AI isn’t going away

This technology is too useful to actually just go away. We’re going to be living with it for the next few decades, at least. LLMs are a fantastically flexible technology. They may not be the best solution for any one problem, in terms of energy efficiency or quality, but they can do an adequate job at a lot of different things. And that’s really attractive, especially in industries where a lot of the work is essentially throw-away.

It’s well-known that 80% of all software projects fail to meet their goals, and that less than 50% of all the code that gets written ever makes it into any finished product. Why even bother to care about whether the code is any good? And social media marketing is based on volume, catching the attention of doom scrollers for 15 seconds at a time. Why wouldn’t you use Sora to create those 15-second advertisements, instead of paying a small army of highly-educated and opinionated creative types to do it?

These days, every industry is one where most work is throw-away. The single-minded focus on quarterly results and shareholder returns have made it impossible to think long-term. If you can’t predict what your priorities are going to be more than 90 days into the future, why expend any effort on making sustainable and repeatable processes for producing anything? I could do a whole blog post on just the collateral damage from the Dodge v. Ford Motor Company decision, and the business-school graduates who’ve had its importance blown all out of proportion by a certain subset of educators. The short version is that there are many things other than share price which are important to running a company properly, and pretending that it really is just one number makes for an extremely unhealthy decision-making process.

Automation costs jobs. Always.

Many years ago, I used to work for a company that did process automation for manufacturing. We primarily worked in the automotive industry, providing manufacturers like Ford and General Motors with process control machines that would allow them to make more car parts, faster, with fewer skilled employees. This was part of the overall contraction of the US manufacturing job market at the time. It’s hard to say with certainty just how many people we put out of work, but as an illustrative point, I know of one production facility that went from 20,000 union workers to just 400 after a full retooling with state-of-the-art robots and other automation. Automating everything raised the minimum quality substantially, but it also put a hard cap on the best quality possible. Because robots don’t care about quality.

At the time, the official story was that the government and industry would re-train those displaced craftsmen to work on the robots, or somehow find a new position in a world that doesn’t value craftsmanship. Unsurprisingly, that didn’t happen. Over the long term, American manufacturing employment did rebound, but not by re-hiring those laid off workers. It was left to their children to find new jobs in a vastly different economy.

AI chatbots are also going to displace a bunch of people from their jobs. As other people have pointed out, the relevant question isn’t even whether the AI can do your job. It’s whether or not an AI salesperson can convince your boss that AI can do your job. And if we’re being fair, I think we can agree that we’re not all doing useful, creatively-engaging work, all the time.

And just like when we were putting lifetime craftsmen out of work in the 1980s, I have already heard AI boosters say that AI will just replace the tedious parts of the job, freeing humans to work on a higher level. That’s just a lie. Nobody, from the AI companies themselves to their customers, believes that there’s enough productive work available for all knowledge workers to command an army of AI bots doing all the actual labor. That’s not a pitch that makes sense to a typical quarterly-focused manager – the real pitch is a 90% (or more) reduction in salary and benefits to those pesky employees.

What should you do about that? I hate to say it, but you probably ought to get familiar enough working with an LLM that you can reasonably claim expertise when you’re looking for a new job. Or seriously think about what it would take for you to retire early. Make a plan for when some snake-oil salesman convinces your boss to fire you. While you’re getting used to working with AI, here are some things to watch out for.

Skill erosion, and draining the expertise pipeline

If you can use AI to do “simple” things that used to be assigned to “junior” employees, where are novices going to get the expertise to become “seniors”? Nobody seems to have an answer for that. We do know that AIs aren’t great at explaining things, and that the people who use them tend to lose the skills that they’re not exercising.

This is a recipe for disaster. I’m sure the AI providers are really excited about a captive population growing ever more-dependent on their tools, but it’s not good news for the rest of us.

A number of educators have already raised the alarm that the use of LLMs is destroying schooling, all the way from elementary school up to graduate studies. Denmark is starting to implement measures to combat cheating with AI in school, and I expect more countries (and in the US, school districts) will do similar things.

Destroying the starter job market and graduating a bunch of kids who can’t think without a crutch, are the kind of long-term consequences that are difficult to rally society to do something about. By the time it’s obvious what the problem is, it’ll be enormously painful to dig our way out of it.

How to avoid losing your edge when the “easy button” is right there, constantly tempting you

It’s going to be difficult, but I have some suggested principles to start with:

• Never use AI to avoid learning how to do something
• Always verify the outputs you get from AI
• Check in with yourself on what you’re choosing to give up in exchange for convenience

Here’s a relevant example from my recent work. I needed to use BigQuery (a system I have no familiarity with), to get the data, to make a business case, to support a proposed change in our product. This was somewhat time-critical, and the documentation for our data models is…not as comprehensive and clear as I might have liked.

I know enough about SQL query language to read it but not to write it quickly, and know very little about this specific variant. So I asked ChatGPT to take an existing query which didn’t quite do what I wanted, and modify it according to the database schemas, to extract the data. I ran it, and it didn’t work.

With a little adjustment, I was able to make it work well enough to spit out the data I needed, in a way that I could further analyze the data elsewhere (in Google Sheets). Good enough for a one-off, and I saved a version of the query which returned the results for the central question. When I need to do this again in a year, the data will be right there.

