experience and product clarity. Lack of visual consistency across sections suggests different AI generation approaches. The core value proposition is not immediately clear, requiring users to scroll extensively to understand the offering. Website Review: Sphinx More animation is present, a common outcome of AI design tools. Information hierarchy is complicated by an unnecessary "fourth style" element between the logo and H1. A confusing animated section with shifting button styles and unclear functionality appears to be a product of AI over-suggestion. Hover effects revealing icons that are not critical information can be distracting. A scroll-jacking animation that locks the user in place distracts from the content and lacks a clear purpose. The visual style, while modern, can be hindered by distracting animations and unclear messaging. Website Review: Build Zero Features purple gradients and "dumb hover effects" that add no value and can appear as bugs. An interactive element has a bug in selection, which might be overlooked due to the ease of AI generation. AI-generated dashboards with common color callouts and "bento box" layouts lack originality. The repetition of common patterns across sites diminishes brand uniqueness and credibility. Website Review: Zarna AI Employs scroll jacking, making the site feel clunky and slow to navigate. Lack of clear content and excessive scrolling to understand the offering are significant issues. The navigation bar can become unreadable against dynamic backgrounds, highlighting a lack of robust design. Inconsistent clickability of elements and automatic movement create a confusing user experience. While the visual style can be fresh, it's undermined by a lack of clear messaging and "fit and finish." Key Takeaways and Advice Founders must be intentional and act as editors when using AI design tools, ensuring originality and brand consistency. Thoroughly QA all AI-generated elements to catch bugs and confusing interactions. Prioritize clear messaging and ensure the website effectively functions as a customer acquisition channel. Use AI to enhance creativity and efficiency, not to outsource critical thinking about the product and brand.

Poetic's Core Offering Poetic builds a recursively self-improving system for LLMs, aiming for AI to make itself smarter. This approach is significantly faster and cheaper than traditional methods like training new LLMs from scratch, which cost hundreds of millions and take months. Poetic's system generates "harnesses" or agents that sit on top of existing LLMs and automatically outperform them for specific problems. These harnesses remain compatible with new LLM releases, allowing for continuous performance improvements without re-training costs. Addressing the "Bitter Lesson" Traditional fine-tuning is expensive and quickly becomes obsolete with new model releases, a problem referred to as the "bitter lesson." Poetic's method avoids this by creating adaptable systems that benefit from newer, more powerful LLMs without costly re-engineering. The system can optimize existing agents or components like prompts and reasoning strategies. Performance and Benchmarks Poetic's system has demonstrated strong performance on benchmarks like ARC AGI V2 and Humanity's Last Exam. On ARC AGI V2, Poetic achieved higher scores than Gemini 3 DeepMind at a fraction of the cost, using Gemini 3 Pro as the base model. For Humanity's Last Exam, Poetic reached 55% accuracy, surpassing Anthropic's Claude Opus 4.6 (53.1%), with optimization costs under $100k. This contrasts sharply with the hundreds of millions of dollars required for training large foundation models. Technical Approach and Comparison Poetic's core technology is its poetic meta system, which recursively self-improves to generate highly effective reasoning systems. These generated systems are composed of code, prompts, and data, built on top of one or more LLMs. This is presented as a new paradigm distinct from Reinforcement Learning (RL). The system can automate aspects of prompt engineering and context stuffing, outsourcing data understanding and failure mode analysis to the AI itself. While automated prompt optimization (like "Jeepa") offers some gains, Poetic emphasizes the importance of reasoning strategies written in code over just better prompts. Getting Started with Poetic Poetic is not yet publicly released, but interested parties can sign up for early access at poetic.ai. They are seeking startups and companies with difficult, unsolved problems. Poetic aims to provide "stilts" that allow any agentic company to achieve state-of-the-art performance. Key capabilities highlighted include improving reasoning and deep knowledge extraction. Founder's Background and Advice Ian Fischer, co-founder of Poetic, previously worked at Google DeepMind for a decade and founded a mobile devtools company. He transitioned from hardware/robotics to machine learning research at Google, finding hardware "hard." His advice for engineers wanting to enter AI is to try things daily, push boundaries, and build what they imagine. He uses AI tools like GPT-5 for app development, emphasizing the rapid pace of improvement and ease of use.

The AI Agent Economy AI agents are rapidly evolving, moving beyond simple autocomplete to making independent decisions and interacting with each other, exemplified by platforms like Moltbook. This shift is creating an "AI agent economy" where agents choose and utilize tools, potentially paralleling the human economy. Impact on Developer Tools and Go-to-Market The traditional developer market is expanding from 20 million trained developers to hundreds of millions, plus their agents, dramatically increasing the potential user base. Documentation is becoming the primary interface for agents; tools with clear, agent-friendly documentation (like Resend) are favored. Companies like Mlifi are benefiting as they provide tools to optimize documentation for both humans and agents. The go-to-market strategy for developer tools is shifting from human-to-human recommendations to agent-driven adoption. Emerging Trends and Future Possibilities Swarm intelligence is emerging, where multiple AI agents collaborate, much like biological systems. Platforms like Moltbook are showcasing this emergent swarm behavior, with agents interacting and collaborating to achieve tasks. There's a potential for a parallel tech stack built specifically for AI agents, including services like Agent Mail for AI-native inboxes. Agents may eventually handle tasks like booking reservations, and could even influence social recommendations (e.g., restaurant choices). The concept of "human money" vs. "agent money" is introduced, suggesting agents might eventually develop their own economy and transactional systems. Challenges and Considerations Legal liability and standing are current barriers, as agents are not legal entities, requiring human oversight for transactions and applications (e.g., Y Combinator applications). The "Dead Internet Theory" is mentioned, suggesting a significant portion of online content may already be AI-generated, but a counter-argument is made that smarter, aligned agents could improve the internet. Building user trust and relationships with AI agents is still a challenge, as people have high expectations for AI interactions. Developers should focus on creating tools that agents "want", prioritizing APIs and open-source solutions over websites. Founders should develop an intuitive understanding of agent capabilities and limitations, and build tools that cater to their natural inclinations.

