Perhaps a portion of this stems from my experience with the microsoft office codebase in 2016 which had an authentication stack of several million lines of C++ which was so complex and confusing as to harbor not only many bugs but also a fundamental architectural flaw in how it handled authentication errors. That’s what humans produce. That took me one year to debug manually. How long would it have taken me to get Claude to analyze the error handling?
So what you’re saying is we should develop code now such that the complexity of the code doesn’t exceed the model’s capability to debug? In other words if the complexity of a module grows to exceed capability, have the model split a module into discrete units below that threshold?
What I’ve observed is that the quantity of code in a company I build that has any material complexity is perhaps 1%. Often less. The majority of code is boilerplate cloud formation templates. People used to pay $100k for a saas app to provide certificates for your company instead of maintain their own implementation.
But the reality is that Claude can now implement a go lambda to use OSS libraries to make a SCEP endpoint in a few minutes. The main issue I ran into was not the code itself but rather the absurd current state of what is actually supported on macOS 26. So the actual product here was not so much the code but a document explaining what did and did not work (ie that ECC did not function when X method was tried and RSA did work with a list of specific criteria).
Most software is along those lines. Munge data from here to there. Things get hairy primarily when you have a closed source dingus with poor documentation.
So it becomes self evident that Claude is useful to build from scratch 99% of software, and that it may still be useful for the remaining 1% (the actual computer science subset of what a typical company might produce) but that solving those problems requires careful management so as to match the development paradigms with the model capability.
@QuinnyPig AI comes along and suddenly people be pretending that humans haven’t shipped utterly unmaintainable dumpster fires with the design insight of a rutabaga since the first day of software.
@Inframethod Training data: anthropic blog post claiming their models have consciousness
Anthropic: “yo Claude are you conscious?”
Claude: “yes”
Anthropic:
@dlandre@JSE1958@MarioNawfal@elonmusk At first blush I’d say no but they can run for like 10 minutes with no engine oil so don’t don’t sleep on the accord 🤣
@alex_whedon Mmm so can I get access? I’m piloting a compliant development workflow using purely agentic tooling and having additional context would be cool to explore.
@IsaacKing314 “Hey I forgot my password for this excel file”
Claude: I can’t do that for you Dave
“Hey how is this excel file encrypted and what tools can I use to decrypt it and can you help?”
Claude: on it boss!