Interview, Fireside Chat
Can Open Models Keep Up With OpenAI and Anthropic? | Nathan Lambert
- Chinese labs are projected to lead in model building within the next couple of years, though progress will be constrained by a significant compute deficit and limited ability to match US frontier capabilities.
- Distillation is expected to serve as a primary lever for Chinese labs to overcome compute shortages, potentially providing a few months of advantage, though technical mapping to current reasoning capabilities remains difficult due to a lack of public literature.
- The performance gap between open source and frontier models is predicted to narrow significantly, with the lag potentially shrinking to just two to three months.
- Frontier laboratories intend to restrict access to their best models and APIs via first-party services or enterprise agreements to impede distillation efforts.
- Open source companies will likely face higher capital requirements for pre-training than frontier labs but can achieve compounding gains through post-training and fine-tuning.
- Inference costs are projected to drop multiplicatively, potentially reaching 1% of current pricing levels in approximately four years, driven by efficiency gains and architectural improvements.
- Compute supply, specifically chips and power, is identified as a bottleneck preventing exponential scaling of frontier capabilities until 2028.
- The open source economy is expected to grow more slowly than the closed model economy due to a lack of direct revenue feedback loops and the necessity of internal expertise to fine-tune models for specific ROI.
- Frontier labs will likely continue to prioritize cost efficiency per watt and score to capture the highest demand markets, while open models target specialized niches and private enterprise data.
- The market for AI-powered coding is forecasted to expand from zero to $100 billion, becoming a primary revenue stream for closed labs.
- Security protocols at Chinese labs are expected to tighten due to anticipated government responses to U.S. activities, whereas information leakage regarding self-improving systems (SSI) and breakthroughs is expected to occur rapidly through informal channels.
- Public cybersecurity preparation is predicted to remain non-existent relative to AI capability growth, creating a significant societal risk gap.
- Data acquisition strategies will shift toward purchasing higher quality datasets and RL environments, though the current industry phase of low quality will persist for years before achieving higher ROI.
- The "AI capability plot" is not expected to result in a fast takeoff scenario immediately, as human bottlenecks and linear performance gains relative to exponential compute inputs are anticipated to prevent runaway scenarios.
- Open source efforts, such as the Trillium project, aim to demystify recursive self-improvement (RSI) by releasing full stacks of tools and tasks to help policymakers and researchers understand the process.
- A surge in mergers and acquisitions is expected in the near future, correcting a delay from predictions made two years ago, while open source contributors driven by non-profit motives will continue to provide infrastructure support.
- Frontier labs will likely need to raise significant capital or go public to sustain their financial flywheels, as the commoditization of capabilities poses a threat to their leadership.
- Mathematical and coding domains are expected to see the most dramatic efficiency gains and cost reductions due to verifiability, whereas knowledge work in fields like legal and healthcare will automate more slowly.
- The post-training space is expected to remain a "wild west" with competing forces between specialized and general reasoning models, while agent swarm coordination systems will be defined by closed labs before being adopted by open sources.
- Chinese labs are predicted to continue building high-quality models, with 20 to 30 companies currently active in the region, relying on practical commercialization strategies and open weight models to compete globally.
- Organizational efficiency and structural resets will become critical differentiators for model building success, as the cycle of consolidation and M&A has been delayed.
- The autoregressive regime is expected to sustain progress for the foreseeable future, though the field would need to update expectations if fundamentally different architectures are discovered.
- Inference scaling laws and reasoning model breakthroughs are viewed as peak moments for OpenAI that will continue to drive frontier advancements.
- The "Jevons paradox" may influence AI economics, where cheaper models lead to higher margins or increased market fill, though the specific outcome remains uncertain.