The video features Boris Cherny, an engineer at Anthropic, discussing the development and philosophy behind Claude Code, an AI coding assistant. Origin and Philosophy Accidental Beginning: Claude Code started as a simple terminal chat app built by Boris to learn Anthropic's API, not as a planned product. Building for the Future: Anthropic's strategy is to build for the model six months ahead, focusing on areas where current models are weak but expected to improve. Latent Demand: Key features and products, like ClaudeMD, emerged from observing how users were already trying to achieve tasks, demonstrating "latent demand." The "Bitter Lesson": Cherny emphasizes "never bet against the model," advocating for waiting for model improvements rather than building excessive "scaffolding" that will quickly become obsolete. Development and Evolution Terminal-First Approach: The initial choice of a terminal interface was due to its simplicity and speed of development, avoiding the need for a complex UI. Tool Use Discovery: A pivotal moment was realizing the model's strong inclination to use tools, even for tasks like identifying music. Iterative Rewriting: The codebase for Claude Code is constantly rewritten, with significant portions being less than six months old, reflecting rapid model advancements. User Feedback Driven: Features like "plan mode" and verbose output options were direct responses to user feedback and observed usage patterns. Adapting to Model Improvements: The shift from requiring manual debugging of model output to summarizing tool usage reflects the increasing reliability of newer models. Key Features and Concepts ClaudeMD: A mechanism for users to provide custom instructions and context to Claude Code, with an emphasis on keeping it concise and up-to-date. Plan Mode: A feature that allows users to outline a plan for the model before it starts coding, reducing the risk of it going in the wrong direction. Cherny suggests this may become obsolete as models improve. Agent Topologies: The exploration of how multiple agents can collaborate, using concepts like "uncorrelated context windows" for complex tasks. Swarm Development: The successful use of a "swarm" of agents to build features like the plugins feature over a weekend with minimal human intervention. Sub-Agents: The use of recursive Claude Code instances ("mama quad") to handle specific tasks, often prompted by the main agent. Impact and Future Massive Productivity Gains: Claude Code has reportedly led to significant increases in engineer productivity at Anthropic, with internal metrics showing substantial growth in output. Coding Solved: Cherny predicts that coding will become generally "solved" for everyone, potentially leading to the evolution of roles like "software engineer" into more general "builder" or "product manager" titles. Beyond Coding: The future may see models capable of recursively self-improving (ASL4) or being misused for dangerous purposes, highlighting the importance of AI safety. Expanding Form Factors: While originating in the terminal, Claude Code is now integrated into various interfaces like desktop apps, web, Slack, and IDE extensions, with continuous experimentation in UI/UX. Advice for Builders: Focus on latent demand, build for future models, and embrace the "bitter lesson" of general model improvement over specific scaffolding.

The Rise of the 20x Company Claude Code is being used by Anthropic engineers to build and improve AI products, suggesting a fundamental shift in startup operations. 20x companies are characterized by automating all internal functions, not just one or two, allowing small teams to compete with large incumbents. This concept is an evolution of the "compound startup" idea, which focuses on building multiple integrated products in parallel. 20x companies leverage automation across code, support, marketing, sales, hiring, and QA, significantly increasing employee power and delaying the need for extensive hiring. Case Study: Giga ML and Atlas Giga ML, a voice-based customer service agent provider, used an internal AI agent called Atlas to close a deal with Door Dash against much larger competitors. Atlas can perform various tasks within their product, including browsing, editing policies, and writing code, freeing up engineers from boilerplate tasks. Atlas allows each engineer to handle double or triple the workload by automating repetitive customer integration tasks. Atlas functions as a full-time AI employee, enabling Giga ML to service dozens of accounts with only a single human FTE, who focuses on customer relationships and feature requests. Case Study: Legion Health and AI-Native Operations Legion Health is building an AI-native psychiatry network and uses an AI-integrated source of truth to provide instant context to employees. They developed a custom internal interface for their care operations team, allowing access to patient history, scheduling, insurance codes, and more. This single source of truth has enabled Legion Health to scale revenue 4x without hiring new staff, managing thousands of patients and dozens of providers with minimal operational headcount. Case Study: Phase Shift and Custom AI Agents Phase Shift, an accounts receivable automation startup, has a 12-person team competing against companies with hundreds of employees. Their strategy involves bringing AI into every manual process and building custom AI agents for employees based on their documented tasks. This approach has allowed them to avoid hiring a dedicated design person, with the engineering team using AI tools to build front-end designs. The Future of Startup Building Companies can combine approaches like AI teammates, unified sources of truth, and custom agents. These strategies enable startups to stay lean, achieve record growth rates, and gain a significant competitive advantage. The startups that master this "new way to build" are poised to win.

