Lenny's Podcast — searchable YouTube transcripts & AI Q&A

What it actually takes to build frontier models | Tulsee Doshi (Head of Product for Gemini)

Tulsee Doshi argues that frontier model development is unpredictable, comparing it to Snape's potions class where the outcome includes surprising 'magic' beyond the ingredients. She advises product managers to build features that are 60% functional today, anticipating that models will improve to support the remaining 40%, and to focus on understanding the 'vibes' of models through hands-on testing and small-scale evals. She also emphasizes the importance of monitoring models for safety and leveraging AI for cyber defense.

A UGC masterclass from the 22-year-old who built Wispr Flow’s growth machine | Hector Serrano

Hector Serrano argues that the most effective paid UGC strategy for B2B SaaS is performance-based compensation, where creators earn a percentage of ad spend or GMV, rather than flat fees. He explains that this model, combined with a tight-knit creator community and open-ended creative freedom, leads to higher-quality ads and better results, citing examples like creators earning $13,000 a month and a viral ad featuring a keyboard in the rain. He also emphasizes the importance of hiring young, hungry creators with a proven social media following over traditional credentials.

How to find product-market fit (but, actually) | Rob Snyder (Author, The Power of Pull)

Rob Snyder argues that product-market fit comes from finding a specific person with a project on their to-do list that is unavoidable, has no good existing options, and is blocked—this is 'pull.' He recommends using the PULL framework to form a hypothesis, then testing it through sales calls where you look for customers who try to buy immediately, not just express interest. He also advises against design partners and pilots, instead selling a service-shaped version of the product early to validate pull and learn what to build.

How do AI models actually work? | The scientist looking inside Claude's mind

Language models like Claude are not understood by their creators; interpretability researchers study their internal activations to decode how they work. The speaker explains that models plan ahead, represent concepts like 'eye' across contexts, and can be manipulated, as shown by the Golden Gate Claude experiment. He argues that anthropomorphizing AI is a useful mental model, and that the field is growing but still small.

Why every company now needs to think like a lab team | Josh Woodward (VP Google Labs)

Every company should operate like a labs team to stay at the frontier of AI capabilities, argues Josh Woodward. He emphasizes tracking what's 'almost possible' and watching for when it becomes possible, then quickly prototyping around it. He also advises looking for people with a high 'unlearning rate' and creating an independent space for experimentation, while noting that labs often fail if they lack independence or are too tightly integrated into existing business units.