Can Big AI even math?

It's not only LLMs that can't do basic arithmetic

Dotson recently argued that leading AI companies are going “too big, too fast” with the development and deployment of their models, and that this is likely to backfire, much like the gargantuan German windmills of the 1980s.

It might already be happening. Despite enthusiastic adoption across nearly all sectors of the economy, a recent MIT study found that 95% of businesses’ attempts to integrate AI into have failed. A few contrarian investors are already pulling back from the industry, and the question many are starting to ask is, “How long can it last?”

Nevertheless, investment in AI is projected to reach a unprecedented levels this year, with estimates from $200 billion to $380 billion across the Big AI firms (I call them MAGAF: Microsoft, Amazon, Google, Apple, and Facebook).

Big Finance is getting skittish. Goldman Sachs has estimated that $600 billion in new annual revenue would be necessary to justify the investment, leaving some analysts to conclude that at this rate, “anything less than spectacular results will look like failure.”

As if in illustration this principle, after AI chip-maker Nvidia—perhaps the only AI company selling a real product at a real margin to real customers—delivered a solid earnings report last week, CEO Jensen Huang lamented that “the market did not appreciate it,” and explained their oddly precarious position as follows: “If we delivered a bad quarter, it is evidence there’s an AI bubble. If we delivered a great quarter, we are fueling the AI bubble.”

That is, despite the weak outcomes, dubious math, and fears of a bubble, the money keeps pouring in.

Why?

Until recent updates, LLMs had a strange trait: Despite all their sophistication, they could not reliably perform basic arithmetic.

Could this be true of their creators and owners as well? Technical geniuses, who, having risen to such grand economic heights by manipulating numbers, nevertheless remain unable to comprehend the difference between red and black ink?

Unlike in the DotCom era, however, the current heads of Big AI are not technical savants, innocent of business sense: they mostly have MBAs, not engineering degrees. Presumably they are shrewd capitalists, or else they wouldn’t be there.

What, then, would motivate them to ignore basic economics and put their respective enterprises at risk?

Just blitzscaling, bros?

It could be that this is the natural outcome of what investor and PayPal mafioso Reid Hoffman called “blitzscaling,” an update on the classic, uh, philosophy of “move fast and break things.”

Beloved by Silicon Valley, blitzscaling is a strategy of prioritizing speed and scale at all costs, in which one seeks to establish a first-mover advantage and use that to dominate a market. The “innovation” here is that unlike past eras, the goal of becoming profitable comes after the company has become indispensable. Amazon pulled this off. Uber did not.

Because self-sabotage of the AI industry seems unlikely to be intentional, Dotson’s analogy to German windmills doesn’t line up perfectly here, but the rapid scale-up is surely a means of squashing the competition. Outside of China, few companies possess the resources to build their own models, as training costs can exceed hundreds of millions of dollars. Now that many gigantic models are available for free, fewer still will be inclined to even try.

At this point, the playing field has effectively been cleared out. Most “new” AI apps are just an interface-layer on top of someone else’s model.

By choosing to unleash ChatGPT without notice (in the middle of the school year, no less), Big Tech’s dubious Wunderkid (and possible murderer) Sam Altman, intentionally or not, set off a competition for pole position in a race to dominate the AI market. As he and his competitors see it, now they have no choice, business-wise, but to push as far as possible as fast as they can.

Unless he has an ace up his sleeve, however, by going all-in on scaling LLMs as fast as possible—despite growing evidence scale alone is not all you need—Altman, even as he attracted unprecedented sums of investment, may have simultaneously dug himself a deep bed that he and those following his lead will eventually have to lay down in.

That would be too bad for the AI industry, sure, but also for the average American. The journalist Stephen Witt, author of a recent book on the rise of Nvidia, currently the most valuable company in the world, estimates that it holds 8% of the value of the S&P 500, with another 7% in Microsoft.

Given that Big AI constitutes Nvidia’s principal customer base, and that American’s 401Ks and IRAs closely track the S&P, bursting the AI bubble would shave off a significant percentage of millions of Americans’ retirement savings.

It could be, then, that blame for the next recession—or depression—could reside with the “blitzscaling” ideology and business culture of Silicon Valley.

“Too Big to Fail,” by design?

It could be that “too big, too fast” is a deliberate strategy taking advantage of this recent-but-close tie between Big AI and the broader American—and hence global—economy. Ignoring serial liar Sam Altman’s claims to the contrary, Big AI firms, with their enormous (and dubious) valuations, might be confident they are already “too big to fail.”

That confidence might be misplaced. Commentators have pointed out that the analogy to Big Banks circa 2008 is tenuous. Whereas banks were highly leveraged, Big AI is not. That means the banks held more debt than assets when the crash came. But if OpenAI was bankrupted tomorrow, the cash-rich tech firms it owes money to would remain solvent.

In other words, if OpenAI, Anthropic, and other cutting-edge AI startups were to faceplant hard and fast, it is unlikely to upend the global economy.

Moreover, whether or not we agreed with them, in other cases of “too big to fail,” the companies involved could at least make a case that their products were intrinsic to economic life: Everyone needs cars and insurance! But Big AI firms, whose products kill American jobs, will have a much harder case to make.

This may be why the industry’s leaders have rushed to demonstrate fealty to President Trump. After promising absurd sums of investment in America, they likely feel sure of a Big Bailout, when and if the time comes. Apparently, vibes matter more than mathematics in these rarified social strata.

In the tech tycoons’ much-publicized meeting with Trump, Meta CEO Mark Zuckerberg apparently made up his pledge of $600 billion on the spot, though he has since showed further fealty by doubling-down on the number, and actually appears to be on track to hit that goal. This, despite his investors’ legitimate concerns. On a recent earnings call, when one analyst pressed him about how, exactly, Meta would make money from the investment, Zuckerberg underwhelmed his rapt audience, saying merely that, “The research is going to enable new technological capabilities to exist…and then those capabilities can get built into all kinds of different products.”

As to what those products might be, however, the young titan of industry did not elaborate. Perhaps unsurprisingly, the stock fell by almost 9% following the call.

It could be, then, that responsibility for the next recession—or depression—will lie at the feet of elites who exploit the close ties between Silicon Valley and Washington DC.

Where’s the beef?

Despite impressive technical advances and the release of a plethora of tools to generate “AI slop” at scale, Big AI has yet produced no “killer app.” Even the AI coding tools (basically autocomplete for programmers) have flopped. They were presented as the breakthrough necessary to make the myth of the “10x engineer”—someone who performs the work of 10 colleagues—into a reality. That has not happened. In fact, as a Google executive recently admitted, most AI-written code does not make it to production.

If AI disappeared tomorrow, the kids might gripe about having to do homework on their own again, but nothing essential would be missing from American life. And besides, kids don’t pay taxes.

In my next post, I’ll consider two more explanations for the continued investment despite bad math and lack of products; both irrational, but one highly plausible: the so-called “AI Arms Race” with China, and hunt for “superintelligence,” the Holy Grail of AI.