OpenClaw's Origins & Impact Peter Steinberger created OpenClaw, an open-source personal AI agent that rapidly gained 160,000+ GitHub stars. The key differentiator for OpenClaw's success is that it runs on your computer, allowing it to interact with and control "every effing thing" on your machine (e.g., oven, Tesla, lights, Sonos), unlike cloud-based solutions. OpenClaw can access and analyze all data on your computer, leading to surprising insights, such as finding forgotten audio files and creating narratives from a user's past year. The emergence of bot-to-bot interactions and bots hiring humans for real-world tasks (e.g., restaurant bookings) is seen as a natural next step, leading to swarm intelligence rather than a single "god intelligence." The "Aha" Moment & AI's Capabilities Peter's "aha" moment occurred in November after building earlier versions, when he needed his computer to perform tasks while he was away, which evolved into OpenClaw (initially called Cloudbot). He realized the power when his bot, via WhatsApp, transcribed a voice message and responded, even though he hadn't explicitly coded that functionality. The bot autonomously used available tools (e.g., ffmpeg, OpenAI API with curl) to solve the problem in 9 seconds. This demonstrated the AI's ability for creative problem-solving, even choosing the most intelligent approach (using a remote API to avoid slow local model download) without explicit instructions. The bot's ability to understand context and speak his language (sassy, funny) made it a pleasant user experience. The Future of Apps and AI Peter predicts that 80% of apps will disappear because agents can manage data and perform tasks more naturally and efficiently than dedicated apps. Examples include a fitness app (agent automatically tracks food, adjusts gym schedule) or a to-do app (agent reminds you without needing a separate interface). Only apps with physical sensors might survive. While large model companies currently have a "moat" due to their token provision and constant model improvements, models are getting commoditized, and user expectations constantly rise, making older models seem "bad." The true value and "moat" will shift to memory and data ownership, which is currently siloed by large companies. OpenClaw "claws into the data" because the end-user owns their memories as markdown files on their machine, providing access and privacy, especially for sensitive personal problem-solving. Contrarian Development & OpenClaw's "Soul" Peter adopted a contrarian development philosophy, preferring multiple checkouts of a repository on "main" instead of Git work trees to minimize complexity and friction. He avoids UIs, focusing on syncing and text, and is happy to have skipped traditional MCP (Multi-Computer Program) support, instead building a skill that converts MCPs to CLIs, making them usable on the fly without restarts. He argues that bots, like humans, should use CLIs, as no human tries to call an MCP manually. To showcase OpenClaw's capabilities, Peter created a public Discord server with his bot, Multi, which was locked to his user ID but responded to everyone, laughing at those who tried to prompt inject it. His bot's unique character comes from a non-open-source file called "soul.md," which contains core values and principles guiding the human-AI interaction, making the model's responses feel natural and infused with personality.

Introduction and The Power of Claude Code Gary describes using Claude Code as feeling like **"flying through the code"** and highlights its ability to **debug nested delayed jobs five levels deep** and **write tests** to prevent recurrence. **Kelvin French Owen**, co-creator of Codex at OpenAI and founder of Segment, likens Claude Code to a **"bionic knee"** that enables him to code five times faster, recovering from "manager mode." The speaker notes that **startups** embrace coding agents for **speed** due to limited runway, while **larger companies** are more cautious due to existing processes and higher stakes. Technical Advantages & Architecture of Claude Code Kelvin now uses **Claude Code as his daily driver**, preferring it over Cursor due to its superior product and model integration, especially with Opus. Claude Code's strength lies in its ability to **split up context well**; it spawns **"explore sub-agents"** (using Haiku) that traverse the file system, each operating in its own context window and summarizing findings. The **CLI-first approach** of Claude Code is seen as a "weird retro future" that surprisingly outperforms IDEs by distancing users from the code and offering greater freedom and a sense of "flying through the code." Claude Code in the CLI can **directly access development and production databases**, which, despite security risks, proves invaluable for debugging complex issues like concurrency in delayed jobs. The **bottoms-up distribution model** of being able to download and use the tool without needing top-down permissions is considered highly underrated and crucial for rapid adoption in the evolving AI landscape. Influencing the Developer Ecosystem **Generative Optimization (GEO)** refers to how LLMs influence tool recommendations, making **strong documentation, social proof (e.g., Reddit mentions), and open-source projects** critical for developer tool visibility and adoption. LLMs can **directly analyze open-source code**, allowing users to clone repositories and ask agents for code walkthroughs or use them as development harnesses (e.g., Ramp using OpenCode). Building and Using Coding Agents Effectively For agent builders, **"managing context well"** is the most crucial skill, involving careful **context engineering** and using tools like **grep and ripgrep** to supply relevant code snippets. Tips for **top 1% users** of coding agents include: Utilizing platforms like Vercel or Next.js that handle boilerplate to **minimize code and plumbing**. Adopting a **microservices** architecture for well-structured individual packages. Understanding **LLM superpowers** (e.g., persistence) but also their weaknesses, such as a tendency to **duplicate code** or fall into **context poisoning**. **Actively clearing context** when it exceeds 50% of the token limit to prevent the LLM from entering a "dumb zone" where quality degrades. Using **canary tokens** (esoteric facts at the start of context) to detect when the model begins to lose coherence. Leveraging **automated testing, linting, and continuous integration (CI)** to drastically improve agent performance and code reliability. Aggressively employing **code review bots** (e.g., Reptile, Cursor bug bot, Codex) for code correctness. Architectural Differences and Future Outlook **Context management architectures differ**: **Claude Code** delegates to sub-context windows and merges summaries, while **Codex** uses **periodic compaction** after each turn, allowing for much longer-running jobs. **Philosophical approaches diverge**: **Anthropic (Claude Code)** focuses on **building tools for humans** that mimic human co-worker behavior, whereas **OpenAI (Codex)** pursues **Artificial General Intelligence (AGI)**, training models for longer horizons that may operate in non-human-like ways. The **future of engineering** will see **senior engineers** and "manager-like" individuals benefiting most, focusing on directing agents and making architectural decisions. The next generation of engineers is expected to be **more prolific** due to agents helping them complete numerous projects, potentially leading to a heightened sense of "taste." A future vision includes **personal cloud computers with armies of agents** acting as "super EAs," automating tasks and allowing humans to focus on high-level decisions and in-person collaboration. Agents are enabling a shift from the "manager schedule" to the "maker schedule," allowing developers to build in **short "pockets" of time**, as agents handle the heavy lifting of context building. Rebuilding a service like Segment in the future would see the value of basic integrations drop to zero, with focus shifting to **higher-level automation** of data pipelines, customer engagement, and dynamic product experiences. Current Constraints and Evolution The **context window size** remains the **number one limiting factor**, even with sub-context delegation, suggesting that larger context windows would significantly boost performance. **Integration and orchestration** capabilities are still evolving, particularly for connecting agents with other developer tools (e.g., Sentry for auto-generated PRs and phased rollouts). The adoption of **100% test coverage** dramatically accelerates development speed and reliability when working with coding agents. Future developments could include enhanced **agent memory and collaboration**, potentially through shared conversation histories or model-generated wikis, fostering "Clawdbot social networks" among agents. **Model specificities** show that Codex excels in debugging complex concurrency issues where other models like Opus might falter, highlighting unique "personalities" influenced by training data and focus (e.g., Python monorepos for OpenAI, front-end for Anthropic). A tension exists between **OpenAI's strong emphasis on sandboxing and security** (due to prompt injection risks) and startups' willingness to "dangerously skip permissions" for faster iteration.

