Comment Re:Can't wait (Score 5, Insightful) 338
And you'll likely end up with a horribly architected and poorly structured solution - but it will be in Java, Rust, or Go or the hot language of the hour instead of COBOL - but by the time it's done, that language will no longer be the hot language of the hour and they will have to hire people out of retirement to fix it.
Giving AI the existing system even as a primary source of direction would probably be a mistake. Better to give it the statutes, regulations, policy manuals and, to a lesser extent, training materials for the current system. Perhaps during the validation phase AI might probe the existing system to create test cases to run on the replacement system. If AI can't do that, the "I" is missing in AI (and it is).
No matter what, given the apparent schedule (even with infinite resources - see The Mythical Man-Month: Essays on Software Engineering published around the time I was in college), the whole idea is nuts unless the scope of the project magically gets redefined to be trivial (perhaps a field office app will get rewritten and DOGE will declare victory). Sort of like "Tesla 'Full Self Driving'" or "ventilators" or Hyperloop, or "much more efficient tunneling" done by The Boring Company (which has every large tunneling project beating down their door to get The Boring Company to do the job - NOT).
Of course 240 months is, technically, "months" but I don't think it qualifies as a "few" months.
Musk has already demonstrated that he has no idea how legacy systems work or the risk of assuming they are "obvious" - such as when he seems to have assumed that because the "Dead" flag for a SSN holder in one of the SS databases wasn't set and the person was over 70 they must be receiving benefits even if they were now over the impossible age of 150 years. Then he kept doubling down on that apparently unwilling to learn or just figuring that if he repeated a lie enough times, the facts would get lost in the chaos and he could claim that he was "never wrong".