A classic. LLMs can help if you know what you’re doing on your own, at the cost of having to burn tokens while basically having to guide it to the result you need. The second you use it for something you’re not an expert in, you enter pure slop territory. The fact is that LLMs are constantly hallucinating but with guardrails that allow them to sometimes hallucinate the correct result.
And yes, I’m in tech and have experience. The only way it’s going to “take” my job is if it crashes the entire economy because the bubble pops or an executive needs to pump some stock from braindead morons on wall street.
I’m a programmer. There is a narrow lane of things I will trust it to assist me with, and none of that without review. I’m not worried about my job in the slightest, barring a significant leap in the tech that I frankly do not see happening any time soon.
All these people building vibe coded products are building houses of cards that are more likely than not security nightmare black boxes.
Claude has a place for simple tasks. It does a great job of being an “advanced find in files” or being a “smart boilerplate generators” but anything remotely more complex and the issues start to show, really quickly. When writing new prototype code, it does very well. But when you have to handle all sorts of edge cases, it doesn’t do it as well. It also doesn’t do a good job with debugging anything more than surface level. Opus does slightly better, but even then gets into the frequent “I found the issue! oh wait, that’s fine. I found the issue! Oh wait, that’s fine” loop over and over again.
And before you ask: Yes, I’ve been using Copilot CLI for work pretty regularly for the past couple months.
Aside: It doesn’t help that the true token costs are off by a factor of 100-1000. Yes, I know general reporting is saying the breakeven is 10x, but…well, you’ll have to trust me that’s not accurate.
Nah
It’s very funny to see the level of denial people are in.
Honey, I’ve seen the horseshit LLMs churn out. My job is safe.
Are you using it properly?
A classic. LLMs can help if you know what you’re doing on your own, at the cost of having to burn tokens while basically having to guide it to the result you need. The second you use it for something you’re not an expert in, you enter pure slop territory. The fact is that LLMs are constantly hallucinating but with guardrails that allow them to sometimes hallucinate the correct result.
And yes, I’m in tech and have experience. The only way it’s going to “take” my job is if it crashes the entire economy because the bubble pops or an executive needs to pump some stock from braindead morons on wall street.
I work in software on automotive safety-critical software. I left LLMs in the trash.
So yes, I am using it properly.
I’m a programmer. There is a narrow lane of things I will trust it to assist me with, and none of that without review. I’m not worried about my job in the slightest, barring a significant leap in the tech that I frankly do not see happening any time soon.
All these people building vibe coded products are building houses of cards that are more likely than not security nightmare black boxes.
Claude has a place for simple tasks. It does a great job of being an “advanced find in files” or being a “smart boilerplate generators” but anything remotely more complex and the issues start to show, really quickly. When writing new prototype code, it does very well. But when you have to handle all sorts of edge cases, it doesn’t do it as well. It also doesn’t do a good job with debugging anything more than surface level. Opus does slightly better, but even then gets into the frequent “I found the issue! oh wait, that’s fine. I found the issue! Oh wait, that’s fine” loop over and over again.
And before you ask: Yes, I’ve been using Copilot CLI for work pretty regularly for the past couple months.
Aside: It doesn’t help that the true token costs are off by a factor of 100-1000. Yes, I know general reporting is saying the breakeven is 10x, but…well, you’ll have to trust me that’s not accurate.