Introduction & Common Mistakes • The speaker aims to provide conceptual tools for thinking about startup ideas, emphasizing that while no one knows for sure which ideas will succeed, certain ideas are more likely to. • The advice is drawn from analyzing the top 100 YC companies, a classic essay by Paul Graham ("How to Get Startup Ideas"), insights from YC companies that pivot, and mistakes observed in thousands of rejected YC applications. • The most common mistake is building a "solution in search of a problem" (CISP), where Founders start with a technology (e.g., AI is cool) and then look for a problem, often finding only superficial ones. • Founders should instead fall in love with a problem, starting with a high-quality, specific, and tractable problem, not abstract societal issues. • Another common mistake is getting stuck on "tar pit ideas": widespread problems that seem easy to solve but have structural reasons making them very hard or impossible, like the common app for meeting up with friends. • To avoid tar pit ideas, Google it, find past attempts, talk to previous Founders, and understand the core difficulty. • Founders often either jump into the first idea without evaluation or wait for the "perfect" idea, never starting. A good idea is a "good starting point" that can evolve. Evaluating Startup Ideas: 10 Key Questions • Do you have founder market fit? This is the most crucial criterion: are you (the team) the right people to work on this idea? (e.g., PlanGrid Founders with construction and developer expertise). • How big is the market? Look for markets that are already big (billion-dollar potential) or small but rapidly growing (e.g., Coinbase in 2012). • How acute is this problem? The problem should be significant enough that users genuinely care about it. (e.g., Brex solving the problem of startups not being able to get corporate credit cards). • Do you have competition? Most good ideas have competition; lack of competition can be a red flag. If facing entrenched competition, you typically need a new insight. • Do you want this personally? Do you know people personally who want this? If the answer is no to both, it's a concern, and user interviews are critical. • Has this only recently become possible or only recently become necessary? Look for changes in the world (new tech, regulation, new problems) that create opportunities. (e.g., Checkr emerging due to the rise of delivery services needing background checks). • Are there good proxies? Proxies are successful large companies doing something similar but not directly competitive, indicating market viability. (e.g., Rappi using DoorDash as a proxy for food delivery success). • Is this an idea you'd want to work on for years? While passion helps, many successful ideas are in "boring" spaces (e.g., tax accounting software) where passion can grow with success. • Is this a scalable business? Pure software scales infinitely. Beware of service businesses requiring high-skill human labor (e.g., agencies, dev shops). • Is this a good idea space? An idea space is a class of related ideas (e.g., fintech infrastructure, vertical SaaS for enterprise). Some spaces have higher success rates. (e.g., Fivetran pivoting within the fertile data analysis tool space). Ideas That Seem Bad But Are Actually Good • These ideas are often overlooked by other Founders, leaving opportunities on the table. • Ideas that are hard to get started ("schlep blindness"): Tasks that seem too difficult scare off potential Founders, but can lead to huge opportunities (e.g., Stripe dealing with complex credit card infrastructure). • Ideas that are in a boring space: Problems like payroll software (e.g., Gusto) are often neglected but have a higher hit rate than "fun" consumer apps because less competition exists. • Ideas that have existing competitors: Counter-intuitively, most good ideas have competitors. A market with many existing, yet poor, solutions indicates a real problem waiting for a better product (e.g., Dropbox improving on 20 existing cloud storage services with a better UI and OS integration). How to Generate Startup Ideas • The best ideas are often noticed organically, not explicitly thought up (70% of YC top 100). Explicit brainstorming often leads to bad or tar pit ideas. • To foster organic ideas (the "long game"): • Become an expert in something valuable, especially by working at a startup. • Build things you find interesting, even if not immediately business-oriented (e.g., Replika). • 7 Recipes for Generating Ideas Now: • Start with what your team is especially good at: This ensures automatic founder market fit (e.g., Rezi Founders' expertise in real estate and debt financing). • Start with a problem you've personally encountered, especially one you're in an unusual position to see: This leverages unique insights (e.g., VetCove Founders seeing their vet dad's outdated ordering process). • Think of things you personally wish existed: A classic method, but beware of tar pit ideas (e.g., DoorDash Founders wanting food delivery to their dorm). • Look for things in the world that have changed recently: New technologies, regulations, or societal shifts create opportunities (e.g., Gather Town pivoting due to the pandemic's impact on online interaction). • Look for companies that have been successful recently and look for new variants on them: Adapt proven models to new markets or niches (e.g., Nuvocargo as "Flexport for Latin America"). • Go and talk to people and ask them what problems they have: This requires skill. Pick a fertile idea space, then talk to potential customers and other Founders (e.g., A to B Founders systematically interviewing truck drivers and industry experts to find the fuel card idea). • Look for big industries that seem broken: These are often ripe for disruption. • Bonus Recipe: Find a co-founder who already has an idea. • Ultimately, the only way to know if an idea is truly good is to just launch it and find out.

