“First time”

The graph seems like it only goes back to April 27 and on that day it was 57% bot…

Maybe "first time on a weekday"? Asit seems it's been above 60% every weekend since they started monitoring it.

I think it’s meant as “for the first time in history..”. Not today in particular, but as a milestone.

A strange issue I’ve found is careless use of AI at my job has lead to many people rolling their own incomplete mini parsers. Think YAML parsers of a frontmatter that expects either `key: value` exactly or treats `item1, item2, …` as a list.

It’s a litmus test I use to see if someone actually glanced over what the AI generated.

This feels like a modern version of people writing regex to validate email addresses: employing a complicated, yet incomplete roll-your-own approach as the wrong solution to a problem

The problem is using AI to “push” the answer to an asker. Unless the company has hidden docs they use for support (which why would that ever benefit them), I could get just as good of an answer if I point my LLM at your docs (“pull” an answer). In fact, the response might be better because I have context set up to tune it to my understanding.

Instead, you (company / support agent) have decided that I should instead have a conversation with an LLM through a worse, more opaque harness to the detriment of all of us.

It’s the fat introduced by the process that annoys me the most. The user of the LLM had an idea, but it got greased up and packaged into something that the average person would create, not a specialist in the domain. It dumbs down everything into a single perspective / way of presenting a topic.

The oversized emdashes are chefs kiss

Entirely accurate, but what an absolute waste of resources across the board.

In really large codebases grep and find timeout. If you operate at that scale you quickly come to realize Claude will not use the tools you built to make searching feasible.

It’s a watershed moment. Basically one of the most controlled applications of an LLM into a robust codebase without regard for the implications of doing so.

Anthropic needed something like this and it must proceed flawlessly. My guess is that nothing will explicitly break. But that’s the difficulty of LLM generated code: nothing breaks. You sit with a codebase that swallows all errors and appears to be working. Silently failing makes debugging performance and behavior much harder.

> Is this your post? We don't fall asleep as easily for Write.as subscribers.

I think one of the things I had forgotten about but sheds some more light in my mind about how this was done is that anthropic bought bun.

The change of tone with the author in the capabilities of Claude. The strategy of merging everything at once instead of a more slow, careful cutover. The “single” author story that every company loves to put forth.