When AI Day of Reckoning?

The world has invested lots in AI over the last few years, and many expect a crash soon. Most attempts to use AI in firms seem to be failing. But is that just what we should expect from early applications? When should we look where for clearer evidence that recent AI is or is not going to justify its investment?

Since 2023, it has been widely reported that LLMs seem quite good at coding, and that this seems their most promising application area. Global spending on software is ~$1-2T/yr, so potential big saving there soon might plausibly justify last year’s ~$0.5T/yr AI investment.

For many decades the demand for software has been large and elastic. Most firms have many software projects they’d like to do, which they don’t do mainly due to cost. And market leaders tend to be firms whose software investments went well. So if the total cost of software fell by a factor of two, total spending on software should more than double. That’s supply and demand. And AI getting a big cut of that increased spending soon might justify its investment.

Of course fielding useful software involves many tasks, including identifying opportunities, securing funding, overseeing projects, defining requirements, marketing, user support, and writing, testing, and maintaining code. Just making code writing much cheaper doesn’t obviously make total software cost much cheaper. Much depends on just how many of these other tasks can be made cheaper as well.

If AI is going to have a big impact, when should we expect to see it? Software projects typically take ~6-9 months from conception to delivery, though orgs can take 1-3 years to reorganize workflows, incentives, etc. to accommodate new techs. Legacy software may not be replaced for up to 3-10 years.

So it seems that one should expect to see substantial AI-driven changes to the scale of the software industry within roughly 3 years of a widespread consensus that AI makes it much cheaper. Which is about now if that consensus happened 3 years ago. Or in about 3 years if that happened one year ago.

The number of U.S. software workers increased by ~50% in the last decade, probably mostly due to falling costs. So we should expect an even faster growth in software spending if AI is in fact causing a big increase in the rate at which its costs fall.

And if we don’t see such a big increase in the next few years, that will suggests that AI does not actually cut software costs nearly as much as advocates hope. Which is of course the usual scenario for hyped new techs. And should lead to a crash.

Of course that’s the short run; we might still plausibly see a “general purpose tech” impact that takes several decades to play out, as we’ve seen previously for techs like steam, electricity, personal computers, and the internet.