The Frontier AI Race is Over: Networks of Smaller Models Beat Centralized AI on Cost and Capability
Andrew Trask argues that centralized AI companies — Fable, Mythos, GPT, Opus — have permanently lost the capability frontier. Using routed/weighted ensembles of cheaper models, anyone can now exceed the accuracy of any single frontier model at lower cost and higher speed.
Key Findings from the Article
- Capability: A differentially private combination of frontier models on Humanity's Last Exam reached the low 50s — higher than any single model. The article shows a chart where an ensemble of GPT and Opus outperforms Fable/Mythos at half the price.
- Speed: OpenRouter's independent speed ratings show open-source models are faster because hosting providers compete on latency.
- Cost: The cheapest way to get Fable/Mythos-level performance is no longer those models — it's an ensemble of GPT-5.5 + Opus + Kimi K2.7, which dropped today and beats Fable on benchmarks.
The Playbook
- Take any frontier AI model (e.g., Fable).
- Find the next-best cheaper frontier model (e.g., Opus or GPT-5.5).
- Ensemble it with a leading open-source model (e.g., Kimi K2.7) and a router.
- Result: a cheaper, more capable system — which can be recursed forever.
Why Centralized AI Cannot Respond: The Hydra Effect
Trask compares centralized AI to 1960s mainframes. Once the internet linked mainframes together, the network was always stronger. Similarly, once you can ensemble any combination of models, no single model can ever catch up — each improvement in a single model only feeds the ensemble.
The article explicitly states: "No single frontier AI system will ever achieve the capability frontier ever again because of how the scaling laws/ensembles work." It predicts the future is 'network-source AI' — networks of neural networks, analogous to the PC+Internet era.
📖 Read the full source: HN AI Agents
👀 See Also
Claude Code v2.1.89 adds deferrable hooks, permission retry, and fixes memory leaks
Claude Code v2.1.89 introduces a 'defer' permission decision for PreToolUse hooks, adds a PermissionDenied hook with retry capability, and fixes critical issues including memory leaks with large JSON inputs and StructuredOutput schema cache failures.
Meta to capture employee mouse movements and keystrokes for AI training
Meta plans to start capturing employee mouse movements and keystrokes for AI training data, according to a Reuters report. The article has generated discussion on Hacker News with 33 points and 7 comments.
LLMs Favor Their Own Outputs in Hiring: 23%–60% Higher Shortlist Rates for AI-Refined Resumes
Large-scale experiment shows LLM-based resume screeners prefer AI-generated resumes 67%–82% of the time, leading to 23%–60% higher shortlist rates for candidates using the same model.
Neuromorphic Ising Machine on FPGA Solves Hard Combinatorial Problems
A neuromorphic Ising machine implemented on an FPGA uses quantum tunneling physics and brain-inspired architecture to solve combinatorial optimization problems like protein folding.