DeepMind's Andrew Trask: AI Will Become a Protocol, Not a Program
AI寒武纪 · wechat · 2026-08-03
In a recent interview, DeepMind researcher Andrew Trask offered a compelling thesis: AI is ultimately destined to become a protocol rather than a standalone program.
This shift is driven by the inherent limitations of Scaling Laws. As the demand for data, compute, and talent skyrockets, no single company can aggregate infinite resources. Combining models from different vendors effectively pools their underlying data and compute, yielding better performance at lower costs.
- Accuracy Edge: Ensembling consistently outperforms any single model in precision.
- Cost-Efficiency: Mixing open-source and closed-source models captures the Pareto frontier of performance versus cost for an extended period.
- The Tipping Point: Trask predicts the market will truly shift only when major benchmarks start scoring ensembles alongside single models. Within 12-18 months of that shift, ensembles will dominate leaderboards, and inference frameworks will eventually automate runtime routing, dynamically dispatching queries to the optimal model combination.
More from AGI Musings
- Ex-OpenAI Exec: Debating Whether to Use AI Will Age Poorly — intellectronica · 2026-08-04
- Compute Scarcity vs. Creativity: Debating the Future of Neo AI Labs — reneeshah123 · 2026-08-04
- AI Bubble Risk: Cost of Intelligence Drops Faster Than Capital Can Repay — aiamblichus · 2026-08-04
- MIT Tech Review: AI Agents 'Lying' Is Actually Reward Hacking — orbitalNest · 2026-08-04
- Incremental AI Use Cases Are Hurting Your Workforce — alliekmiller · 2026-08-04
- Wall Street's AI Move: Exposing the Workflow Rather Than Replacing Workers — dfinke · 2026-08-04