LLM tools have become very rapidly widely talked about among all people involved in software. In amongst this there is a lot of fear. LLMs have reached the tipping point on the automation curve where they're both cheap and good enough that avoiding them all together just doesn't make practical sense.

Understandably, many people are worried that their skills are soon to be useless and they might be better off going into a different career sooner than later. I won't deny that this might happen. The machines may outstrip us in intellectual and creative capacity. But if this happens your problems —and society’s— will be so bizzare relative to all previous human affairs that it's hard to say what you should do to be best placed afterwards. So instead I'm going to focus on what to do before that, while current trends continue, by talking about previous instances of automation.

When power looms were invented the quality of the fabric they produced was to put it mildly, shit. Power loom fabric was thinner, stiffer, less durable, and had a limited range of design complexity of patterns it could accomodate. The work of handloom weavers continued to be softer, more plush, more durable, and accomodate extremely complex design patterns. It continued to be this way for a long time. In some respect, forever actually. Handmade clothes still bear the reputation of being more durable, more comfortable and so on, and they have that reputation rightly. So why did the power loom win ?

The power loom "won" because the unit price of the fabric a power loom produced was significantly lower than handloomed fabric, and the fabric was good enough. This is why now almost everyone on earth in their day to day lives wears clothing made from machine made fabric of lower quality. It's cheaper. It's cheaper to buy a shirt which will come apart in a few years than to pay for a skilled practictioner to make one by hand.

Unless you are in the upper deciles of wealth you simply cannot afford anymore to buy handmade clothes, or if you can it will only be a few items of your wardrobe. And so the vast majority of fabric on earth is produced by engineers operating power looms, not handloomists.

All This Has Happened Before, But It Will Not Happen Again

Following the fabrics analogy: at different points in time, a profession may find that it is at first an "engineer operating labour saving marvel" then later find out that over time it has instead become a handloom artisan, able to create a product far greater than anything else on the market but only viably at a pricepoint that is too high to stomach for the vast majority of the market.

Before there was a general-purpose computer, a “computer” was a person. Compute is a verb, close in meaning to calculate. A computer was someone who computed. Then we invented a machine that computes and called that machine a computer. There were many complex computations that needed to be done before the invention of the computer machine. In it's absensce Computers, the people, filled whole office blocks. Day by day they executed mostly repetitive algorithmic steps to turn some value into another.

Before machine computing the movement, storage, and processing of information of all kind was carried out by hand, again by huge teams of people and filing cabinets filling whole office blocks. This includes all bureaucracy. Because these tasks were executed by people those people could intelligently skip steps, route around processes, use their intuitions, do research, and all the other functions of which human beings are capable, in the execution of their aims. The service they performed was very good.

The computer machine did away with the human computers. But that computer machine could not intelligently skip steps, route around processes, use intuition, do research, or any of the other functions that Human Computers did before it. Therefore the software solution was worse along many dimensions than the skyscrapers full of human beings and filing cabinets at executing on the information processing aim that replaced them. But like the enigneer with the powerloom and the handloomist the unit cost was low and the result was good enough. That's all, merely good enough.

With the progress of LLMs it may not be true anymore that a skyscraper of humans executing some bureaucratic process would produce a better result than a computing machine, but it was true for the entire period of software engineering until now at least. If you find the automated bureaucracy is lacking today your only recourse is to find the human beings still left in the system who can use their intuition, intelligently bypass processes, do research, and so on, to achieve the desired outcome despite the automation.

Since the invention of the general purpose computer, if you want to process information at you scale hire an engineer to operate it. This person is like the handloomist, an expert. They understand in great detail how the computer works, how it can be used, what is possible. But now, or very soon, the LLM will produce software which is good enough at a unit cost far below the software engineer. So what will happen to the software engineer ? Do they become the handloom artisan: niche, high quality, and mostly displaced?

How Does One Make Software ?

In order to create a piece of software you need to know, or be capable of finding out, what it is you want to build. When the software engineer created programs to replace human computers, they first needed to understand the process those humans followed. This trend continued. Software engineers, or those that manage them, spend a significant proportion of their time figuring out in detail what exactly it is they want to achieve. Broad and vague goals don't cut it. If your problem is of any complexity it must be broken down into smaller chunks and details on how exactly we will instruct the computer must emerge. This is often called Requirements Gathering. I have a different name for it. Design.

