The Modern 'It's Compiling': Waiting on AI Rate Limits
Description
This is a four-panel comic strip, stylistically a black-and-white stick-figure drawing, based on the famous XKCD 'Compiling' meme (number 303). The top of the image has a large title: 'THE #1 PROGRAMMER EXCUSE FOR LEGITIMATELY SLACKING OFF:'. Below this, a red-backgrounded warning message is prominently displayed, reading: 'Message limit reached for Claude 3.7 Sonnet until 3:00 PM. You may still be able to continue on Claude 3.5 Haiku'. The comic itself depicts a manager-like figure opening an office door and shouting 'HEY! GET BACK TO WORK!' at two programmers. The programmers are shown standing on their rolling office chairs, engaged in a sword fight. One programmer yells back the excuse, 'Claude limit'. Upon hearing this, the manager figure, now outside the closed door, says 'OH. CARRY ON.'. The meme cleverly substitutes the original's classic excuse of 'compiling' with the very modern problem of hitting usage limits on a large language model (LLM). For an experienced developer, this is highly relatable; waiting for an AI API to become available is a new form of legitimate, unavoidable downtime, just as waiting for a long compilation was in the past. It humorously captures the evolution of developer workflows and their associated blockers
Comments
26Comment deleted
My productivity is now measured in tokens per minute, and my main blocker isn't the CI/CD pipeline, it's the AI's rate-limiting window
We used to schedule chair-jousts for 45-minute C++ link times; now we call it “exponential backoff on Claude’s rate limit” and invoice it as AI resilience testing
Remember when 'the compiler is still running' was our go-to excuse? Now it's 'I've hit my Claude rate limit' - at least this one comes with an official error message and a timestamp. The real irony is we're so dependent on AI to write our code that when it's down, we genuinely can't remember how we used to do this job without it
When your sprint velocity is literally throttled by Anthropic's token buckets, and management can't argue with a 429 status code. It's the modern equivalent of 'my code's compiling' - except now it's 'Claude's rate-limited' and you're genuinely blocked. The beauty is that non-technical managers understand AI hype enough to accept it as legitimate, but not enough to question why you're not using the fallback model. Peak developer efficiency: finding the one external dependency that's both critical to your workflow and conveniently unreliable enough to justify extended coffee breaks
Funny how compiling used to buy five minutes, but Claude's rate-limit reset at 3pm buys the whole afternoon -- and it still shows up in the OKRs as AI-driven productivity
Claude rate limit: the only backpressure mechanism that justifies throttling your commit velocity without a single SRE alert
We’re no longer compile-bound; we’re token-bucket-bound - the sprint halts until the LLM’s rate limiter refills at 3pm
Real programmers: Comment deleted
How terrible do you have to be in your job to exceed the message limit? 😭😭 Comment deleted
100x engineers use llms 100x more often Comment deleted
Skill issue Comment deleted
no correlation there Comment deleted
True, the title was slacking off Comment deleted
I'm an Android programmer and I almost never use GPT for Android or Kotlin. Even for Gradle - still very rare. What eats up my limit (on those rare occasions when it does) is bash and python (which I never learned properly) - for various dev and CI tools and tasks 🤷♂️ And that is the beauty of these new AI tools: I'd never embark myself on those tasks if it wasn't for ability to get help from AI Comment deleted
Shell is a shoot yourself in a foot language unless you precisely know what you are doing. Like half of the scripts I see have problems in them. Whitespace handling mostly, but other stuff too. I can't imagine LLM getting that right. GIGO If it's for one-off stuff, sure. If it's something you distribute, that's gonna be *fun*. Comment deleted
Define "distribute" Comment deleted
Runs on someone else's files/data. Comment deleted
Well, as I wrote: for various dev and CI tools and tasks We don't ship these or sell them - all used internally in dev-teams or within the company. And they almost always have single responsibility and stable format of input and output, because they are, e.g. linked to their neighbouring tools up- and downstream of our continuous-integration/delivery/whatever. In general I guess we're good 😅 Comment deleted
Yeah. I think the worst you can expect is "oops our test suite was failing all along and no one noticed". It should be sanboxed enough to not nuke anything important. The weird thing to me is that shell isn't even complicated and it's the one thing that's certainly well worth learning from all the obscure UNIX tools. It just made some tradeoffs that are highly unusual from today's POV but functionally it's really simple. Good source is https://mywiki.wooledge.org/ Also you want to be aware if you are writing (POSIX) sh, bash, or some other flavor and what the difference is. Mostly to avoid the old error-prone constructs that are available there for compatibility. Comment deleted
Thank god my fellow colleagues, this company's toolchains and processes are better than that 😅 I wouldn't want to work anywhere near company, where this is possible due to some 70-lines *.sh script being written by careless coder, with or without AI tools 💯 But I hear you - actually learning a tool that you use is crusial. The issue in my case is that I don't use it regularly. More like a once a month or two - for several days (until I get it to do what I need). Imagine you go to a barber once a month and they does not speak your language, but only, let's say, Portugese. You wouldn't learn it to just chat with them once a month. You would learn basic words like "Hi", "Bye" and "Thanks". If you actually would need to give them a big motivational speech 😅 you'd use Google Translate or whatever and show them the translation on the screen. That is what I'm doing with GPT here. Comment deleted
Yeah. That's vast majority of shell use judging by the scripts I see. Try things that look about right and the first thing that works gets used. And once in a while it erases someone's /usr because variable was unset or gives out arbitrary command execution capability. The amount of professional sysadmins that I see using shell like this though is concerning to say the least though. I usually try to push people to use zsh and enable setopt no_unset because that disables the two most dangerous "features" of classic sh. But as it's not something that's installed by default in many places I don't think I've had much success. Comment deleted
There are so many great tutorials/courses out there. Just learn it if you need it. Comment deleted
Imo AI is good for when you do know what you're doing and what you want but you're stuck at actually getting started. It won't give you a finished product first try but with a bit of back and forth you'll get a usable "first draft" as it were Comment deleted
not if you work at anthropic Comment deleted
You may still be able to contribute to Haiku OS. Comment deleted
Back in the day it was "Stack overflow is down" Comment deleted