The Hidden Risk Inside General AI
The much-watched OpenAI boardroom drama has wrapped up, with things nearly back to where they started. ChatGPT just turned one, so it's a good moment to take stock of the AI industry a year into its explosive growth.
Let's start with the OpenAI boardroom saga — the media has already covered the details extensively, so here's a quick recap for anyone who missed it: OpenAI CEO Sam Altman was abruptly fired by the board, which accused him of not being consistently candid. In the end, he came back, backed by investor Microsoft. For a few days the story kept flip-flopping — one moment he'd successfully returned, the next the board had named a new CEO and his comeback had failed. The back-and-forth played out like farce. Of course, he ultimately did return, which makes sense given Microsoft's $13 billion sitting on the table. But the whole episode is a genuinely worrying case study in corporate governance.
Safety Meets the Uncanny: AI Inches Closer to Human-Like Responses
OpenAI's core problem traces back to its origins as a nonprofit: a nonprofit board ended up governing arguably the most important company in Silicon Valley. The company succeeded so fast that the board's caliber never caught up with the company's weight, and the result felt more like a college club's governance structure than anything befitting a firm of OpenAI's stature. That's on Altman — the hybrid structure was his attempt to balance running Silicon Valley's most important company with staying true to an AI mission, but in practice it proved extremely fragile.
This whole episode also surfaced a deeper question: the risk posed by general AI (AGI). The spark that lit the board coup was AI safety. Ilya Sutskever, OpenAI's chief scientist and a board member himself, apparently believed Altman was too cavalier about AI safety. AGI is the holy grail of artificial intelligence, but most researchers roll their eyes at the topic, because with today's technology it's still fairly far off — a bit like worrying now whether a grade-schooler will get into fights once he's in college. It's a stretch. But this time, Ilya thought the problem was serious.
As a rule, when a smart person asks a dumb-sounding question, it means the issue isn't actually simple and deserves to be taken seriously. This safety-first camp believes what's currently being built is already edging into AGI territory, and that developers need to slow down and get safety right. But it looks like they've already lost the fight to the move-fast camp.
This whole debate exists because AI is now edging closer to human-like responses in certain domains, which brings us to "emergence" — the phenomenon everyone in this AI wave keeps talking about. ChatGPT gives us a very intuitive feel for emergence: it's just predicting the next word over and over, and then suddenly you realize, wait, this thing actually has a decent brain! Discussions of emergence usually describe that gut-level sense that an unremarkable AI application has suddenly acquired a spark of something almost sentient. In engineering terms, it means that once an AI model crosses a certain scale, accuracy jumps sharply, or once compute crosses some threshold, an application suddenly becomes genuinely useful.
This is especially noticeable in speech recognition — quantitative change producing qualitative change. Of course, there are also researchers who argue in papers that emergence doesn't really exist at all, that the underlying improvement is actually linear. That counter-argument exists too.
But emergence is real on a gut level, because even if the underlying change is linear, once you cross a certain threshold, people still feel like the thing has suddenly come alive. If something can answer questions convincingly enough to pass for human, then for practical purposes, it basically is a person. So really, the whole crisis boils down to OpenAI's lead being a little too large right now (I'd love to have that kind of problem). They've bet the most on this path, so they may genuinely be capable of doing things no one else can yet. But after the chaos at OpenAI, I suspect they'll proceed a bit more cautiously from here.
So what are other AI companies doing in the shadow of OpenAI's march toward AGI? A good number of generative AI startups are pushing into more vertical niches — video, image translation, and the like. But OpenAI also has its own portfolio of companies it has invested in directly, and those get a huge edge in tapping OpenAI's resources. The other thing many generative AI companies have discovered is that building something usable is easy, but building AI that actually impresses investors still means buying a mountain of Nvidia chips and training models the hard way.
The Generative AI Arms Race — Jensen Huang Is the Real Winner
So how many chips do you actually need to buy? To match Microsoft and Facebook, you'd need around 150,000 H100-class chips. Even a database company like Oracle bought 50,000. And the price? Over NT$1 million per chip. AI chips are expensive and hard to get — a mere supply commitment has itself become something of value.
One data-center company, CoreWeave, managed to raise $2.3 billion in debt using its Nvidia H100 supply commitment as collateral. (To be fair, CoreWeave has been almost religiously loyal to Nvidia for years — back when the GPU market was in a slump, they still kept buying, and the payoff was this supply commitment.)
But that's not the only headache for generative AI startups. YouTube has already rolled out its own AI tools for creators, and ByteDance's editing tools (ByteDance owns TikTok) now come with AI built in, so the big platforms are staking out the major use cases too. Still, founders are a fearless bunch, and generative AI startups worldwide keep piling in undeterred, because being the one-in-a-thousand winner still pays off enormously. For investors, though, there's really only one sure winner in sight: Nvidia. For every dollar you put into a generative AI startup, maybe 40 cents ends up going straight to Nvidia anyway. You really have to admire Jensen Huang's business foresight.