Latest Interviews

We’re still in the early innings with AI.

YC-backed ventures like DoorDash, Instacart, and Uber exemplify smartphone-driven success stories that launched roughly four years after the iPhone's release. These examples demonstrate that determining which ideas will succeed and where value will accumulate requires a significant period of maturation before market dominance can be achieved.

Gmail creator Paul Buchheit on the very first version of Google’s “Did you mean?” feature

Paul Buchheit

A former Google engineer designed the original "Did you mean" spell correction feature after analyzing query logs that revealed roughly one-third of searches contained errors. Although the initial implementation used a standard spell correction library, it frequently produced incorrect suggestions, such as recommending the fish "Turbot" for "TurboTax." These early accuracy failures highlighted the complexity of auto-correction and drove subsequent refinements to the system's reliability.

Sales pre-PMF should be done by the founders.

Early-stage B2B startup founders must personally drive sales efforts prior to achieving product-market fit to validate their business models effectively. Because this pre-PMF selling requires entrepreneurial vision and a tight feedback loop with product development, external sales hires often fail to replicate these critical dynamics. Relying on external sales teams before validation is identified as a suboptimal strategy that risks the venture's ability to succeed later.

When you’re pre-product market fit, sales is a job for the founders.

The event asserts that early-stage B2B founders possess the unique capability to sell their products before achieving product-market fit, rendering external sales hires premature and ineffective. Success in this phase requires founders to leverage their vision for high-volume experimentation and establish tight feedback loops between sales efforts and product development. This distinction highlights pre-PMF sales as an exclusively entrepreneurial role, fundamentally different from the specialized functions assumed after a company validates its market.

The artistry is actually in that interface between the human and the technology itself.

The speaker argues that while artificial intelligence will significantly assist with backend development and API writing, it cannot fully automate the creative interface between humans and technology where true software artistry resides. By leveraging strongly typed systems and natural language to translate requirements into specifications, AI serves as a powerful tool that augments human productivity rather than replacing it. Ultimately, the event emphasizes that defining the fundamental scope of products remains a critical human responsibility that transcends mere code generation.

You should probably still learn how to code.

A recent analysis of Large Language Models reveals that their logical development stems from processing code repositories like GitHub, providing empirical proof for the long-suspected link between programming and increased human intelligence. While the ultimate goal of computing is to eventually eliminate the need for manual coding, experts argue that learning to write code remains essential today because the cognitive challenge of programming itself enhances intellectual capacity. This emerging data transforms a historical practitioner belief into a validated conclusion, urging continued education in coding despite advancements in AI-generated software.

Better models, better startups.

B2B companies are leveraging advanced AI models to achieve productivity scales where a single employee performs the equivalent work of ten, while simultaneously upselling premium features to drive year-over-year revenue growth. End-users focus on functional utility rather than model architecture, allowing vendors to incrementally charge for enhanced capabilities as technology improves. This strategy has proven highly effective, with a Y Combinator cohort demonstrating rapid expansion from $6 million to over $30 million in annual revenue within a single batch cycle.

It only really matters if you can find a handful of users that use your product habitually.

Early-stage founders are advised to prioritize identifying users who habitually integrate their product into daily workflows over pursuing vanity metrics like total signups. Success is defined by the retention of a core user base that returns repeatedly, rendering other performance data insignificant without this foundational evidence of habitual usage. Consequently, high acquisition numbers remain secondary unless they demonstrate tangible integration into the users' regular routines.

RFS: AI to build enterprise software

Enterprise software development is undergoing a paradigm shift as AI integration replaces costly manual customization with a single, dynamically standardized codebase. This technological transition threatens incumbent firms by rendering their reliance on large sales teams and unique engineering solutions uncompetitive. Consequently, a new hiring initiative targets individual contributors eager to leverage AI for writing enterprise software in the sector's most lucrative market.

Probably not.

Founders of free consumer apps are advised to avoid paid user acquisition because it typically generates only temporary metric spikes while failing to address underlying growth limitations. Sustainable scaling instead relies on discovering non-obvious, low-cost distribution channels that bypass the need for expensive marketing campaigns. Paid spending is reserved exclusively for strictly bounded experiments with predefined learning objectives, serving as a tool for insight rather than a strategy for long-term revenue generation.

Focusing on data quality over quantity, Metalware built a foundation model with less compute.

Metalware, a hardware design firm without in-house PhD expertise, launched a project to build a specialized AI co-pilot by curating high-quality textbook figures and hardware data rather than relying on volume. By leveraging this refined dataset, the company successfully trained a functional model using the significantly smaller GPT-2.5 architecture instead of resource-intensive alternatives like GPT-4. This strategic pivot demonstrated that constrained tasks with superior data quality can effectively replace the need for massive computational resources while delivering practical results.

To be able to create something different, you have to be somewhat contrarian.

The event posits that the unlikable traits of founders, such as confrontational candor and an innate compulsion to fix broken systems, are essential contrarian mindsets required to disrupt the status quo. While this critical behavior often leads to conflict with authority and creates difficulties as an employee, it serves as a prerequisite for identifying operational failures and driving innovation. Ultimately, the presentation argues that the same dissatisfaction with inefficiency which marks a "shitty employee" is the defining characteristic of a successful founder.

Find the one narrow thing that you can do well, and carve out something that's great.

The presentation outlines a strategy where founders dominate specific, narrow verticals to achieve organic revenue targets of $10 to $20 million before pursuing broader expansion. This approach prioritizes generating $100 million annually through cash flow over seeking immediate external capital, allowing founders to maintain significant equity ownership. By avoiding one-size-fits-all products and focusing on distinct sub-segments, companies secure a stable foundation that creates strategic options for future scaling.

RFS: Machine learning to simulate the physical world

Software tools relying on physics-based models for complex tasks like weather prediction and drug discovery face severe computational bottlenecks that require supercomputers and weeks of processing time. AI models function as efficient general function approximators that solve these same physics problems in seconds on standard hardware, eliminating the need for massive computational resources. This initiative targets founders seeking to enter markets that were previously infeasible by leveraging this technology to reduce prediction latency from days to seconds.

RFS: Better Enterprise glue

This event addresses the inefficiency of enterprise custom integration code, dubbed "dark matter," by demonstrating how large language models can automatically generate and maintain specific "glue" code for unique business systems. The discussion highlights how LLMs can disrupt the multi-billion dollar ecosystem built by vendors like Oracle, Salesforce, and NetSuite to support these manual development processes. Additionally, the presentation serves as a recruitment signal, explicitly inviting skilled professionals to apply for roles dedicated to solving this automation challenge.