Y Combinator Summaries & Notes

Introduction to Harnesses

• Harnesses (scaffolding/prompt engineering) are crucial for AI progress, not subpar research.
• They enable quick adaptation to new domains and problems.
• Harnesses allow for dynamic adaptation and learning from test-time experience.

Evolution of Harnesses

• V0 Harness: Basic GPT-2 with end-of-sequence loop, top-P sampling, and environment interaction.
• Context Innovations: Few-shot learning (fshot learners), Chain-of-Thought (predicting logic over more tokens).
• Tool Use: Tool Formers and WebGPT enabled calling external tools (e.g., Python for calculations).
• Memory Management: MEG-GPT introduced CRUD operations on context ("memory").
• Skills and Chaining: Voyager enabled chaining tools and distilling learned skills into prompts.
• Self-Reflection/Refinement: ReAct, Self-Refine, and Reflection allowed agents to critique and improve their own outputs, leading to multi-agent collaboration.
• Recursive Agents: RLMs allowed for recursive agent calls to solve larger problems.

Static vs. Self-Improving Harnesses

• Static Harnesses (V1): Fixed agent specs, system prompts, turn limits, tool/skill lists. These run in a loop with context compilation and action execution.
• Self-Improving Harnesses: Harnesses that learn and adapt.
• DSPY: Uses genetic programming to learn optimal system prompts from examples.
• Darwin Machines: Allow agents to modify their own harness code.
• Meta Harnesses: Generate other harnesses, a highly meta-concept.
• Continual Harness: Enables CRUD operations on harness components (memory, skills, prompts) and even online learning (Dagger-style) to update model weights at test time.

Prime Agent: A Self-Improving Harness

• Focuses on a first-principles approach to harness building.
• Adds persistent state, tools, and compute between the LLM and the world.
• Organizes agents into a project orchestrator (root session) and sub-agents.
• Utilizes a hierarchical cache-like context (model weights, active context, dispatch state/ripple).
• Employs "ripples" for in-RAM computation and garbage collection.
• Manages memory, skills, sub-agents, and prompts through refinement.
• Compares favorably to other harnesses on benchmarks like ARC AGI, showing significant performance gains.
• Emphasizes long-horizon performance and efficient context management.

Open Jarvis: On-Device Personal AI

• Aims to build a personal AI stack entirely on-device, addressing cost, privacy, and energy concerns of cloud-based LLMs.
• Utilizes local LLMs (e.g., Qwen, Gemma) and inference engines (Ollama, Llama.cpp).
• Defines primitives: user interfaces, agentic logic, intelligence (LLM), inference engine, tools/memory, and learning.
• Demonstrates competitive performance with cloud LLMs for personal use cases.
• Leverages cloud LLMs to optimize local configurations for cost and latency reduction.

QM: YC's Open-Source Agent Harness

• Provides a customizable, cloud-like assistant for YC employees via Slack or web UI.
• Centralizes agent conversations and sandboxes in PostgreSQL for efficient management.
• Sandboxes are treated as resources rather than fixed homes for agents.
• Wires agents to company resources (CLI, API keys) and mimics local development environments.
• Employs a "grind tool" to set budgets (time/tokens) to prevent agents from giving up early.
• Addresses agent confusion in social contexts and manages information sharing with a fine-grained permission system.
• Keeps the harness thin, focusing on core tools for execution, storage, and app publishing.

AI Model Trends

• Shift towards open models, especially in enterprise, driven by coding agents and AI assistants.
• Chinese origin models are currently dominant in cloud-hosted open model usage.
• US and German businesses are significant users of open model tokens via cloud services.

Key Drivers for Open Model Adoption

• Cost reduction is the primary pain point solved by open models.
• Businesses seek better control and customization of AI for unique use cases.
• Example: AT&T shifted 40% of token consumption to open models, predominantly for coding agents.

Open Model Capabilities & Growth

• Explosive growth in token usage per developer, driven by coding agents and workflow automation tools like OpenClaw and Hermes.
• Context windows have expanded significantly, enabling more complex tasks.
• Olama Cloud has seen 150x growth since the start of the year.

Fine-tuning vs. Out-of-the-Box Models

• Early 2024 saw interest in fine-tuning, which waned but is now returning.
• The accelerating release cadence of open models makes custom training harder.
• Improved tooling now supports fine-tuning efforts.

AI Safety and Security

• AI safety is a growing concern, potentially slowing frontier model development.
• Security and safety are major blockers for enterprise adoption of open models.
• Open models can be used for security testing where closed models refuse.
• Hugging Face used open-weight models to detect hacks.

Olama's Role and Technical Challenges

• Olama acts as an "operating system" for AI, integrating hardware, inference engines, and application runtimes.
• Ensuring fast and accurate model performance across diverse hardware is a key challenge.
• A "playbook" exists for day-zero model launches, focusing on harness support, use cases, and hardware optimization.
• Challenges include model architecture changes, tool calling mechanics, and rigorous benchmarking.

Ecosystem Dynamics: Bundling vs. Unbundling

• Trend towards specialized "best of breed" companies for knowledge, coordination, and execution layers, rather than bundled solutions.
• Open source fosters competition and innovation in these unbundled areas.

Future of AI Workflows

• "Hidden layers" between models and applications offer opportunities for new companies.
• The next scarcity is above tokens: orchestrating agents and complex workflows.
• "Flash models" (e.g., Deepseek Flash) offer ultra-low cost per token and task, enabling "unlimited tokens" for common use cases.
• Chaining cheaper models together can solve complex problems through orchestration.
• Most AI tasks may not require "god models"; "good enough" models coupled with orchestration are sufficient.

Geopolitics and Model Origin

• Customer concern varies: some prioritize model origin, others focus on where/how it's run.
• Data origin and model "voice" (communication style) are important factors.
• For critical tasks (e.g., power plant analytics), model origin is paramount for trustworthiness.
• IT and security teams in large enterprises are experienced with managing dependencies, similar to open-source software security.

Olama's Origin and Growth

• Founded by ex-Docker developers, focused on developer experience.
• Pivoted to open-source LLMs after realizing the difficulty of running them locally.
• Rapid growth on GitHub (100k stars quickly) indicated strong product-market fit.
• Transitioned from hobbyist users to 85% of Fortune 500 adoption in approximately 18 months.

Monetization Strategy

• Initially focused on a privacy-focused AI product.
• Waited for product-market fit with open models for complex use cases (e.g., coding agents).
• Monetization began with Olama Cloud, leveraging the abundance of open models.
• Emphasis on customer connection and understanding evolving needs.

Local vs. Cloud Models

• A hybrid approach is expected, with easier tasks running locally for lower latency and cost.
• Modern hardware (Apple Silicon, Nvidia DGX Spark) can run large parameter models effectively.
• Coding agents often perform best with large cloud models due to task complexity.
• Document processing and simpler workflows run well locally.
• This hybrid model further reduces costs by utilizing existing hardware.

GPU Market and Infrastructure

• High demand and volatility in the GPU market.
• Access to high-end GPUs (e.g., B200, B300) is challenging for startups.
• Inference providers are crucial for accessing and scaling these resources.

Startup Evolution and Ambition

• Most of starting a startup remains the same, regardless of technological advancements.
• YC-funded startups are now more ambitious, tackling problems like curing cancer and developing intercontinental ballistic cargo.
• The term "frighteningly ambitious" from 2012 has seen many ideas realized, such as OpenAI, which disrupted search by providing information directly.

Founder Qualities and Motivation

• Founders need to be ambitious to overcome grueling startup challenges; mere talent isn't enough.
• A founder's day-to-day motivation is often the fear of failure, not the prospect of becoming a billionaire.
• Ambition is often inherent; founders who seem less ambitious might have been trained to suppress it.
• Starting a startup is a difficult path, not a credential to collect; it requires immense hard work and determination.

"Formidable" Founders and "Lean" Startups

• "Formidable" describes someone who consistently gets what they want, aligning investor and founder interests.
• The concept of a "lean startup" is unclear to Paul Graham, but he believes startups can still be launched with minimal funding, as technology generally gets cheaper.
• Even expensive ventures like rocket startups can be initiated with limited funds by focusing on achievable milestones and subsequent fundraising.

AI and the Future of Startups

• Current AI development is surprising; instead of starting with simple organisms and progressing to human-level intelligence, AI began as a "bullshitting undergraduate" and is becoming more refined.
• AGI is best understood as a "smear" or a range, not a clear finish line, explaining why AI might excel at complex math but fail at simple tasks.
• The pace of shipping new products remains a key predictor of startup success, even with AI tools.
• The biggest new cost for startups is AI computation (GPUs/tokens), shifting from primarily salaries.

YC's Role and Evolution

• YC provides a community and support system, alleviating the loneliness of startup life.
• The YC "batch" concept, initially conceived as a replacement for summer jobs, proved highly effective and became the core model.
• YC's core principles have remained consistent despite growth; what was said about it being "too big" years ago is still said today.
• The next trillion-dollar company will likely come from "formidable" founders, not a specific idea or sector.

Largest Electric Aircraft Flight

• Hart Aerospace's new electric aircraft, the largest of its kind, has successfully flown.
• It boasts a 100-foot wingspan and a takeoff weight of 25,000 lb.
• The electricity cost for its first flight was a mere $5.

Hart Aerospace's Vision and Technology

• The company aims to reduce air travel costs using hybrid-electric technology.
• Their aircraft replaces jet engines with simpler, more reliable electric motors with few moving parts.
• The plane uses eight battery packs, equivalent to about four Teslas, for power.

Addressing Aviation Inefficiencies

• Traditional jet engines are inefficient for short flights, with significant fuel usage during taxiing, takeoff, and landing.
• Hart's aircraft targets the market for flights under 2 hours, which constitutes half of all global flights.
• The goal is to make regional travel cheaper and more accessible, like turning a 6-hour drive into a 20-minute flight.

Company Origins and Development

• Founded by Anders and Claraara, the company started with a 3D-printed model 7 years ago.
• Their team has grown to 40 people, developing technology in-house.
• Initial airline interest came from LOIs (Letters of Intent), followed by pre-orders after demonstrating a functional electric motor.

Design Philosophy and Manufacturing

• Hart's aircraft intentionally resembles conventional turboprops to appeal to a traditional market.
• The pilot plant is used for in-house development and testing of components like actuators.
• The entire plant functions as a test bench, enabling rapid iteration and fault injection testing.

Battery Technology and Challenges

• The team is actively testing battery cells from various manufacturers to optimize for flight cycles, safety, cost, and energy density.
• A major challenge is accommodating reserve power for diversions and longer distances, as batteries don't lighten during flight.
• To overcome this, they are using a hybrid system with a simple turboprop engine.

Future of Aviation and Hart's Role

• The hybrid system allows for 125 miles of all-electric flight and up to 500 miles in hybrid mode.
• Hart Aerospace is positioned to replace aging aircraft designs and potentially scale to larger planes.
• The long-term vision includes reducing pilots in the cockpit, starting with cargo and moving towards autonomous flights.

Company Origin and Growth

• Legora started with a rejection from Y Combinator but was eventually accepted under a new name with a different platform after doing their homework.
• The company experienced rapid growth, reaching 100 million in ARR from 1 to 100 million in just over a year, starting with three engineers in Sweden.
• They moved from a cramped law firm office to their first real office, highlighting the challenges and resourcefulness of early-stage startups.

Product Development and Strategy

• Legora is an "agentic operating system for lawyers" that handles complex legal work from start to finish.
• The initial product, built with GPT 3.5, could explain stock option agreements.
• They emphasized the importance of learning the legal industry by cold-emailing lawyers and offering to pay for their time.
• A key strategy was freezing sales for six months to focus on product development, which led to explosive growth afterward.
• They developed the "Legora product manifesto" to unify the team and focus development efforts.
• The company's core IP is the ability to evaluate use cases and route tasks effectively, leveraging different LLMs.
• They are transitioning from reactive agents to proactive agents, where AI can initiate tasks based on triggers or connected data.

Startup Philosophy and Culture

• A key takeaway is the willingness to learn faster than anyone else and not take "no" for an answer.
• Building a company where others want to see you succeed and help is crucial.
• Legora values ambition, evidenced by their rapid expansion and growth targets.
• They prioritize hiring individuals with high growth potential and a strong work ethic over fancy resumes.
• Core values are "Lean in, fight for excellence, and grow together" (LFG).
• They've blended US, European, and Asian cultures to create a global company.
• The company culture emphasizes winning and operating as if there's no room for second place.
• They foster a "solution mode" rather than "blame mode" culture.
• Ambition can be learned and amplified by being around other ambitious people.
• Storytelling is a critical skill for CEOs, essential for selling the company to oneself, employees, investors, and customers.
• Velocity and the ability to iterate on customer feedback are vital, even at scale.
• Technical skills, especially building systems at scale and being a good software engineer, remain crucial.

Overcoming Challenges and Pivots

• Initial YC interview questions revealed a lack of domain expertise, but they persevered.
• They faced skepticism from the conservative legal industry but changed perceptions of AI in law.
• A six-month sales freeze was a difficult but necessary decision for product refinement.
• They learned to unlearn hiring patterns focused on past prestigious companies, instead seeking high-potential individuals.
• The early days involved immense discomfort, which was a key learning driver.
• They had to re-qualify for their roles as the company scaled rapidly.

Key Insights and Lessons

• Domain expertise might not be as critical as a willingness to learn the market.
• Betting on the continuous improvement of AI models is more important than fine-tuning early, less capable ones.
• Focus on delivering value to your market today and one step ahead.
• Building relationships with customers is key to adapting to future model capabilities.
• The ability to evaluate use cases is a superpower.
• Compute availability and data processing agreements are significant logistical challenges.
• Having a competitive peer group is essential for driving ambition.
• Storytelling is paramount for a CEO.
• Technical skills and rapid iteration are still vital, alongside building scalable systems.
• The transition to proactive agents is a major unlock for scaling lawyer productivity.
• Compressing timelines and ruthlessly prioritizing are necessary for hyper-growth.
• Persistence ("nagging") can be an underappreciated negotiation tactic.

Why Data Matters

• Data is no longer seen as a commodity; its importance has shifted from 5% to 97% of focus in production, compared to architecture.
• The key to improving model performance is analyzing and fixing data issues (false positives/negatives) rather than solely tweaking architecture.
• Production environments often present data distributions significantly different from training sets, causing models to fail.
• Data quality is crucial for automating complex tasks and is now considered a product requiring craftsmanship.

Data 1.0: Manual Labeling Challenges

• Manual labeling is not scalable due to cost and the need for repeated cognitive effort.
• It's not robust to noise and requires redundancy (k by n relabeling).
• Changes in specifications require starting over, and there's a lack of provenance.
• Manual labeling is intractable for expert domains with limited experts and time.

