Compression is when you feed a large language model something big that you want to make small. Summarise this book. Give me the gist of this meeting. Large language models are generally pretty good at this, which makes sense given that they themselves are kind of like compressed artifacts.

Transformation is when large language models convert from one format into another. Turn this audio into text. Turn this jumble of data into structured JSON. A large language model can handle these tasks pretty well. There’ll probably be a few errors so make sure that’s not a deal-breaker.

Expansion is when you give a large language model a prompt to generate something from scratch. An image. A presentation. An email. A poem. This is where slop lives. The output inevitably betrays its origins, glistening with a sheen of mediocrity.

Source: Threat models

This is a great model to think about LLM jobs. It’s also a great way to think about our own jobs and things that require creativity.

Programming is something of an exception to the efficacy of large language models in general. Instead of relying on the subjectivity of painting, poetry, or prose, programming can be objectively tested. Throw enough money at the worst people in the world and they’ll give you tokens you can use to get the machines to test their own output. So you can get a large language model to create something reasonably good from scratch as long as that something is code.

If you had asked me about the threat model of large language models two years ago, I probably would’ve been worried for artists, writers, and musicians. I thought that software had enough inherent complexity to be relatively safe.

Now my opinion has completely reversed. Software is almost certainly the killer app for large language models.

This resonates.

Related follow from the post

I see AI generated text and most of the time, I think it’s rubbish. It’s dull, it’s derivative, it always sounds like a thousand other things I’ve read before. Because the AI has been trained on those thousands of things, all now easy to find on the internet.

But: do I think AI is quite good at making simple software, or basic web tools? Well, yeah, I have tried it for that, and I thought: “Hmm yeah this isn’t too shabby.”

I have not been a professional software engineer for nearly two decades, yet this resonates.

I have a feeling that everyone likes using AI tools to try doing someone else’s profession. They’re much less keen when someone else uses it for their profession. I fall into the same trap as everyone else. I recognise, and admit to, my own bias.

I want to be mindful about the Gell-Mann Amnesia effect when interpreting this. I hear from true? software engineers where the models fail today. However, by its very nature of being able to test and verify, I can accept the thesis that software can be the area that LLMs continue to thrive.

This doesn’t mean the impact is any lower. Software ate the world. Then cloud came and made it easier to develop low incremental margin businesses. LLMs are now further decreasing the cost of software. So, I am increasingly holding the opinion that this is an incremental / sustaining innovation. It still has big implications.

The following is now theoretically possible:

Resultantly, functions that never justified their business optimization can be improved. These are the types of functions that were previously outsourced, either because the company couldn’t justify putting their own teams on, or there were always higher order bits to focus.

This opens up an entire sector of work that was previously squared off in the land of limited amount of software / services that could leverage economies of scale.