I could definitely have done this by reading the documentation for BigQuery, experimenting in the query builder, and gradually iterating over time. But this is only the second time I’ve needed to do this in the 4 years I’ve been working for this company.

I went through our actual interface to verify the data for one customer, and it was exactly right, so I know the custom query is doing the same thing as our production code.

It made sense to take the easy route here, since the likely alternative was asking someone else to do it for me anyway. If making queries and reports ever becomes a regular thing for me, I will certainly invest some time in learning it from the bottom up.

The bubble is real, and the crash is going to be pretty bad

There is no way to sustain the investment of money, energy, and water going into the AI boom. In their IPO prospectus, SpaceX valued their addressable market for AI services at $26.5 trillion. That’s…actually completely insane. The US Gross Domestic Product is only $32 trillion. The whole world’s economic output is on the order of $100 trillion. So SpaceX thinks it can capture 25% of all economic activity in the world with their AI products. Whenever I remember that this is something that a company put into their S-1, and that nominally-serious bankers signed off on it, it cracks me up.

Unless you could replace every single skilled worker on the planet with an AI-controlled robot, there’s just no way that SpaceX is going to be 25% of the world economy. And SpaceX is not the only AI company out there planning to be the largest enterprise in the history of humanity.

Collectively, AI companies have bought out essentially all computer memory production for next year. This means there isn’t memory available at the (former) market price for products like smartphones and personal computers. There’s also no memory for all sorts of other products like routers and firewalls. Even products with no memory have to compete for production capacity with factories now switching to DRAM production to chase the gold rush. Memory is on track to have increased in price by 400% over the last two years, after literally decades of ever-decreasing prices.

When you look at the state of “the stock market”, it’s sobering to realize how much of the various stock indexes is made up of the same few companies. The theory that you can “diversify” your investments easily by just buying some index funds doesn’t hold up well when you see that the S&P 500, NASDAQ top 100, and even the Dow Jones “Industrial” Average, all have Nvidia, Apple, Microsoft, Google, and Amazon as the stocks that make up a majority of their value. A major stumble on the part of any one of them could move the stock market as a whole very easily.

I’m not your financial advisor, but now might be an excellent time to actually diversify, into investments that aren’t strongly correlated to the 6 largest US tech companies. Yes, you will lose out on some of the insane growth these stocks are currently having. But then, you also won’t need to try to time the bubble popping. The S&P 500 lost 60% of its value in the 2008 financial crisis. The NASDAQ-100 dropped by 70% when the dot-com bubble burst. Click here to watch How Much Would an AI Crash Destroy? by Patrick Boyle, who actually does know something about markets.

The Great Data Center Build-out is already stalling, even in Texas. Not only do voters hate the noise and disruption of data centers, the utility companies can’t produce and deliver enough power to support them, either. And we are still early in the process of building out all of this capacity that’s being financed on credit, largely via a crazy shell-game of mutual investments between the data center providers, the AI companies, and their common investors. At some point, all of this money has to be paid back, with interest. It’s not clear to me how much of this data center capacity is going to get built, much less used, but it doesn’t feel very stable either way.

And holy shit – the environmental impact is going to be dire. I just watched the latest Pivot to AI episode (not publicly available yet, but I’ll link it here when it is), and he had a fantastic interview with Ketan Joshi, where they talk about how the AI datacenter buildout is single-handedly reversing the transition to clean energy in several countries. And how the bubble popping won’t actually improve that situation much. Burning additional fossil fuels is kind of baked into the next few decades, regardless.

And finally, let’s talk about productivity

From my previous anecdotes, you might be thinking that there is something to the idea that a skilled person can get actual productivity gains from working with LLM assistance. And this is true, but again – it’s not true to the extent the AI techbros would want you to believe. There is no replacing an entire engineering team with a single “super-prompter”. There is no way a non-technician is going to make a viable, scalable commercial product with no prior experience or training. And at this point, they’ve been selling us this story for literally 3 years, with the ongoing promise that if it’s not possible today, then the next big model version will make all of the current limitations obsolete. If you’re still buying it at this point, I have a bridge to sell you.

My experience so far is that you can probably just about double the output of a competent engineer with AI assistance. It’s a bit of a sliding scale, of course. For the things that the AI is really good at, it can be much higher. A lot of the work that programmers do is either mechanical changes in a lot of places, or debugging problems from scanning through log files full of clues. For this sort of thing, the AI assistance can be invaluable. You can have an agent search through the entire source code, find all of the references to something, and plumb a data model change all the way up through the UI in no time at all.

But it’s easy to extrapolate from that to “everything is going to be so much faster”, and I just don’t see that in the work I’m doing. No matter how much detail I provide in my planning documents, or how much course-correction I do along the way, some part of the work needs someone who actually understands the problem, the technology, and the future trajectory of the product, to make good decisions about overall design.

I think it’s possible, maybe even likely, that there will be significant improvements in the ability of AI technologies to do this sort of work eventually. I am extremely skeptical that “more parameters, more context, and more loops” is going to get us very far, though. I think OpenAI, Anthropic, and the rest are committed to this idea, and they are going to drive down this dead end until they’ve burned the entire planet to a crisp, if we let them.

I don’t want to write about this any more

I’m just about done writing about AI. For now. I think.

I might have one more essay in me, a sort of lament on how AI is destroying a job that I used to love. But I believe I’ve covered most of what I wanted to say, and I hope you found it interesting following me on this journey.