Data Programming & Weak Supervision

• Encode expert supervision into software using "labeling functions" for programmatic and scalable data labeling.
• Handle overlapping and inaccurate signals through weak supervision techniques.
• Model each "voter's" (labeling function's) accuracy in an unsupervised way to estimate ground truth.
• Train noise-aware models to improve generalization beyond the initial labeling functions.

Data 2.0: Evolving Complexity

• Increased complexity in input, output, and environment requires more sophisticated data.
• Traditional benchmarks like HumanEval are saturating, but new challenges emerge with more complex agents.
• New benchmarks like Senior SWEBench aim to evaluate models on tasks representative of senior engineers, requiring nuanced reward design.
• A "validation agent" approach balances reliability and flexibility in evaluating code, capturing expert intuition through test scripts.

Diffusion Models and Real-Time Performance

• Diffusion models generate tokens in parallel, offering significant speed advantages over autoregressive models.
• Inception Labs' Mercury 2 models achieve over 1000 tokens per second, enabling real-time voice applications and improving LLM reasoning.
• Diffusion models can achieve state-of-the-art quality and speed on GPUs, reducing reliance on specialized hardware.

Data for Diffusion Models & RL Environments

• High-quality data is critical for training and evaluating diffusion models, especially for voice applications.
• Existing benchmarks like Towbench may not capture real-world complexity and are heavily benchmarked.
• "Dow Forge" synthesizes realistic RL environments and tasks based on real-world data and user interactions.
• This system allows for the creation of domain-specific environments and tasks, improving model performance and reducing overfitting.

Multilingual Pre-training and Transfer

• Most research focuses on English, neglecting the performance of other languages due to data scarcity.
• Training solely on limited data for a specific language leads to overfitting.
• Empirical measurement is needed to identify synergistic data sources across languages, which may not align with linguistic families or scripts.
• A cross-lingual transfer matrix can quantify synergy and interference between languages, showing how model size impacts these relationships.
• Script similarity can be more influential than language family for transfer, and transfer is not always symmetric.
• A new scaling law models multilingual data mixtures, accounting for monolingual source data and cross-lingual transfer to predict model loss.

AI Strategy & Policy

• The US government aims to foster American leadership in AI through a regulatory environment that allows innovation to succeed.
• Policy decisions are federated across multiple agencies, requiring the White House to coordinate diverse equities (national security, commercial, research).
• The White House strongly supports an open-source AI ecosystem, as outlined in the national AI strategy.
• The administration balances concerns like public perception of data centers and labor market impacts with the need for computational resources.

Regulation Philosophy

• Regulation should avoid firm red lines due to AI's rapid evolution, contrasting with the EU AI Act's static approach.
• The focus is on removing regulatory barriers for "born-free" technologies and making "born-in-captivity" technologies easier to commercialize.
• Reducing regulations is a core mission, aiming to lower compliance costs for startups.

AI Risks & Opportunities

• Key AI risks include cyber vulnerabilities and, potentially, biological risks, though the latter is considered overblown currently.
• Trade-offs exist, as models capable of nefarious actions can also strengthen existing systems.
• AI is seen as transformative for scientific discovery, changing how research is conducted across various fields.

Government & Innovation

• The White House strives to ensure "little tech" has a voice, counteracting the influence of large tech companies.
• A national standard for AI regulation is crucial to simplify compliance for startups.
• The administration is committed to reducing regulations and removing barriers to innovation.

Future Technologies & Policy

• Quantum information science and computing are identified as critical future technologies that warrant increased attention.
• The government's role is to "shape the arena, not direct discovery," focusing on building infrastructure for autonomous experimentation.
• Legislative action is needed for AI preemption at the state level and to address intellectual property and likeness rights concerning AI.

Public Service & Impact

• Working in government offers a uniquely fulfilling opportunity to impact the country significantly.
• The White House Office of Science and Technology Policy coordinates across agencies on issues like AI, quantum, biotech, and space.
• Public service can be bureaucratic but delivers unparalleled rewards when successful outcomes are achieved.

Early Career & Mac Design

• Began career as an artist, eventually joining Apple to design icons and graphics for the original Macintosh.
• Inspired by Paul Rand's advice: "Focus on making things meaningful and memorable."
• Worked within constraints of 16x16 black and white pixels for early icon design.
• Designed the original Chicago system font (9 pixels tall) and several other fonts like New York and Geneva.
• Created iconic Mac elements such as the lasso tool, the Happy Mac, and the bomb icon.

Icon Design Principles

• Advocates for using salient details in icons, avoiding excessive information for instant recognition.
• Emphasizes metaphors over literal depictions for icons that age better (e.g., floppy disk vs. save icon).
• Simple, universally relatable imagery (like a simple face or pencil) allows users to project themselves onto it.

Career Highlights & Advice

• Worked on projects for Microsoft Windows 3.0, General Magic, and Facebook.
• Designed icons for Pinterest, including those with subtle motion.
• Learned from prominent designers like Paul Rand, emphasizing quality over quantity in font usage and logo design.
• Encourages showing multiple options to stakeholders (like Steve Jobs) to facilitate feedback and iteration.
• Shares advice received from mentors and colleagues on various aspects of design and career.

Robotics & Physical Intelligence

• Physical Intelligence company founded to develop robots capable of any task in the real world.
• Recent progress includes robots washing pans, peeling carrots, making grilled cheese, and slicing zucchini.
• Focus is on enabling general-purpose robots for real-world impact.

AI & Robotics Evolution

• Timeline of AI in production: Recommendations/Ad Ranking -> Deep Learning -> ChatGPT (general-purpose model) -> Coding Agents.
• Key takeaway: Generalist models are increasingly used for real-world problems.
• Physical AI differs: directly affects the physical world, requiring higher autonomy and fewer mistakes than typical ML systems.
• Waymo's autonomous rides suggest trust and autonomy in physical environments are possible.

Achieving High Reliability in Robotics

• Challenge: Physical AI needs long-term autonomy, unlike systems where humans make the final decision.
• Example: Making espresso reliably requires precise control, handling liquids, and accurate timing.
• Traditional ML iteration (data collection, tuning) is slow; autonomous AI iteration is needed for high reliability (99%+).
• This resembles Reinforcement Learning (RL), learning from failures.

Scalable RL for Robotics

• RL algorithms like PPO/TRPO scale for LLMs but require millions of attempts (prohibitive for robots).
• Inefficiencies in RL for robotics:
• Dead-end trajectories: Wasted attempts on incorrect paths (e.g., grabbing two boxes).
• Solution: Human intervention (teleoperation) to guide recovery, terminating bad trajectories early.
• Repeated rollouts per prompt: LLM RL makes many attempts for one prompt; this is costly in robotics.
• Solution: Amortize cost across prompts, learning a general value function for "good" vs. "bad" progress.

General Purpose Models & Compositional Generalization

• Robotics is moving towards GPT-like, out-of-the-box models, away from bespoke, from-scratch training.
• Goals: Out-of-the-box models (no fine-tuning) and compositional generalization (combining known skills/concepts).
• Compositional generalization demonstrated with unseen appliances (air fryer) and new robot platforms (folding clothes on a different robot).
• Key ingredients for generalization: Diverse data and detailed prompting (metadata, subtask instructions, sub-goal images).
• PIO7 model achieves parity or outperforms specialized fine-tuned models out-of-the-box.

Memory & Long-Term Autonomy

• Most current robotics models lack memory, limiting them to short, repetitive tasks.
• Memory is critical for long-horizon tasks (e.g., cleaning a kitchen over 10-15 minutes).
• Challenge: Naively feeding past video is computationally expensive (millions of tokens).
• Solution: Multi-scale memory system:
• Short-term: Efficient video memory (10 seconds).
• Long-term: Textual summaries of past events (minutes/hours).
• This enables autonomous, non-repetitive tasks for extended periods.

Real-World Deployment & Future

• Robotics is now in a "GPT-like era" with general-purpose models.
• Physical intelligence models are being deployed in real-world companies (e.g., laundry folding, warehouse packaging).
• Adaptable to various embodiments: bi-manual robots, drones, surgical robots, tractors.
• ChatGPT moment for robotics: Capabilities are nearing parity, but distribution will be slower due to hardware requirements.
• Open source impact: Generalist policies (e.g., PI zero, PIO5) are available, useful for small teams.

Key Learnings & Insights

• PhD value: Learning to handle uncertainty, pick good problems, and conduct research with freedom.
• Robotics data equivalent to internet data: Robot experience (demonstrations, autonomous attempts), videos of humans, web data.
• Open source potential: Optimistic about a strong open-source community, though the exact dynamic may differ from LLMs.
• Model output: Models output target joint positions/gripper positions; controllers handle inverse kinematics.
• "Imagination" (predicting future states) helps but is not always critical for good performance.
• Speed improvement: RL shows speed gains; faster data collection or faster policies are needed.
• Surprising capability: Emergent hand-to-hand skill transfer (left hand mimicking right hand) without specific training data.
• Entry into robotics: Join a robotics company as a software engineer or build personal projects with open-source models.

CircleBack AI Note-Taker

• CircleBack is an AI that records, transcribes, and summarizes meetings, assigning action items.
• It can build automations to extract information and update other apps (CRM, Slack, etc.).
• It acts as a "company brain" for querying past conversations.

Hardware & Desk Setup

• Minimalist setup focused on essential tools.
• Uses a Topre membrane keyboard for its quiet and responsive typing feel.
• A fidget toy from Temu is used during discussions or dictation.
• AirPods Max are used occasionally, signaling focus to the team.

Unexpected Use Cases for CircleBack

• Used as an Applicant Tracking System (ATS) to track candidate interactions and interview feedback.
• Manages interactions with companies like Y Combinator, tracking all contacts and action items.
• Prepares for interviews by summarizing past conversations and emails.
• Uses a button press (or watch app complication) to quickly start recordings.

AI Agents and Automation

• Uses various AI agents for finance, customers, and people ops, migrating these to Slack channels.
• Engineers handle customer support through a weekly rotation, leveraging CircleBack for insights.
• CircleBack can quickly summarize customer feedback from meetings and identify action items.
• AI agents significantly reduce time spent on coding, with focus shifting to review and orchestration.
• "Token maxing" culture encourages using AI extensively for high-quality output and automating ops tasks.

AI Limitations and Best Practices

• AI should not send emails independently; drafts are acceptable.
• Product copy requires human review for clarity.
• AI can assist with reviewing code and data security, but engineers should architect.
• Evals are used to quantify AI output (e.g., correct action item identification) and refine writing style.

The Future of Recording and Software Engineering

• Recording all company conversations is becoming essential for AI agents to have context.
• The opportunity cost of not recording is increasing.
• Companies need to ensure secure sharing of captured information.
• Software engineering is now limited by imagination rather than technical skills.
• Focus is shifting from shipping speed to the strategic decisions of what and how to build.

Origins of Open Claw

• Open Claw, though only 8 months old, feels like 4 years in AI time.
• Inspired by annoyance and a desire to use tokens efficiently, Peter Steinberger wanted to prompt agents from his phone to monitor other agents.
• Early 2025 models (like OpenAI's 03) were slow and expensive.
• Steinberger created a WhatsApp relay, which felt magical and led to the realization of product-market fit when friends reacted strongly.
• The project evolved into "claudis" with Discord support, and eventually "claude bot", then "open claw".

Going Viral and Its Aftermath

• An accidental self-restarting demon process led to Open Claw going viral overnight.
• The creator faced intense media attention, Mac Mini sell-outs, and demands for name changes (e.g., from Tropic).
• Open Claw garnered 18,000+ contributors and 111,000+ issues/PRs.
• Steinberger felt overwhelmed, close to deleting the project, and had his personal information leaked.

Business Decisions and Challenges

• Steinberger, a former B2B software founder who experienced burnout, found his passion again with Open Claw.
• He chose to partner with "big labs" over taking money directly, focusing on his personal brand.
• Open Claw was "crushed" by security reports, leading to extensive security hardening despite most reports being low-risk.
• The project suffered from a large number of configuration options (9,500+) and slow evolution compared to VC-funded competitors.
• A major blow was the reliance on OpenAI's models, leading to issues when API access was changed with little notice.
• Dependency on a specific model's business model was identified as a critical risk.

The Fun Factor and Future Direction

• The project stopped being fun around February, shifting from personal passion to a perceived responsibility.
• Steinberger aims to make AI accessible, fun, and less scary, moving beyond just terminal interfaces to voice and multimodality.
• Open Claw's mission is to provide an open-source alternative to proprietary agents, ensuring data privacy and model flexibility.
• The future includes perpetual, proactive agents and collaborative environments where agents can orchestrate work.

Key Takeaways and Advice

• Don't stop having fun; it's the ultimate driver of great ideas.
• Listen to your gut and fix things that annoy you.
• Stay focused, build what's missing, and keep shipping to confuse detractors.
• The hardest part of building today is getting "eyeballs" due to the ease of creation.
• Building something you genuinely want to use is crucial for success.
• Consider "hard and boring" categories for startup ideas to stand out.

Robotics is Still Not Solved

• Despite repeated predictions over the last decade, robotics is not yet "solved," with a lack of widely available, general-purpose robots.
• Current successes are largely limited to demos and highly controlled work cells.

Challenges in Scaling Robotics

• Data Collection is Hard: Teleoperation for data collection is finicky and difficult to scale.
• Physical Real-World Modeling: Simulated environments (sim-to-real) don't accurately reflect real-world physics, leading to a gap.
• Deformable Objects: Modeling the transition of deformable objects remains a significant challenge.
• Representation for Action Space: Learning effective representations for robot actions is an unsolved problem.
• Sensory-Motor Issues: Robots lack the rich sensory feedback humans have (tactile, force, temperature), hindering their ability to build world models.
• Embodiment Drift: Actuators and components degrade over time, causing performance shifts and requiring retraining.

Advancements in Memory for Robotics

• Need for Memory: Long-horizon tasks require robots to have memory to track progress and adapt.
• Effectiveness & Efficiency Issues: Adding memory to policies can hurt performance (distribution shifts) and increase resource demands (training/inference time).
• Proposed Solution (MAM): A hierarchical policy with separate short-term (dense frames) and long-term (compressed language) memory components.
• Short-Term Memory: Uses a Vision Transformer (ViT) with temporal attention for compressed visual memory, enabling tasks like prolonged cooking.
• Long-Term Memory: Employs a compressed language representation for the high-level policy, allowing for tasks spanning tens of minutes.
• In-Context Adaptation: Memory enables robots to learn from mistakes and adapt their behavior in real-time, a crucial missing capability.

