Teaching Everyone to Fish for Tokens

Newsletter · Watch/Read

• Nathan Lambert frames Nvidia’s $26 billion investment in “open-source recipe” models (like Nemotron, providing full training data/code) as a strategy to “teach everyone to fish for tokens,” thereby creating massive demand for their inference chips and preventing intelligence monopolization by closed-model providers.
• He distinguishes between “open-source recipe” models (full training stack, analogous to Linux) and “open-weight models” (just weights/inference code, analogous to specific software versions). Lambert posits two futures for open AI: either Nvidia’s strategy works, or open models diverge into a “long-tail ecosystem” focused on efficiency and specialization for niche enterprise tasks, which he considers the more likely outcome.
• The article notes a trend where training base models for general agentic reasoning is becoming increasingly complex and opaque, leading to less interest in full open-source training and a greater focus on post-training/finetuning existing open-weight models (e.g., DeepSeek V4 Flash) for specific tasks.
• The open-source ecosystem is in an “existential window,” dependent on Nvidia’s financing, where its economic viability must prove itself within a few years, or it risks being outpaced by the capital intensiveness of frontier model development.