Schmidt's Accidental AI Confessions

Something interesting happened recently: former Google CEO Eric Schmidt sat down for a Stanford class interview and let slip far more truth than he meant to (he claims he didn't realize he was being recorded), and the video was briefly pulled. That kind of accidental leak got AI enthusiasts everywhere buzzing, and once I took a closer look, I understood why — wow, is he really allowed to say all this?

Actually, these startup-focused courses are extremely common in Silicon Valley. The students are top performers or budding founders, and their questions are sharp — which is exactly why entrepreneurs so often end up blurting out uncomfortable truths in front of them.

Schmidt once held the most powerful CEO seat in Silicon Valley, but that was years ago; these days he's essentially semi-retired from public life. Semi-retired or not, plenty of organizations still bring him in as an advisor, so the range of topics he touches on is vast — if I tried to list everything he covered, I'd blow through my entire column's word count. So I'll just pick three points I found most interesting.

First: where are large models and open-source models headed? Schmidt believes large models will see the strong get stronger. Since he's also an investor in France's pride and joy, Mistral, anyone who uses open models knows the company. But he pointed out that because building models has gotten so expensive, there's now a real chance this open-source company ends up going closed anyway.

From Schmidt's vantage point, the industry will move toward closed-source models over the long run, for a simple reason: building models already costs an enormous amount of money, and the costs keep climbing — to a degree that's hard for us to even imagine, routinely in the tens of billions of dollars, potentially even $300 billion.

Large Models Will See the Strong Get Stronger — AI's Power Demand Is Hard to Even Estimate

Consider that Taiwan's entire annual GDP is only about $800 billion! Which brings up another issue: investment at this scale in AI comes with staggering electricity demand, which is why Schmidt half-joked that the U.S. needs to stay on good terms with Canada (for its hydroelectric power). That echoes what Taiwan's Minister of Economic Affairs Kuo Jyh-huei has said — once you factor in AI's power consumption, it becomes very hard to say whether there's enough electricity to go around.

There's also a common debate: some people cite MIT economist Daron Acemoglu's research from this past May, which argued AI's economic benefits will be limited over the next decade — and economists keep making similar arguments (not sure why economists in particular seem to have it in for AI). Schmidt had a good answer for this, drawing on the historical experience of electrifying factories:

When factories first switched from steam power to electric motors, efficiency barely improved. The reason was simple: the old factory layout was built around a single giant steam engine powering everything, and because everyone feared the drive shafts connecting to that power source would snap, all the equipment had to be clustered close to the engine. So when electric motors were introduced, people just ripped out the steam engine and dropped a motor in the same spot. The resulting factory looked almost identical to the old one.

Google Prizes Work-Life Balance — Startups Grind Around the Clock Instead

So for a good long while, these factories simply ran on an electrified version of the old steam-engine layout, and productivity gains were modest. It wasn't until roughly 30 years later that people realized the real advantage of electric motors: they can be big or small and placed anywhere — you no longer need a single central power source. That's what finally produced factories built along completely different lines, laid out in ways that made assembly lines possible and multiplied productivity many times over.

But at the outset, people struggled to imagine what electricity could actually be used for, so pouring new wine into old bottles was inevitable. That's actually a great answer to the question of AI productivity: AI is undoubtedly revolutionary, and if its benefits look limited for now, it's simply because we're still living in a pre-AI society — replacing all that old logic takes time.

One last interesting point: Schmidt also talked about how far ahead various major players are. Someone asked him why, if Google invented the Transformer, it was OpenAI that ended up transforming the industry.

He gave a subtly pointed critique: Google places a relatively high value on work-life balance, while competing startups work around the clock — Schmidt even mentioned visiting TSMC(台積電) in Taiwan and being deeply impressed by how hard its employees work.

Schmidt believes that in industries with strong network effects, time is everything, and it takes borderline-crazy ideas to win. Microsoft's agreement with OpenAI, for instance, was a wild idea — essentially outsourcing a piece of your own core competitive advantage — yet it worked.

To wrap up, here's a simple summary of Schmidt's conclusions:

1. Will large models see the strong get stronger? Definitely. Will the mainstream shift toward closed-source models? In the short term, yes.

2. Does AI have real economic benefits? Yes, and enormous ones — we just don't necessarily know how to capture them yet.

3. What kind of effort does AI require? Round-the-clock work; the winners will be a bit obsessive.

Beyond these points, Schmidt's interview is well worth reading in full.

All told, after hearing this insider leak from a semi-retired Silicon Valley veteran, it's remarkable how much he still knows — imagine how much more front-line executives like Sam Altman or Elon Musk must know. It also leaves you with the sense that the AI revolution has barely begun, and there's a lot more in store than we can currently imagine.