The hottest running app has nothing to do with speed | E2303

Episode Summary

Executive Summary: The episode opens with Interval, a gamified running app that turns exercise into territorial competition on a live global map, then pivots to Verge Labs, which is using AI and uniquely curated human brain data to improve drug discovery and patient matching. The discussion emphasizes that the biggest breakthroughs come from the right data, not just bigger models, and that both consumer apps and biotech can benefit from AI-driven feedback loops.

Main Topics: Interval: gamified running as territorial competition (Priority: 5/5): Louis Phillips explains Interval as a running app where users claim map territory by finishing runs, creating personal, competitive motivation through stolen territory notifications and local leaderboards. Retention, density, and game design challenges (Priority: 4/5): Jason and Louis discuss why the app needs density and freshness to stay engaging, and how arenas, onion-skin progression, and time-based challenges are meant to solve ephemerality and speed-related limitations. Growth strategy: organic content plus paid ads (Priority: 4/5): Louis describes how social-media-led growth, talking-head explanations, and Meta ads helped scale the app, while a third-party ads team improved predictability and unit economics. Verge Labs: AI for drug discovery and patient matching (Priority: 5/5): Alice Zhang outlines Verge’s evolution from drug-target discovery to a broader platform that predicts which patients will respond to therapies, using large human brain datasets and multimodal AI. Why brain tissue is the 'LiDAR of neuroscience' (Priority: 5/5): Alice argues that direct brain tissue is the ground truth for neurological disease, unlike proxy data such as blood or imaging, and is essential for anchoring AI world models in biology. Scaling data, not just parameters (Priority: 4/5): The conversation stresses that in bio AI, the main gains so far have come from collecting the right data and modalities rather than simply increasing model size or compute.

Key Arguments: Gamified competition increases motivation because territory loss feels personal and immediate, making users more likely to run or reclaim territory. Interval works best when there is enough local density, so the experience is strongest in cities with many active users. Freshness and replayability need new mechanics, such as arenas and daily resets, because pure territory capture can become ephemeral. Social-media-first growth can work for apps if founders are willing to create public-facing, explanatory content and 'suck publicly.' Paid Meta ads brought predictability to Interval’s acquisition funnel after relying too much on volatile organic growth. In biotech, the biggest bottleneck is not model architecture alone but access to high-quality, disease-relevant, multimodal human data. Brain tissue from deceased donors provides molecular ground truth that proxy measures cannot offer, enabling better patient-response prediction. Transformer-based multimodal models can fuse incomplete data from different patients and infer missing modalities, which traditional ML could not do as effectively. Verge’s transition from making its own drugs to selling platform insights and licensing targets is a better fit for its data advantage. The long-term value of AI in medicine is reducing drug-development failure rates, which can lower costs and enable more personalized treatment.

Data Points: Team size at Interval: 5 total, including 3 developers - Louis says the company kept the team lean while building and scaling the app. App downloads: about 1 million - Louis says organic growth and social media helped reach this scale. Instagram followers: about 100,000 - Louis cites this as part of the app’s marketing success. Cost per trial start: about $12 - Louis says Meta ads currently cost this amount for Interval. Average customer lifetime value: about $17 months - Louis describes the app’s current retention/lifetime value economics. Annual product price: about $60 per year - Jason infers the subscription price and Louis confirms it. Brain dataset size: over 12,000 human brains - Alice says Verge Labs built one of the field’s largest brain datasets. Patient count in dataset: 6,000 patients - Alice pairs the brain data with clinical and multimodal information. Tissue-bank partnerships: more than 24 - Alice describes the global sourcing infrastructure for brain samples. Model representation size: 512-dimensional vector - Alice explains how each patient is encoded in the model. Partnership upfront value: $25 million to $42 million upfront - Alice says major pharma partnerships are structured this way. Potential milestone value: up to $700 million to $800 million each - She describes total deal value for large partnerships. Lilly target validation rate: 83% - Alice says targets from the Lilly partnership validated in wet-lab experiments. Lilly internalization: 2 targets - Alice says Lilly optioned two AI-derived targets into its ALS pipeline. Cost to develop a drug: about $5 billion on average - Alice uses this figure to explain why better prediction matters. Failure rate in drug development: 9 out of 10 attempts fail - Alice cites this as the main source of R&D cost inflation.

Pivotal Quotes: "You're taking the competitive spirit. You're taking the slot machine nature of apps and smartphones, and you're using it for good." — Jason: Jason summarizes the appeal of Interval as constructive gamification. "We created Interval, which is a gamified running app. You run around the block and you claim territory on a live global map." — Louis Phillips: Louis gives the core product description early in the interview. "Brain tissue is like the LiDAR of neuroscience." — Alice Zhang: Alice explains why direct brain data is essential for accurate disease modeling.

Implications: The episode suggests that the next wave of consumer apps and healthcare AI will come from combining strong incentives with high-quality data. For startups, lean teams and smart acquisition can still scale. For biotech, the winning moat is data infrastructure, not just model size.