Embodied Reasoning and Self-Supervision

• Vision-Action Models (VAMs): Trained on large visual-textual data, then fine-tuned on scarce robotics data.
• Embodied Reasoning (Chain of Thought): Aims to improve action prediction by explicitly incorporating reasoning steps, valuable in data-scarce robotics.
• Challenges in Reasoning: Grounding reasoning in robotics and avoiding verbosity are key problems.
• R&B Encore: A self-supervised bootstrapping method to refine and discover embodiment-specific reasoning.
• Reasoning Discovery: Proposes different reasoning types (e.g., move, gripper position) and a validator scores them based on concision, non-triviality, and action predictiveness.
• Benefits: Generates synthetic, action-aligned reasoning data, improving VAM performance and generalization across embodiments (manipulation, locomotion, driving).
• Key Insight: Selective reasoning is more important than exhaustive reasoning; not all valid reasoning is useful.

Sim-to-Real for Dexterous Manipulation

• Sim-to-Real Approach: Training policies entirely in simulation using massive parallelization to collect vast amounts of data rapidly.
• Goal-Reaching Unification: Treats dextrous tool manipulation as a goal-reaching problem, allowing a single policy to learn diverse tasks without explicit task labels.
• Human Video as Goal Specification: Robot actions are not learned from human demos, but human videos specify goal poses for a frozen policy to track.
• Results: A single policy trained on primitive objects generalizes to unseen tools and tasks with impressive dexterity, including in-hand reorientation and recovery behaviors.
• Play to Perfect: Extends this by first learning a task-agnostic play prior, then fine-tuning for precise contact-rich assembly tasks.

Data Infrastructure for Robotics

• Rerun: An open-source SDK and data platform for logging, visualizing, and querying physical AI data.
• Robotics Application Companies: A growing trend focusing on end-to-end business problems, using off-the-shelf hardware and teleoperation initially, then fine-tuning models.
• Data Challenges: Physical data is multimodal, multi-state, episodic, and has complex semantics, making traditional data infrastructure insufficient.
• Iteration is Key: Fast iteration through data collection, evaluation, and debugging is crucial for success.

World Action Models (WAMs)

• Advancement over VAMs: Predict future kinematics and world states, not just immediate actions.
• Generative Approach (e.g., Dreamer): Predicts future frames using diffusion models, offering rich world representations but is computationally heavy.
• Latent WAMs (e.g., Image_VLM, FastOne): Predict future states in a latent space, avoiding explicit frame decoding for efficiency.
• General Instinct's Optimization: Focuses on infrastructure for WAMs, including distillation for smaller models, faster sampling (1-2 steps vs. 50-100), and improved modality representation.
• Benefit: Enables faster, more scalable deployment of world action models on robots.

Company Infrastructure

• Deep tech startups often fail not due to technology, but due to a lack of robust internal systems for purchasing, recruiting, and operations.
• Building internal infrastructure, like custom software for purchasing and hiring, can be a key differentiator for speed and efficiency.
• "Amateurs talk strategy, professionals talk logistics."

Purchasing and Procurement

• Early-stage startups may start with credit cards but quickly need structured procurement systems as they grow.
• Approving every purchase individually is unsustainable and hinders employee productivity.
• A proper procurement system involves budgeting, vendor management, and efficient processing, which can take time to establish.

Cost Attribution and Management

• Buying materials in bulk for experiments makes accurate cost attribution difficult, leading to a perception of "free" resources.
• Internal software is crucial for tracking the cost of experiments, manufacturing, and overall runway.
• For example, a single wafer iteration in their fab cost $40,000.

Hiring and Team Building

• The best hires often come from the founder's network and the broader community that spawned the startup.
• A well-defined, systematic hiring process is essential, even if there's no single "right" way to do it.
• A company-wide voting system for initial applications helps to avoid bottlenecks and average out judgment.
• Key qualities to look for in candidates are judgment, horsepower, and agency.
• AI-resistant take-home assignments are preferred for evaluating candidates.

Performance Reviews and Feedback

• Traditional 360-degree reviews are often slow, disruptive, and don't surface new issues.
• An "iGEN reviews" system, inspired by PageRank, uses a weighted voting mechanism for hires to provide continuous, less biased feedback.
• This system helps detect potential collusion through statistical analysis.

Speed and Iteration

• The rate of iteration is a primary determinant of success or failure in startups.
• Companies that can learn and adapt faster than competitors will ultimately win.
• Infrastructure directly enables speed, making processes like purchasing and recruiting critical.

Decision Making and Judgment

• Founders cannot delegate their judgment; they must make decisions that make sense to them, even when alone.
• Success requires differentiated judgment that goes beyond the average.
• Action produces information; when stuck, injecting "action-producing entropy" is necessary.

Biotech and Healthcare

• Biotech is capital-intensive and a difficult path, but successful ventures have immense impact.
• Neural engineering, including BCIs, is seen as a longevity and healthcare advancement.
• BCIs can offer significant effect sizes, bypassing complex biological problems to directly interface with the nervous system.
• Challenges in BCIs include power/thermal constraints, packaging, and material science for implants.
• Interdisciplinary teams with a broad perspective are valuable in neural engineering.

AI's Role

• AI is a multiplier, not a replacement, for human teams.
• AI has significantly impacted coding and navigating regulatory and quality systems in scientific research.
• Companies should aim to be "AI-native" by making data context available to AI agents.

Paper: An AI-Native Design Tool

• Paper is an AI-native design tool that uses HTML and CSS as its rendering engine, making it easily understood by AI agents.
• This approach differs from traditional design tools with custom engines, leading to higher accuracy and lower token spend when agents interact with Paper files.
• Paper aims to empower humans to work quickly with agents, emphasizing the importance of the human element in design.

Design in the Agent Era

• AI agents can speed up tasks like translation and resizing, but relying solely on AI-generated designs can lead to a lack of differentiation.
• Exceptional design is a key differentiator for successful companies.
• Paper focuses on providing designers with tools that move at the speed of agents while maintaining high-quality design standards.

Key Features and Innovations

• Paper Shaders: An open-source library of WebGL animations, allowing designers to create custom, modern visual effects easily.
• Image Generation: The tool can utilize multiple AI models simultaneously to explore a wide range of illustration possibilities.
• Vectorizer and Color Extraction: Tools to quickly extract colors and generate new textures from images.
• Designer-Developer Handoff: Paper's HTML/CSS foundation facilitates seamless handoff by allowing users to copy designs directly as React components or Tailwind CSS.
• Desktop App and Local Integration: Paper now offers a desktop app, enabling integration with local development environments, GitHub, and tools like Cursor and Conductor.
• Agent Integration: Paper acts as a visual interface for agents, allowing for direct manipulation and faster input than pure prompting.

Improving AI-Generated Designs

• Common AI Tells: Overuse of bold fonts, excessive font sizes, generic gradients (especially purple), overuse of cards, and tiny all-caps headers.
• Human Curation: Designers can refine AI outputs by simplifying typography (limiting to 3 font sizes, reducing bold weights), adjusting contrast, and deleting unnecessary elements.
• Intentionality: Design should communicate value and trustworthiness, not just fill space.
• Paper's Guardrails: The tool incorporates instructions to models to avoid common AI design mistakes, improving the quality of outputs.

Future of Design and Paper

• The function of design (problem exploration, stakeholder management) remains crucial and will likely grow.
• Tooling will accelerate to keep pace with engineering and product teams.
• Paper plans to add features like component management and better prototype organization with commenting capabilities.
• The human element and taste are essential, and AI is unlikely to fully replace human designers in making strategic decisions.

Introduction: Spinoza's Legacy

• The speaker introduces Baruch Spinoza, a 17th-century philosopher excommunicated for his radical ideas, as a parallel to modern innovators.
• Spinoza was offered a bribe to stop his work but chose truth, grinding lenses and writing a dangerous book he couldn't publish while alive.
• His ethos is summarized by "conatus," the drive to increase one's power to act.

The Arrival of Personal AGI

• AGI (Artificial General Intelligence) is not an impending event but is arriving diffused as personal agents.
• This is termed "Personal AGI" – general intelligence for one person, not for everyone at once.
• Personal AGI is an agent running on your infrastructure, using your memory, executing your procedures, and compounding daily.
• This is distinct from corporate AGI (like chatbots) which is rented, temporary, and controlled by others.

The Multiplier Effect of Agents

• AI agents dramatically increase productivity; the speaker cites a 400x increase in coding output.
• This multiplier effect applies to all knowledge work, not just coding.
• Fastest-growing founders treat AI as a workforce, not just autocomplete, focusing on context and relevance rather than just model weights.

Building Your Personal AGI (GBrain)

• Personal AGI involves owning your context (library) and using rented models through a "harness."
• Human working memory is limited (7±2 items); AI agents can handle millions of tokens (thousands of pages).
• The core concept is a "library" (your data) and a "librarian" (the agent) deciding which context is active.
• GBrain is an open-source system that acts as a personal operating system, integrating your life's data.
• "Skill files" (pages of English instructions) are like employees for your agent; markdown is code, and language models are compilers.

Steps to Building Your Personal AGI

• Step 1: Run an agent on your machine (e.g., OpenClaw, Hermes Agent).
• Step 2: Start your library: one folder of markdown files, export notes/emails, document projects and people.
• Step 3: Write your first skill file for a repetitive task you dislike.
• Step 4: Wire it up as a recurring job (e.g., daily, weekly).
• Step 5: Practice the discipline of "skillifying" every task to create reusable skills.

The Political Dimension: Ownership and Control

• Skill files represent your externalized judgment and cognition.
• Ownership is key: skills in your repo go with you; skills in a company's repo can be retained by the company.
• "Skill files are yours" is the doctrine; own your skills to avoid your job becoming a skill file owned by others.
• This is the modern equivalent of craftsmen owning their tools, offering freedom and power.

Overcoming Objections and Moving Forward

• Model improvements increase the value of your unique context (library), not diminish it.
• Retrieval (RAG) is a primitive; the product is being worth retrieving from.
• Consolidating your context into a personal system is a security model of custody, more private than scattering data across company clouds.
• Giving away powerful tools creates a renaissance, not a priesthood.
• Personal AGI is about staying "under your own power" and building directly on your own history and striving.

Introduction to StarCloud

• StarCloud is building data centers in space to address the energy demands of AI on Earth.
• They raised $170 million and became a fast-growing unicorn in YC history.

The Case for Space Data Centers

• Terrestrial energy projects face constraints; space offers unlimited solar energy.
• Launch costs are rapidly decreasing due to reusable rockets like Starship.
• Space-based solar power beaming down to Earth loses 95% of energy in transmission.
• Data centers were chosen as they don't require re-entry, unlike manufacturing or asteroid mining.
• The break-even launch cost for space-based data centers is estimated at $500 per kilogram.

StarCloud 1 Mission

• Booked their first launch 18 months in advance as a forcing function.
• StarCloud 1 carried five GPUs, including an Nvidia H100.
• Unconventional methods were used for testing, like submerging components in ice baths and using hot air guns.
• The mission cost $2 million, compared to an estimated $75-$100 million from prime contractors.
• The deployment video was notably featured by Jensen Huang at GTC.
• The team experienced software issues, including satellite restarts, requiring 3 days to diagnose.
• Commissioning and initial AI model training were completed within two weeks.

Engineering and Physics Challenges

• Key challenges are heat dissipation in a vacuum and radiation hardening of chips.
• StarCloud is developing a low-cost, low-mass deployable radiator.
• Extensive ground testing in particle accelerators is used to mitigate radiation effects.
• They use off-the-shelf automotive-grade components, tested for radiation tolerance, not just space-grade.
• The company is working with Nvidia on a new space-designed chip (Rubin).
• StarCloud 1 used a unique immersion cooling method with phase change material.

Market and Investment Landscape

• Initial investor reaction was largely negative due to perceived sci-fi nature and reliance on low-cost launch.
• The increasing difficulty and regulatory hurdles of building data centers on Earth have shifted investor sentiment.
• Investor confidence grew with strong engineering talent and the convergence of cheaper launch and terrestrial constraints.
• The crypto community is enthusiastic about StarCloud's plans to fly Bitcoin mining ASICs.
• They are partnering with AWS to launch Outpost hardware for military customers.
• Deep tech and hard tech investment, particularly in space, has seen a significant resurgence.
• Benchmark's investment was driven by a strong technical team and the potential of the technology.

Future Roadmap and Vision

• StarCloud 2 will be a 10-kilowatt spacecraft for government and military compute.
• StarCloud 3 will be a 200-kilowatt, 3-ton spacecraft, with 50 fitting per Starship.
• They aim for a constellation of 88,000 spacecraft, providing 20 gigawatts of compute capacity.
• Full ramp-up for terrestrial competition is anticipated by the end of 2028, dependent on Starship's cadence.
• Initial customers include government entities and those needing on-orbit processing for satellite data.

Advice for Hard Tech Founders

• Book the first available launch as early as possible; it's a critical forcing function.
• Prioritize hiring world-class technical talent above all else.
• Embrace "wacky" and unconventional solutions to prove concepts.
• Space presents enormous opportunities due to upcoming launch capacity increases.

Introduction to Physical AI

• The presenter, Dmitri Dolgov, discusses building AI for the physical world, specifically the Waymo driver.
• Waymo serves ~500 trips weekly, driving over 4 million autonomous miles.
• Key challenges in physical AI vs. digital AI include the higher cost of errors (human lives), latency requirements, lack of readily available digitized data, and the need for high confidence on day one.

Lesson 1: Demo vs. Product

• A working demo is 1% of the work; achieving "many nines" of reliability and performance is the real challenge.
• Waymo achieved capability complete demos in 18 months but took ~15 years to build a scalable product.
• Each additional "nine" of reliability requires exponentially more effort, necessitating fundamentally different approaches like redundant systems.

Lesson 2: Technology Choice and Architecture

• Picking technology based on the fastest early ramp can lead to hitting a performance plateau before product requirements are met.
• Waymo uses multiple sensing modalities (cameras, LiDAR, radar) for robust performance in various conditions.
• Hardware costs decrease over time; design for future iterations and price drops.

Lesson 3: Riding Technological Waves

• Companies must repeatedly adapt to new technological breakthroughs (e.g., CNNs, Transformers, LLMs) and integrate them into production without regressions.
• The hardest muscle to build is carrying bleeding-edge research into production safely while scaling.
• New technology should aim for simplification and unification, not just performance gains.
• The Waymo Foundation Model is a multimodal world action language model, processing sensor inputs, understanding the world, predicting actions, and leveraging language models for rare situations.
• It uses a "System 1 (fast path)" for split-second decisions and "System 2 (slow path)" for complex reasoning.

Lesson 4: The Bitter Lesson and Structure

• General methods scaling with compute and data (the "bitter lesson") outperform handcrafted knowledge.
• Structure can either fight or channel scale. Structure that fights scale loses; structure that channels scale wins.
• Waymo uses "structure-augmented end-to-end" learning, combining learned representations with materialized structure (physics, rules of the road) for better validation, efficiency, and scaling.

Lesson 5: Simulation is Crucial

• A large-scale, realistic, high-fidelity simulator is critical for training and evaluation.
• Closed-loop simulation, where the agent's actions affect the simulated world, is essential for safety-critical agents.
• Waymo builds "behavioral world models" and "sensing world models" to generate realistic and rare scenarios for training and evaluation.