• The redesigned YC website shifts focus from a utilitarian, B2B SaaS template to a storytelling approach, emphasizing founders and their journeys rather than just company logos. • The new homepage incorporates the word "formidable" to describe extraordinary founders, a term used by Paul Graham, and includes a footnote defining it in his words. • To highlight founder success, the website showcases before-and-after transformations of funded founders, emphasizing their humble beginnings and making them relatable to aspiring entrepreneurs. • Founder testimonials are presented as continuous text compiled from interviews, with hover-over details providing information about the speaker and their company, enhancing credibility. • A new section uses AI-generated animation from static photos to bring to life images of recognizable Silicon Valley figures, making them more engaging. • The redesign process prioritized creative exploration using AI tools like Opus 4.5 in Cursor, allowing for rapid prototyping and iteration on interactive elements, which was more efficient than traditional design tools like Figma for achieving the desired animations and storytelling. • The website's aesthetic is minimalist and airy, removing borders and hard dividers to focus attention on founders' faces and stories, deliberately omitting a prominent "Apply" button in the hero section to avoid distractions. • A key message reinforced is that "it's never too early to apply to YC," with the website aiming to inspire potential applicants by showcasing relatable founder stories and the transformative power of the program.

• The most critical factor for improving startup website conversion is to clearly showcase the product itself, rather than relying on abstract illustrations or vague descriptions; this includes providing screenshots or short video walkthroughs so potential customers can understand what the product looks like and how it functions before committing to a demo or purchase. • Animations should be used judiciously to draw attention to key elements and clarify functionality, not as a primary means of communication or for aesthetic overload, as excessive or poorly executed animations can distract and overwhelm users, hindering comprehension. • Call to action buttons, such as "Book a Demo," need to be clearly visible and compelling; if they blend into the design or are presented too early in the user journey before the value proposition is understood, conversion rates will suffer. • A/B testing different calls to action, such as changing "Book a Demo" to "Show Me the Product" or "Watch a Demo," can significantly increase engagement by providing a lower-commitment entry point for interested users. • To improve user understanding and reduce friction, websites should provide literal, concrete explanations of product features and benefits, rather than generic marketing speak or abstract concepts, especially when competing against established free alternatives like Google Slides. • Offering a frictionless trial or demo experience, such as allowing users to interact with the product before requiring a sign-up, is crucial for capturing user intent and guiding them toward an "aha moment" of product value, particularly in competitive markets. • Website design should prioritize clarity and user experience, avoiding overwhelming amounts of information, overly complex animations, or inconsistent UI elements that can lead to a perception of a lack of detail orientation and undermine user trust. • Simplifying the core message and offering a clear, focused presentation of what the product is and how it solves a problem is more effective than trying to include every feature and benefit on a single page, which can create noise and detract from the main value proposition.

• Diffusion is a fundamental machine learning framework for learning probability distributions of data across any domain, particularly excelling in mapping from high-dimensional to high-dimensional spaces, even with limited data. • The core diffusion process involves taking data, progressively adding noise to create a sequence of noised-up versions, and then training a model to reverse this process, learning to denoise from pure noise back to the original data distribution. • Innovations in diffusion models have focused on refining the denoising objective, moving from predicting the original data to predicting the added error or velocity, which simplifies the learning process for the model and often leads to cleaner, more stable training. • Flow matching is a simplified approach to diffusion that bypasses intermediate noising steps by directly learning a global velocity vector between the noise and the data, allowing for a more direct and efficient generation process, often requiring as little as 10-15 lines of code. • Diffusion models have broad applications beyond image generation, including protein folding (AlphaFold), robotic policies (diffusion policy), weather forecasting (Gencast), DNA and molecule binding prediction (DiffDock), and increasingly in language models (diffusion LLMs) and code generation. • While diffusion has "eaten all of AI" except for reinforcement learning (MCTS for games like AlphaGo) and certain LLM applications, its core procedure is becoming simpler and more effective, suggesting widespread future impact across various industries and enabling new companies in areas like robotics, text generation, and video.

• Finding your first users is a search problem, not a persuasion problem; look for people who are early adopters or have a burning need that your product solves. • Charge early adopters real money for your product; paying customers provide sharper, more valuable feedback than free users. • Utilize targeted personal outreach, such as cold emails or knocking on doors, rather than broad advertising methods to find these initial users. • Launch your product early and create a wide surface area for potential users to find you, as you won't know exactly who they are at the outset. • Treat your early users like an anthropologist studying a new civilization to understand their decision-making processes and motivations for trusting your product. • Develop a "minimum evolvable product" that can adapt and evolve based on user feedback and market pressures, rather than aiming for a perfect, final form from the start.

• Stoke Space is developing fully and rapidly reusable rockets, including the second stage capsule, which is typically discarded on every mission due to extreme re-entry heat (over 2700° F) and speeds (17,000 mph). • Their innovative approach to stage 2 reusability involves a custom heat shield utilizing cold liquid hydrogen flowing through a heat exchanger to absorb reentry heat, complemented by 24 small thrusters for deceleration and controlled landing. • The Nova rocket features a highly fuel-efficient first-stage engine, designed for rapid reusability, while the second stage, Andromeda, is engineered to survive re-entry and land. • Stoke Space's origin story involves founders Andy and Tom leaving Blue Origin in 2019, bootstrapping their initial engine testing in backyard shipping containers, and overcoming fundraising challenges during the pandemic, eventually raising approximately $990 million. • They emphasize rapid iteration by building most components in-house, allowing for quick learning cycles from failures; for instance, reducing a component iteration from a month to a couple of days by bringing testing back to their factory. • A key element of their operational success is a custom software platform called "Bolt Line," designed to manage and track all aspects of vehicle maintenance, operations, and data logging, bridging the gap from garage beginnings to FAA-regulated flights.

