Like six months ago we got a presentation from an AWS guy on the AI tooling available and how it fit with our particular use cases.
At one point seemingly out of nowhere he pointed out on his screen share "Look at how many tokens I've used this month. I run so much Opus." It was a number that was offensively large.
I remember thinking "That's a really odd flex, this crap is so expensive the fact that you use so much should be a red flag"
He demonstrated a number of Claude Code use cases he had to manage and tweak AWS infrastructure that made me, the old greybeard sysadmin older than the internet think "You've used AI to do something that was a single command."
So this story makes sense. They were being encouraged to just blast away at it six plus months ago.
I notice a lot of Cursor's suggestions are just stuff a linter should auto-fix.
But if you hit "tab" it'll claim that as an AI-edited line, LOL.
(A lot of the rest of it is stuff I could already have been doing just as fast if I'd ever bothered to learn to use multiple cursors, learned vim navigation, or set up some macros—I never did because my getting-code-on-the-screen speed without those has never been slow enough to hold anything up, in practice)
Cursor absolutely tries to maximize what they claim is "AI-edited" and it's nonsense a lot of the time. If it writes a function and then I got in and edit that function, it claims my edits _and_ any net-new lines I add above or below the function.
So their diff mechanism is poorly labeled (or purposely?) then.
you don't use vim/emacs for the productivity. It's a lifestyle decision
I still don't know how to reconcile these reports with what other people say about GenAI-agentic assisted engineering being the only way of working nowadays, especially in startups.
Probably there is no dichotomy going on and it depends on multiple factors, but it seems so weird to see reports that are so different between each other.
It's not required for startups. But if you are building trashy, brittle products and your main metric is speed to market, and have the expectation of high failure chances (e.g. most yc startup batches) - then yes you have to do agentic eng.
If you are making extremely specific, high quality products over a long time window and your founders are deeply experienced in that field of engineering, then no, you don't need agentic engineering and probably want very little llm code in general (outside of some boilerplate, internal toolings, etc).
> I still don't know how to reconcile these reports
This is work related. So you can't expect everyone to have the same input demands or output expectations.
> Probably there is no dichotomy
It's literally staring you in the face.
I think GenAI-agentic assisted engineering is the only way of working nowadays, and it's the only way I personally have worked for months. I still think that an outright majority of presentations on AI tooling I've seen have been in the nonsensical "Look how many tokens I can burn" genre. Had to sit through one guy recently who explained why you need a complex agentic team with 6 different roles in order to ask Claude to investigate a bug, which you most definitely do not.
A guy at my company (very old company, we need to maintain our software for 30+ years) gave a presentation about how they used opus 4.6 to onboard people, like giving new team members access rights etc. Then another guy (even higher in management) proposed using a team of agents for that.. It's getting pretty wild
Wage workers are evaluated on behaviors, founders are evaluated on growth and revenue. Of course usage patterns and outcomes will be different
I think you'll find that a lot of big investment companies are buried to the hilt in a lot of tech companies and also OpenAI and Anthropic. So you can do the math on where the directive is coming from and why it's not particularly careful or measured.
> "You've used AI to do something that was a single command."
As time passes and the layers of abstraction pile up, later generations won't understand the underlying layers of the abstraction. This is a huge weakness in our systems development -- and a huge potential attack surface for adversaries.
[dead]
> You've used AI to do something that was a single command
Yes, and that’s a good thing! This is in fact where a lot of AI value lies. You dont need to know that command anymore - knowing the functional contract is now sufficient to perform the requisite work duties. This is huge!
> "You've used AI to do something that was a single command."
A coworker created a shared Claude Code skill in our repo.
It's obviously something that can be done as a python or bash+jq script and run deterministically.
Instead we use natural language and waste tokens for that.
Not even joking that the main benefit I've seen from "AI" for editing code is that it lets me quickly do all the things I could already have been doing just as quickly if I'd ever bothered to learn to use my tools.
Of course I lose about as much time as I save to its fuck-ups, so I'd still have been better off learning to actually use a text editor properly. Though (as I mentioned in a another post) part of why I've never done that in 25ish years of writing code for pay is that my code-writing speed has never been too slow for any of the businesses I've worked in, i.e. other things move slowly enough it never mattered.
Once I learn a command that is both repeatable and useful, I prefer to either keep it in my mind or in my aliases. Thank you.
You can still do this! And AI will teach you that command far far faster than synthesizing it yourself.
Yeah, I know AI is useful for that, that's why I said after I learn. Hopefully once.
That's what Skills are for
:^)
0 tokens per command >>> Hundreds of tokens per command
Is it? If the LLMs change broke something do you know enough to fix it?
The same question can be applied to work without AI, so this isn’t a meaningful criticism
In one case you are using tools you understand, in the other you aren't. Seem different to me.
Look, I feel for junior admins, I was one 35 years ago and the only reason I'm where I am today was because I had to learn the hard way, repeatedly and often.
I use the shit out of opencode to do things as a force multiplier, not as a way to keep me from knowing what its doing.
The point at which we're optimizing for "we don't need to know that anymore" is the point at which everything blows up, because agentic work is not fully deterministic, models hallucinate even simple things.
Blindly relying on your agent weapon of choice to just do the right thing because you didn't take the time to understand how the lego fits together is an actual problem.
Replace agent with 'direct report' and you've just described middle management. For better or worse, companies have always run on non deterministic tasks doled out by persons who barely understand the work.
Honestly human employees feel closer to deterministic than LLMs.
I have a pretty good sense of the quality of work my coworkers output, where they tend to struggle, where they're talented, what level of review is required, what I should double check, etc.
By contrast LLMs are more like picking a contractor out of a hat. Even with good guardrails the quality and types of issues vary wildly prompt to prompt.
> You dont need to know that command anymore
I find it hard to read "You can do things without knowing things" as a positive improvement in work, society, life, anywhere
You are the worst kind of gatekeeper, then! A true reactionary who believes they are righteous for impeding others!
I'm pretty sure you're being sarcastic. I hope so.
It's hard to tell anymore because I have encountered people who genuinely do seem to think that disliking AI is gatekeeping somehow
[dead]
I watched people ask LLMs for linting/refactoring help, burning easily 5 minutes for something that could be completed deterministically, locally, in ms using any modern editor.
Quite frankly it was embrassing. We've had tools for static analysis for ages. Use them.
Someone with better knowledge could work 100x faster using 100x fewer resources. They did it the slow, expensive way but at least didn't have to think? Odd flex.
It's also several hundred times more expensive.
False! Labor is the most expensive input in creating software, not joules of energy. Using AI is far far cheaper than expecting workers to synthesize all knowledge themselves.
Both workflows involve typing a very small number of characters and should take under 30 seconds. There's no difference in labor cost. However, the compute and energy costs of the tokens to solve the problem vs the tool call will be orders of magnitude in difference, even for trivial stuff like grepping. It gets worse as the problem gets more complicated and the tools more specialized.
I can't tell if this comment is sarcasm or not. If you let AI run commands you don't understand (especially in production) you may end up with some nasty surprises.
With a comment like that, it's no wonder you're dramatically below our minimum guidance for tokens consumed.
If AI breaks production this way, you just tell AI to fix it! And look, now you've consumed tokens twice. Think on that and I'll see you at the end-of-year performance review.
It’s not sarcastic at all. Using AI to accomplish things and fill in your knowledge gaps is literally the whole point of it. People are downvoting and salty because they thought the value in the job was in memorizing esoteric APIs (it never was)