Y Combinator

In this episode, Boris Cherny discusses the capabilities of Opus 5 and its integration with Claude Code, highlighting several key advancements. Opus 5 can run autonomously for days or months without external scaffolding, and it features a three-layer defense against prompt injection, including a mechanistic interpretability classifier, making it effectively unprompt-injectable. The Claude Code team deleted 80% of the system prompt for Opus 5, finding that a simpler prompt can make the model slightly more intelligent. Cherny advises users to give high-level instructions rather than specific steps, as the model now handles autonomy well. A notable example is the Bun team using Claude Code to rewrite their JavaScript runtime from Zig to Rust in 11 days with a single prompt, a task that would have taken top engineers over a year. Cherny introduces the concept of "product overhang"—re-testing older problems with new models—and notes that capabilities like drawing with OpenCV emerge without explicit training. He emphasizes a shift from prompt engineering to context engineering, and that coding is solved for common tasks but not for deep systems or distributed systems. The best users adopt an empirical mindset, letting go of past assumptions and experimenting freely. For students, Cherny recommends learning programming by solving real problems and building products, starting with something personal before scaling to something others want.

In this episode, Jensen Huang reflects on the mindset and key decisions that built NVIDIA, beginning with the company’s near-failure when its foundational 3D graphics technology proved “absolutely wrong.” He describes how confronting that reality and learning OpenGL from textbooks saved the company and established a lasting principle: technology changes, but the ability to learn any new domain is what matters. Huang argues that great companies are built on a unique perspective—NVIDIA’s was that it should accelerate algorithm domains, not just build chips—and that CEOs should shape their organizations to fit their own strengths, like an F1 car built for its driver. He predicts systems thinking will become the most critical skill as AI automates low-level tasks, and he counters fears about job destruction by asserting that AI eliminates tasks, not jobs, citing growth in software engineering and radiology roles despite automation. Huang also notes that generative video convinced him robotics articulation was imminent, and he calls this the greatest time in 60 years to start a company due to a complete technology reset. Throughout, he emphasizes an optimistic, almost naive mindset—“how hard can it be?”—alongside continuous learning and resilience as the most important qualities for entrepreneurial success.

The episode explores how AI is reshaping startup dynamics, particularly by rapidly lowering the cost of intelligence—potentially 10x per year—and making AI-assisted engineers far more effective than those without. While these claims are flagged as needing research, the discussion emphasizes that founders should focus on execution rather than current token costs. A central theme is the enduring value of co-founders, not for complementary skills (which AI can now help bridge), but for emotional support and resilience; solo founders statistically perform worse, and a co-founder’s primary role is to provide motivation and stability. The episode advises joining established startup hubs, launching early and often despite rejection, and treating startups as empirical exercises with rapid hypothesis-testing cycles. Founders should set ambitious goals and overwhelm their biggest bottleneck every two weeks. Distribution is highlighted as increasingly critical—software is no longer the hard part—and early-stage founders should differentiate through unscalable, white-glove service rather than competing on scale. The discussion also notes that while VC funding can accelerate growth, AI may reduce capital needs, and that founders should avoid using AI as a black box, instead engaging in a self-learning cycle to continuously improve.

In this episode, the guest shares lessons from his entrepreneurial journey, including the acquisition of Wit.ai by Facebook and the challenges of founding a chatbot company in 2002 that was two decades ahead of its time. He argues that raising $1.2 billion creates immense expectations that can be more damaging than equity dilution, and that founder strengths should dictate whether to enter a market early or wait for readiness. For AI advancement, he identifies three critical bottlenecks: talent, data, and compute—noting that securing GPUs remains extremely difficult even with ample funding. He also observes a shift in the AI community, where the once-contrarian view that large language models do not lead to artificial general intelligence is gaining mainstream acceptance, validating his startup’s focus on world models. The episode covers the operational responsibilities of a CEO at a capital-intensive AI startup, including managing researchers and maintaining a culture of careful spending despite raising the largest seed round in Europe at $1.2 billion.