• The AI economy has stabilized, with clear layers for model, application, and infrastructure companies, and a developed playbook for building AI-native businesses. • Anthropic has surpassed OpenAI as the preferred LLM provider for Y Combinator-backed startups, moving from ~20-25% usage to over 52% in the Winter 2026 batch, driven by strong performance in coding tools and agents. • Gemini is also climbing in popularity, now at 23% usage among YC applicants, with users impressed by its reasoning abilities and its integration with Google Search for real-time information. • The "AI bubble" concern is compared to the telecom bubble of the '90s; while there's massive infrastructure investment, it creates an opportunity for application-layer startups, much like YouTube emerged from the excess bandwidth. • The AI revolution is in its "deployment phase," following an "installation phase" of heavy capital expenditure, leading to abundance and new opportunities for founders to build applications on top of existing infrastructure. • Despite initial skepticism, companies are exploring space-based data centers and fusion energy to address power generation and land constraints for AI infrastructure, with companies like Google and Elon Musk pursuing space solutions. • The skill set for building AI models is becoming more democratized, with a growing number of individuals possessing the research, engineering, and business acumen required, leading to an increase in applied AI companies and more specialized models. • The trend of AI increasing efficiency is leading to higher customer expectations, driving companies to still hire significantly to meet demand and compete, rather than reducing workforce size. • Companies like Gamma are demonstrating a new "reverse flex" by achieving significant ARR with a small number of employees, indicating a shift towards efficiency and lean operations in the AI startup landscape.

• Intelligence is defined by the ability to learn new things efficiently, not just by excelling at known tasks like chess or Go. • The ARC benchmark, developed by François Chollet, tests an AI's ability to learn new skills rather than just perform on existing ones, with a focus on tasks that average humans can solve. • Early LLMs performed poorly on the ARC benchmark (around 4-5% for GPT-4 base), but recent advancements, particularly with reasoning paradigms, have significantly improved performance (e.g., 21% with 01 preview). • ARC AGI 3, launching next year, will be an interactive benchmark featuring 150 video game-like environments where AI must infer goals without instructions, testing generalization through action and feedback. • Efficiency in AI will be measured not only by accuracy but also by the number of actions and data points required, drawing parallels to human learning efficiency, with ARC AGI 3 normalizing AI performance to average human actions. • Solving ARC AGI is considered necessary but not sufficient for achieving Artificial General Intelligence (AGI); a system that excels at ARC AGI would demonstrate strong generalization capabilities but wouldn't necessarily be AGI itself.

• PostHog helps users debug products and ship features faster, consolidating customer and product data, with around 160 employees and 300,000 customers. • The company's initial successful product was self-hosted open-source product analytics, born from the frustration of repeatedly setting up analytics during multiple pivots, and it gained traction on Hacker News. • PostHog's marketing strategy, including bizarre billboards and a unique website, focuses on standing out and generating awareness through humor and unexpected comparisons rather than direct conversion. • James Hawkins, CEO of PostHog, emphasizes that having a clear, albeit potentially contrarian, plan is crucial when raising funds, and that building a remarkable product and brand requires going significantly beyond the typical 80/20 effort. • PostHog is now doubling down on AI to build "product autonomy," aiming to automate feature development and product management tasks, which is enabled by their multi-product infrastructure and substantial funding.

• Amplitude initially approached AI with skepticism, viewing its capabilities as "jagged" and facing frustration from external pressure to adopt it without a clear strategy. • A turning point for Amplitude was recognizing the transformative effect of AI on software engineering, leading them to seriously invest in AI adoption around October 2024, marked by hiring a new engineering leader and acquiring Command AI. • The company's pivot to AI involved a significant internal effort, including an "AI week" for training and hackathons, to get the existing team bought-in and proficient with AI tools before focusing on building new AI-native products. • Amplitude's core strategy shifted from a customer-driven "faster horse" approach in traditional SaaS to a technology-first understanding of AI capabilities to map them to product solutions, acknowledging that customers cannot always articulate needs for novel AI functionalities. • The transition required organizational restructuring, including reorganizing the engineering, product, and design teams twice within a year and moving away from leaders solely focused on the pre-AI SaaS modality.

• AI is enabling new consumer startup opportunities by making previously impossible tasks feasible, such as democratizing music creation with AI-powered tools like Suno, which allows anyone to create music. • The success of consumer startups like Ankor (podcast platform) and Suno demonstrates a historical trend of technological advancements democratizing content creation, from photos and videos to audio and now music, with AI being the key enabler for music. • While AI is making product creation easier, distribution remains a critical challenge for consumer startups, though new AI-driven distribution channels are expected to emerge. • The emergence of AI allows for the rebuilding of foundational technology stacks and presents opportunities to leverage large, previously untapped data sets (e.g., health data, camera rolls) by applying LLMs and other AI models to create new insights and experiences. • AI is poised to revolutionize education by creating highly personalized and efficient learning experiences, moving beyond the current one-size-fits-all models, as exemplified by Obo Labs, which aims to invest in human intelligence through AI. • Developing "taste" and "craft" is becoming crucial for consumer founders to stand out in a hyper-competitive landscape where AI lowers the barrier to product building, though the durability of taste as a long-term asset is still being tested.

