Tag: AI

Posted on
by
The main entrance of the Federal Reserve Bank of Chicago

I don’t usually worry much about investment bubbles. There have been a lot of them over the past few hundred years, and most of them (railroads, telegraph, dotcom…) were expensive disasters largely only for the people who invested in them. Some though, such as the Great Financial Crisis of 2007–2009, were expensive disasters for lots of other people as well. So it’s worth thinking a bit about whether the current AI bubble is of the former sort or the latter—and how to protect your finances in either case.

Bad just for investors

One big difference between bubbles that are going to be wretched for everybody when they pop and those that’ll end up mostly okay except for the foolish investor’s portfolio, is whether the excess investment got spent on something of enduring value.

For example, railroad lines got enormously overbuilt in the 1840s in the UK and in the 1880s in the US, leading in both cases to a stock market bubble, followed by a stock market crash and a banking panic. But (and this is my point), the enormously overbuilt railroads were of some value. As the firms went bankrupt, the people who had over-invested lost a lot of money, but the railroad tracks, rights-of-way, and rolling stock all still existed. The new firms that got those assets, free of the excess debt, were often viable firms that went on to be successes—hiring workers, providing transportation, and eventually providing a return to the new investors. The people who got screwed were the old investors. (And not even all of them, as the original investors often saw the overbuilding happening early and sold out just as the clueless people who knew nothing about running a railroad, but just saw stocks soaring and wanted to get in on it, started piling in.)

Much the same was true of part of the dotcom bubble. A lot of money got spent on a lot of things. To the extent that it was spent on buying right-of-way and burying fiber, there was something of enduring value that ended up owned by somebody, making it one of the less-bad bubbles.

The key to avoiding catastrophe in bubbles of this sort is largely just a matter of not investing in the bubble yourself.

Bad for the economy

But some bubbles have produced horrible, wretched, prolonged difficulties for the whole economy. The other part of the dotcom bubble, besides the dark fiber build-out, was the bubble in companies with no profits and no prospect of ever having profits, whose stock prices went up 10x based on nothing but a story that sounded compelling until you thought about it for 10 seconds. As usual, that ended up being very bad for the people who invested in those companies, but it also was bad for the whole economy, because when those firms went bankrupt, they left behind nothing of enduring value.

The result was that the imagined wealth of those companies just vanished. The stock market went down, which was bad for (almost) everybody, and it produced a general economic malaise, because post-dotcom crash it became hard even for legit companies with real assets, a real profit, and a real business plan for growth, to raise money, which made actually producing that growth much harder.

Really bad for the economy

There is, however a step beyond just pouring a bunch of money into a bubble that doesn’t actually produce anything of enduring value, like a fiber optic network or a railroad. That’s when the money is raised with leverage (i.e. debt).

The 1929 stock market crash was a rather drastic example. People invested in stocks not because there was an underlying business that was worth what the investors were paying for it, but purely because the stocks were going up. That might have been okay in other times, but stock brokers had recently started allowing ordinary people (as opposed to just rich people) to buy on margin—where you just put up a fraction of the price of the stock you want to buy, and the broker lends you the rest.

In the 1920s you could buy on 90% margin, where you only put down 10% of the price of the shares. That meant that, if the stock price went down by just 10% your whole investment was wiped out, and the broker would sell you out to raise money to pay off (most of) the loan. And of course, all those sales into a falling market produced more losses, leading to the crash.

Since the 1930s you could only buy stocks on 50% margin, making it much less likely that your broker will sell you out into the teeth of a general stock market crash—although it can still happen.

Bubbles with leverage

A great example of a bubble with leverage is the Great Financial Crises of 2007. (Most people date it from 2008, because that’s when Lehman Brothers collapsed. I date it from 2007 because that’s when my former employer closed the site where I worked and I ended up retiring rather earlier than I’d planned.)

