Don't be a meat proxy
73 points by antonmedv
73 points by antonmedv
don't you feel that we're being pushed to meatproxying, or some form of higher order meatproxying? That seems to be the only way to collect the productivity gains that OAI/Anthropic are trying to sell us.
Sure, but I'm also pushed to wear Nike and watch Netflix and eat McDonald's, and I do none of these things ever.
But you are either tracked or "punished" for not consuming them. :(
Tracked? Probably. I ignore it. I am not rich or a celebrity and have nothing really worth stealing.
Punished? Not at all. I feel better for not doing it. I feel free. "Brands" are toxic, and by avoiding them, life is improved.
I'll buy and wear T-shirts for bands I like, because I like them. I'll also wear free T-shirts with products I like or respect, but I do not and will not pay for them. As for clothes with clothing brands on, I'd rather go naked.
That's not what happens with workplace-mandated vibe coding though.
TBH I'd be looking for a new job, ASAP.
Good luck finding one. I am job hunting currently, and in every interview I've had so far a point has been made about how "we don't write code manually any more".
Definitely frustrating. Hang in there. I'm at least thankful there seems to be a trend of companies really into LLM usage that are willing to state this up front in the job description. "We use LLMs in our development daily, and we expect you to as well." Sweet, thanks for not wasting my time.
I think there's a material difference between being a meat proxy (great term) as defined in this article and working at a company where most code is written by agents.
Most of the code I produce is written by agents now, and I feel completely engaged with the process. I'm also careful not to forward agent slop onto other humans: slop is allowed in branches, but by the time it needs review by someone else or lands on main it needs to not be a waste of time for another human to look at it.
What this means is still shaping up. I expect a lot of coding agent users haven't found a great balance yet - I'm still trying to find the right balance myself.
Fundamentally though I believe that being a meat proxy or a slop canon is a choice, and so is choosing to produce lower quality work faster using AI. We can chose differently. I think employers that try to enforce the lower-quality-but-faster thing will soon be outcompeted by companies that chose higher-quality-and-faster.
That's an incredible way to miss the point, your shot landed on a whole different planet to the barn wall I've put up.
I don't want to drink leaded gasoline
Have you considered that guzzling leaded gasoline is not necessary? You can just tastefully sip it through a straw! And moreover, we are yet to figure out the health implications, and I just love the extra performance my car gets!
Anyways.
I think employers that try to enforce the lower-quality-but-faster thing will soon be outcompeted by companies that chose higher-quality-and-faster.
The low-quality-but-faster has already been winning for pretty much as long as software existed, why would you expect that to change now that you can go even lower, and even faster, and also reduce head counts?
The only way this can change, is if the whole economic system that's incentivizing this way of doing business collapses, or is forced to change by regulatory pressure.
The low-quality-but-faster has already been winning for pretty much as long as software existed, why would you expect that to change now that you can go even lower, and even faster, and also reduce head counts?
Because it's so much cheaper now to do things well.
Teams used to skip out on automated tests and good documentation and experimental prototypes and load tests and security reviews because each of those things took significant time, and software engineering is fundamentally about deciding how to invest the limited time resource that is available to your team.
The time needed for all of those things has dropped - and the value provided by those things increases the faster you want to go.
Smart companies can chose to invest their resources in practices that help them ship better software, faster.
(And yeah, if your starting position is that using coding agents is equivalent to drinking gasoline I'm afraid this profession may be a frustrating place for you over the next few years.)
Smart companies can chose to invest their resources in practices that help them ship better software, faster.
This just feels out of touch with how large companies operate and the incentives that actually lead to a lack of testing or quality. Everyone can see that the most common response is to ship slop faster.
Do you consider Microsoft or Meta or Google or Amazon to be "smart companies?" Do you think some upstart is going to outcompete them because they have a thousand AI-generated unit tests?
Edit: To expand on this - developers don't write tests because they're not something that the developer or company ascribes as much value to as other things they could be doing, like adding more and more features. If they suddenly have more time, they're going to spend it doing more of the things they value, not going back to do the things they don't.
Everyone can see that the most common response is to ship slop faster.
That's why I said "Smart companies can chose", not "Smart companies will chose".
The best developers I've worked with value tests. They're the developers who land "senior engineer" roles at good companies.
I think there is a sizable chunk of the top levels of our craft who understand what good software engineering look like and are ready to use coding agents to help them invest more in those practices.
And honestly, saying "yeah but every company is awful and will write bad software regardless" doesn't mean WE can't choose to write good software.
What's with this 'WE' business? I don't run the company. Heck, I've had a "proof-of-concept" I was writing put into production by my manager without my knowledge.
