When I talk to other software engineers, there seem to be two camps developing when it comes to LLM assisted coding. Some are over the moon happy about it, and say it improves their productivity by an order of magnitude. Others are fairly indifferent to slightly negative about it, saying it is more of a nuisance than actually helping, inventing APIs that sure would be nice if they existed etc.
Finding myself in the latter camp, I just can't deal with the unpredictability of the tools. Both in terms of reproducibility, but also in terms of continuousness. Some tasks it does well on, but it's near impossible to tell whether it will do the same for a similar, related task. To me, the one advantage and disadvantage computers have is that they do exactly as told, and a probabilistic tool breaks that and just feels buggy.
The other thing I can't deal with is that it seemingly doesn't lend itself to iterative, local approaches. Every query is a new thing, that may discard everything it had before, and you can't really hone in on the thing you want.
Add to that that it often suggests things that might be reasonable, but aren't the way I would have done it, and I like doing things the way I would have done it, so I'm going to replace the entire suggestion wholesale, and I end up with a tool that feels at best only marginally useful, if not detrimental to writing code.
Sophie Schmieg /
npub10q…r4hg8
2024-06-01 07:59:25
Author Public Key
npub10qaqswg07tsrztrauf69ylup6qj50kef7e4ut7t0vganygk0t2rsyr4hg8Seen on
wss://relay.nostr.bandPublished at
2024-06-01 07:59:25Event JSON
{
"id": "d89be2610052dd81850a33bf2d590d41ee2a4853a0ce994dd3a6d3880be97bfe",
"pubkey": "783a08390ff2e0312c7de274527f81d02547db29f66bc5f96f623b3222cf5a87",
"created_at": 1717228765,
"kind": 1,
"tags": [
[
"proxy",
"https://infosec.exchange/@sophieschmieg/112540304348931248",
"web"
],
[
"proxy",
"https://infosec.exchange/users/sophieschmieg/statuses/112540304348931248",
"activitypub"
],
[
"L",
"pink.momostr"
],
[
"l",
"pink.momostr.activitypub:https://infosec.exchange/users/sophieschmieg/statuses/112540304348931248",
"pink.momostr"
]
],
"content": "When I talk to other software engineers, there seem to be two camps developing when it comes to LLM assisted coding. Some are over the moon happy about it, and say it improves their productivity by an order of magnitude. Others are fairly indifferent to slightly negative about it, saying it is more of a nuisance than actually helping, inventing APIs that sure would be nice if they existed etc. \nFinding myself in the latter camp, I just can't deal with the unpredictability of the tools. Both in terms of reproducibility, but also in terms of continuousness. Some tasks it does well on, but it's near impossible to tell whether it will do the same for a similar, related task. To me, the one advantage and disadvantage computers have is that they do exactly as told, and a probabilistic tool breaks that and just feels buggy.\nThe other thing I can't deal with is that it seemingly doesn't lend itself to iterative, local approaches. Every query is a new thing, that may discard everything it had before, and you can't really hone in on the thing you want. \nAdd to that that it often suggests things that might be reasonable, but aren't the way I would have done it, and I like doing things the way I would have done it, so I'm going to replace the entire suggestion wholesale, and I end up with a tool that feels at best only marginally useful, if not detrimental to writing code.",
"sig": "167bc10712b9cb38cc0fa249173b85fcd538db0a95c765ce47eee27aa58ff9bbaa83a428f76d92c1c22a54fa2ef1e8145f07d57780fa5db676510e3519377f1e"
}