PhotoRoom’s founders describe how deep, repeated focus on narrower market segments—first photo, then e-commerce photo—drove successive 10x growth, a strategy they argue is not a limit on ambition but a path to it. Y Combinator transformed their vague billion-dollar dream into a concrete, tangible goal, with one founder noting their ambition shifted from $1 billion to at least $10 billion after seeing the playbook and meeting founders of multi-billion-dollar companies. During YC’s remote Demo Day, PhotoRoom grew from $0 to $1 million in revenue, becoming the fastest-growing company in the batch and reinforcing their belief that they could achieve unicorn scale. To maintain global ambition, the Paris-based company uses English as its internal language, filtering for employees comfortable operating on a world stage. The founders stress the need for co-founders to explicitly align on ambition level to avoid a disconnect that slows growth. As the company scales, they rank all projects by potential revenue impact and conduct retrospectives on failed initiatives to learn from mistakes. They also advocate building a minimum viable version (V0) of any project to test impact quickly, noting that about half of projects fail to meet expectations and that AI now allows what once took months to be done in days. PhotoRoom has reached 300 million downloads and 20 million active users, serving major clients like Amazon and Uber, and has been called the biggest YC company with a European headquarters.

Stanislas Polu, co-founder of Dust, argues that building a model-agnostic AI platform is a strategic necessity, as it decouples the product from any single AI provider and allows flexibility to switch between models as they improve. He left OpenAI after three years, giving up stock options worth more than Dust’s entire valuation at Series B, because he preferred the tangible impact of product development over research. Dust focuses on applying LLMs to the workplace, a thesis Polu believes is already transforming work for developers and will make work feel unrecognizable within two to three years. He acknowledges that building in France adds friction for fundraising and hiring but views it as a worthwhile trade-off for sovereignty. Polu argues that verticalized AI products must develop network effects to remain defensible as intelligence becomes commoditized. He notes that pricing pressure has forced Dust to shift from flat to credit-based models, as agent loops and token consumption have exploded, compressing margins. He claims AI labs like Anthropic and OpenAI capture massive margins at the token level, but believes open-source models catching up will eventually pressure those margins, while product-layer companies can still capture margin through differentiation.

Paul Copplestone, CEO of Supabase, discusses the company's rapid growth, emphasizing that money alone cannot solve fundamental startup problems like product-market fit. He shares his entrepreneurial journey, noting that Supabase was incubated as a side project while he helped a co-founder, and its spark came from migrating from Firebase to PostgreSQL, which gained traction on Hacker News. A key insight is that by late 2024, AI coding agents like Bolt and Lovable were driving 60-90% of database launches, forcing Supabase to create a unified platform for both AI agents and enterprise needs. This led to an acceleration in growth after Christmas 2024, with users rising from 6.5 million to 10 million, along with improved conversion and activation rates—all without marketing spend, as the company prefers giving free databases to developers over advertising. Copplestone also outlines Supabase's vision for self-driving databases that automate operations like patching and security, arguing that the operate stage is a harder and more valuable problem to solve than the easily replicable build stage.

James Hawkins argues that ambitious startup ideas are paradoxically easier to sell because investors are drawn to massive, unrealized upside potential rather than downside protection. He illustrates this through PostHog’s pivot from open-source product analytics to building "self-driving software" that autonomously identifies and fixes engineering problems, creating a recursive AI loop for product improvement. The company is also developing a desktop tool to capture engineer intent, aiming to outperform human product managers by synthesizing data from sales calls, meetings, and user behavior. A key theme is that the more ambitious the idea, the more critical it is to ship quickly, as ambitious projects are more likely to go wrong. Hawkins credits Y Combinator with normalizing high ambition and crushing imposter syndrome, noting that inside YC, even early-stage companies felt chaotic and underwhelming, which was liberating. He contrasts this with European culture, where declaring ambitious goals is often frowned upon. A verified claim is that YC’s main value was normalizing high ambition and boosting confidence, rather than just providing a network. Other claims—such as a computer outperforming a product manager or the importance of shipping speed for ambitious ideas—remain unverified but are central to the episode’s philosophical and psychological themes about ambition, execution, and the human side of building startups.