@iris_meredith’s latest has me considering in more detail the quality (control) argument for/against the adoption of LLMs.
The analogy of LLMs to the previous mechanization of forms of labour is often made; curiously it’s a position I’ve most often seen from leftist anti-LLM critics, although on the face of it it would seem to support a pro-inevitability narrative.
The classic example from history is the mechanization of the production of textiles. One of the arguments of the historical Luddites was that industrially produced textiles were lower quality than hand-made ones. This may have been true (maybe especially in the early days of factories when Luddites were active) but factories – the ones that wanted to stay in business, anyway – did do quality control.
Some combination of the vastly lower cost of industrially produced textiles plus the quality being good enough to make clothes meant that factories drove individual hand-weavers out of business. But in order for the quality to be good enough while maintaining the lower cost, the time required by, and therefore the additional cost of, quality checking still had to be lower than the hand-weavers’ methods. Fortunately, it was very quick, because flaws were quite easy to spot.
[Interlude about how individual workers bore the cost of the flaws in the product their employers’ machines had spun and weaved omitted. Insert your own Marxist analysis here!]
Now, if you do quality control on your LLM output – ignoring the people who don’t bother with this – is your output really faster at the end? And even if it is slightly faster and still within acceptable quality bounds, does the total economic cost after accounting for your wage plus the cost of LLM compute work out to less than the cost if you had simply done the same work without LLM assistance?
For software development – the space where LLM inevitability discourse is most widespread – I’ve doubted for a while that LLM coding would really be faster if quality checking time is included. Studies are basically showing this to be true.
Maybe the output quality of LLMs will still continue to increase, to a dramatic extent (although actually we seem to have long passed the point of diminishing returns). Maybe the cost of GPU time will also continue to fall (this one seems more likely).
At the very least, software development shops that are investing heavily into LLM development and creating vastly more software as a result are very unwisely betting on Jevons’s paradox before the efficiency increase has actually happened. And even if the quality of output gets better to the point that developers can actually empirically work faster with an LLM than without, and the economic efficiency of LLM use improves: to be a real benefit it has to improve enough that the cost of the GPU time plus the labour cost of the developer using the LLM is lower than the labour cost of the developer alone, working without the LLM. Even sacrificing some quality while remaining within acceptable bounds, I don’t think you’re going to get there.
In a free market economy, ‘getting left behind’ by an ‘inevitable’ technology simply won’t happen unless this condition is satisfied. Yes, the market can stay irrational for a long time, but ultimately management and clients do not care what technology you use to make the software as long as it gets made, meets the requirements, and is cost effective. If you can be more cost effective without an LLM – which seems to be the case at least for teams which work on maintaining large projects collaboratively – then at least outside of a hype cycle, you will continue to win.
There are other ways out of the problem that overall efficiency seems to be worse when team use LLMs, but this is quite long enough already so I’ll just leave it at a single remark. Namely, let’s just take it as a given that the ‘pencils down’ approach of not even checking the code quality will, at sufficient scale, always lead to the same problems that early vibe coders had: eventually the sheer architectural mess of the purely vibe coded software will be so bad that the LLM can’t work with it in any reasonable context window and can’t add one new feature or fix one bug without introducing ten more.