The First Software Jobs AI Will Replace Are... 

You're onto something.

Basically very constrained systems, and searching in a database and doing pattern matching. Constrained: data analysis, transportation, manufacturing, financial services Search: customer service, healthcare diagnostics, legal and compliance

Very simple legal documents can be generated using AI, but not any contract I would use in a business I own or manage. For example, a typical non-disclosure agreement is two to three pages. The non-disclosure agreement I wrote is seven pages plus a signature page. I write contracts with the same level of detail, solid logic, and "error handling" to retain most of the agreement if any one provision is held to be unenforceable. My NDA was submitted for review by one of the top law firms in the region and I received only the comment that the agreement would not be binding without a quid pro quo, a fact I was already aware of. People who accept mediocrity in all things are placing too much confidence in generative AI. The current state of AI limits it from replacing all but menial tasks like customer service. As for manufacturing, that industry has been using increasingly sophisticated robotic controls for decades to automate discrete processes and to maximize the efficient use of material, equipment, and labor to ensure the right products are made at the right time to meet customer demand. Likewise, the transportation and logistics industry has also been using algorithms which optimize efficiency based on cost and schedule constraints. Generic AI models are not going to be more efficient than dedicated software applications which have been refined over the past 40+ years. There are legal issues yet to be considered about using AI to replace healthcare diagnostics, financial services, and legal/compliance. No underwriter is going to issue professional liability coverage for even the most advanced AI currently available, and no business would put it to any of these uses without such insurance.

"we will never fly to the moon or have more than 640kb in our computers"

@@HoD999x LLMs aren’t the only AI techniques. Maybe a sufficiently complex LLM could gain enough emergent properties to do it. However, I’m placing my money on different techniques that do the things mentioned in my comment more efficiently, simply, and cost effectively.

Separation of concerns already exists, not everybody's a full stack developer. I don't see why the programming paradigms would stay static when given the ability to have specialist agents perform the tasks for far less money.

I use Copilot with Scala, and it works how I'd expect it to: autocompletions that follow the codebase style or in a functional style. I think LLMs will still be able to generalize no matter the source language.

@ Maybe Smalltalk training datasets help in that case.

I use Copilot with both Java and Clojure. The Clojure is output is definitely worse - usually not as I would want it. Even for Java, it tends to have bad variable names, so I have to go through and rename things.

@@bobbycrosby9765 We get used to magic really fast. Soon we will say: Why do I have to push that button to generate my project?

Must have learned that from me!

@@lodrnr ha ha

They can write if they are told to write.

@@danilomenoli thats a recipe for reward hacking

neither do humans....