The division between a problem statement and its solution is in my view often artificial. What and How are inextricably linked. To me this seems obvious as reliably and repeatably requirements gathering and design fold back into one another throughout a project. Design work uncovers requirements; therefore design is part of the requirements gathering process. After all if you did not uncover all of your requirements, the requirements gathering process is definitionally incomplete.

Today, if you follow software engineers on social media who are adopting LLMs, you can find many examples of people getting stuck. They provide requirements, spend tokens to produce software, only to find they're left with some kind of mess: glaring security flaws (see the fiasco's with moltbook), inflexible code which is hard to evolve, or a product which simply does not do what they intended.

But not everyone is having this problem. Who are they ?

One group doing especially well appears to be experienced software engineers, especially those with managerial or system-design experience. The more experience, the better. They know what they want and how to express it. They can specify requirements in detail and anticipate what a good solution should look like.

When such people use LLMs seriously, they often already understand most of the path from beginning to end. Under those conditions, an LLM that generates valid code at small to medium scope is less likely to go badly wrong, because the operator catches and redirects errors quickly or never stumbles down the bad path at all.

Only in cases where a genuinely novel algorithm is required will they falter. But they are capable of recognising when they are in this position and have the skills to design it themselves or find the right expert.

So coding with LLMs produces the greatest dividends for the same person who was already the best software engineer. This is wholly different to the handloomist, or almost any profession that has been almost entirely automated away. In the absence of a genuine artificial general intelligence take-off, which I by all means do not rule out, an enterprise that needs software solutions still needs experts in general purpose computers to cover truly novel problems.

So the unit cost of software is sure to crater, like the power loom it will win the market with "good enough" software. The demand for niche, high complexity, or high performance software will likely still accomodate a class of handloomist software engineers. Unlike historical handloomists, those experts themselves may also be significantly productivity-boosted by LLMs.

Put this way the shift seems more like early shifts in software engineering such as compilers, higher level languages, integrated development environments, and the like. Powerful tools which accelerate efforts in a manner more multiplicatively than additively. The hard part has always been knowing what you want to do and how to do it (from the bottom to the top of the stack). If anything the premium on that knowldge intensifies.

Only if these systems become genuine active learners, or the pace at which new models improve continues to accelerate, will we see a genuine step shift like the first industrial revolution.

How Much Software Can You Eat Anyway ?

In the example of fabric production, and the automation of other physical products, demand is bounded. You can only have so many clothes. Land is scarce, therefore storage space is scare. In reality most people reach the limit on how many articles of clothing they desire well before they run out of storage space.

Similary mechanical agriculture is bounded: people can only eat so much.

Software is not like fabric or food. Software mostly provides services. Consumer demand for information services is far less bounded. It's limited on the consumer end by knowledge and curiosity more than anything else.

You may think that it's limited on the consumers time as well. There are only so many hours in the day through which to recieve information. But imagine that you had a genie who can do any cognitive task indefinitely, at high quality and also he can produce copies of himself. What would you do with him ? Sure you'd have him fill out your tax returns, market research for that car or house you're even idly thinking about buying. But he could think about anything, anything at all.

He could be modelling each of your competitor enterprises, down to simulations of individual employees. If you had a free genie you would likely have him performing many more tasks than this, including ones that would appear esoteric to us now, even if you never intended to see the output. Simply on the off chance that you might want it in the future, and you could task the genie with the task of modelling and monitoring yourself in order to give the results to you if you ever do want it without you needing to request it. Costing you none of your time at all.

LLMs are more like the genie than our current information services. They are bounded by compute and the energy production to operate that compute. There are physical limits on energy production. Doing any physical work produces waste heat. Any body can only radiate up to a certain amount of heat before it becomes a runaway hothouse. But at current trends that limit on earth will only be hit well after you have died. In the medium term it's likely that the unit cost of LLMs will remain so low that demand for information services is free to explode.

And if it continues to hold that effective use of LLMs depends on knowing what you want and how to build it, that knowledge will continue to be economically valuable.

For now at least, knowledge is still power, so don't stop learning.