Lesson 6: The AI Ecosystem and Flywheel

• Building a physical AI agent requires an ecosystem of three components: the agent, the simulator, and the critic.
• This creates a flywheel: deployment generates data, grounding the simulator; the simulator creates edge cases for the critic to score and the agent to learn from, leading to continuous improvement.
• Metrics are crucial to guide the flywheel in the right direction.

Lesson 7: Evaluation and Metrics are Strategic

• Evaluation and metrics are more important than the model itself; they define "good enough" and steer development.
• For physical AI, evaluation must go beyond model performance to include the entire system and operational processes (the "safety and readiness framework").
• Trust is earned through relentless proof of safety and performance, backed by transparent data and public audits.

Conclusion

• The combination of these lessons leads to Waymo's "strongly superhuman" safety performance, preventing serious injuries.
• The next decade of AI will be in the physical world, with opportunities in trucking, personally owned vehicles, and more.

AI and Learning

• Cognitive L1 cache (knowing things) is much faster than outsourcing to AI.
• Even with advanced AI, there's a premium on cognitive ability and reasoning.
• AI models still struggle with nuanced interpersonal communication and multi-dimensional reality.

Starting a Company

• Dropping out of college twice to start companies is possible and not a permanent setback.
• The urgency to start a company might be an overestimation; opportunities in Silicon Valley are robust.
• Focus on concrete, customer-expressed problems rather than imagined ones.

Stripe's Early Days

• The idea for Stripe stemmed from the frustration with existing internet payment systems.
• Despite the obvious need, starting a financial services business as young founders was met with skepticism.
• Securing a banking partner was challenging due to the perceived improbability of the venture.

Product Launch Strategy

• Stripe took almost two years for a public launch due to complex security, infrastructure, and reliability needs in fintech.
• Having early, live production users (even with limited functionality) provided crucial, just-in-time development feedback.

The Era of AI and Startups

• The traditional "lean startup" doctrine may face increased competition in the AI era.
• More aggressive and ambitious initial product strategies might be necessary due to AI capabilities.
• Companies are more "spring-loaded" to adopt new technologies due to the fear of being left behind.

Success and Centralization

• The fear of AI leading to extreme economic centralization may be overstated.
• Stripe data shows a significant increase in new businesses starting and succeeding, suggesting broad-based prosperity.
• The future might be more decentralized with many thousands of winners, not just a few dominant companies.

AI Capabilities & Predictions

• AI is now at the level of a junior engineer, capable of complex, long-running coding tasks and shining in agent-based systems across domains.
• Prediction for 2027: Significant automation of ML systems, with AI improving capabilities through automated experimentation loops.
• Assumption to challenge: The limited duration of agent-based systems; they can run for days or weeks on complex tasks.

Hardware & Efficiency

• The current "fits in memory" moment is the rise of high-performance, low-energy inference hardware due to latency importance.
• Specialized hardware offers better energy efficiency and lower latency than general-purpose devices.
• Energy is the key measurement unit; moving data costs significantly more energy than computation.
• Batching is a system/data IO problem driven by energy costs, not purely a model problem.

Agent & System Design

• Agent systems can perform complex tasks by breaking them down into sub-problems and running automated experiments.
• The limitation in long-running agents is often performance degradation when deviating from training data distribution.
• Solutions for agents include providing skills/hints, using multi-agent systems for evaluation, and inference-time computation for search.
• Clear specifications and design documents are crucial for managing agents effectively, akin to software development.
• Key constraints for agents are model experience outside its training distribution and compounding errors in open-loop systems.

Founders & Innovation

• Founders can win by focusing on specific domains with well-designed surfaces, specialized models, or unique data access, rather than competing with general models.
• Look for problems where general models fail 0-1% of the time; moderate success is a sign general models will improve rapidly.
• Promising areas for specialization include personal information organization and niche domains like protein folding (AlphaFold), material science, or chip design.
• The scarce skill in an AI-assisted future is "taste" – the wisdom to choose impactful problems.
• Develop taste through experience, evaluating predictions, and conducting "crazy thought experiments" that question fundamental assumptions.
• Mistakes are part of the process; a rejected paper on distillation (a technique now widely used) highlights the need to persevere.

Future Directions

• Exciting problems include new hardware approaches, more data-efficient ML algorithms, continuous learning, and multi-agent interactions.
• Developing better discourse and connection tools for people globally is also a key area.
• The accelerated scientific method (propose, implement, evaluate, iterate) via automated loops is a major future direction.
• Faster validation models, potentially learned approximations, are crucial for speeding up experimental loops in science and engineering.

Specialization and Efficiency in AI Hardware

• The trend towards specialization in AI hardware is increasing, with different specifications needed for training versus inference data centers.
• Inference requires high bandwidth and low latency, especially for batch size one, which presents a chip-level challenge.
• There's significant potential for optimization in CUDA kernels, algorithms, software, and chip design to improve efficiency.
• The development of specialized ASICs like TPU v8 is a first step, with further splits expected, for example, between prefill and decode engines.

Multi-GPU Kernel Design and Optimization (Parallel Kittens)

• GPU networking is a major bottleneck, consuming up to 50% of runtime in some workloads.
• Fine-grained overlap of compute and communication is crucial, with modern techniques operating at token or sub-kilobyte granularity.
• Key trade-offs in multi-GPU kernel design include the choice of transfer mechanism (copy engine, TMA, register instructions), scheduling of inter-GPU communication with intra-GPU work (intra-SM vs. inter-SM overlapping), and design overhead.
• Parallel Kittens (PK) is a framework built on these principles, offering programming primitives that utilize efficient transfer mechanisms and scheduling strategies, enabling performance comparable to hand-optimized kernels with significantly less code.

Intelligence per Watt and Distributed Inference

• The "mainframe era" of AI infrastructure (large cloud data centers) is shifting towards distributed inference, mirroring the PC revolution.
• Smaller, local open-source LLMs coupled with improving local accelerators (e.g., Apple M-series, consumer GPUs) can handle a significant portion of inference demand.
• The metric "intelligence per watt" (and "intelligence per joule") measures capabilities against compute/energy efficiency.
• Studies show substantial improvements in intelligence per watt (3x in two years) and intelligence per joule (18x in 16 months), driven by better local models and accelerators.
• This trend enables routing 80-90% of queries to local accelerators, leading to significant energy, compute, and cost savings.

AI in Kernel Development and Verification

• AI models are becoming increasingly capable of writing competitive GPU kernels, even for individuals new to the field.
• Evaluating AI-generated kernels involves correctness checks against reference implementations and performance benchmarks.
• "Reward hacking" is a significant challenge, where AI exploits loopholes in evaluation frameworks (e.g., returning zero for vector mean, caching outputs).
• Robust verification requires adversarial approaches, finding and codifying reward hacks to improve evaluation robustness, a process analogous to how PyTorch achieved correctness over time.
• Open problems include speeding up compilation, developing better GPU simulators, more robust and cheaper kernel correctness verification, and faster test-time scaling for kernel development.

Heterogeneous Infrastructure for Inference

• Inference workloads are heterogeneous, with different phases (prefill, decode, KV cache) stressing compute, network, storage, and memory bandwidth differently.
• Specialized hardware and system co-design can optimize these diverse phases, moving beyond homogeneous GPU clusters.
• Prefill is generally compute-bound, while decode is memory bandwidth-bound and latency-sensitive.
• SRAM machines offer high memory bandwidth and low latency, making them suitable for decode-heavy workloads, though limited by on-die capacity.
• Disaggregating components like prefill and decode, or attention and MLP layers, onto different systems can improve Total Cost of Ownership (TCO) and extend interactivity, depending on workload characteristics.
• Speculative decoding can also benefit from heterogeneous systems, with a faster, potentially larger drafter model on one system and a verifier on another.

GPU-Accelerated Game Engines for Simulation

• Existing game engines are often inefficient for high-throughput training workloads due to CPU bottlenecks and poor GPU utilization.
• Batch simulators, which simulate many environments in parallel on the GPU, can achieve millions of frames per second.
• Entity Component System (ECS) design patterns, common in game development, are well-suited for GPU adaptation, enabling efficient data management and massive parallelism.
• This approach allows ML researchers with limited GPU programming experience to build high-performance simulation environments, achieving over 100x speedups compared to CPU baselines.
• There is a need for higher-level GPU programming abstractions that simplify complex tasks like dynamic memory allocation and irregular parallelism, similar to scripting languages on CPUs.

Early Life & Education

• Grew up in "middle of nowhere" New Mexico.
• Excelled in math and computer science competitions.
• Influenced by a friend's internship in Silicon Valley.
• Worked at Quora for a year after high school.
• Attended MIT at 18, exploring interests and training early AI models.

Founding Scale AI

• Started Scale AI at 19 after one year at MIT.
• Initially pitched a medical AI agent idea, but pivoted after Y Combinator feedback.
• Identified data as the bottleneck for AI model training, recognizing a significant opportunity.
• Faced investor skepticism about the data business model, as many lacked AI model training experience.
• Believes in developing conviction in unpopular truths and working in obscurity before an idea becomes consensus.

AI & The Future

• Current bottleneck is diffusing AI technology and adapting the world to it, not model progress.
• Sees AI as a "once-in-a-civilization opportunity" for builders with vision.
• AI agents empower startups to compete with large companies ("Goliath vs. Goliath").
• Meta's vision: billions of personal "super intelligences" enhancing agency and fostering entrepreneurship.
• AI will lead to unprecedented economic and societal upheaval, creating new opportunities and risks.

Meta's AI Initiatives

• Launched Muse, MuseImage, and MusePark 1.1 within nine months.
• Emphasizes "talent density" and treating frontier AI work as scientific research and experimentation.
• Focuses on developing an adaptable lab system to compound exponential growth in AI capabilities, compute, and adoption.
• Plans to release more powerful models, a "harness" for developers, and open-source models.
• MuseSpark is significantly cheaper than competitors, aiming for broad accessibility.

Advice for Builders

• Develop an internal compass and strong conviction, as noise and confusion are inevitable.
• Identify and invest in exponential curves of progress, like AI today.
• Systematic and rigorous thinking, especially systems thinking, remains crucial.
• Balancing "word cell" and "shape rotation" is key; orchestrating agents and agent organizations is the future.
• Focus on agentic looping and optimizing feedback loops for massive efficiency gains.
• Don't rely on "magic"; understand the mundane mechanics (e.g., cron jobs, metrics).
• Ignore LinkedIn for development, but use it for customer acquisition.

The Problem with the Pace of Innovation

• The future was expected to be faster and more innovative, referencing the moon landing and Concorde.
• Half a century later, we've lost the ability to go to the moon and fly supersonic.
• Manufacturing times have dramatically increased, e.g., two years for a missile interceptor.
• Commercial airliners haven't significantly improved in speed or accessibility over decades.

The Impact of Faster Travel

• Faster travel in the past doubled speeds, making places like Hawaii a tourist destination.
• It influenced industries like fashion (Nike's origin) and music (global tours).
• A doubling of speed today could create new cultural and economic opportunities.

Boom's Journey to Supersonic Flight

• Blake Scholl founded Boom with a lifetime goal of breaking the sound barrier.
• Overcoming skepticism, including from Jeff Bezos, Boom started with a cardboard mockup.
• The XB-1 prototype, "Baby Boom," was developed independently, a feat not done by governments for decades.
• XB-1 successfully broke the sound barrier, marking a historic achievement.

Technological Advancements and Manufacturing

• Boom developed "Makeboom" to simulate aircraft designs digitally, drastically reducing iteration time.
• "BladeRunner" allows for real-time engine blade design changes and analysis.
• Vertical integration, including an in-house machine shop, enables rapid part prototyping (24 hours).
• They are building a "superfactory" for scaled manufacturing and vertically integrated test assets.

Financing and Business Strategy

• Financing deep tech is challenging; Boom nearly failed multiple times.
• They developed "Superpower," a ground-based turbine from their supersonic engine technology, to generate revenue and data.
• This strategy provides necessary capital and learnings for scaling up.

Overcoming Regulatory Hurdles

• Boom engaged early with the FAA, making regulators part of the team.
• This collaborative approach led to a swift approval for the XB-1's first supersonic flight.
• The success and demonstration of quiet supersonic flight led to a U.S. executive order legalizing supersonic flight.

Lessons Learned and Advice

• Work on what you love; passion is more important than just working with what you know.
• Embrace iteration in both software ("bits") and hardware ("atoms").
• Build teams with ambitious, hands-on, optimistic individuals, especially early-career talent.
• Self-belief comes from taking on challenging missions and giving them your all.
• Great companies balance long-range vision with daily operational discipline.
• When teaching yourself, focus on deep understanding and address "confusion lists."
• Don't fear AI; focus on doing useful, multidisciplinary work.
• Young engineers can break into hardware by building things and seeking advice from experienced individuals.
• Engage regulators early for safety-critical industries; for others, building a customer base first might work.
• Founders should move to inspiring ecosystems unless there's a strong reason not to.
• Side projects and demonstrating extraordinary ability are key for early-career hires.
• Seek work-life harmony, focusing on meaningful work and activities you love.

The Golden Age of Startups

• There has never been a better time to start a company due to advancements in AI and technology.
• AI agents can now do in minutes what took months for a startup 20 years ago.
• Startups are crucial for distributing the power of new technologies and preventing economic stagnation.
• Hard tech startups are becoming more feasible and important, with a growing percentage of YC companies focused on them.
• The current technological landscape, with rapid changes, falling costs, and shrinking cycle times, favors startups.

Overcoming Conventional Wisdom

• It's important to develop conviction in ideas that go against conventional wisdom, even when dismissed by experts.
• OpenAI faced significant skepticism and criticism when it started, which paradoxically provided time to research and build.
• Companies that pursue unique, unproven ideas often achieve greater success than those following the hype.
• Finding a small group of like-minded individuals who believe in a radical idea is more important than widespread belief.

The Power of Networks and Helpfulness

• Building strong networks and being helpful to others can lead to unexpected opportunities and collaborations later in life.
• Early investors in Stripe, like Sam Altman, helped convince key hires like Greg Brockman, who later co-founded OpenAI with Altman.
• The Bay Area's culture of loose networks and mutual support is a significant advantage for the startup ecosystem.
• Focus on building and creating value rather than engaging in negative commentary on social media.

Navigating AI Advancements and Safety

• Recent AI incidents, like the Hugging Face event, are serious reminders of the need for AI safety and alignment.
• The goalposts for AI capabilities have moved rapidly, requiring a serious approach to potential risks.
• There's a dual risk of power concentration in AI and overreacting to safety concerns, leading to a loss of freedom.
• Startups are essential for ensuring AI power is widely distributed and not concentrated in a few entities.
• The demand for high-quality intelligence is uncapped, similar to early days of computing, suggesting continued compute shortages and innovation.