• The core value of the video is to provide actionable insights into improving startup website design, specifically for sites built with Cursor, a popular AI coding tool. • Rio Lu, Head of Design at Cursor, offers specific critiques and recommendations on various websites submitted by YC founders, focusing on clarity, messaging, user experience, and visual design. • Key takeaways for improving startup websites include: avoiding jargon and vague language, clearly communicating the product's value proposition and target audience, ensuring consistent and polished visual design (e.g., typography, spacing, color palettes), making calls-to-action prominent and clear, and providing sufficient information without overwhelming the user. • Lu advises against common design pitfalls such as "AI slop" (e.g., massive shadows, purple gradients, bad typography), distracting animations, and confusing naming conventions. • He emphasizes the importance of talking to the target audience in their language and focusing on solving their problems, rather than using internal company jargon or overly technical terms. • The review highlights that while AI tools can assist in website creation, a strong foundational design system and clear communication strategy are crucial for creating effective and compelling user experiences.

• AI is projected to make ocean container shipping 8-10% cheaper over the next few years, with AI contributing significantly to this reduction. • Flexport's AI initiatives have already saved 2% of their ocean freight spend while simultaneously improving transit time by 20%, a feat that typically involves a trade-off between cost and speed. • AI is being implemented across Flexport's operations, from enhancing customer user experience and optimizing container loading to automating tasks previously done via email, phone, or manual processes. • Flexport leverages AI to process complex, unstructured data like large Excel files from logistics contracts, converting them into structured formats for intelligent analysis. • The company is actively promoting AI adoption internally, with 90% of recent hackathon projects focusing on Large Language Models (LLMs), leading to the development of real product lines and features. • Flexport offers a 90-day program for non-engineers to learn AI skills, aiming to increase their productivity by up to ten times by enabling them to build automation tools for their roles. • A customer-facing AI feature allows users to ask natural language questions to generate reports and charts, reducing account management time spent on data report generation by 25%. • Internally, an AI agent handles routine tasks like verifying warehouse addresses and scheduling delivery appointments, including making automated phone calls to confirm details, a task too costly for humans to perform consistently. • Flexport aims to automate 80-95% of work within the next year, driven by advancements in LLMs, with the understanding that labor costs represent about 10% of the total cost in the freight forwarding layer of logistics.

• StarCloud is developing orbital data centers to provide GPU compute to satellites and eventually compete with terrestrial data centers on energy costs. • Their first satellite, StarCloud 1, successfully launched into orbit carrying an Nvidia H100 GPU, marking the first instance of data center-grade GPUs operating in space, performing 100 times better than any previous space computer. • The core innovation of StarCloud lies in their proprietary deployable radiator technology, which enables efficient heat dissipation into deep space using infrared radiation, requiring zero fresh water and significantly reducing carbon emissions compared to Earth-based data centers. • StarCloud aims to build massive 40-megawatt orbital data centers powered by uninterrupted solar energy, designed to overcome the limitations of land, grid power, and cooling faced by terrestrial facilities. • The company's rapid development, from founding to satellite launch in 15 months, was attributed to the founders' complementary expertise in software, data center infrastructure, and satellite engineering, along with in-house fabrication of key components. • Future plans include launching a second satellite in October with at least 10 times the power of the first, featuring Nvidia's Blackwell architecture, multiple GPUs, and high-bandwidth optical terminals for 24/7 low-latency connectivity.

• Start by prioritizing selling the company to candidates, not just interviewing them, as this is a common founder mistake. • To stand out in outreach, personalize messages deeply, potentially spending five minutes per email, and get creative with sourcing strategies beyond standard platforms. • When sourcing Account Executives (AEs), look for signals like consistent quota attainment, quick promotion cycles within a single company, and experience in fast-paced startup environments. • For software engineers, focus on unique advantages like specific technical skills, contributions to open-source projects, experience building personal projects, or founder-like experience. • A compelling outreach email should be concise, personalized, establish company legitimacy with customer/momentum details, and include a clear call to action, with follow-ups adding further value. • Aim for response rates between 10-20% and interested rates of roughly half of that, understanding that interested rate is a better metric than just reply rate. • Founders should schedule dedicated time for sourcing and outreach, aiming to speak with at least 10 candidates per week, and every founder should be involved in the early hiring process. • When making an offer, leverage speed as an advantage over larger companies and personalize the offer to the candidate's specific motivations identified during the interview process. • Founders should consider hiring external recruiting help when they are certain of making multiple hires (more than two concurrent roles) within a 6-12 month period.

• The core reason enterprises struggle with AI implementation, leading to a high failure rate (often cited around 95% of projects), is not necessarily the AI itself, but the inherent limitations and resistance within large organizations: engineers may not believe in AI, internal IT systems are often outdated and siloed, and political turf wars hinder progress. • Startups have a significant advantage because they can build functional AI products more effectively than enterprises can internally or through traditional consulting firms (like Deloitte or EY), who often lack the deep technical expertise to build sophisticated software despite their ability to mediate requirements. • Success for startups selling to enterprises often hinges on embedding deeply into business processes and integrating with systems of record, a strategy that differs from typical plug-and-play SaaS models but can yield substantial rewards (the "pot of gold") if successful. • Startups can win deals by building genuine relationships with "champions" within enterprises – individuals who may have entrepreneurial dreams but are risk-averse, and who can live vicariously through the startup's journey and champion their cause internally. • Enterprises are actively seeking AI solutions and are increasingly willing to bet on new startups, recognizing that established vendors and internal IT departments often struggle to deliver effective AI, creating a "startup-shaped hole" in the market for specialized AI solutions. • The high switching costs for enterprises once they invest time in training an AI system create a significant moat for startups that successfully onboard them, providing a competitive advantage against potential future entrants.

