This is possibly one of the more cursed single sentences I've ever seen in a job posting
-
Asta [AMP]replied to Cassandra Granade 🏳️⚧️ last edited by
@[email protected] @[email protected] Perhaps that's why they think this will "totally work": why, LLMs are so "natural" and human like! Unlike that awful machine transpilation (which, you know, is auditable and not stochastic and actually works). So they, perhaps foolishly, think of this output as being maintainable?*
* they probably don't think any of this, but I wonder if that's somewhere in their brains as re: how this all works. -
Asta [AMP]replied to Cassandra Granade 🏳️⚧️ last edited by
@[email protected] @[email protected] @[email protected] oh, well. There's that too. I dunno much about COBOL so I don't have that level of insight into the problem.
-
@aud @xris @xgranade there's that for sure. the idea that functions and code blocks should always have well-defined boundaries, not porous edges, had to be invented. any attempt to translate to a modern language needs to grapple with that tension. in our other thread we wrote about our experience writing a mechanical translator for a language called DB/C...........
-
@aud @xgranade it scans as a reasonable choice to me. It has a few desirable properties:
1) It's not COBOL (read as: you can find literally anybody to maintain it)
2) It has a *lot* of operational experience in "this absolutely must not fall over" contexts. See (1), not only can you find a lot of people to work on it, you can find a lot of people whose entire careers are organized around critical support for keeping applications in it afloat.
3) It's perfectly adequate. Not sexy, but fine.
-
Cassandra Granade 🏳️⚧️replied to SpookJ 👻 last edited by
-
Cassandra Granade 🏳️⚧️replied to Irenes (many) last edited by
@ireneista @aud @xris Makes a lot of sense. You could pretty easily prove and test your transpiler, but if you make a single change to the emitted code then that all goes out the window.
-
@[email protected] @[email protected] @[email protected]
where the risk of making ANY change is just enormous
yeah, that kind of gets into why I poked fun of this in the first place: I do not, at all, trust the people posting this to take this risk seriously.
Even if the stochastic element is small, you really just shouldn't go around and introduce a stochastic element in a scenario where determinism is critical. It's just... wrong from the start. -
SpookJ 👻replied to Cassandra Granade 🏳️⚧️ last edited by
@xgranade @aud oh, yea, vote of no confidence in the particular choice to transpile it.
It's a billion-dollar industry if such a tool existed, but… well, it doesn't exist yet for a reason…
And the LLM thing just signals "we are absolutely not serious about this". One can only assume they don't understand how that scans to the target market.
-
@aud @xris @xgranade absolutely agreed. it is fundamentally not possible for LLMs in their current form to do anything about reliability or testing their output against reality, and none of the lines of research we're aware of that are getting venture capital or institutional investment would change that, even if successful
-
Asta [AMP]replied to Cassandra Granade 🏳️⚧️ last edited by
@[email protected] @[email protected] @[email protected] Right? And, well, actually, even the testing is probably difficult without introducing the idea of "reasonable error"... and considering the "reasonable error" for these types of systems is "0. precisely 0. it's a fucking airplane."
welp -
-
@[email protected] @[email protected] @[email protected] agreed. Since this is how they wrote their pitch to developers interested in applying, though, I suspect they may well go bankrupt before realizing that.
-
@[email protected] @[email protected] @[email protected] "we are improving drug delivery through injections by adding noise to our hand motions!"
-
@aud @glyph @xgranade I almost wonder if the target market is actually phase one/investigatory contracts: banking on a stronger bid (because you may trick the fool on the other side with Today's Hokum) and never planning to go beyond that.
COBOL always makes me think "IRS" out of the gate, but I guess there's *some* bank/etc. code out there that might be under the kind of credulous management that might toss a few million at this kind of "maybe it'll work"
Probably giving them too much credit…
-
@[email protected] @[email protected] @[email protected] Honestly, I suspect half the startups throwing "LLM" or "AI" in their company descriptions are basically just an elaborate funding scam in one form or another. I'm sure some business types without a foundation in the field they're attempting to wade into genuinely don't realize the technology literally can't even; I'm sure others will quietly pivot away from LLMs for it (it's possible they're not even seriously considering them); and then some others probably have some sort of money loop where the money goes to contracting all their friends or who knows what the fuck.
-
@[email protected] @[email protected] @[email protected] @[email protected] I should just learn COBOL and then get funding to create COBOL 2: COBOLLIN' since it seems like there's a lot of money to rake in there.
I mean how hard could it be* -
Cassandra Granade 🏳️⚧️replied to Asta [AMP] last edited by
@aud @SnoopJ @glyph @ireneista COBOL has classes now.