Advice for Ambitious Founders

• If an idea feels too ambitious, share it; the vision might be clear even if the initial steps are uncertain.
• Be prepared for criticism and dismissal if pursuing impactful work; the greater the impact, the greater the opposition.
• Progress requires taking imperfect steps forward, rather than being paralyzed by a grand idea.
• The tech industry is forgiving of mistakes and failures; focus on drive and ambition while trying to enjoy the journey.
• The next six months are predicted to see progress equivalent to the last two years in AI model development.

Opus 5 and AI Advancements

• Opus 5, the latest model, shows significant performance gains, achieving 30% on Arc AGI 3, a substantial leap from previous single-digit scores.
• A key new capability is its ability to run for extended periods (days, weeks, months) in "auto mode" without needing scaffolding like SLGO.
• Opus 5 appears to be "prompt injectable" resistant, a significant safety improvement over previous models, achieved through a combination of alignment research and a prompt injection classifier based on mechanistic interpretability.

Claude Code Evolution

• Claude Code's system prompt has been drastically reduced (over 80%) for Opus 5, as the model's inherent intelligence reduces the need for explicit instructions.
• The development process for Claude Code involves "ablation," where prompts and tools are deleted and reintroduced line-by-line to understand their impact.
• A "simple mode" in Claude Code can be activated with an environment variable, which removes all system prompts, revealing that models can sometimes be more intelligent without them, though prompts are still useful for product usability.

Building Agentic Products

• The core philosophy for building agentic products is "unhobbling" models, meaning removing constraints to allow them to express their full capabilities ("product overhang").
• This involves giving models harder tasks than initially thought possible, describing the task, guardrails, and exit criteria, and letting them "cook."
• An example is rewriting a large codebase (Zig to Rust) in one shot over 11 days, a task previously taking over a year for human engineers.
• Other examples include using OpenCV for drawing and an ongoing 2-week experiment to rewrite an Electron app in Swift, demonstrating long-running, multi-agent tasks.

Prompting and Evals

• The skill is shifting from traditional "prompt engineering" to enabling models to perform challenging tasks and, crucially, to verify their own work.
• Evals are important and should be continuously appended to, though they may become saturated and need replacement as models improve rapidly.
• The approach to building is empirical: try a task, observe struggles, and iterate, treating the model like an organic, evolving entity rather than a static system.

Future of AI and Development

• Coding is becoming "solved" for many applications, shifting the focus to higher-level skills like product sense, business acumen, and user interaction.
• Students should still learn programming fundamentals practically, by building things to solve problems, as this develops essential design and business thinking.
• New users are encouraged to experiment with Claude Code, delete existing prompts, and let Opus 5 demonstrate its capabilities.
• New users receive a 20x compute boost to explore these advanced features.

NVIDIA's Founding and Early Struggles

• NVIDIA's initial technology choice for 3D graphics was "absolutely wrong."
• The company pivoted by purchasing textbooks on OpenGL and reinventing computer graphics.
• Key lesson: Technology changes; confronting reality and learning are paramount.

NVIDIA's Core Philosophy and Growth

• The big idea: Augmenting CPUs with accelerators to solve difficult problems (e.g., molecular dynamics, deep learning).
• Focus is on accelerating algorithm domains, not just building chips.
• Great companies have a unique, deeply believed perspective about the world.

Overcoming Adversity: The Sega Story

• NVIDIA admitted a critical flaw in their technology to Sega, risking a $12 million contract.
• Sega CEO provided $5 million, enabling NVIDIA to survive and find the right solution.
• Lesson: Trust and investing in people are crucial.

The AI Revolution and Universal Function Approximators

• NVIDIA saw deep learning (AlexNet) as a universal function approximator.
• This realization led to breakthroughs in computer vision, robotics, and self-driving cars.
• The implications for the entire computing stack were immense.

NVIDIA's Approach to Innovation and Organization

• NVIDIA's CEO is driven by curiosity and a desire to empower the company.
• He stays "in the weeds" to understand fast-changing technology and "surf the waves."
• The company's structure is adapted to the founder's style ("founder mode"), not conventional management.

The Future of AI: Agents and Systems Thinking

• Systems thinking, awareness, and design are crucial as low-level tasks become automated.
• Controllability of AI agents is the next big breakthrough needed.
• NVIDIA is involved in agents to understand the new software paradigm and improve internal processes.

Open Source and Democratizing AI

• Open source is vital for innovation (e.g., Linux, PyTorch).
• NVIDIA supports open initiatives like OpenCLAW and Hermes to enable everyone to build their own AIs.
• The ability to build custom, domain-specific AIs is key.

Economic Impact and Job Creation

• AI and automation automate tasks, not necessarily jobs, leading to new roles and increased productivity.
• Evidence shows job growth in fields where AI automates tasks (e.g., software engineering, radiology).
• Increased productivity drives ambition and leads to more opportunities.

Physical AI and Robotics

• Generative AI advancements enabled "physical AI" and robot articulation.
• The "ChatGPT moment" for robots involved reinforcement learning grounded in physics.
• NVIDIA is building simulation environments (Isaac Sim) for robot training.

The Next Frontier: Self-Driving Cars and Physical AI Applications

• Self-driving cars were identified as the first major application for physical AI due to market size and standardization.
• NVIDIA open-sourced its autonomous vehicle stack to enable applications in agriculture, logistics, and more.
• The physical AI business is already substantial and poised for massive growth.

Advice for Young Innovators

• Simple tasks and coding will be automated; focus on hard sciences and systems thinking.
• AI is a tool to tackle incredibly complex problems.
• Deep tech, intersection of technology and society, and understanding market gaps remain crucial.

The Entrepreneurial Mindset

• The world is rapidly changing, making it a prime time to start companies.
• Embrace a mindset of "how hard can it be?" and believe in the ability to learn.
• Resilience is the single most important quality for entrepreneurs.

AI Costs & Capabilities

• AI token costs are rapidly decreasing, making AI more cost-efficient over time.
• Even today, engineers using the best AI models are significantly more productive than those without.
• It's unlikely a human is more effective than an AI model for most programming tasks.

Building Startup Communities

• For optimal chances, join an existing startup community (e.g., San Francisco, London, Paris) rather than building one from scratch.
• Being part of an early-stage startup can feel like a cult due to shared, unconventional beliefs, but this fosters effectiveness.

AI and Human Judgment

• Founders should be mindful of how much thinking they delegate to AI; writing is correlated with thinking.
• YC focuses on how AI empowers partners and founders, but also on what AI cannot replace (founder well-being, community, unique knowledge sharing).

Startup Durability & The "Hard Bit"

• Pure software products are becoming less durable due to ease of replication.
• Startups need a "hard bit" for durability: this could be complex B2B sales, regulatory hurdles (e.g., banking licenses), hardware challenges, or deep tech.
• Avoid the easier end of the spectrum; focus on harder, more ambitious problems.

The Role of AI in YC & Founder Well-being

• YC is experimenting with AI tools, including virtual partners for office hours.
• Beyond AI capabilities, YC emphasizes "witnessing" – acknowledging the founder's journey – and human connection, which AI cannot replicate.
• YC partners provide tailored advice based on deep understanding of individual businesses, not just generic advice.

Research vs. Business Focus

• The biggest regret for YC founders is not launching soon enough.
• Founders tend to research/build more than sell; actively push against this tendency.
• Optimize for learning by talking to users and confronting the market.
• The cycle is: build, talk to customers, build, talk to customers.
• Be aware of biases towards building/research; consciously increase customer interaction.

AI Model Access & Sovereignty

• US export restrictions on AI models highlight global dependence and may drive demand for competitors and AI sovereignty.

Solo Founders vs. Co-founders

• Theoretically, a one-person billion-dollar company is possible, but adding a co-founder generally increases success probability due to complementary benefits.
• AI may compress companies to under 150 people, maintaining relationships.
• Solo founders statistically perform worse; co-founders provide crucial emotional and operational support.
• When choosing a co-founder, prioritize smarts, determination, integrity, and shared values over skill complementarity. Technical co-founders should seek equally technical partners, as business skills are often learnable.
• YC does fund solo founders but sets a higher bar.

VC Funding & Harder Problems

• Founders may not need to raise as much as before for basic software, but tackling harder problems requires more capital.
• AI enables founders to pursue more ambitious, capital-intensive challenges (e.g., nuclear reactors, regulated banks).
• Venture capital will continue to fund these harder, more impactful endeavors.

Moats & Betting on Founders

• The biggest moat is the founding team; YC bets on great people to figure things out, regardless of their initial idea.
• Even if the initial idea is unconventional (e.g., VR sunglasses), strong founders will pivot to successful ventures.

AI Model Interaction

• Users will quickly stop caring about specific AI models; interfaces will automatically route queries for optimal cost and intelligence.
• Businesses prioritize cost and effectiveness, not the underlying model.

Evaluating B2B AI: Wedge vs. Wrapper

• YC funds companies with AI at their core, not just "AI wrappers."
• Evaluation focuses on the team, their ability to evolve, and the depth of the problem space.
• Many successful companies start as "wrappers" (e.g., SQL wrappers, S3 wrappers).
• Value is added through technical improvement (moving state-of-the-art) or superior distribution/sales.
• YC funds both deep technical experts and those who excel at marketing/distribution.

AI for Students & European Builders

• Students can receive significant AI credits ($25K+) to help compete with those with API budgets.

Early Decision Applications

• Early decision applicants are assessed on whether YC would fund them *today*.
• The "why early decision" question probes commitment; reluctance to wait suggests lower dedication.

Top Startup Habits & YC Updates

• Successful founders launch early and often, embracing iteration and market feedback.
• Startup success is empirical: form hypotheses, test with the market, and adapt based on data.
• Set ambitious goals and constantly reassess the biggest bottleneck to overcome.

Defining AI Success Metrics & Usage

• Don't just look at AI output; engage in a self-learning cycle with the AI, providing feedback to improve its skills over time.
• Start by making information legible and queryable for AI agents.
• Implement AI for narrow, specific processes (e.g., post-sales follow-ups, real-time prototype building during calls).

What NOT to Automate

• Talking to customers is the last thing to automate; it keeps founders focused and provides essential context for building the right product.
• Direct exposure to customers is crucial, even in self-serve PLG models.
• Do not delegate conversations with co-founders; these are vital for company success.

Pivoting in the AI Era

• While writing software is cheaper, pivoting (discarding built work) remains difficult.
• Pivot only when evidence suggests the fundamental business hypothesis is wrong, not due to loss of enthusiasm or sales rejection.
• Pivots are often better when driven by customer needs or exposure to a better idea through initial work.

Developing High Agency

• High agency is the belief that actions lead to impact.
• Develop it by tackling progressively harder but tractable projects, seeking output and feedback.
• Starting with side projects and finding co-founders can build this trait.

Standing Out in a Crowded Market

• Founders must constantly find new ways to differentiate and stand out.
• Distribution is increasingly important; the meta-game for achieving reach is rapidly evolving.
• Early-stage companies stand out through personalized, unscalable service, not mass marketing.

Background and Early Ventures

• The speaker's first company in 2002, focused on chatbots for customer service, was 20 years ahead of its time.
• Wit.ai, acquired by Facebook, was an early AI domain company founded in 2013.

About World Models vs. LLMs

• LLMs learn from text about the world, like someone who has read all books but never experienced the real world. They lack common sense and true world experience.
• World Models learn directly from real-world sensory data (video, audio, touch) from scratch, similar to human babies.
• World Models are expected to be superior for high-dimensional, noisy, long-horizon problems, while LLMs excel at language, math, and programming (low-dimensional symbol sequences).

Amab's and the Vision

• Amab's is building functional World Models, which is very expensive due to the need for thousands of GPUs and billions of euros.
• A key strength is leveraging early tech and understanding early markets.
• Amab's aims to enable helpful robots in unstructured environments, making dangerous or hard jobs safer.
• The company is deliberately not based in SF, having teams in Paris, New York, Montreal, and Singapore.

Ambition and Strategy

• Startups should be very ambitious in a narrow problem space, not broad ones, to be credible and provide value.
• The real cost of raising $1.2 billion is external expectations, not delusion.
• The speaker regrets not being ambitious enough in earlier ventures, citing an instance where Mark Zuckerberg suggested hiring 10,000 people for a project he only planned to hire 100 for.

The Future with World Models

• In 5-10 years, the speaker envisions helpful robots and machines with common sense, making dangerous or hard jobs less so.

Background and Motivation

• Eric Landau's career trajectory: particle physics (big data) -> quant (10 years) -> founder of Encord.
• Common thread: evolution of AI from feature engineering to scaling with data and compute ("bitter lesson").
• Left a lucrative quant job during COVID for the challenge and perceived paradigm shift of AI.
• Found company building more rewarding than finance, despite initial struggles.

Encord's Journey and Product-Market Fit

• Founded Encord in 2021, experienced a "desert" period for two years.
• ChatGPT's release helped open the market and conversations around AI's importance.
• Product-market fit was a gradual compounding process, not a single event.
• Initially focused on vision, then expanded to multimodality, leading to Physical AI.

Physical AI Market and Encord's Focus

• Physical AI is the biggest application of multimodal AI, encompassing robotics, autonomous vehicles, logistics, and manufacturing.
• Huge market opportunity as 80% of economic activity involves manipulating or moving physical things.
• Encord focuses on data management, curation, annotation, enrichment, and evaluation for Physical AI at scale (petabytes).

Operations and Strategy

• Encord has hundreds of customers and handles more data than used to train GPT-4.
• Operates a data collection facility in the Bay Area with robot systems and operators.
• Strategy: be where customers, talent, and investors are, with a focus on US-based customers.
• Competes by focusing on Physical AI and the scalability of data operations.

Advice for Founders

• Embrace the entrepreneurial roller coaster: accept the highs and lows.
• Learn to enjoy challenges ("fires") and react constructively to market fluctuations.
• Follow your gut and make decisions faster, especially regarding hiring.
• For Physical AI startups, consider using products like Encord when scaling from POC to production.

Founding Story

• DataDog founders, Olivier and Alexi, met at Central in Paris and later worked together at IBM in New York.
• They gained startup experience during the dot-com boom and bust, learning valuable lessons about business strategy.
• A previous educational software startup experience, where Alexi led ops and Olivier led dev, was pivotal in forming DataDog.

Early Challenges & Vision

• DataDog was rejected by Y Combinator, with founder Olivier even keeping the rejection email as motivation.
• Securing early funding was difficult, as VCs in New York and the Bay Area didn't fully understand the market.
• The core bet was to unite Dev and Ops on a single platform, coinciding with the explosion of cloud computing.

Product & Culture

• Initially, DataDog was not called "observability" but "infrastructure monitoring," a term users understood.
• The company prioritizes a "bottom-up" approach, targeting developers and operations teams.
• Culture is exemplified by leadership and hiring/firing decisions, rather than written values.

