Y Combinator

Gary Tan, CEO of Y Combinator, describes how he returned to coding after a 13-year hiatus using Claude Code and AI-assisted development tools, shipping hundreds of thousands of lines of code while running YC full-time. He built Gary's List (a politically-focused blogging and research platform) and GStack (a collection of AI prompting skills), achieving what he estimates is 400x his previous coding productivity. He argues we are entering a 'personal AI' revolution analogous to the personal computer era, where individuals must choose between controlling their own AI tools or ceding that control to corporations.

Y Combinator is hosting Startup School Paris on June 29th, a one-day event bringing together builders and hackers from France and Europe. The event will feature founders from notable tech companies and provide direct access to YC partners and alumni. It is positioned as a celebration and cultivation of the next generation of European entrepreneurs.

Harshil Mathur, co-founder of Razorpay, discusses the company's journey from a side project to India's largest payments platform. He covers early struggles including a near-death experience when their bank partner pulled out, the strategic bets that drove explosive growth, and lessons on founder mode, capital efficiency, and adapting to AI.

Current GPUs are poorly optimized for agentic AI workloads, achieving only 30-40% peak utilization due to the bursty, multi-modal nature of agent execution loops. Purpose-built inference silicon designed around the agent loop itself represents a significant hardware opportunity. The speaker argues that compiler design, not just chip architecture, will be the critical differentiator for whoever builds this next.

The transcript argues that AI is transforming scientific discovery by enabling closed-loop autonomous research systems. Frontier AI models have reached PhD-level performance on scientific benchmarks and are beginning to run full design-make-test-analyze cycles in fields like drug discovery and materials science. The speaker predicts that the most impactful companies will build 'AI-native discovery engines' rather than simple research co-pilots.

The transcript argues that AI-native companies succeed by making their entire organization 'queryable' — capturing all meetings, tickets, and interactions into a unified AI layer. This transforms companies from open-loop to closed-loop systems, enabling continuous monitoring and adjustment. The speaker sees a major opportunity to build the connective infrastructure that makes this possible by default.

The transcript argues that AI has dramatically reduced software development costs, making legacy SaaS companies vulnerable to disruption. Startups are encouraged to build AI-native challengers targeting even the most entrenched enterprise software markets. The speaker frames this moment as analogous to the cloud transition that created the last generation of great software companies.

Adi Oltean, co-founder of Star Cloud, advocates for developing industrial capabilities in space, focusing on extracting raw materials and 3D printing structures on the moon. He highlights the efficiency advantages of lunar manufacturing and encourages related startups to apply to Y Combinator.

The transcript argues that AI agents represent the next trillion users of the internet and require purpose-built software infrastructure. Unlike humans, agents need machine-readable interfaces such as APIs, MCPs, and CLIs rather than visual UIs. The biggest startup opportunity lies not in building agents, but in building the software agents depend on.

The transcript discusses the significant gap in hardware iteration speed between the US and China, particularly Shenzhen's ecosystem. A few US startups are beginning to address this, but the overall infrastructure stack remains incomplete. The speaker signals strong investment interest in startups that can dramatically accelerate hardware development cycles.

YC visiting partner Francois Shaard discusses two 2025 AI papers—Hierarchical Reasoning Models (HRM) and Tiny Recursive Models (TRM)—that demonstrate recursion at inference time as a powerful alternative to simply scaling model size. A 7-million parameter TRM outperforms much larger LLMs on reasoning benchmarks like ARC Prize by leveraging recursive hidden states instead of chain-of-thought token generation. The conversation contrasts these approaches with traditional LLMs and RNNs, exploring why recursion addresses fundamental limitations in transformer reasoning.