That was a particularly bad bubble. A whole lot of money was raised, with leverage, to buy housing. But very little of the money ended up being spent to build more housing (which would have been something of enduring value that would have lasted through the subsequent collapse). Instead, the money was spent bidding up the prices of existing housing, which then fell in value after the bubble popped.

So we had two of the classic producers of bad bubbles: Nothing of enduring value created, and leverage. The whole things was made even worse by the structure of the leverage in question.

This is getting rather far from my main point, so I won’t go into much details, but to raise the large amount of money that was going into houses, the rules on housing market leverage were being eased over a period of time. It used to be that you had to put 20% down on a house. Then you still had to put 20% down, but only half of it had to be cash, with the other half being funded with a second mortgage on the property (at a higher interest rate). Then they started letting people put just 3% down. Then they started letting people with good credit put nothing down. Then they started letting people with no credit put nothing down. At the same time, “structured finance” obscured just how risky all those mortgages were, meaning that when the bubble went pop lots of “mortgage-backed securities” ended up being worth zero.

Which kind is the AI bubble?

This brings us to the current AI bubble. A whole lot of money is pouring into building two things:

  • Data centers (buildings filled with computer chips of the sort used to train and run AI models)
  • Large language models (non-physical things that are basically just a bunch of numeric weights of a bunch of tokens which can be used to produce streams of plausible-sounding text)

Each of those may have some enduring value.

Data centers will have some. They will probably have a lot less than a network of fiber optic cables, which can be buried and will have value for decades with minimal cost or maintenance. Since newer, faster chips are coming out all the time, a data center is well behind the cutting edge as soon as it’s finished. Plus, training or running an AI model runs those chips hard, meaning that they probably only last a couple of years (due to thermal damage on top of regular aging).

Large language models probably have even less enduring value, because so many people are training new ones all the time. People are always trying to make them bigger (trained on more data) while also making them smaller (so they can run without a giant data center). All that means that your two-year-old LLM probably isn’t worth what you paid to build it, and a four-year-old LLM probably isn’t worth anything.

That’s how things looked a year or so ago—a perfect example of a bubble that would burn the people who sank money into it, but leave the broader economy untouched.

Sadly, that’s been changing.

First, the structure of the leverage has been changing. It used to be rich people and rich companies were building data centers and hiring software engineers to build LLMs. But lately that’s been getting screwy. Those large companies are raising off balance-sheet money with Special Purpose Vehicles (small companies that big companies create and provide some capital to, that then borrow a bunch of money to make something, with the loans collateralized by the things they’re making—but importantly, not an obligation of the big company that created them). Any particular SPV can blow up, if it turns out that the things it built don’t earn enough to pay the interest on the money the borrowed to build them. And large numbers of SPVs can blow up if financial conditions change to make it harder for all the SPVs to roll over their debts as they constantly have to keep their data centers running.

Second, they’re also engaging in weird circular investing and spending arrangements, where company A buys stock in company B which then turns around and pays all that money back to company A to buy chips, letting company A treat it as both income and an investment, while company B can pretend it got its chips for free.

Finally, there’s all the non-financial obstacles that may well throw a wrench into the whole thing. The fact that LLMs are all built on copyright violations. The fact that running data centers requires huge amounts of power and water (that has to be produced and paid for). The fact that producing that water and power brings with it horrible environmental impacts.

What to do

So, if AI is a bubble, and its one of the bad sort that will produce a panic and a recession when it pops, what should you do?

There are a lot of little things you can do that will help. I wrote an article with suggestions at Wise Bread called Are your finances fragile? It talks about what financial moves you can take to put yourself in a better position if there’s a general financial crisis. (If you’re interested in my writing about this stuff more broadly, I wrote a overview of my perspectives on personal finance and frugality called What I’ve been trying to say, that includes a bunch of links to other of my posts at Wise Bread.)

Besides that general advice, there are also a few things to strictly avoid. In particular, strictly avoid thinking that you can find some very clever investment strategy that lets you make money off the popping of the bubble. Yes, after the fact there will be some investments that make a lot of money, but no amount of keen insight will let you find and make those investments, as opposed to the thousands of very reasonable-seeming investments that will blow up just like all the rest.