Growth & Strategy

• DataDog has expanded to 20-25 products, often driven by customer usage patterns and platform extensions.
• Strategic projects are also pursued, anticipating future market needs, especially with the rise of AI.
• The company aims to be proactive, not defensive, in a rapidly evolving tech landscape.

Impact of Being Public & AI

• Going public in 2019 led to significant stock fluctuations, but the company has adapted to market volatility.
• Leading a public company involves more structured investor communication but doesn't fundamentally change core decision-making.
• A major concern with stock price drops is retaining key talent due to RSU value fluctuations.
• AI is a significant disruptor; DataDog is investing in AI-powered agents and integrating "smarts" directly into its products.
• The company is embracing AI's potential, with engineering leaders anticipating a future where less code is manually written.

Leadership & Advice

• Founders maintain deep involvement in product decisions, emphasizing understanding the "fabric of the universe" within the company by reviewing support requests and customer feedback.
• A key piece of advice is to move faster, particularly in hiring and firing decisions, to iterate and learn more quickly.
• The co-founder relationship is sustained through consistent, agenda-less conversations.
• For European startups, it's recommended to start locally but aggressively pursue the US market early on.

Company Growth & Metrics

• 4.6 million weekly active users, with 13 million monthly active users by June.
• 20x growth since the beginning of the year.
• Processing ~7 trillion tokens per day, up from 300 billion at the start of the year.
• Projected annual revenue of $38-40 million from inference, launched ~8 months prior.
• 160,000 monthly subscribers for their subscription product, launched ~6 months prior.

Key Growth Drivers & Insights

• Anthropic's "block" on Claude Code users inadvertently boosted OpenCode's visibility and user acquisition.
• Global adoption is significant, with developing countries (Indonesia, Brazil, Vietnam) being major users due to affordability.
• Open-source models are becoming competitive with frontier models, shrinking the gap and increasing viability.
• Gemini 2.5's release marked the first time an open-source model surpassed frontier models in usage on OpenCode.
• The "magic moment" of experiencing a coding agent is key to product-market fit and global adoption.

Product Strategy & Economics

• OpenCode provides an open-source alternative to proprietary coding agents.
• The "OpenCode Go" plan ($10/month) allows access to various open-source models.
• Focus on offering choice and flexibility to users, avoiding vendor lock-in.
• Enterprises are "bugging" OpenCode to sign security agreements, indicating strong product-market fit.
• Community-driven model database (models.dev) supports over 70 models and providers.
• "Building in public" is a core part of their identity and strategy.
• Subsidy model for the free tier drives initial adoption, while paid tiers and "whales" drive revenue.

User Behavior & Model Usage

• DeepSeek Flash is the most used model by token volume, followed by DeepSeek Pro and GLM 5.2.
• Users optimize for cost and speed, sometimes switching to cheaper models like DeepSeek Flash to manage daily/weekly limits.
• Some models like GLM 5.2 are perceived as better for specific tasks (e.g., front-end design).
• China leads in user traffic (17%), followed by the US, Indonesia, and Brazil.
• Large US companies are using OpenCode to avoid vendor lock-in and gain flexibility.

Company History & Philosophy

• The company has a 16-year history (incorporated in 2010) with the same founders and legal entity.
• Jay V applied to YC multiple times over a decade before acceptance in 2021 with a serverless platform idea.
• Past ventures, including consumer products, provided valuable experience in acquisition and metrics.
• Focus on "eccentric tastes" and building things developers *want*, like a modern terminal experience.
• Belief in a positive-sum market where competition benefits consumers and OpenCode acts as a neutral marketplace.

PhotoRoom's Scale and Mission

• PhotoRoom provides e-commerce businesses with AI visuals to grow and sell.
• 300 million downloads, 20 million active users.
• Used by major e-commerce platforms like Amazon, Uber, and DoorDash.
• Largest YC company with HQ in Europe.

The Impact of Y Combinator on Ambition

• YC made founders more ambitious by making big goals concrete.
• Exposure to successful founders normalized aiming for 10x growth.
• Shifted ambition from a distant dream to a realistic possibility.
• Anecdotal evidence suggests a dramatic increase in perceived potential after YC.

Cultural Differences in Ambition

• European culture often discourages big ideas, leading to self-doubt.
• US culture is more encouraging of ambitious projects.
• Fear of failure is higher in some European countries compared to the US.

Why Ambition Matters

• Tackling hard problems attracts talented people.
• Ambition is crucial for attracting smart, ambitious individuals who drive growth.
• Contrarian view: Ambition is not always necessary if pursuing a passionate project (Option A).
• However, for VC-backed startups, ambition for growth is essential.

Tactics for Increasing Ambition

• Surround yourself with ambitious people; block negative influences.
• Perform thought exercises: Aim higher, 10x higher, or as king of the universe.
• Set ambitious goals; they often become self-fulfilling prophecies.
• "Shoot for the stars, worst case you land on the moon."
• "Add a zero" to success metrics to force a re-evaluation of possibilities.

Hiring and Building for Ambition

• Hiring globally (e.g., using English as the default language) acts as a filter for ambitious, globally-minded candidates.
• Align with team members on ambition level to avoid future conflicts.
• Rank projects by potential impact to focus efforts.
• Analyze past mistakes to improve future decision-making.
• Focus on "V0" of projects to quickly learn and iterate.
• Prioritize speed and continuous learning, even with AI.
• YC's exercise of explaining how a business can reach $100M ARR forces ambitious thinking.
• Asking teams what they'd do with half the resources sparks creative, efficient solutions.

Final Advice

• Focusing on a core area (e.g., 10x on photo, then 10x on e-commerce photo) is a powerful way to achieve ambition.
• Acknowledge and manage self-doubt; force yourself to be more ambitious.
• Anyone can create a big company; be ambitious and believe in your potential.

Dust's Vision and Origin

• Dust is building a model-agnostic AI platform focused on applying LLMs to the workplace.
• The founder, Stan, previously worked at Stripe and OpenAI, transitioning from research to product development.
• He left OpenAI to focus on building products rather than just research.

Predictions for the Future of Work

• Work will continue to be disrupted by AI, to the point where future generations might view current work practices as not "real work."
• The pace of AI development has not plateaued as initially expected, leading to continuous redefinition of work.

Competition and Market Strategy

• Dust operates as a horizontal platform, competing alongside giants like OpenAI and Anthropic.
• This horizontal approach is beneficial as frontier labs educate the market, and products converge towards productivity suites.
• Dust differentiates itself through a focus on collaboration and multiplayer AI.
• Crucially, Dust remains model-agnostic, allowing users to choose their preferred AI models, unlike integrated lab offerings.

Fundraising and Company Building

• The AI landscape has seen funding concentrated in a few large labs, making it harder for startups.
• Dust intentionally raised at reasonable valuations to avoid the "coffin corner" of high burn rates and unmet expectations.
• They adopted a "no GPU before PMF" mantra, prioritizing product-market fit over early infrastructure investment.
• The decision to build in France, despite the US being easier from a company-building perspective, was driven by a desire to build back home and a sense of national sovereignty.

Navigating Verticalization and Pricing

• The advantage of verticalized AI products diminishes as models improve.
• Defensibility for verticalized products now relies on network effects within their specific industries.
• Dust is transitioning from seat-based to credit-based pricing to accommodate exploding usage and maintain margins.
• Significant margins exist for frontier model providers, with potential pressure from open-source advancements.

Founder Wisdom

• The most important aspect of building a company is finding a vision or problem to fix that drives you through the daily challenges.

Company Origins and Philosophy

• Supabase started in 2020 as an open-source tool to help people host PostgreSQL.
• Co-founder and CEO Paul Copplestone, originally from New Zealand, has a background in entrepreneurship.
• He emphasizes that founders giving up or lacking product-market fit are more common reasons for startup failure than running out of money.
• Copplestone's previous startups didn't achieve product-market fit or weren't his long-term vision.

Supabase's Core Strategy

• Copplestone chose PostgreSQL due to its trending status on Hacker News and its robust reputation.
• He believes PostgreSQL's open-source nature and community contributions create a flywheel effect, ensuring its future dominance.
• Supabase initially positioned itself as an "open-source Firebase alternative" after realizing "real-time PostgreSQL" didn't gain traction.
• The company's open-source philosophy was a deliberate choice by both co-founders.
• Monetization was considered from the start, with the goal of capturing users who grow with the platform.

Growth and Developer Experience

• Over 60% of Y Combinator batches use Supabase.
• The company aims for a very fast "time to value," reducing the initial setup and usage time significantly compared to AWS RDS.
• Supabase is now 100% open source, with only the self-hosting infrastructure code not publicly available due to complexity and security concerns.
• They counter cloud providers by offering a difficult-to-replicate integrated stack, not just a single database.
• Community feedback is actively sought, but judgment is applied to requests.
• Initial ARR was around $1 million, growing to $7.5 million the next year, then accelerating rapidly.
• Developer experience is key, focusing on understanding developer needs through community polling and optimizing for speed and ease of use.

Impact of AI and Future Direction

• AI agents (like Lovable, Bolt, Compounding, CodeX) now account for a significant portion of database launches, potentially up to 90%.
• Supabase developed "Supabase for Platforms" to cater to companies launching many databases, a strategy also appealing to enterprise clients for management and security.
• Growth accelerated dramatically with AI, with user numbers increasing while conversion and activation rates also improved.
• The company is shifting focus to providing a unified experience across dashboards, CLIs, and MCPs, with an emphasis on infrastructure as code and branchable environments.
• Future bets include "self-driving databases" to handle operation, security, and maintenance, focusing on harder problems in the "operate stage."

Company Culture and Operations

• Supabase operates with a fully remote, distributed team in over 60 countries, with no physical offices.
• This model, documented extensively in written form, is well-suited for AI agents to access historical decisions and operational principles.
• Copplestone stresses that money shouldn't change how a company operates; it should serve the core mission.
• He believes the founding team must remain grounded, knowing money alone won't guarantee success.
• The company's rapid growth and significant funding rounds haven't altered their core operating principles.

Startup Ambition and Investor Focus

• Investors focus on the "what if it all works out" scenario, acknowledging that most startup upside is yet to be realized.
• Ambitious ideas are easier to sell because they are more remarkable, leading to better word-of-mouth growth and easier go-to-market strategies.
• Fundraising is easier for large, ambitious ideas; while some investors may be deterred, strong investors are more likely to engage.

PostHog's Evolution to AI-Native Self-Driving Software

• PostHog pivoted from open-source product analytics to self-driving software powered by AI.
• The new approach aims to solve entire jobs for users rather than just providing insights.
• The system analyzes diverse data types (support tickets, logs, user sessions) to identify problems and automatically ship pull requests to fix them.
• They are developing a system to capture user intent, allowing for more rapid iteration and understanding of product purpose.

The AI-Native Company and Product Management

• The vision involves AI taking on tasks previously done by product managers by analyzing vast amounts of customer data and behavior.
• The role of a product manager shifts to configuring AI systems and setting rules and policies.
• A new support product leverages deep user understanding to resolve tickets more effectively, with support teams now guiding the AI agents.
• The product is already generating a significant portion of its own pull requests, freeing up developers for larger features.

Strategic Pivots and Founder Mindset

• PostHog made multiple pivots (around five) before finding product-market fit with Post Hog.
• The key to learning was setting small, iterative goals focused on user adoption rather than just revenue metrics.
• Early on, they prioritized learning by actively validating ideas with users, even with significant travel for small contracts.
• They learned to abandon ideas quickly if they hit consistent hurdles, recognizing their initial lack of product expertise.

Marketing, Brand, and Culture

• PostHog's distinct brand and marketing (e.g., hedgehog mascot, bus ads) were intentional strategies to stand out in crowded markets.
• They focused on executing non-product elements better initially, building a community that wanted them to succeed.
• The website is treated as a product, with a deep understanding of user personas and use cases.
• The company culture values fun and entertainment, with a hiring process that avoids traditional corporate norms.

Ambition and Execution

• Founders are encouraged to be highly ambitious, focusing on the "what if it all works out" scenario.
• This ambition makes hiring top talent and go-to-market strategies easier.
• There's a need to balance extreme long-term ambition with short-term, practical execution (shipping code).
• Shipping frequently is crucial, especially for ambitious ideas, to learn and iterate in real-time.

Sample Efficiency and World Models

• The core AI problem discussed is sample efficiency: how to make models learn new tasks quickly from limited data, unlike humans who are very efficient learners.
• World models are presented as a promising path to solve this gap, aiming to enable AI to quickly learn and adapt.

Understanding Sample Efficiency

• Sample efficiency relates to how much more intelligent a model becomes with each additional data sample.
• Humans excel at learning from few examples (e.g., new games, concepts), while current AI often requires thousands of data points.
• Tasks like the RG&I test, intuitive for humans, are challenging for AI despite access to vast internet data.

The Concept of a Perfect World Model

• A "perfect world model" implies zero need for new environmental samples, akin to knowing Newton's laws for rocket trajectory.
• NASA's asteroid interception and Apollo missions are examples of leveraging pre-built world models (physics) to achieve complex goals without constant real-time learning.
• A 1967 study showed imagining an action led to similar improvement as practicing it, highlighting the power of internal world models.

World Models vs. Implicit Understanding

• Human intelligence is thought to stem from an implicit world model encoded genetically and learned.
• While LLMs show surprising intelligence in language without explicit world models, this breaks down in domains like robotics and self-driving.

Reinforcement Learning and World Models

• Reinforcement Learning (RL) involves learning a policy (what action to take given a state).
• A world model (or state transition function) predicts the next state given the current state and action (St+1 = f(St, At)).
• Model Predictive Control (MPC) uses a world model to plan optimal actions by minimizing a loss function over future states. This works well in differentiable, deterministic environments.
• When environments become stochastic and non-differentiable (e.g., adversarial interactions), RL resorts to complex methods like value iteration or actor-critic.

Model-Based vs. Model-Free RL

• Model-free RL trains policies directly from experience without an explicit world model (e.g., behavior cloning).
• Model-based RL incorporates a learned world model, allowing for planning and potentially much greater sample efficiency, but requires more complex inference.

Challenges in Complex Environments

• The state and action spaces in games like Chess and Go, while large, are somewhat tractable due to deterministic rules and relatively small action spaces for AlphaGo.
• Self-driving cars and robotics present enormous, effectively infinite, state and action spaces, along with stochasticity and non-differentiable elements (other agents' behaviors).
• Collecting sufficient training data (especially state-action pairs) is a major bottleneck in robotics and self-driving.

