[jrsj]

Yeah I really don’t get why people keep hyping AI like this. It really doesn’t make things go that much faster. At best you’re able to generate prototypes more quickly + get better autocomplete. Nothing particularly revolutionary there.

Anyone claiming a generalized 100x, 10x, or even 2x productivity gain is either delusional or trying to sell you something. Possibly both.

The companies saying they are reducing the size of their workforce because of gains they’re getting from AI are probably just telling investors what they want to hear while cutting costs for the same reason they always have.

[tyre]

I felt this way until Claude Code. It works much, much better in large codebases than anything else I've tried. It implements smaller features, including ones with FE + API changes and tests for each, pretty well. I'm going to try cloning our main repo multiple times to get it working on multiple branches at once.

[JadeRoller]

100%. I just tried it the other day. Game changer

[jrsj]

I will definitely check this out, thanks.

[logicchains]

>Anyone claiming a generalized 100x, 10x, or even 2x productivity gain is either delusional or trying to sell you something

I don't understand how anyone who spends a couple hours or more per day coding new functionality couldn't at least double their productivity with LLMs, unless their organisation prohibits LLM usage. Even just limiting the LLM to writing unit tests would still save that much time.

[jvanderbot]

The thing is, tab complete using LLMs is really great. But I still read it, then press tab, then press enter, then type a few chars, then wait.

Sometimes I get 1 good line from that. Sometimes I get 30. Usually I get 10 bad lines and have to type a bit more to coax out 8 good ones.

It just looks faster but typing was NEVER the bottleneck for coding.

Where it really flies, though, is building tooling around a well known API. FML if I ever have to write AWS CDK or AWS API calls without an LLM again. You're looking at ages of reading through really bad docs to get it going.

For that, which is a 1% task of convenience for most jobs, I can use most LLM output verbatim. But that's like I said less than 1% of the job, and only then when the core software is done.

[pama]

Did you notice how much better things are today (eg Claude Sonnet 3.7) than they were 1 year ago? Don’t you expect things will not improve in the next year? Even R1, a public weights model, can add huge value when left to code in a loop.

[bccdee]

> how things were 1 year ago

Not a substantial productivity multiplier.

> how things are today

Somewhat better than before, but still not a substantial productivity multiplier.

> Don’t you expect things will not improve in the next year?

I expect they'll be marginally better than they are now, but still not a substantial productivity multiplier.

"A huge paradigm shift is just around the corner" is a very popular narrative & it almost never bears out.

[moltar]

Hm, I’m a CDK pro (4y of full time experience). I used all LLMs, except latest Claude model. All were bad in my estimation and just got in the way. I don’t use them for CDK code anymore.

[jvanderbot]

Yeah! That's exactly the thing. It's passable for novices and bad for experts. But I don't need expert level CDK I need an instance to start up. Hate it or love it that's all I need.

[wkirby]

The bottleneck is not putting code on the hard drive, or turning my thoughts into code — the productivity bottleneck is thinking and frankly no LLM is thinking better than an average developer.