Along about the end of the Great Financial Crisis I wrote an article called Investing for Collapse, which explains why any such effort is pointless. It holds up pretty well, I think.

Short version? Avoid debt. Keep your fixed expenses as low as possible. Build a diversified investment portfolio that limits your exposure to the most obviously stupid investments, but doesn’t do anything too weird or wacky in an effort to get them to zero—it’s pointless, and will probably do more harm than good.

Good luck when the AI bubble pops!

Posted on
by

I’m pretty down on AI. I view Large Language Models as a very expensive way to generate strings of words. Other things that were called AI when they were purely speculative (for example, chess-playing computers), quit being called AI once the thing actually existed.

An open notebook and a pen, with a clock behind them.

I know there are other people who find AI useful as a way to minimize the amount of work they need to do, although again I’m very dubious—as soon as you let AI do something for you without double-checking everything it does, it’s going to screw up badly. (That is, you are going to use an expensive, environmentally destructive, copyright-thieving tool to screw yourself up badly.)

It’s from that perspective that I can see this odd, narrow use of an LLM perhaps useful. (I actually got this idea from a fitness YouTuber, but he said he was going to delete his video, so I don’t see much value in linking to it.)

His suggestion was to use an LLM as a journaling tool. Each day, in the morning or in the evening, dictate a journal entry to an LLM. Do a kind of a brain dump of all the things that are worrying you or exciting you. List the things you need to do and the things you’re expecting other people to do. List your good ideas. List your long-term plans and your progress on your long-term plans. List your successes, your failures, your anxieties, interesting facts you came upon during the day, and so on. Then, once you’ve finished your brain dump, ask the LLM to organize those lists: Give me a list of things to do today. Give me a list of blind spots I haven’t been thinking of. Suggest a plan of action for addressing my issues. Tell me if there’s any easy way to solve multiple problems with a single action.

Now, I think it’s very unlikely that an LLM is going to come up with anything genuinely insightful in response to these prompts. But here’s the thing: Your journal isn’t going to either. The value of journaling is that you’re regularly thinking about this stuff, and you’re giving yourself a chance to deal with your stresses in a compartmented way that makes them less likely to spill over into areas of your life where they’re more likely to be harmful.

The things LLMs do—like generating summaries, making lists, expanding on ideas, suggesting alternatives—are really risky in any use case where getting it right matters. We all know the stories of people who used an LLM to write a legal brief, and the LLM hallucinated citations to cases, complete with footnotes, that were all lies. But I think in this case, none of those things matter much.

You aren’t required to have footnotes in your journal, or even to tell the truth. If you ask your LLM-supported journal to suggest blind spots that you’ve missed, and its suggestions are all either obvious or completely off-point, it costs you nothing but the 10 seconds you waste reading the list. If you say, “What are the top five things I need to get done today?” and it gives you a list that doesn’t include the very most important thing, you’re probably going to notice and get that thing done anyway.

Without having tried it, I can imagine that this might be a useful tool, or at least a harmless one. If all having the LLM in the background does is provide a bit of novelty that gets you started journaling again, even just that seems worthwhile.

One risk to keep in mind: If you’re telling an LLM things that are confidential—business secrets, personal secrets, other people’s secrets—you can’t have any confidence that the LLM isn’t going to feed all that info right into its training data and spit it out to some other LLM user. But pretty soon we’ll all be able to train up our own personal journaling LLM that doesn’t share its info with others, so I don’t think that’s a long-term problem.

Of course, I’m eliding the ethical problem of LLMs having been trained on stolen texts, but I don’t think that’s something the users are obliged to try to fix on their own. Rather, the creators of LLMs should be forced to turn their revenues over to the copyright holders of the people whose texts were stolen.

If you’ve been journaling with an LLM, I’d be interested to hear about it. And if you know of an adequately powered LLM that can be run on a personal computer without sharing any data off the machine, I’d be interested in hearing about that too.