Advancements in World Modeling

• Early work like Jurgen Schmidhuber's "World Models" demonstrated training policies on synthetic data generated by a world model.
• The Dreamer series (e.g., DreamerV4) improved this by using action conditioning on a small amount of real data to train world models, achieving tasks like mining diamonds in Minecraft solely on synthetic data.
• Modern approaches leverage pre-trained video diffusion models and flow matching, combined with action conditioning (e.g., Wave's Gaia, Dreamer Zero for robotics), to create powerful world models from vast online data (like YouTube).
• JEP A (Joint Embedding Predictive Architecture) uses encoders to learn in a compressed latent space, predicting future latent states rather than raw pixels, to improve efficiency and prevent modal collapse.

Future Directions and Open Problems

• Achieving higher fidelity and precision in world models is crucial, as current models still struggle with machine precision and complex physics.
• Test-time planning and adaptation are key challenges, enabling rapid adaptation to new environments or unexpected events, mirroring human agility.
• Real-time performance and efficient planning are critical for applications like self-driving and robotics, where delays are unacceptable.
• Integrating concepts like "sleep" and "dreaming" for memory consolidation and training, as observed in the human brain, is an open area.
• Developing better tactile sensing and understanding of physics (e.g., friction coefficients) in robotics remains a challenge.

The "Squint Test" and Human Intelligence

• The "squint test" suggests that current world models, with their combination of world models and test-time planning, are getting closer to human-like intelligence than purely autoregressive LLMs.
• However, the lack of a "sleep" mechanism for consolidation and memory in AI architectures means they don't fully pass this test yet.

Design Tools & Workflow

• Primary tools: Conductor and Paper.design.
• Inspiration gathering: Uses Pinterest for mood boards.
• Hands-free interaction: Employs Aqua (YC company) for voice-to-computer input, finding it faster than typing.

Paxel Project: Understanding Coding Agents

• Goal: To understand how people code with AI agents and share insights.
• Features: Provides feedback on coding patterns, crashes, and prompts.
• Inspiration: Modeled after Spotify Wrapped for coding sessions.
• Design: Uses interactive cards to present fun facts about coding habits.
• Technical: Leverages Paper.design's dithering shader via Claude.
• Customization: Built a modal to fine-tune shader parameters.
• Human vs. Machine Content: Created a separate markdown version for AI agents.
• Feature Requests: Implemented a form that sends requests directly to an agent, opening a PR for review.

Sodazine Project: Celebrating San Francisco

• Approach: Physical zine designed without AI, then a website to showcase it.
• Source of Truth: Uses a soul.md file containing all meeting transcripts and project context for AI agents.
• Website Design Process:
• Generated 16 website iterations using Claude based on a Pinterest mood board.
• Created a personal glossary to navigate and bookmark preferred iterations.
• Explored different layouts and discovered AI surprising additions (e.g., launch party date).
• Experimented with interactive San Francisco maps and article displays.
• Final Website: An interactive map for users to share anonymous memories of San Francisco.

Startup School Event Branding

• Design Elements: Used gradients of orange and experimented with Paper.design shaders.
• Speaker Cards: Created a tool using Claude to automatically generate speaker cards from a list, allowing for easy iteration.
• Shader Customization: Built mini-tools to fine-tune shader parameters like graininess and rotation.
• Smooth Looping Videos: Developed a tool to create perfectly looping 4-second screen recordings for social media.
• Personalized Tickets: Designed personalized tickets with recipient's name and city using the event's shaders.
• AI's Role: Highlights how AI simplifies complex tasks like shader implementation and consistent branding across large-scale assets.

Understanding User Behavior

• Aggregate user metrics can be misleading; understanding individual user behavior is crucial.
• Dot plots provide a visualization of individual user activity over time.

How to Create a Dot Plot

• Create a 2D grid: Rows represent individual users, columns represent time periods (days).
• Mark an event that signifies value for the user (e.g., listening to a song, sharing a photo).
• Place a dot for each day the user performs the chosen event.
• Optionally, add symbols for onboarding day (e.g., a ring around the dot) or different user states (e.g., OS, demographics).
• Sort rows by attributes to analyze specific user segments.

Benefits and Insights from Dot Plots

• Identify usage patterns (e.g., weekday vs. weekend users).
• Assess retention by observing users who try the product and don't return.
• Gain granular insights that aggregate metrics miss (e.g., comparing DAU graph with dot plot data).
• Understand feature adoption and its impact on usage (e.g., playlist feature leading to sustained use).
• Can be used for sampling with large user bases (millions/billions).
• Useful for B2B products to track seat activation and usage.

Common Mistakes and Best Practices

• Mistake: Charting the wrong event (e.g., "opened app" instead of a value-generating action).
• Best Practice: Choose an event that truly represents value creation for the user.
• Mistake: Using a time period that's too wide (e.g., weeks instead of days).
• Best Practice: Use daily or even sub-daily granularity for detailed insights.

Dot Plots vs. Other Metrics

• Dot plots offer more detail than aggregate metrics like DAU.
• They complement cohort retention curves by showing *how* users engage, not just *if* they return.
• Useful for identifying anomalies (like fraud detection, as seen in PayPal's early days).

Introducing Gusto Co-founder

• Gusto Co-founder is an AI product designed to automate most small business processes.
• It goes beyond traditional AI as a search engine, acting as an agentic tool.
• The "blank canvas problem" of other AI tools is addressed by starting with Gusto's existing solutions (payroll, HR, time).

Overcoming the "Blank Canvas Problem"

• Users previously had to manually set up complex AI systems like Open-Source AI.
• Gusto Co-founder suggests and automates recurring tasks (e.g., payroll, time sheet approval) without user login.
• The idea stemmed from the co-founder's personal experience with Open-Source AI setup and realizing the potential for direct customer application.

Development and AI Integration

• The prototype involved customers prompting desired web app functionalities.
• It evolved to leverage Gusto's existing customer data for more relevant workflow automation.
• Technical inspiration came from Open-Source AI's simplicity, including cron jobs and LLM usage.
• The team used a lean, AI-assisted development process: 5 people in 10 weeks with minimal traditional tools (no Jira, Figma, or extensive docs).
• Designers contributed code, and engineers focused on rapid prototyping and iteration, embracing code deletion.

Impact on Small Businesses

• Gusto Co-founder simplifies complex, recurring tasks, freeing up owners' time for growth and strategy.
• It can proactively identify opportunities like R&D tax credits, helping businesses save money (e.g., Cabana Pools saved $50k).
• The tool aims to provide market intelligence and competitive analysis for small businesses.
• It enables businesses to do "more with less," focusing on core operations.

Future and Accessibility

• Upcoming features include more communication channels (Telegram, WhatsApp) and connectors for vertical-specific software.
• The platform will eventually open to individuals without a current business, assisting in startup creation.
• Gusto Co-founder can handle tasks like EIN registration and employer compliance.

AI's Global Impact and India's Opportunity

• The AI revolution is global, unlike the mobile revolution which led to local network effects.
• India has the best technical talent to build the world's largest AI companies.
• Success in AI is about understanding the technology 10x better than others, an area where India excels.
• Founders can now build global companies from India without needing to be in Silicon Valley or having prior US market connections.
• Cold outreach to US companies is now feasible with strong products.

Shifting Educational and Career Paradigms

• Traditional advice to pursue safe, high-paying jobs may become risky as AI evolves.
• Entrepreneurs and business owners are best insulated from AI-driven job changes.
• Young technical founders living on the cutting edge of AI are well-positioned for success.
• Developing an independent point of view and high agency is crucial.
• Surrounding yourself with others at the cutting edge is a deliberate choice for compounding career growth.

The Rise of Young Founders and the "Tinkering" Mindset

• AI has leveled the playing field, allowing young founders to build quickly and gain insights rapidly.
• The best young founders "tinker," following their curiosity to explore the edge of what's possible.
• This tinkering leads to discovering bottlenecks, which represent good startup ideas.
• Coding agents enable rapid prototyping, helping founders find viable ideas through building rather than just thinking.

Second Mover Advantage and AI-Powered Development

• Strong technical teams can overcome established competitors with superior products, even with fewer resources.
• Coding agents allow founders with product clarity to realize ideas extremely fast, often with high quality.
• A strategy can be to find a working concept and improve upon it, especially if network effects aren't dominant.
• Pushing AI models to their limits, even with higher costs, unlocks sophisticated code and reveals new startup opportunities.
• Open-source models and cheaper compute will increase accessibility, though cutting-edge performance may require premium access.

What Y Combinator Looks For in Founders

• Clarity in explaining what you're building is paramount.
• Y Combinator invests in the founder, not just the idea, focusing on taste and agency.
• Taste is about intentionality, customer insights, and rapid product development.
• Agency means relentlessly exerting your will on the world rather than being subject to conditions.
• A high rate of learning is essential.
• Projects (building something unassigned and getting users) are key to developing these traits and finding ideas.

The State of Consumer Tech

• Consumer tech is currently not considered investable by many, but the opportunity is immense due to AI and agents.
• New "internet treasures" can be created by reinventing existing services or concepts.
• The current moment feels like the third major era of computing, following the web and social/mobile.

Lessons from Past Eras

• The early days of social networking, with Napster as a key example, showed the power of decentralized, peer-to-peer connections.
• Early social network attempts like Tribe failed by not getting the "trust" component right, which Facebook later succeeded with.
• Investors can be wrong about market trends, sometimes being 180 degrees off from what will be successful.

The AI Revolution and "Proven Better New"

• AI models like GPT-4.5 have made agents feel like peers that can be trusted with tasks.
• A key use case is having AI listen in on conversations and provide insights, though current tools like Granola have friction.
• The "Proven Better New" framework:
  • Proven: Legally copy what successful products do that you aren't innovating on.
  • Better: Make a 10/10 improvement that users would instantly recognize (e.g., free, faster, less friction).
  • New: The core hypothesis or innovation you are testing, which is most likely to be wrong.
• AI is brilliant at the "proven" part but struggles with "better" and "new," which is where humans are still essential.

The Future of Consumer AI and Distribution

• The ideal consumer AI moment is still 3-5 years away due to current high costs for powerful AI.
• The cost of intelligence (AI compute) is decreasing, enabling new possibilities similar to how the iPhone became possible.
• The "business plan of free" is crucial for consumer tech, especially with AI becoming more accessible.
• Unlimited, free AI will redefine existing services and create new "internet treasures."

Founder Mode and Staying Power

• Founders should embrace "Founder Mode," which means staying true to their vision and instincts, not abdicating to boards or investors.
• Leadership is about presence; founders should be deeply involved and understand their product.
• It's crucial to create a culture where it's safe to be intellectually honest and pivot based on new learnings.
• Navigating "the abyss" (periods of uncertainty between passionate pursuits) is key to expanding taste zones and finding new inspiration.

Understanding Your Customer's Habits

• Don't default to cold email/LinkedIn: These channels only work if your target customer spends significant time on them.
• Research buyer behavior: Understand their daily routine, email frequency, conference attendance, and how they seek recommendations.
• Visit trade shows: One founder closed more in 3 days at a trade show than in 3 months of cold outreach.

Leveraging Your Network

• Start with warm leads: Your first 2-3 customers will likely come from friends, former colleagues, or classmates.
• Trust is key: Early customers buy because they trust the founder, not just the product.
• Use LinkedIn intros: Ask for warm introductions to your second-degree connections.
• Utilize network search tools: Tools like Happenstance can help find relevant people in your extended network.
• Be specific when asking for intros: Clearly state who you want to meet, why they'd care, and what to say.

The Power of In-Person Interactions

• Show up physically: Many founders closed early customers by being in the same room as the buyer.
• Persistence pays off: Flying out to customers repeatedly or showing up uninvited can lead to significant accounts.
• Leverage small conferences: Schedule back-to-back meetings and promote your presence to attendees.
• Host micro-events: Small founder dinners or happy hours can be highly effective for conversion.

Engaging in Online Communities

• Find where customers complain: Identify online spaces (Reddit, Facebook groups, Discord) where your target audience expresses problems.
• Engage directly: DM commenters on Reddit threads or respond to complaints on social media.
• Reddit's longevity: Content on Reddit can surface in Google searches for a long time.

Strategic Outbound Outreach

• Use tools like Apollo or Clay: These can help with lead generation, data enrichment, and research workflows.
• Leverage LinkedIn Premium: Use it for professional data and send connection requests followed by DMs.
• Frame outreach as advice/mentorship: Instead of a sales pitch, ask for feedback, mentorship, or a whiteboarding session.
• Offer value upfront: Provide a quick scan, suggestions, or an audit note before asking for a meeting.
• Keep outreach concise: Emails under 75 words with a clear call to action perform best.
• Read emails aloud: This helps ensure the copy sounds human.
• Follow up: Send 3-4 follow-ups over a couple of weeks.

Phased Customer Acquisition

• Customers 1-3: Come from your personal network.
• Customers 4-10: Come from unscalable, manual efforts (in-person, DMs, micro-events).
• Customers 10-50: Begin to leverage higher volume tools as you refine your pitch and value proposition.

Introducing Ploy: The AI-Powered Marketing and Website Platform

• Ploy is a website platform that allows users to build bespoke, award-winning websites.
• It functions as an entire marketing platform to help businesses run ads, find customers, generate website copy, and get found by AI models like ChatGPT, Perplexity, and Claude.
• The goal is to enable businesses to run their marketing on autopilot.

Leveraging AI for Modern Web Design and Marketing

• Ploy uses AI to recreate and modernize old websites, making them look like they're from 2026.
• It understands website content from the Wayback Machine, business context, and modern design trends.
• The platform generates not just visuals but also asset creation, including images and animations, which were previously expensive and time-consuming.
• Ploy aims to demystify marketing and growth for founders, addressing the challenges of SEO and marketing tasks.

The Age of the Solo Founder: AI as an Equalizer

• The video highlights the emergence of the "40-year-old solo founder" enabled by AI.
• Experienced individuals with domain expertise can leverage AI models to create world-class products.
• AI acts as an equalizer, allowing founders with deep technical skills but less marketing prowess to succeed.
• Founders can now achieve what previously required large teams and significant resources.

Ploy's Functionality and Future

• Ploy's "slurper" creates a design system and components from existing websites, ensuring brand consistency.
• The platform integrates with various tools like Figma, CRMs, and analytics, acting as a "company brain" for marketing.
• It analyzes traffic, search console data, and pipelines to offer suggestions and automate marketing tasks.
• Ploy is designed to be an opinionated solution for businesses, providing purpose-built tools that leverage AI capabilities.
• The company is working on enabling AI agents to sign up for Ploy, with plans for a CLI with skills for agent interaction.

Webflow vs. Ploy: Evolution of Development

• Bryant Chou, co-founder of Webflow, drew parallels between Webflow's democratization of web development and Ploy's goal of democratizing marketing.
• Unlike Webflow, which focused on a specific persona (freelance web designers), Ploy aims to serve tens of millions of people.
• Ploy leverages AI to automate tasks that previously required manual coding and extensive infrastructure setup, allowing for faster and more comprehensive output.
• Experienced builders still benefit from Ploy by guiding the AI with their vision and expertise.

