Dull the Cutting Edge with AI

AI can bust out a lot of code. But where will actual improvement come from?

Past

A large-language model, LLM, is created with a ton of ingested data. For coding, where does that data come from today? Source code, docs, tutorials, questions and answers, forum conversations, chat logs. These are artifacts of programmers doing things, creating things, pushing things forward.

These are things that have already happened. In the past, a programmer wrote something, created a pattern, wrote an explanation. AI thinks up nothing new. It knows only about existing connections and can imply connections in the gaps between existing knowledge.

Present

If we accept the line that AI is the new, best way to code, we'll get what has been done in the past. It will not be progress. It will not be better.

LLMs tend toward the statical mean, the most-likely next word in a stream of associated words. This is mediocrity, settling at the middle.

Well, what if we get better at prompting? We'll be able to get the mediocre output faster.

When will the AI get better? When the input gets better, but we're too busy prompting on old input. Garbage out, garbage back in.

The improvements on AI will slow over time. It'll have less to ingest. It'll be at peak knowledge without fresh insights.

Future

Is this our future? I hope not. It sounds lame.

I guess there is at least one good thing about it. We get rid of a kind of churn -- progress for progress' sake. I like that. But no progress, when you're stuck in what feels like mediocrity, feels like a bad place to stay.

In the mid-term, we might satiate our need for novelty and progress with AI itself -- a wealth of new and changing tools and a feverish appetite to use them at every chance.

In the long-term, it's hard to see how we're not awakened, probably after some hard realizations that we've been lulled asleep. The future is to be made, and it will be made by humans, the real agents, the real creators.

In the future, an LLM kiddie will awaken. He, and hopefully many like him, will want to break out of that horrible existence, where he has traded in code creation for code prompting. ICanHazCode? They will release something new upon the world. Other people be excited again (by something that's not just an AI tool) and will contribute, enhancing, trying it before AIs know what it is, creating new examples, new docs, new tutorials, asking new questions. AI can't do this.

Then AI can ingest it. Then AI will have "learned" something. It'll "know" something new. The cutting edge will be resharpened.

But while we're in era of thinking that AI is the gateway to the future, we'll slow and dull the cutting edge of real innovation and creativity.