The Dawn of a New AI Era
There's a famous story about the "Gordian Knot": Alexander the Great came across a knot in a certain city that was said to be absolutely unsolvable — it had no visible end at all, harder to untangle than your iPhone earbuds — and whoever managed to undo it was destined to become the ruler of all Asia. Alexander promptly drew his sword and sliced the knot in two. Problem solved in about two seconds.
Well, swap out "Alexander the Great" for "AI" and you basically get today's mainstream approach to artificial intelligence. Strictly speaking, Alexander never actually untied the knot — and today's AI models don't have anything resembling genuine consciousness either. The GPT model taking the world by storm right now, and models like it, all share roughly the same origin story: brute-force the problem with an astronomical amount of compute. Thanks to Nvidia and the other GPU makers, even as graphics cards get more expensive individually, compute overall keeps getting cheaper — which is exactly why pulling this off keeps becoming more feasible.
These underlying algorithms didn't just appear overnight, either. What makes ChatGPT genuinely astonishing is that it built a great general-purpose interface for accessing the underlying model's output, and wrapping a conversational language bot around the front of it makes it hit completely differently. Think of it this way: monkeys typing away in some room somewhere doesn't move you, because you never see it. But watch a monkey type out Shakespeare right in front of you, and that's terrifying. A well-trained AI may already be excellent under the hood, but out of sight is out of mind. Add a chat interface — a genuinely universal input-output layer — and suddenly it really does feel like watching a monkey speak, and that's shocking.
Chasing the Holy Grail of Artificial General Intelligence
Because ChatGPT's answers feel so close to what we'd recognize as consciousness, plenty of people now see it as proof that artificial general intelligence (AGI) might someday really exist. For years there's been an ongoing debate about whether the era of AGI will ever actually arrive. AGI is basically the thing sci-fi movie protagonists are always up against — typically some machine voice that's far smarter than we are, capable of giving a correct answer to any question (and, invariably, scheming to wipe out humanity). By definition, AGI is artificial intelligence that matches or exceeds human intelligence across the board — it's the holy grail of AI research.
Along the current path, there are two competing schools of thought. One holds that the concept is hard, but the execution is easy: keep piling on more compute, keep pushing the system forward, and eventually you break through to something that looks close enough to AGI — and once we humans can no longer reliably tell the difference, you might as well just call it AGI.
The other school holds the opposite: the concept is easy, but the execution is hard. Piling on more compute is useful, sure, but what you end up with is still just simulated consciousness, with countless unresolved issues waiting downstream. So unless there's a genuinely new technical breakthrough, the current technical path will eventually hit a stage where it perpetually approaches the finish line without ever actually reaching AGI. There's a similar debate happening right now around self-driving cars, and based on the lesson there — that full self-driving keeps getting harder, not easier, the closer you get — I lean toward believing the concept-is-easy-execution-is-hard camp. We'll likely need an entirely new technical path before we can reach genuine AGI.
Two Possibilities, Both Capable of Creating New Economic Growth
But we don't need to wait for AGI at all — the AI we already have is plenty usable in the near term. It's a bit like the infrastructure showed up ahead of schedule and we're still figuring out what to actually do with it: the highway is built, but what exactly are we going to drive on it? So going forward, AI splits into two separate tracks. One belongs to the mega-players who can afford a seat at the table: pouring in astronomical compute, building models, and searching for new technical breakthroughs. The other track is about figuring out how to actually use this stuff — and that "how to use AI" question is absolutely going to be a new engine of economic growth over the next few years. Countless applications built on top of GPT will launch this year alone, and companies that put it to good use will pick up a real edge.
Beyond the AGI debate, GPT has also stirred up what might sound like an alarmist concern: the birth of a new version of the digital divide.
Based on our own company's testing with OpenAI's compute pricing, the cost of asking questions and getting answers is currently very affordable for the average person. But if AI compute eventually spreads into every tiny corner of life the way chips have, the cumulative cost of all that usage could add up to something substantial. If people without money can't afford access to enough AI compute, they'll end up like a free-to-play gamer getting stomped by a whale who's dumped a fortune into premium gear — except this time you're not just losing at the starting line, you're losing across dozens of dimensions at once. Even Buddhist hell only has 18 levels! Right now, the gap in access to compute may not be all that visible, but if a handful of killer applications emerge down the road, that gap could widen dramatically.
Take smartphones as an example. Today, the gap between the latest iPhone and some ancient Android handset is mostly just "the photos don't look as good." But once you add, say, AI compute power fused with a brain-computer interface into the mix, the gap suddenly becomes a question of "smart or not smart" — and that could be an enormous divide. It's genuinely hard to imagine right now, much the same way it was hard to imagine, back in 1995 when Bill Gates was talking about the "information superhighway," just how deeply internet access would eventually become woven into every layer of daily life. In the same way, it's hard for us today to imagine how much individuals will eventually need access to AI compute in the future.
There's no settled answer to this yet, but just as internet access eventually came to be treated as a basic right, the right to call on AI compute may well come to be seen as part of the same basic human-rights package in the future — with public institutions expected to guarantee sufficient access to AI capability. That, too, could turn into another growth area for AI applications. Either way, current AI compute still runs on a semiconductor foundation, so Taiwan is bound to keep playing a genuinely important role in global AI.