Models don’t get better just by time passing. The specific reasons for why they’ve been getting better don’t necessarily look like they’ll extend indefinitely into the future.

If the last-mile problems of things like autonomous vehicles have been anything to go by, it seems the last mile problems of entrusting your entire business operations to complete black box software, or software written by a novices talking to complete black box, will be infinitely worse.

There's plenty of low-code, no-code solutions around, and yet still lots of software. The slice of the pie will change, but it's very hard to see it being eliminated entirely.

Ultimately it's going to come down to "do I feel like I can trust this?" and with little to no way to be certain you can completely trust it, that's going to be a harder and harder sell as risk increases with the size, complexity, and value of the business processes being managed.

Even if seniors still do the last mile, that’s a significant reduction from the full commute they were paid for previously. Are you saying seniors should concede this?

No, I think there's cases in which if you can't do the last mile, you aren't entrusted to do the first either.

It's astounding that basically every response you give is just "it can't happen to me". The actual answer to the OP's question for you is a resounding: "nothing".

More or less. I'm just going to enjoy the ride of what becomes of the industry until I either retire or just go do something else with my life.

I already had my panic moment about this like 2 years ago and have calmed down since then. I feel a lot more at peace than I did when I was thinking AGI was right around the corner and trying to plan for a future where no matter what course of action I took to learn new skills to outpace the AI I was already too late with my pivot because the AI would also become capable at doing whatever I was deciding to switch to on a shorter timeline than It would take me to become proficient in it.

If you at all imagine a world where the need for human software engineers legit goes away, then there isn't likely much for you to do in that world anyway. Maybe except become a plumber.

I don't think AGI is this magic silver bullet Silicon Valley thinks it is. I think, like everything in engineering, it's a trade-off, and I think it comes with a very unpalatable caveat. How do you learn once you've run out of data? By making mistakes. End of the day I think it's in the same boat as us, only at least we can be held accountable.

Compilers weren't trusted in the early days either.

LLMs are not the right tool for the job, so even if some marginal bits can be improved, it fundamentally cannot "get better".

The job of a software engineer is to first understand the real needs of the customer/user, which requires a level of knowledge and understanding of the real world that LLMs simply don't have and will never do because that is simply not what it does.

The second part of a software engineer's job is to translate those needs into a functional program. Here the issue is that natural languages are simply not precise enough for the kind of logic that is involved. This is why we invented programming languages rather than use plain English. And since the interface of LLMs is by definition human (natural) languages, it fundamentally will always have the same flaws.

Any precise enough interface for this second part of the job will by definition just be some higher level programming language, which not only involves an expert of the tool anyway, but also is unrealistic given the amount of efforts that we already collectively invest into getting more productive languages and frameworks to replace ourselves.

> Models get better with time, not worse

A decade (almost?) ago, people were saying "look at how self-driving cars improved in the last 2 years: in 3-5 years we won't need to learn how to drive anymore". And yet...