Ex-OpenAI Staffer Starts RL Data Firm, Calls Lab Valuations Delusional
AI寒武纪 · wechat · 2026-07-30
After an eight-month stint at OpenAI, Andrew Ho has left to launch a new company dedicated to producing high-quality reinforcement learning (RL) datasets. Drawing from his internal experience, he pointed out that current LLMs have terrible generalization and are severely uneven in their capabilities, lacking true code universality despite massive industry investments.
Startup Focus & Data Bottlenecks
- He argued that designing RL datasets is an art form, and most data vendors lack an understanding of large-scale training realities. Demand for targeted, high-quality data from major labs will explode, potentially requiring over $100 billion in acquisitions.
- The startup's first products focus on biology and statistical reasoning, aiming to solve datasets for long-term scientific reasoning. His research shows frontier models struggle with complex data analysis (e.g., scoring only around 30% on GeneBench-Pro).
- The second product targets scientists' daily workflows, such as querying models with images. He emphasized that current multimodal models are still inadequate for these tasks, requiring heavy investment to build foundational capabilities.
Critique of Frontier Lab Valuations
- He is extremely bearish on frontier AI lab valuations, calling primary market enthusiasm irrational and predicting a brutal reality check during IPOs.
- Labs are burning massive cash. Supporting a trillion-dollar valuation would require halting R&D and netting hundreds of billions annually, which is unrealistic. Facing cheap competitors like Qwen or Kimi, leaders are forced to endlessly fund next-gen model training.
- He heavily doubts the assumption that scaling models will trigger autonomous AI-driven economic booms, insisting AI progress will be slow and bottlenecked by data. Even if capabilities were frozen today, fully integrating LLMs into human life would take 20 years. He explicitly stated he would not buy OpenAI or Anthropic stock at their current valuations.
Related event: Ex-OpenAI Staffer Leaves to Build High-Quality RL Data Firm(3 posts)→
More from AGI Musings
- Over 1,000 AI Researchers Call for Brakes on Superintelligence — anderssandberg · 2026-07-31
- KOL Notes Claude's Narrative Bias, Calls Grok a Necessary Counterbalance — jamesdouma · 2026-07-31
- Peter Diamandis: We Are One AI-Discovered Material Away From Changing Civilization — PeterDiamandis · 2026-07-31
- Apple Hits $5T Market Cap as AI Spending Narrative Flips — asymco · 2026-07-31
- Analyst: AI Fundamentals Unchanged Amidst Severe Compute Deficit and Low Penetration — BenBajarin · 2026-07-31
- Measuring Intelligence Per Watt Reveals Lack of Frontier Pricing Moat — ajratner · 2026-07-31