LLM Code Generation

Some public discourse around LLMs is quite polarising. This is usually a sign of some emotional arguments making it in to a real discussion.

There are a few real issues with using LLMs for certain tasks, and there are others that are under used. Like most things the answer is not clear cut. When thinking about problems with LLMs, I am usually less worried about how people are applying it.

Ultimately, the positive cases are quite over blown, and the negatives are under played. The reality is that most are focused on what LLMs can do, and few are thinking about whether we should.

LLMs are good at generating tedious code

However the tedious code is usually the stuff you shouldn’t write. Extract the tedium so the next person to make a change doesn’t need an LLM to change it.

LLMs remove friction in discovery

Moments of friction are what trigger learning, growth and enjoyment. Just like how short form content is alluring, but hinders the users ability to give attention to long form content.

This often comes up when asked why I don’t use , with all the new features. The reality is that when your editor has a refactoring button, you never feel the pain of having to change 20 files with similar code. It’s a good indicator of when something needs a little more work. I don’t need LLMs to tell me how a project works, because I can read the code, and the WTF moments stick in my mind when I am debugging an outage months later.

It finds better results to my queries

This is an indictment of how bad search engines have become, and not how good LLMs are at searching.

Security

LLMs offer up a massive surface area for attack vectors.

Privacy

Gapping holes in privacy are already being found and it’s still early days. ^1

Energy

Energy costs are extremely under represented in arguments regarding LLMs. We had this problem with Crypto, and it’s the same here. The real cost of asking an LLM “what is the best type of toilet paper?” has been externalised. Externalised in terms of the subscription cost vs real costs, and externalised from a raw energy point of view, as environmental impact is not included in any cost representation.

And so…

LLMs need a “cost of proof” style improvement in efficiency and a first principles investigation on it’s 2nd and 3rd order impacts on society as a whole.

I don’t use LLMs very much yet, but every month or two, I give the latest round of tools a good college try. Nothing has jumped out as being worth the cost yet (not the price). So for now I will continue to work as I have done for the past 20 years. Using Unix, lightweight programs, reading reference manuals and thinking with my own electrified meat slab when I have problem.

ai

references

1. https://cse.engin.umich.edu/stories/power-hungry-ai-researchers-evaluate-energy-consumption-across-models