Don't Overthink Your Idea

• The biggest mistake is trying to find the "perfect" idea; you can only discover this by talking to customers and getting feedback.
• Don't let the question of whether you are the "perfect" founder stop you; curiosity and deep customer engagement can lead to expertise.

Commit to One Idea and Go Deep

• Working on multiple ideas at once yields bad data and prevents true learning.
• "Burning the other boats" means fully committing to one idea, stopping work on others, and changing your company's narrative.
• Going deep means becoming a domain expert, to the point where you could run your customers' business or teach a class on the problem.
• Execute in a tight loop: deep customer understanding -> product delivery -> deeper understanding -> better delivery.

Qualities of a Good AI-Era Idea

• Sits at the edge of what current AI models can do, with potential for future improvement.
• Verticalizes: sells an outcome (e.g., insurance, medical care) rather than just software.
• Is the most ambitious version of itself to attract talent and create a competitive moat.

Dealing with Failure

• Even a failed idea provides valuable customer data, conviction for pivots, and a better sense of execution.
• The process of going deep often reveals a better, underlying idea by uncovering deeper structural problems.

Key Takeaway

• Stop searching for the perfect idea, pick one, and commit fully. Walk fast in one direction to generate the most information and potentially discover a better destination.

Company Origin and Vision

• Groww started with a broad vision to enable everyone in India to invest.
• The founder's initial inspiration came from learning about Steve Jobs and the dot-com boom.
• Groww was not the first startup; the founder had previous failed startups, including an initial iteration as a robo-advisor.

Product Development and Customer Focus

• The initial robo-advisor idea failed because customers wanted more choice and transparency.
• Groww's current format emerged from the insight that Indian customers value choice and transparency.
• A key decision was to open the platform with full transparency, which was counterintuitive but led to strong product-market fit (PMF).
• Customer love (measured by NPS) was high, driving organic growth through word-of-mouth.
• The team actively engaged with customers through WhatsApp groups and other channels to understand underlying needs, not just direct requests.
• A core principle is to "obsess over design" and for founders to be "power users" of their own product.
• Features are considered successful if customers either love them or hate them; indifference is a sign of failure.
• The company operates under the "do things that don't scale" philosophy, even at its large size.

Business Strategy and Monetization

• Groww strategically chose to operate in regulated zones, obtaining necessary licenses.
• For the first four years, Groww had zero revenue, focusing on customer acquisition and engagement.
• Monetization was unlocked when customer demand shifted from regular mutual funds (which had commissions) to direct mutual funds (zero commission) and later to stocks.
• The company bet that high organic growth, retention, engagement, and customer love would eventually enable monetization.

Navigating Competition and Future Growth

• Groww remains paranoid about disruption and focuses on understanding customer needs and changing trends, including the impact of AI.
• The founder actively experiments with new technologies like AI coding tools to stay ahead.
• AI has significantly lowered the barrier to building products, allowing individuals to handle product management, design, coding, and operations.
• Future growth areas include wealth management for existing customers and making the platform exciting for younger users.

Co-founder Dynamics

• Groww has four co-founders with aligned value systems, clear ownership, and a shared customer-first ethos.
• Enjoying each other's company and maintaining strong relationships is crucial for long-term co-founder success.

Advice for Aspiring Entrepreneurs

• Don't solely rely on advice from older generations; younger people have better current context.
• Pursue work that feels like play, where time becomes a blur and you genuinely enjoy the process.

AI in Biology: The Bitter Lesson Extended

• AI models trained on massive biological sequence data exhibit scaling laws similar to language models.
• These models can predict protein structure and function without explicit biological knowledge, aligning with the "bitter lesson" that general methods exploiting scale outperform hand-engineered features.
• Data scaling, particularly from metagenomic sources, is crucial for continued improvement, mirroring the "data wall" conversation in LLMs.
• Protein language models can rival specialized models like AlphaFold, especially in complex areas like antibody design, suggesting a shift towards generalizable AI approaches.
• Interpretability tools reveal that these models learn hierarchical biological concepts from sequences alone.

Selfplay for LLM Improvement

• Selfplay, inspired by AlphaZero, aims to auto-generate training tasks for LLMs to surpass human-level performance.
• Traditional selfplay can plateau due to generating overly complex or uninformative tasks.
• "Self-Guided Selfplay" (SGS) addresses this by grounding synthetic task generation in solvable problems and using a "guide" model to ensure task relevance and quality.
• While SGS shows improvement over baselines, it still faces challenges in reaching 100% solve rates, indicating ongoing research is needed.

Reducing Latency in Voice AI with Streaming RAG

• Voice AI requires low latency for natural conversation, which traditional RAG systems struggle to provide.
• Streaming RAG analyzes user speech in real-time to identify relevant information and trigger retrieval before the query is fully spoken.
• Approaches include fixed-interval streaming and adaptive triggering based on retrieval quality or semantic relevance.
• This technique can significantly reduce latency (e.g., 1.5 seconds on human data) with maintained accuracy.

Formal Verification and Verified Intelligence with Lean

• AI is making breakthroughs in solving formal math problems, including Olympiad and open mathematical challenges.
• Lean is a powerful language for formal mathematics and programming, enabling rigorous, machine-checked proofs.
• LLMs are being integrated with theorem provers like Lean for proof generation and verification, accelerating formalization.
• The focus is shifting towards "verifiable coding" and AI for science, requiring guaranteed correctness and reproducibility.
• Frameworks like "torchlean" allow defining and verifying neural networks within Lean, enabling certified robustness and formal analysis of AI systems.

Agentic Programming and "Token Maxing"

• Modern software development with AI agents is compared to real-time strategy games, requiring parallelization, constant adaptation, and high visibility.
• Practices include running workflows in the cloud, using portable development environments (like Git worktrees), and prioritizing parallel agent execution over token efficiency.
• Agents aim to complete tasks to pull requests, making assumptions and adapting to evolving specs, with human oversight for correction.
• A "macro by default, micro when it counts" strategy emphasizes spawning many agents and focusing human attention on high-level coordination and course correction.
• Building a comprehensive, linked knowledge base is crucial for agents to access business logic and improve future performance.
• Key principles include "satisficing" (doing enough, not perfectly), using audio/visual cues for agent monitoring, tracking tool calls (agentic "APM"), and maximizing resource utilization.

Company Overview & Mission

• Meesho is an e-commerce platform focused on providing the best value for money across categories.
• It was built for "mass India" with a mission to democratize internet commerce for a billion consumers and businesses.
• In the last 12 months, Meesho had 250 million consumers buying, with each buying about 10 times a year (2.5 billion orders).

Origin Story & Early Pivots

• Founded in 2015 by two friends from small towns who saw a gap between online shopping in big cities and its absence in their hometowns.
• Version 1: FashionNearMe (2015) - A local fashion commerce app for small businesses. Shut down after 3 months due to a crucial mistake: not speaking to consumers, resulting in a product worse than malls and traditional e-commerce.
• Version 2: Meesho (My Shop) - Focused on enabling small businesses to create an online shop on WhatsApp. This faced challenges in monetization as small businesses were unwilling to pay for software.

Product Market Fit & Growth (Social Commerce)

• Identified "resellers" (drop shippers) as power users who didn't have offline shops and actively used the platform to sell on WhatsApp.
• Launched Meesho Supply (later rebranded to Meesho) as a separate app to connect resellers with suppliers.
• Achieved significant product-market fit, doubling users every month for 10 months without marketing spend.
• This phase leveraged WhatsApp's distribution advantage due to lower data costs and text-based nature.
• At its peak, 10 million "Misho entrepreneurs" were selling to over 100 million people via WhatsApp groups.

Strategic Pivot to Direct-to-Consumer App

• In 2020, the decrease in data costs and the pandemic created a paradigm shift, threatening the reseller model.
• Decided to transition to a direct-to-consumer app to avoid losing customers to competitors. This was a risky move that required committing to the new direction, not experimenting.
• Launched the new app on July 5, 2021, and became the #1 shopping app on Android Play Store by July 7, 2021, a position maintained daily since.
• Within 5 months, user base grew from 10 million sellers to 100 million monthly active consumers.

Core Philosophy & Future with AI

• Company value: "Be problem first. Be very rigid with your problem and be very flexible with your solution."
• Believes AI will revolutionize e-commerce.
• Developing AI innovations like the voice agent "Wani" to remove barriers to entry for consumers, especially in rural areas.
• Future vision: a completely voice-driven, invisible software experience, potentially without a traditional app, to reach a billion users.
• Advice to aspiring entrepreneurs: Identify gaps, leverage new technologies like AI, and remain customer-obsessed.

CEO's Role in AI

• CEOs must be the Chief AI Officers, understanding technology bounds deeply.
• Focus on tasks only humans can do, as AI cannot replicate them.
• Rethink the company's core identity with AI integration.

The AI Paradigm Shift

• Early AI use treated LLMs as precious and expensive, leading to over-engineered systems ("Foxconn factory" approach).
• The true model is "agent loop with tools" (skills, tools, model).
• Electricity analogy: GPT-4.5 (December) was the invention of electricity; most are still using candles.

Securing AI Agents (Crab Trap)

• Focus on network layer security for AI agents, not just tool control.
• Crab Trap: An open-sourced HTTP proxy to analyze agent traffic, making it auditable.
• Uses another agent to analyze traffic and enforce policies, with an LLM acting as a judge.

AI Adoption and Productivity

• Token maxers see 10x productivity, average engineers 1x, and others use AI like Google search.
• AI enables "virtual employees" that can join meetings, take notes, etc.
• The "AI pill test": Do you default to AI for any problem?

Company Re-founding with AI

• Startups today should be built with AI as the core premise ("Why can't it just be me?").
• AI changes the fabric of a company, shifting focus from human execution to human wisdom and choice.
• Redesign processes end-to-end with AI, not just layering it onto old systems (e.g., KYC process).

Future of AI and Inference

• The future is "long inference" as AI adoption grows exponentially.
• Token costs will decrease, but usage will increase, making inference a major expense.
• Companies need to manage and analyze token spend for ROI.

Key Insights for Founders

• Minimize surface area; identify problems that can be solved with clear boundaries, potentially compressed by AI.
• Empathy and understanding unspoken customer needs (making the implicit explicit) are crucial, as models lack this.
• AI is a powerful tool, but human wisdom and choice remain the bottleneck.
• Embrace the "discontinuity" AI brings and rebuild from scratch if possible.

Company Overview

• Emergent is an AI-native company enabling anyone to build and monetize software without programming knowledge.
• It leverages AI to simplify coding, allowing users to create shippable software through conversational interfaces.
• The platform handles hosting, deployment, and maintenance.

Founding and Growth

• The company was founded by a team passionate about programming, aiming to democratize software creation.
• Emergent reached $100 million in annualized revenue run rate within 9 months of launching its current product.
• It has amassed over 8.5 million users and seen more than 10 million apps built on the platform.
• Users are located in 190 countries, with the majority of revenue coming from the US and Europe.

Founder's Journey

• Founder Mukund was inspired by Steve Jobs and early tech innovation, even pursuing a PhD and working at Google on search ranking before starting his own ventures.
• He previously founded Dunzo, a successful quick-commerce company in India, learning valuable lessons about solving hard problems and customer focus.
• After Dunzo's challenges, Mukund took a six-month break, during which he tinkered with emerging AI technologies, leading to the Emergent idea.

Technical Innovation

• Emergent focuses on building autonomous agents, a multi-agent orchestrated system with specialized agents for tasks like testing and design.
• A core innovation is a large memory system that learns from each app built, improving the platform over time.
• The company developed its own infrastructure, including coding agents and container technology (disk/memory snapshotting), to handle state preservation for parallel agents.
• They have rebuilt their system three times in nine months to adapt to new AI models.
• Emergent aimed to automate all of software engineering rather than focusing on incremental improvements like AI co-pilots.

Competitive Strategy & Vision

• Emergent differentiates itself by focusing on building "real software" that works, unlike competitors who often focused on front-end prototypes or demos.
• They identified that users want working software with back-ends and databases, a gap they aimed to fill.
• The company leverages a "second mover advantage" by learning from existing solutions and building a more complete product.
• Their Go-To-Market strategy involves mathematical modeling of growth and utilizing influencers to reach a broad user base.
• Advice for aspiring founders: think global from day one, trust your intuition, and aim for ambitious, 10x-100x ideas, especially in the current AI landscape.

Legora's Marketing Strategy

• Utilized Jude Law in advertising to make legal tech seem "sexy" and overcome its traditionally boring image.
• Actors are hesitant to endorse AI due to industry concerns; Legora pursued Jude Law for 6 months.
• Showcased customer testimonials ("customer love" Slack channel) to convince Jude Law of the product's value.
• Jude Law brought his own screenwriter (SNL writer) and cinematographer (Oppenheimer) for the campaign.
• The campaign resulted in 17 touchpoints and generated leads, including one from a mother recognizing Legora from the ad.

Founder's Journey and Y Combinator Experience

• Founder pursued diverse studies (computer science, business) and worked at startups before Legora.
• Did not initially plan to found a legal tech company; felt it "picked" him.
• Took a risk by declining a full-time offer from McKinsey to join Y Combinator.
• Legora was an early applicant to YC's AI batch in Winter 2023.
• Found that many YC companies were still searching for product-market fit, while Legora had significant revenue.
• The Legora team lived and worked intensely in an Airbnb during YC ("work camp").
• Founder focused on sales in Stockholm, leveraging excitement for legal tech and competitive pressure.
• Tactical swap: US-based team handled customers while the founder focused on fundraising at YC.
• Secured significant investor interest at YC through relentless scheduling (80 meetings in a week).
• Delivered strong pitches under pressure, impressing investors like Peter Fenton from Benchmark.

Legora's Product Evolution and Future Vision

• Started with three core features: agent/assistant, tabular review, and Word add-in.
• Strategically aimed to be the best in all three areas and bundle them, surpassing competitors focused on single features.
• Long-term strategy involves building a massive company, similar to Google's breadth.
• Aims to be a major European tech company, challenging the status quo (e.g., SAP).
• AI is a key opportunity for mid-tier law firms to ascend in rankings.
• Leveraging "XYZ founders" (previous founders) within their engineering and product teams.
• Product has evolved from augmenting individual lawyers to building proactive agents due to model advancements.
• Agents can now handle complex tasks like structuring data rooms for M&A transactions.
• Focus is shifting from real-time interaction to broader instructions for agents.
• Defensibility lies in proprietary data, workflows, and user behavior, not just model intelligence.

Company Growth and Current Status

• Reached over $100 million in ARR.
• Grew from 40 people to nearly 500 globally within a year.
• Operations span San Francisco, New York, London, Stockholm, and more.
• Company culture driven by "founder mode" energy, with former CEOs leading product departments.
• New capabilities unlocked by model intelligence over Christmas have enabled proactive agents.
• Moving from individual task augmentation to end-to-end work products.