Latest / The Emerging Markets Podcast with Fexingo: Developing Economies, Growth, and Investment / How AI Distillation Is Reshaping Emerging Market Tech
Transcript
- Lucas: So there's this term taking over every tech conference right now: distillation. It was all over the news this morning — from Silicon Valley to DC, the tech world is suddenly obsessed with it. And for emerging markets, it might be the most important AI development since the transformer. Luna: Distillation — that's where you train a smaller, cheaper model to mimic a much larger one, right? Like copying the knowledge without needing the same compute power. Lucas: Exactly. You take a huge model like GPT-4 or Gemini, run it on a bunch of prompts, capture its outputs, and then train a much smaller model on that data. The smaller model can achieve 80 to 90 percent of the performance for a fraction of the cost. Luna: And the appeal for emerging markets is pretty obvious — you don't need a fortune in GPUs to build a competitive model. Lucas: Right. And this is happening right as the export controls on advanced chips tighten. The US dollar index hit 120.5 this week, which makes it even more expensive for companies in developing economies to buy hardware. So distillation is becoming a workaround. Luna: Are we seeing real examples? I know China has been pushing distillations of their own models for a while. Lucas: Absolutely. Several Chinese labs have released distilled versions of their large models. For instance, Alibaba's Qwen team put out a distilled 7-billion-parameter model that beats some open-source alternatives on benchmarks. And it runs on a single consumer GPU. Luna: So they're basically doing more with less. That feels like a classic emerging market move — innovate under constraints. Lucas: Exactly. And it's not just China. India's seen a surge in open-source distillation projects. There's a startup called Sarvam AI that distilled a model for Hindi and other Indic languages. They used a much smaller dataset than the original model needed, because they focused on quality over quantity. Luna: But is there a catch? I mean, if you're copying the output of a model that might have biases or errors, you're inheriting those problems, right? Lucas: That's the big caveat. Distillation doesn't correct the original model's flaws. It can even amplify them if the student model overfits to certain patterns. Plus, there's a ceiling — you can't surpass the teacher model's knowledge. So for frontier applications, you still need the big iron. Luna: But for a lot of use cases in emerging markets — local language chatbots, agricultural advice, education — 80 percent of GPT-4 is plenty. Especially if it runs on a laptop. Lucas: That's the key. And the economics are compelling. Training a frontier model costs tens of millions. Distilling a decent model can cost a few hundred thousand. That opens the door to a much wider set of players. Luna: But doesn't this depend on having access to the teacher model? If the biggest tech companies keep their models behind APIs, can you really distill them? I remember some terms of service prohibit that. Lucas: They do, but enforcement is tricky. And many of the best open-source models — like Meta's Llama — are fully available for distillation. So you don't need to break any rules. Plus, some emerging market governments are funding their own distillation projects using open models. Luna: Are we seeing any pushback from the US? I mean, if distillation effectively bypasses chip export controls, that could become a geopolitical flashpoint. Lucas: It's a gray area. The export controls target hardware, not the knowledge in the model weights. So a Chinese company can legally download Llama and distill it in Shanghai. The US can't stop that. But some lawmakers are calling for restrictions on model weights themselves. That's a whole new regulatory frontier. Luna: That would be huge. If the US starts controlling the export of powerful AI models, distillation becomes even more important — it's the only way to get access. Lucas: Exactly. And that's why this moment matters. Look at the trade deficit numbers we just saw — the US current account deficit widened to over 226 billion dollars in the first quarter. The strong dollar is making US exports less competitive, but it also makes US tech products more expensive. Distillation is a way for emerging markets to decouple from that cost. Luna: So in a weird way, the same forces that squeeze emerging economies — strong dollar, trade tensions, chip controls — are also pushing them to innovate faster in AI. Lucas: Absolutely. And the venture capital flows reflect that. We're seeing more deals in Indian and Southeast Asian AI startups that focus on efficient models. The total investment in emerging market AI this year is on track to surpass last year's, even as global VC slows. Luna: Which ETFs or stocks capture this? I mean, an investor who believes in this trend — how do they play it? Lucas: Hard to buy pure distillation. But the KraneShares CSI China Internet ETF, KWEB, gives exposure to Alibaba, Baidu, and others that are building distilled models. It's trading around 26 dollars. And the broader emerging market ETF, EEM, is at 63.33. But these are broad plays. The real action is in private companies. Luna: What about the quality issue long-term? If everyone is distilling the same few teacher models, doesn't that reduce diversity in AI? We could end up with a monoculture of knowledge. Lucas: That's a real risk. If all distilled models inherit the same biases from a handful of US or Chinese base models, local perspectives get lost. Some researchers argue that true innovation requires training from scratch on local data. But distillation is faster and cheaper, so it's winning for now. Luna: I wonder if we'll see a backlash — maybe some emerging market regulators require that models trained on local data get preference in public services. Lucas: We're already seeing hints. India's AI mission is funding foundational models trained on Indian languages, not just distilled from English models. And Indonesia has talked about a national AI stack. So distillation might be a stepping stone, not the final destination. Luna: So the key takeaway is that distillation is democratizing AI for emerging markets, but it's also a temporary shortcut. Lucas: Exactly right. It's a way to catch up without buying into the most expensive hardware. But the long-term winners will be those who use distillation as a launchpad, not a crutch. And if you're following this space, keep an eye on the open-source community — that's where the next big distillation breakthrough might come from. Luna: And if walking through the economy with us has made something click for you, we'd love it if you considered supporting the show. It's listener-supported and ad-free — you can find us at buy me a coffee dot com slash fexingo. Just a small way to keep these conversations going. Lucas: Yeah, that really helps us stay independent and focused on what matters. Now, back to AI distillation — one more data point. The strong dollar means that us based cloud computing costs are rising for emerging market users. So distillation isn't just about chip access; it's about lowering the cost of inference, too. Luna: That's a good point. If inference costs remain high due to the strong dollar, smaller models become even more attractive for local deployment. Lucas: Absolutely. And we're already seeing companies like ByteDance and Tencent deploying distilled models for real-time translation and content moderation — tasks that used to require massive server farms. The shift is real. Luna: Lots to watch. Thanks for breaking it down, Lucas. Lucas: My pleasure. Until next time.