Following Open AI's GPT-3, startups in China, South Korea, Israel, and Germany are building general purpose AI language tools for non-English languages
Will Knight / Wired : Tweets: @willknight and @ilparone Tweets: Will Knight / @willknight : My latest story for @WIRED is about the global rush to build powerful GPT3-type language models. The programs already reflect different cultural perspectives, and they may underpin countless future AI products. https://www.wired.com/... Jarno M. Koponen / @ilparone : Large-scale language models: if you don't build it, you can't control it. “If Europe doesn't have the technical competence to build these systems, then we're relegated to being users of something from the US or China.” -@JonasAndrulis https://www.wired.com/... by @willknight #ai
Context & Ripple Effects
OpenAI's GPT-3 demonstration reframed large language models from research curiosities into the substrate for future products — and immediately exposed who controls them. Within weeks, the Beijing Academy of Artificial Intelligence, a government-sponsored lab, was working on a GPT-3-style model, while over 500 researchers rallied behind an open-source large language model explicitly designed to keep the capability out of any single company's hands.
This story extends that arc from state labs to commercial startups: builders in China, South Korea, Israel, and Germany are training general-purpose language tools for non-English markets, arguing — per Aleph's Jonas Andrulis — that without homegrown technical competence, Europe and others become mere users of US or Chinese systems.
First-order effects
- Non-English speakers gain locally tuned alternatives to English-centric models, addressing a gap researchers later confirmed when they found ChatGPT-class chatbots significantly less capable outside English (the documented capability gap).
- US model providers lose their default position in markets where startups train on Korean language and culture directly, as South Korea's competitive startup scene demonstrates (Korea's domestic-language push).
Second-order effects
- The open-source effort and state-backed labs give smaller countries a fallback path: if proprietary frontier access narrows, national projects can fork capability rather than negotiate for API terms set abroad.
- Compute demand from dozens of national-scale training runs tightens the market for advanced chips, pulling hardware supply chains into the same geopolitical contest as the models themselves.
Third-order effects
- If every major economy treats language models as infrastructure, the industry splits along access-sovereignty lines: a few frontier providers plus a layer of nationally aligned models whose values and cultural assumptions differ by design.
- Control over model weights and APIs becomes a regulatory subject in its own right, since governments that cannot build these systems cannot shape what their citizens' AI products reflect.
The trend: Frontier-model capability is diffusing from a handful of US labs toward sovereign, language-specific builds, making model access itself a matter of national strategy.