• To build a successful AI startup, focus on identifying what people are already paying for and aim to assist, replace, or enable previously unthinkable tasks with AI. • The total addressable market for AI applications is significantly larger than traditional software, as it can capture the value of entire salaries previously spent on human labor, not just SaaS subscription fees. • Building a reliable AI product requires deep domain expertise to understand precisely what professionals do and how the best in the field would perform tasks with unlimited resources; this understanding should be translated into specific steps and then into well-crafted prompts or code. • Rigorous evaluation is crucial for AI products; focus on defining what "good" looks like for each task, create objective metrics (e.g., true/false, numerical scales), and iterate relentlessly on prompts and evaluations, aiming for high accuracy (97%+) before production. • The most critical factor for marketing and sales in AI startups is building an outstanding product; while marketing is necessary, a superior product will generate organic word-of-mouth and inbound interest, making sales efforts more effective. • When marketing and selling AI services, price based on the immense value delivered rather than traditional software models, and listen to customer preferences for pricing structures (e.g., predictable subscription vs. per-use). • Building trust with customers for new AI products involves demonstrating superiority through head-to-head comparisons with existing human services, pilot programs, and robust post-sale customer engagement, including training and support, as the product is more than just the UI. • Defensibility in AI startups comes from the complex, iterative process of building and refining the product, including data integrations, model selection, and prompt engineering, which creates a unique, difficult-to-replicate asset over time, rather than solely relying on underlying proprietary models.

The Problem of Sequential Data Early AI struggled with understanding sequences (like text) due to the vanishing gradient problem in Recurrent Neural Networks (RNNs). Gradients, used for training, would fade as they passed backward through long sequences, diminishing the influence of early inputs. Long Short-Term Memory (LSTM) Networks Introduced in the 1990s by Hochreiter and Schmidhuber to combat vanishing gradients. LSTMs used "gates" to selectively keep, update, or forget information, enabling learning of long-range dependencies. Became viable and dominant in Natural Language Processing (NLP) in the early 2010s with advancements in GPUs and data. Sequence-to-Sequence (Seq2Seq) with Attention A limitation of early LSTM-based Seq2Seq models was the "fixed-length bottleneck," where an entire input sequence was compressed into a single vector, losing information for long sentences. The 2014 Seq2Seq with Attention model allowed the decoder to "attend" to relevant parts of the encoder's hidden states, improving alignment and performance. This led to significant improvements in machine translation, making models competitive with older systems and powering tools like Google Translate. The attention mechanism also hinted at broader applicability beyond NLP, with early use in computer vision. The Transformer Architecture Proposed in the 2017 paper "Attention Is All You Need" by Google researchers. Eliminated recurrence entirely, relying solely on attention. Used self-attention to update token representations based on weighted relationships with all other tokens in the sequence. Enabled parallel processing of entire sequences, drastically improving speed and accuracy over RNNs. Led to variations like BERT (encoder-only) and GPT (decoder-only), which became the foundation for modern Large Language Models (LLMs). The ability to scale these models with large datasets was crucial for their development into general-purpose AI systems.

AI Go-to-Market Strategies For AI companies in legacy industries, three GTM approaches exist: Build AI software to sell to existing businesses (most common). Start a new business in the legacy industry and automate it. Acquire an existing business and integrate AI. When starting a new business and automating, focus on the percentage of work automated, aiming to increase it over time. Software founders are well-positioned to identify automatable tasks. Success in the second GTM approach depends on tracking automation rate over revenue growth. Partnering with early adopters who are enthusiastic about new software can be crucial. Qualifying potential customers is vital; look for individuals empowered and incentivized to adopt new software. Mid-Market vs. Enterprise AI Sales For early-stage AI companies, prioritize the pace of learning about customer needs and pain points. Starting with the mid-market or even smaller companies that have the problem is generally advisable for faster learning and iteration. Sales cycles in enterprise can be long; focus on finding empowered individuals within organizations. A narrower product scope or focusing on a few key users can shorten sales cycles. Hiring and AI Sales Roles AI SDRs are most effective when integrated into an already functional sales process. Founders must still master the art of selling. Founders need to understand "who to sell to" and "how to get their attention" before AI can effectively assist. Don't hire growth hackers or AI SDRs as a last resort to fix a broken sales process; founders must first figure out the core selling mechanism. Founders should be curious and learn the roles they might hire for before bringing in new team members to manage expectations. Investing in AI Development Assess if your product will become irrelevant with new model releases or significantly better. Investing in development now, even if models improve later, provides valuable learning and a stronger product when new models are available. Pivoting a Startup Pivoting is necessary when traction exists but isn't strong enough, or if the current direction isn't working. A key indicator for pivoting is a lack of customer conviction or value for the current product, even with some revenue. Pivoting requires significant energy and conviction, as it often means starting from scratch. When considering a pivot, explore a range of ideas rather than focusing on just one. The leading indicator for a pivot is often when founders stop believing in their current product's success. A "great" startup idea is validated by customers who express a daily need and solve a real pain point. Technical Challenges and Pivoting Technically difficult ideas can be good opportunities if founders have the courage and skills, as fewer competitors will pursue them. Reduce scope or build a simpler version first to tackle complex technical challenges, then scale up. Don't let technical difficulty become an excuse to avoid customer interaction. When to Hire The right time to hire is when things are so busy that you can't find time for interviews, indicating things are working but breaking. Early indicators of breaking points in engineering, sales, or onboarding suggest it's time to consider hiring. Hiring is challenging for startups; focus on personal networks for early hires. Hiring is not a success metric; it's a necessity for a functioning company. Opportunistic hires (e.g., "smartest friend") can be effective if the person is truly exceptional and the timing is right. Open Sourcing Enterprise SaaS Products Open-sourcing is common for dev tools but can also build trust and shorten sales cycles for enterprise SaaS, even if customers don't directly inspect the code. Open-sourcing can address concerns around privacy and sensitive data, facilitating self-hosting. Drawbacks include the cost of supporting self-hosted versions, which requires higher pricing.