I'm also reminded of this quote: "Consistent mediocrity, delivered on a large scale, is much more profitable than anything on a small scale, no matter how efficient it might be."
Maybe it's due to seniority, but at every company I've ever worked at I've had the agency to shape the way I work.
As long as value was being delivered, whether I wrote tests for every line of code / used the preferred editor / introduced new open source libraries (after carefully considering the license) as opposed to NIH were mostly left up to me and my peers.
I guess you're right, but as you probably know, that goest against the zeitgeist.
It's much harder to resist this than go with it. And there are places where this is a mandate, not an option, and people cannot really decide not to use it; If perf review is coming, and everybody else is using it, it's hard not to do it.
that goest against the zeitgeist.
That is almost my hobby.
As I said above: if my job mandated AI, I'd be looking for a new job ASAP.
Not only from taste, but also because that company is utterly doomed, and soon, and any sane worker needs an exit route.
Doomed is the right word. LLMs are crazy good at destroying companies from the inside.
In medium/large organisations for each person creating value with LLMs there are more destroying value. Bob used to only waste one person time with LLMs he can waste the entire organisations time. Anyone can use an LLM make a bad ideas sound like a good idea. I imagine this will lead to insane amounts of productivity loss as the entire organisation ends up pivoting to follow the bad idea of a mediocre VP etc.
Icing on the cake is you pay 1-10 million a month to wreck your own org.
I'd like to believe that corporations that don't do what I think they should are doomed. But in my experience large enough corporations can absorb a large amount of silliness without any real consequences. Meta is doing great despite basically no product innovation and all the money they've wasted on unsuccessful side projects.
Maybe there is some cynicism in me, but I think for some sort of discussion in the corporate world, an LLM answering in my stead may not be that bad.
You know the kind. Your non-technical project manager poke you on Team on a Friday afternoon, and start asking you if you will be done by today, it is obvious that you won't because there are still several days of work remaining. So the PM ask you to justify why you are not done, and no matter what your answer, the discussion go downhill because the "the team must commit to finish what it promise to deliver", and he can't understand that you can't think faster than you are doing now.
Nothing I would like more than to plug an LLM on all the classic book on project management, on the nature of probability, and the notes I took on my current issue; and let that LLM very patiently explain to my manager the nature of uncertainty and of risk management and how no matter what someone promise (or more exactly, have been forced to promise) reality always win in the end, and reading through my note, demonstrating that I have in fact been working hard, and despite that hard work, there is still no chance whatsoever the issue will be resolved within 2 hours. Which will likely result in a request for overtime but that would have happened anyway.
Depending on how much of a corporate drone the PM is, that may or may not go over his head, but at least I wouldn't lose my time (and my temper) talking to a brick wall.
I agree–but I do think at the minimum some redundancy/verbosity should be trimmed. Nothing worse than reading a page long message that could have been a sentence. The effort required to trim LLM output down to its core feels not very high, but gives high returns.
I am wondering if pasting LLM output is often a modern version of "let me google that for you". The person answering thinks that the first person should have looked up the information themselves. A variant of this applies to code review: if I link to some reference documentation or section of the programming guidelines I am pointing the author to resources they should have consulted in the first place.
To me it feels different in intent than "googling that for you". Pasting a result/link from google has always been, in my experience, a subtle insult: look how helpless you are, this took my 5 seconds and you could've done it yourself (hence websites like https://letmegooglethat.com/?q=testing).
Repasting slop on the other hand doesn't seem to carry that in my experience, people genuinely think they're helping. It baffles me a bit.
Exactly. In my experience it's not a passive aggressive move to post LLM output. People who tend to do this really believe they are helping.
I think the comparison to a Google search is apt though. Maybe I'll start suggesting that as a way to kindly ask people to stop doing it. "Would you feel satisfied if I sent you a link to Google search results?"
To be fair, link to Google results could show another person something completely different due to search bubbling! Copy-pasted text is not supposed to do that.
(LLM output with a promise that it was checked and it doesn't contain absurd hallucinations could in fact be useful, if the promise is credible)
same can be said for languages with managed memory, or that compile to some VM-bytecode. You dont control it! You dont even understand what it compiled into! Who cares now? AI coding is the new Java from 90s. You just go and make stuff 40x faster and simpler!
AI is a lot less reliable.
I thought this would be about how we are migrating from doing thing with our own hands to asking something else to do it.
But more related to the article, I notice this a lot at work. Instead of explaining why a change was proposed with our own words and imperfections, we paste a long stream from an AI that even mentions every line changed. I admit I do it when I'm too lazy or feel incapable of explaining with more details what happens in a pull request, but at least I try to cut some redundant content.