Y Combinator
All the world is changing around technology and you may contribute a line of code. What will yours be? Subscribe for startup ...
François Chollet: ARC-AGI-3, Beyond Deep Learning & A New Approach To ML
Francois Chollet discusses NDIA and ARK’s approach to AGI, arguing for a shift from purely deep learning toward symbolic and agentic paradigms that enable verifiable, data-efficient learning. He traces the history of benchmarks (ARC/ARK), emphasizes the importance of compounding foundations, and envisions ARK4/ARK5 as continual-learning, minimally supervised systems that ultimately automate complex tasks with human-level efficiency.
How AI Is Unlocking Millions Of New Builders
Emergent’s founders recount building a platform for AI-powered coding agents, detailing their shift from a heavy engineering focus to enabling nontechnical users to create production apps. They discuss the evolution of their tech stack, verification, and infra, lessons from rapid growth, and strategies for distribution, productization, and hiring across a global team. The conversation also covers market dynamics between first- and second-mover advantages, the rise of agent autonomy, and tangible user examples (e.g., toolmakers, CRM-like apps, and domain-specific solutions) that illustrate the platform’s potential impact.
The AI Agent Economy Is Here
The video dives into a surge of AI agent capabilities, highlighting OpenClaw and the Moltbook ecosystem as examples of agents interacting autonomously and reshaping how developers and CEOs work. It discusses the shift from human-driven tool selection to agent-driven economies, the rapid rise of developer tools and documentation tailored for agents, and the looming need to rethink startup culture, YC parameters, and the potential for agent-enabled services (like email, databases, and restaurant reservations) to redefine work and value.
The Powerful Alternative To Fine-Tuning
Ian discusses the rapid evolution of AI, introducing Poetic and its recursive self-improving meta-system that builds high-level automation (harnesses) on top of language models to solve hard problems more cheaply and quickly. He contrasts this with traditional model-centric approaches, shares early results (e.g., cost efficiency, performance gains), and invites startups to explore using Poetic to put AI “on stilts.” Topics/Themes: recursive self-improvement, harnesses and meta-systems, cost/efficiency, importance of data and prompts, model landscape (frontier models vs. fine-tuning), and advice for engineers and startups entering applied AI. Key Points: - AI progress is accelerating; practice with AI every day. - Poetic aims to automate the creation of problem-solving systems that outperform underlying LLMs. - Recursive self-improvement can be faster and cheaper than other approaches. - Frontline models (Gemini, Claude, OpenAI) are competitors in the sense of being baselines; Poetic’s approach leverages higher-level orchestration to exceed them without full retraining. - The current cycle emphasizes cost-effective optimization (e.g., under $100k for a run) and automated tuning rather than large-scale retraining. - Poetic sees value in building scalable “stilt” systems that let any startup deploy capable agents without needing to reinvent the wheel. - The company has demonstrated strong results on tasks like humanity’s last exam and internal benchmarks, suggesting practical benefits for real-world problems. - There is an emphasis on not just better prompts but on code-driven, automated strategies (the poetry/meta-system) to generate robust reasoning and context management. - For startups, there’s an invitation to sign up for early access and to share problems to tailor Poetic’s capabilities. - The broader arc contrasts RL-based and data-centric paths with a new paradigm that relies on automated system generation and meta-level optimization.
Common Mistakes With Vibe Coded Websites
A design-review-style exploration of AI-generated and AI-assisted websites, weighing the good, bad, and the messy in current AI design trends. The discussion highlights common patterns (purple gradients, hover reveals, animated scrolls), the tension between speed and originality, and practical guidance on maintaining clear hierarchy, branding, and user experience while using AI as a tool rather than a replacement for thoughtful design and thorough QA.
The Future Of Brain-Computer Interfaces
The conversation with Max Hodak covers the rapid, non-insensitive progress in brain-computer interfaces, highlighting retinal prostheses as a near-term, high-impact example and exploring broader future possibilities like biohybrid systems and high-bandwidth brain-to-brain communication. Hodak discusses current clinical trials, technical hurdles (surgery, safety, signal decoding), and the evolving landscape where AI, neuroscience, and consumer tech converge, emphasizing that medicine may shift from single-product solutions to multi-application platforms and that society must grapple with ethics, accessibility, and long-term implications of enhanced human capabilities.
Inside The Startup Reinventing The $6 Trillion Chemical Manufacturing Industry
The video traces Solugen’s journey from a PVC-pipe hydrogen peroxide prototype to a billion-dollar, US-based chemical manufacturer powered by biology-driven chemistry. It explains their chematic processing, which merges enzymes with metal catalysts to create cleaner, safer, scalable production, and highlights their startup approach—low-cost, customer-focused, YC-backed, and relentlessly iterative—leading to multi-site manufacturing and plans for diverse assets in the future.
Get daily AI recaps from
Y Combinator in your inbox
Get AI-powered summaries delivered to your inbox. Save hours every week while staying fully informed.