Posted on
by

Clearly the move here is to come up with a large language model that writes really good prose that advocates for the policies that you want.

The draft says the department must greatly expand its use of artificial intelligence to help draft documents, and to undertake “policy development and review” and “operational planning.”

—https://www.nytimes.com/live/2025/04/20/us/trump-news#trump-state-department-overhaul

No doubt other people thought of this before me. I wonder how far along they are?

Posted on
by

As someone who’s been paying attention to AI since the 1970s, I’ve noticed the same pattern over and over: People will say, “It takes real intelligence to do X (win at chess, say), so doing that successfully will mean we’ve got AI.” Then someone will do that, and people will look at how it’s done and say, “Well, but it’s just using Y (deep lookup tables and lots of fast board evaluations, say). That’s not really AI.”

For the first time (somewhat later than I expected), I just heard someone doing the same thing with large language models. “It’s just predicting the next word based on frequencies in its training data. That’s not really AI.”

Happens every time.

Posted on
by

I just thought of a possibly actually useful use-case for large language models (what’s being called AI these days): Generating metadata for your photo library.

This is useful, because almost nobody is willing to generate their own metadata for photos. Most people have vast libraries with literally nothing but the date, time, and location captured by their phone or camera, the image itself, and details of the capture (exposure time, ISO, etc.).

Using the date, time, and location info, together with the image itself, AI could:

  • Write a brief description of the image.
  • Tell you where it was taken from (not just the latitude and longitude, but the name of the place where you were standing).
  • Look up if an event were underway at that place and time and say what it was (county fair, protest march).
  • Tell you any number of arbitrary things, like if there was something going on with the weather at that time (blizzard, wind chill advisory)—but only if it was interesting.

I know Google Photos can already do some of this. I don’t think it writes metadata for you, but it will find all of your photos that were taken in St. Croix, for example. (I’d heard that it could locate all your photos of a particular sculpture, but it didn’t work for the sculpture I just tried to find.) In any case, an LLM running on your own computer, saving the data to your photo library, would have all kinds of advantages. There are the obvious privacy advantages, but also sharing advantages—the metadata (or a subset that you selected) would be available to be included when you shared the image with a friend.

Posted on
by

There’s a lot of talk these days about the risks of AI, with many suggestions that it should be “regulated,” but with little specificity of what regulations would be appropriate. As usual, anybody who has an AI loves the idea of some sort of regulation, which would serve as a barrier to entry for competitors.

I have a suggestion that avoids that trap, minimizes the harm of regulation, and yet sharply constrains the opportunities for AIs to do bad stuff. It’s also easy to implement, because it requires little or no new legislation.

It’s very easy: enforce copyright laws.

Any firm that uses or makes available a large language model AI should be required to identify every copyrighted text used in training the model, and then share with all the copyright holders any revenues that the use or availability of the AI brings in.

This burdens existing AIs whose creators thought they were getting all their content for free by scraping the web for it, while giving a big leg up to any AIs that are simply trained on a corpus of text that the AI owner has the rights to. (I read about a physician who had been answering patient questions by email for twenty years training an AI on his numerous messages. He’d be fine.) That seems all to the good.

As to how much to pay the copyright holders, I think the publishing model of the past couple hundred years provides a good guide. Roughly speaking, book publishing contracts proved half the profits to the writer—but because it’s too easy to game the expenses side of the business to make the profits disappear, the contracts are written to provide something more like 10% to 15% of the gross revenues. That would probably be a reasonable place to start.

The huge cost of actually identifying each copyrighted text used, and finding the copyright owner is very much part of the desired outcome here: We don’t want people pointing at that difficulty and then saying, “Well obviously we should be able to just steal their work because it’s too much trouble to figure out who they are and divvy up the relatively small amount of money they’re due.” Making firms go through the process would provide a salutary lesson for others tempted to steal copyrighted material.