Sequoia Capital
Showing 1–15 of 181 transcripts.
- 55 min
Parallel’s Parag Agrawal: Building a New Web for AI Agents
Parag Agrawal, Sonya Huang, Andrew Reed
Parallel, founded by ex-Twitter CEO Parag Agrawal, is building a "web systems" infrastructure that replaces human click data with direct agent feedback to optimize search indexing and ranking for software agents. The company has launched a specialized search agent product and secured a strategic partnership with Google Cloud to serve as the primary grounding provider for enterprise AI, achieving retrieval speeds of 200 milliseconds for top-relevant tokens from a trillion-page web. By shifting from traditional advertising economics to a "Shapley value" model for content attribution, Parallel aims to monetize high-quality data extraction for workflows ranging from financial modeling to autonomous agent triggering.
- 54 min
Rich Sutton and Khurram Javed: Why AI Models Stop Learning, and How to Start It Again
Rich Sutton, Khurram Javed, Sonya Huang, Alfred Lin
Rich Sutton and Oak Lab advocate for "continual learning" as the essential default for true intelligence, critiquing current Large Language Models for freezing weights and relying on finite human-curated data rather than adapting through real-world experience. To overcome the barrier of catastrophic forgetting, the lab proposes a 12-step research agenda utilizing individual step-size optimization and meta-learning to enable neural networks to continuously update their internal world models. This approach aims to create energy-efficient, self-maintaining agents capable of forming general abstractions across diverse domains, moving the field beyond static training paradigms toward systems that evolve alongside their environments.
- 22 min
Continual Learning: How AI Agents Get Better With Every Use | Arjun Karanam, Trajectory
Arjun Karanam, Ronak, Gabe, Harrison, Nico, Harvey
Trajectory, co-founded by Arjun and Ronak, addresses the lack of accumulated experience in AI by building a platform that enables models to continuously learn from the 100 trillion daily tokens generated by real-world agent interactions. The company utilizes a dual-learning architecture combining differential privacy with reinforcement learning on user-corrected traces, allowing organizations to transition from static models to systems that compound capability through automated post-training and flexible harness optimization. By abstracting complex training parameters into a 15-minute workflow, Trajectory empowers enterprises to retain ownership of their specialized models while refining agent performance directly against production traffic.
- 24 min
When to Build Your Own Agent Harness | Harrison Chase, LangChain
The framework defines autonomous agents as systems built from three owned components: the model, context, and a prioritized harness that orchestrates data flow through an iterative LLM loop. Organizations can customize this harness via middleware for domain-specific optimizations or maintain off-the-shelf versions for in-distribution tasks, ensuring compatibility through dynamic model profiles. Continuous improvement is driven by a flywheel where trace data from evaluations using the Harbor benchmark feeds into an automated engine that identifies failures and suggests prompt, code, or context fixes.
- 26 min
RL Environments Explained: How AI Agents Learn Real-World Work | Brendan Foody, Mercor
Mercore has expanded its revenue run rate to $2 billion by transitioning the AI data market from basic crowdsourcing to high-skilled "agentic data" services that enable frontier labs to build complex reinforcement learning environments. The company leverages expert networks of lawyers, engineers, and doctors to create realistic simulated worlds with precise human-verified rubrics, demonstrating a fivefold increase in model performance on specific legal tasks during recent training. As the primary data vendor for major application layer companies, Mercore addresses the industry's need for ultra-long horizon tasks and social dynamics evaluations that synthetic models cannot yet self-generate.
- 28 min
Post-Training Is How You Keep Your Taste | Fireworks CEO Lin Qiao
Fireworks CEO Linh Nguyen advocates for a strategic industry shift from relying on rented APIs to owning intelligence through deep model customization, enabling companies to preserve unique business judgment while reducing inference costs by five to ten times. This approach utilizes a structured lifecycle of data curation, fine-tuning, and serving loops to transition from generic prompting to specialized models, as demonstrated by success stories like Cursor and niche vertical leaders in healthcare and security. Ultimately, post-training is positioned as the critical mechanism for startups to scale after product-market fit by converting proprietary user data into unclonable domain expertise before high API expenses threaten unit economics.
- 29 min
How Harvey Built a Research Lab on a Budget | Gabe Pereyra
Harvey, Gabe Pereyra, Brendan, Julio, Ross, Brock
Harvey differentiates itself from well-funded frontier labs by leveraging an application-layer strategy that combines synthetic data generation guided by domain experts with post-training on open-source models. The company builds specialized legal benchmarks and utilizes infrastructure partnerships to train agents on complex tasks like contract negotiation without exposing sensitive client information. By deploying these capabilities across multiple vendors and product surfaces, Harvey aims to solve organizational productivity challenges while mitigating the performance gaps inherent in current long-context environments.
