Bankers and (AI) Models

A unique characteristic of the AI business and technical revolution we’re living through is its dizzying evolutionary speed. Everything is better, faster, and cheaper, not just by millimeters, but by leaps and bounds. Successive phases of any new advanced technology used to happen over years. Pick up a popular computing magazine any year from, say, 1986 through 2010, and you’d read something like: “in the past eighteen months, the technology leapfrogged an entire generation.” At the time, that was impressive, understandably so. Things were happening fast, and we liked that.

In 2026, though, transformations and changes of that scale happen every couple of weeks, and this is making people nervous. When I started reading about AI again last year, I was aware that much of what I’d learned about it in the Nineties was not only superseded, but in some cases outright reversed. Now, I and everyone else are dealing with the fact that so much of what we learned about it up through April 2026 has not only been superseded, but in some cases has been outright reversed.

Up through this spring, and to a lesser degree even now, the business model, the flat assumption was that AI would be largely a subscription service, following the pattern of phones and pay television. The major companies supply the service at a loss, aiming to build market share, knowing that if AI does become the next cable or cell phone habit, they’ll eventually make the hefty investment worth it. That’s why they need the data centers that have become so controversial.

But what if they won't be needed?

Imagine if, in the early years of the auto industry, as startup car companies rushed to build huge assembly plants all over the Midwest, someone made an invention or discovery that somehow allowed us to homebrew the cars without building the factories. It would have been massively disruptive.

Since the spring, open-source local AI models have become the rage, especially Chinese ones. It’s not just the hobbyist stuff I enjoy: small businesses have been using them to minimize paid AI use. This will take some business away from central, data center-based cloud AI. How much? No one really knows yet. Put it this way: Cloud AI is better and always will be ahead of local. But right now, at home, you can securely and privately run the equals of the finest cloud AI in the world only a year ago.

In The New York Times, Zeynep Tufekci had a good column, noting that the kind of AI almost everyone knows is connectionist, not rules-based, so it’s not possible to be absolutely sure what’s happening inside the software. We’ve all seen articles and TV segments about “jailbreaking,” using clever or deceptive prompts to get a program to overrun its so-called safety guardrails. One technique is overloading its context window, effectively degrading its ability to remember all of the rules prohibiting certain matters from being discussed, like explosives or neurotoxins. Another is familiar to Star Trek fans: using paradoxes to paralyze a system with confusion.

For that matter, we’ve all seen movies like The Forbin Project and 2001: A Space Odyssey. It’s a serious concern, even if it’s easy to exaggerate. But even if AI doesn’t want to take over the world, bad human decisions regarding trillion-dollar investments can crash Wall Street. The smartest futures are hard to predict. Harder than it seems.

What’s one of the flashiest quality challenges for AI? “Cinematic level” video generation. OpenAI pulled its release of Sora partly because of Hollywood concerns that it was too good, too capable of mimicking human actors.

As we’ve noted here, China has fewer inhibitions, and its cloud AI Seedance series generates realistic pictures for very little money. How much money? About 23 cents per second. For a 100-second-long trailer, $23. At that price, you’re getting full motion and decent 720-line HD quality, nowhere near theatrical level. Still, it’s amazing—100 minutes at that intermediate level would cost all of $1380. That’s not exactly big money. If anyone can do it—and now, anyone can—films might have to distinguish themselves on, gasp, writing. A feature film, of sorts, for the cost of a moped. That’s a different media world.

Of course, more money spent in the cloud will yield better quality AI. I had a post called Screwtape’s Screening Room to ironically “admire” and point out the fine and devious workmanship of successful propaganda. In the past year, one such team has emerged in Iran, Explosive Media, makers of very popular anti-American, antisemitic web videos lampooning Trump and Netanyahu as sad-sack Lego characters. They’ve had hundreds of millions of views. Their methods are simple, using very ordinary tools of cloud AI. If the Western internet shuts off, I’m sure they could run DeepSeek or Alibaba AI. But to my knowledge, they don’t run local.

A considerably more salubrious 2026 attempt at political influencing via AI video generation was considerably closer to home for me: Spencer Pratt’s audacious web video campaign for mayor of Los Angeles. He came out of nowhere and charmed much of the country (though, alas, not enough of the city). With no slight to Spencer’s overall instincts and guidance, the video was largely the product of Charles Curran of Maniac Studios. He used some Chinese software, but apparently it was mostly OpenAI and Anthropic, along with a host of L.A. boutique “post houses.” It empowered a vastly underfunded campaign.

We should be clear: Most of that benefit comes from cloud AI, which anyone can access. If you’re loaded with cash, sure, get in line to buy yourself a local rack of Nvidia’s Grace Blackwell GPUs. But Elon can back up his inferences with something like 250,000 top-level Nvidia cards. That’s tough to compete with (although OpenAI/Microsoft, Gemini/Deep Mind, Anthropic, and Meta grimly do so anyway). After a tumultuous year for corporate AI, there’s a bit of a price war going on. These guys are supplying the most advanced research in the world, and at a loss for the time being.

Maybe financial critics of the industry are right; this can’t go on forever. As customers, though, why not take full advantage of it now? As tycoon D.D. Harriman said in Robert A. Heinlein’s The Man Who Sold the Moon”, “It’s raining soup—get a bucket”.

So with that understanding, what’s having a tame AI at home worth, and what does it cost? I’m reading more about systems and configurations I don’t need or covet—well, okay, “covet” maybe but fundamentally don’t need—but do find interesting nonetheless. The tokens are “free," the equipment and electricity you still pay for. Enthusiasts in any field do tend to overspend. But you don’t have to. Of course, it depends on your needs. I’ve seen amazing video made with a $1300 Intel Arc 32gb graphics card.

Step up to the plate, MiniMax H3. It can do much of what current Belle-of-the-Ball Seedance 2.5 can, but it runs locally on your own machine. Best of all, it doesn’t just run locally. It can even run on consumer equipment, like my 16gb Mac. No elaborate graphics channel needed. There are, of course, quality tradeoffs.

On the other hand, create a flawless, 4K theatrical quality 30-second clip of the streets outside of the Empire State Building in 1944? Prompting even the virtual “extras” down to each detail as if they were human actors? Now you’re definitely talking cloud. This calls for SuperGrok Imagine or Google Veo. You’re on, AI Kling. Feel free to do yo’ thing. Because, local AI or cloud, when it comes to potential material to play with as images... oy, have I got material. I bet you do too.