AI Evolution Mirrors Biology—Open Source Speeds Progress 1,000× (Daniel Povey)

Daniel Povey, chief speech scientist at Xiaomi and the "father of Kaldi," likens the progress of artificial intelligence to biological evolution: a continual trial‑and‑error process where each new idea must be "copied" and the speed of evolution depends on the time required for that copying.

Core Insights

AI development follows a pattern of long stagnation periods punctuated by sudden breakthroughs, mirroring the "punctuated equilibrium" observed in nature.

Open‑source projects act as the primary accelerator of AI evolution; without them, research speed would be roughly a thousand times slower because every company would need to reinvent the wheel.

Generation time for a new AI idea has shrunk from about two years to roughly six months, thanks to tools like PyTorch that enable near‑perfect reproducibility of published "formulas".

Large organizations should pursue a two‑legged strategy: continue to exploit industry‑leading models such as Transformers for current products while allocating resources to exploratory research that may uncover the next disruptive breakthrough.

Zapformer: A New Universal Sound Model

Povey introduces Zapformer, a successor to the previously released Zipformer, highlighting three major advances:

From "human voice" to "all sounds": the model expands from focusing solely on speech to handling a wide range of audio sources, becoming a universal sound foundation.

From structural tweaks to a new theory: incorporating Gradient Flow theory into the design improves speech‑recognition accuracy by 10‑15% over the already strong Zipformer baseline.

From specialized optimization to robust generalization: removing the Dropout layer and adopting the TransformAdam optimizer enhances scalability on massive datasets, delivering faster convergence and greater stability.

While Zapformer is not claimed to be the next "big hit," it embodies several interesting ideas that illustrate the benefits of open‑source collaboration and theoretical innovation.

Diversity and Balance in AI Evolution

Povey stresses that AI research must maintain a balance between "generalist" and "specialist" approaches, preserving multiple model architectures and task explorations. Just as ecosystems retain both specialist species (e.g., pandas) and adaptable generalists (e.g., mice), AI should keep diverse pathways because it is impossible to predict which will ultimately dominate.

In summary, Povey argues that open‑source is the essential catalyst for rapid AI evolution, that large firms should simultaneously exploit current state‑of‑the‑art models and invest in exploratory research, and that the Zapformer prototype demonstrates concrete gains achievable through theory‑driven, open‑source development.