Did OpenAI Steal the Navier-Stokes Solution? | E2335

What It Covers

Five VCs — Jason Calacanis, Yohi Nakajima, Ben Lear, and Rebecca Lynn — debate AI doomerism following an Anthropic researcher's viral quit thread, examine OpenAI's alleged theft of NYU mathematician Tristan Buckmaster's Navier-Stokes research, and assess Meta's Muse agent launch, capital-as-weapon startup strategies, and AI's workforce displacement timeline.

Key Questions Answered

• •Research Data Security: Never run proprietary or competitive research through frontier model APIs like ChatGPT or Claude. OpenAI's own timeline confirms it began prompting on the Navier-Stokes problem only after hearing rumors of Buckmaster's work — and AI agents now scrape preprints and partial solutions across the web, meaning even unpublished drafts can accelerate a competitor's discovery before you publish.
• •On-Premise AI Deployment: Regulated industries and IP-sensitive companies should deploy containerized, on-premise models rather than cloud-based frontier APIs. Solutions like Go.ai (formerly Abacus) ship a physical server that runs models locally, enabling HIPAA and finance compliance without sending data externally. Covenant Labs offers an alternative: encrypting open-source LLM inputs so GPU providers never see unscrambled data during inference.
• •AI Adoption Forcing Function: To accelerate team-wide AI adoption, implement a mandatory policy requiring AI as the first attempt at any problem-solving task. Calacanis reports Claude login rates at his firm climbed from 20% to 70% after enforcement, with the top two AI users winning 10 of 12 monthly performance bonuses and delivering three-to-ten times the output of non-users.
• •Hiring Filter via Paid AI Project: Screen candidates by offering a $500 paid project requiring them to build something — a dashboard, agent, or deal memo — using AI tools. Calacanis found roughly 50% of otherwise qualified candidates refused or failed to engage, immediately disqualifying them. The remaining hires showed no hesitation with AI tooling, which he treats as a non-negotiable baseline competency.
• •Frontier Model Competitive Risk: Startups building on top of frontier model APIs face a structural threat: Anthropic launched Claude Code directly competing with Cursor, its largest customer, and Claude Design competing with Figma. Founders should assume frontier labs will enter the application layer as token revenue proves insufficient against hundred-billion-dollar data center buildouts, and should migrate toward open-source or verticalized models.
• •Capital-as-Weapon Math for Acquirers: For companies valued at $2–4 trillion, acquiring ten startups at $10–30 billion each represents only 1% of market cap per deal. If one acquisition returns $300 billion — as YouTube and Instagram did for Google and Meta — the entire portfolio cost is covered. This math justifies aggressive M&A even with low hit rates, and explains valuations like Cursor at $60 billion and Scale AI at $14–15 billion.