- 17 min
How Companies Are Building Their Own Intelligence | Sonya Huang, Sequoia Capital
Approximately 80 portfolio company founders and AI leaders convened to strategize the adoption of Sovereign AI, a framework defined by vertical integration where organizations own model weights rather than relying on external APIs. The event combined high-level market analysis with technical workshops led by industry experts to outline a four-step roadmap for building custom intelligence capabilities. Participants explored critical architectural decisions regarding cost efficiency, latency reduction, and the necessity of dedicated research labs to leverage open-weight models for proprietary domain performance.
- 47 min
Chai Discovery's Bitter Lesson: Drug Design Is Another Scaling Problem
Josh Meier, Matt McPartlon, Pat Grady, Sonali Singh
Chai Discovery is industrializing drug discovery by deploying a simplified, AI-driven infrastructure that replaces traditional trial-and-error with scalable molecule design, partnering with major pharmaceutical firms like Eli Lilly and Pfizer rather than managing a full internal pipeline. The company's second-generation model has achieved a 15% binding success rate through diffusion-based generation and a continuous feedback loop from wet-lab experiments, effectively targeting historically undruggable biological structures. By shifting the industry toward computational "last in class" solutions, Chai aims to accelerate development timelines from months to days while ensuring extreme safety and manufacturability at the molecular generation stage.
- 1 min
The Problem With Testing AI Architectures at Small Scale | Jerry Tworek, Core Automation
The event critiques the industry's conventional practice of validating model architectures on small datasets, arguing that this approach fails to capture the true potential of scalable systems. It specifically highlights that reinforcement learning models require a minimum threshold of compute to exhibit meaningful capabilities, suggesting that lower-scale testing often yields uninteresting results. Consequently, the organization advocates for a paradigm shift toward large-scale initial evaluations to ensure architectural research accurately reflects future performance.
- 2 min
The Most Automated AI Lab Isn't Removing Humans | Jerry Tworek, Core Automation
Aiming to become the world's most autonomous laboratory, the organization is deploying AI-driven coding agents as foundational infrastructure to drastically accelerate research iteration and data gathering. Rather than adapting legacy structures to these tools, the company is building native workflows that prioritize human agency and allow individual researchers to scale their output through automation's force-multiplying effects. This strategic pivot transforms the research environment, enabling single agents to execute significantly larger volumes of work and test ideas with unprecedented speed.
- 49 min
Building the Automated AGI Lab: Core Automation's Jerry Tworek and Rohan Anil
Jerry Tworek, Rohan Anil, Sonya Huang, Pat Grady
Founded by Jerry Liu and Rohan Ramachandran, Core Automation aims to replace static Transformer models with a new class of AI systems capable of continual, test-time learning. The organization is building an automated research lab designed to overcome architectural bottlenecks by developing hardware-efficient kernels and enabling models to autonomously optimize their own code. Success for the venture is defined by the system's ability to self-improve without human intervention, effectively extending the team's operations while bypassing the diminishing returns of current scaling methods.
- 1h 9m
The Philosopher CEO | Clay Co-Founder Kareem Amin
Clay Co-Founder Kareem Amin, Kareem Amin, Varun, Nikolai, Mike Vernal, Pete Caputa, Jack Dorsey, David Solomon, Tarek
Clay CEO Kareem guides his company through a "long strange trip" of pivoting from broad ambitions to the specific niche of Go-to-Market Engineering, where he prioritizes radical transparency and usage-based pricing to align with technical users. By rejecting standard management dogmas in favor of "momentum detective" leadership and "deep work" practices, Kareem fosters a culture of devotion that withstands pricing backlash and drives a unique product philosophy valuing flexibility over simplicity. This strategy successfully established a new market category while navigating a five-year period of ambiguity to achieve product-market fit without relying on external validation.
- 1 min
Every CIO will have to answer for every token | Factory's Matan Grinberg
The factory router introduces a dynamic model selection strategy that optimizes enterprise AI efficiency by routing distinct tasks to specialized, cost-effective models rather than relying on a single frontier solution. This framework enables CIOs to address the critical need for granular token allocation, assigning specific capabilities such as lightweight generation for "vibe coding" or custom fine-tuned models for legacy COBOL maintenance. By matching model performance to precise organizational workflows, the approach allows for complex multi-model validation pipelines while avoiding the inefficiency of blanket usage caps.
- 1 min
90% of AI tokens will be asynchronous | Matan Grinberg, Factory
The event outlines a critical shift in AI consumption from fragile, prompt-dependent synchronous models to robust asynchronous workflows where autonomous agents initiate tasks without human triggers. Industry forecasts predict that within 12 to 24 months, 90% of tokens will be generated by these autonomous systems, marking the transition from current "co-pilot" tools to fully "agent-native" operations. This evolution promises to decouple revenue growth from active human intervention by enabling systems to independently identify customer signals and execute first-pass solutions.