We automated everything except knowing what's going on • Hacker News Board

Re: We automated everything except knowing what's going on

#71

Post by bluGill » Wed, Mar 04, 2026, 11:41 AM UTC

aetherson wrote: post #68

bluGill wrote: post #57

Earlier quoted context omitted.

I don't know. AI has an understand of some really complex things, but it also does some really stupid things. Depending on which it did most recently for me I change my answer. The question is does AI understand well enough to maintain that thing for whatever maintenance I need to do in the future?

People also do some really stupid things, I'll just throw out. But it also kind of sounds like you're just saying that there's a scalar of "get a little better" before you'd trust AI with something.

Also I'm reserving the right to discover other areas of concern. But yes overall I'm waiting to see if it gets better by enough.

Re: We automated everything except knowing what's going on

#72

Post by Traubenfuchs » Wed, Mar 04, 2026, 12:20 PM UTC

simgt wrote: post #35

Traubenfuchs wrote: post #13

This person has never worked with several decades old government, bank or tax (return) code where all that's ever done is edge cases and implementations of new laws and capabilities being forever bolted on to each other. Systems that were half migrated from a PL/I / Cobol mess to java 7 by Accenture, until the money ran out with the result being that both systems now exist forever and have to be integrated into each…

I agree with all of this, but that's assuming we've reached a plateau. Maybe Claude 6.3 will be able to churn through 10M lines of Java and Cobol, tidy it and convert it to Rust. Or maybe not, but so far the scaling laws are holding.

Oh I am sure with the right prompts and the right chunking and subagents this is already possible today. But converting the old cobol mess to java code that is "most likely doing the same things" is just a fraction of the work.

You need to run those systems in parallel, integrate them in each other or have an abstraction/two-rails layer in front of them, test, test, test and test again, all of which takes months, while new features now need to be added to both systems, whether by AI or some average developer schmuck!

Do you have any idea how much any of this costs? Mind boggling amounts of money. Infrastructure costs will at least DOUBLE for several MONTHS. Meanwhile you will need extra manpower and AI credits for MONTHS on top of your increased existing costs for maintenance and new features.

Welcome to a new hell! LLM can't solve this!

Re: We automated everything except knowing what's going on

#73

Post by simgt » Wed, Mar 04, 2026, 1:32 PM UTC

Traubenfuchs wrote: post #72

simgt wrote: post #35

Earlier quoted context omitted.

I agree with all of this, but that's assuming we've reached a plateau. Maybe Claude 6.3 will be able to churn through 10M lines of Java and Cobol, tidy it and convert it to Rust. Or maybe not, but so far the scaling laws are holding.

Oh I am sure with the right prompts and the right chunking and subagents this is already possible today. But converting the old cobol mess to java code that is "most likely doing the same things" is just a fraction of the work. You need to run those systems in parallel, integrate them in each other or have an abstraction/two-rails layer in front of them, test, test, test and test again, all of which takes months, whi…

You don't need to convince me that this is hell, in particular with all the managers who have never written a single line of code being now convinced that programming is completely solved.