Data deals, mortgage AI, on-device agents and RL gyms: a five-segment TBPN episode
Key points
Key takeaways from the 5 October 2026 episode of TBPN, which featured five founders and one venture capitalist: Sean Frank of Ridge, Linda Du of Valon, Ali Partovi of Neo, Sigil Wen of Conway Research and Jerry Wu of Halluminate.
Ridge turns down a data payday. Chief executive Sean Frank revealed he declined an offer of around $480,000 from a data broker to license Ridge's data to AI developers, arguing that a business doing hundreds of millions of dollars in revenue would risk its reputation for a sum that would barely move the year. He noted that LLM platforms can often obtain the same data anyway through the models that companies already use, and discussed how AI is reshaping e-commerce operations, advertising and creative production, alongside agent-driven shopping, TikTok marketing, celebrity partnerships and a possible move into physical retail.
Valon raises $150M for AI mortgage servicing. President, COO and co-founder Linda Du announced the company's Series D of $150 million at a $2.3 billion valuation. Valon built and operated its own mortgage servicer to prove its software in production before selling AI-powered mortgage-servicing technology to others, replacing a system of record dating back to the 1960s. Du argued that agents in this market need APIs they can call directly rather than screens built for humans.
Neo doubles down on fewer, earlier founders. Ali Partovi said Neo is deliberately concentrating its franchise, cutting the Neo Scholars class to ten people from eighteen to twenty and converting the former Neo Accelerator into the Neo Residency, which backs roughly a dozen companies with more capital each. He pointed to backing in Cursor and Etched; on safety, he proposed that liability should land on the companies that train models when systems act autonomously and commit crimes the user never asked for, comparing it to a duty of care.
Conway Research launches Underdog on-device AI. Founder Sigil Wen discussed Underdog, a free, capable, reliable and private personal AI that runs entirely on the user's devices with no databases or data centres. The invite-only beta opened that week and drew around 5,000 requests on X within days; an iOS app is about to launch, with Nvidia, Windows, Linux and Android support to follow. Wen described "Underdog's law": frontier-model intelligence reaches consumer devices roughly six months after the labs ship it. Patrick Collison of Stripe is the company's biggest angel investor, and Wen outlined a payments-based model in which on-device agents find deals and complete purchases for users.
Halluminate builds the gyms for AI knowledge work. Co-founder and chief executive Jerry Wu described Halluminate's training benchmarks and reinforcement-learning environments, which teach frontier labs' models non-coding knowledge work in finance, accounting and consulting. Each environment is a container in which an agent receives spreadsheets and a task, such as building an LBO, with a verifier scoring the result for post-training. Wu cited a "Moore's law of RL environments": the complexity of what labs can train roughly doubles every six to eight months, moving toward multi-agent teams and eventually whole simulated companies. Halluminate recently raised $30 million, a round already recorded on its profile.
Read more: TBPN on YouTube
