If you have had interns who can write better code than Opus 4.5 I would very much like to hire them.

I didn't say the platform was the novel aspect. And I'm getting pretty idiomatic code actually, just based on a bit of example code that shows it how. It's rather good at extrapolating.

PRs where somebody who clearly doesn't know the tech being used well enough, or enough about how the complex app they're working on really works, thus isn't able to determine a good design from a bad one for the feature they're working on, but has AI*-assisted themselves to something which "works", can become an absolute death spiral.

I wasted so much work time trying to steer one of these towards the light, which is very demotivating when design and "why did you do this?" questions are responded to with nothing but another flurry of commits. Even taking the time to fully understand the problem and suggest an alternative design which would fix most of the major issues did nothing (nothing useful must have emerged when that was fed into the coin slot...)

Since I started the review, I ended up becoming the "blocker" for this feature when people started asking why it wasn't landed yet (because I also have my own work to do), to the point where I just hit Approve because I knew it wouldn't work at all for the even more complex use cases I needed to implement in that area soon, so I could just fix/rewrite it then.

From my own experience, the sooner you accept code from an LLM the worse a time you're going to have. If wasn't a good solution or even was the wrong solution from the get-go, no amount of churning away at the code with an LLM will fix it. If you _don't know_ how to fix it yourself, you can't suddenly go from reporting your great progress in stand-ups to "I have nothing" - maybe backwards progress is one of those new paradigms we'll have to accept?

Here is a sample

We are talking about a "stupid" tool that parses a google sheet and makes calls to a third-party API

So there is one google sheet per team, with one column per person

One line per day

And each day, someone is in charge of the duty

The tool grabs the data from the sheet and configures pagerduty so that alerts go to the right person

Very basic, no cleverness needed, really straightforward actually

So we have 1 person that wrote the code, with AI. Then we have a second person that checked the code (with AI). Then the shit comes to my desk. To see this kind of cruft:

  def create_headers(api_token: str) -> dict:
    """Create headers for PagerDuty API requests.

    Args:
        api_token: PagerDuty API token.

    Returns:
        Headers dictionary.
    """
    return {
        "Accept": "application/vnd.pagerduty+json;version=2",
        "Authorization": f"Token token={api_token}",
        "Content-Type": "application/json",
    }

  def delete_override(
    base_url: str,
    schedule_id: str,
    override_id: str,
    api_token: str,
  ) -> None:
    """Delete an override from a schedule.

    Args:
        base_url: PagerDuty API base URL.
        schedule_id: ID of the schedule.
        override_id: ID of the override to delete.
        api_token: PagerDuty API token.
    """
    headers = create_headers(api_token)

    override_url = f"{base_url}/schedules/{schedule_id}/overrides/{override_id}"
    response = requests.delete(override_url, headers=headers, timeout=60)
    response.raise_for_status()

These people fail to produce anything meaningful, this is not really a surprise given their failure to do sane things with such a basic topic

Does AI brings good idea: obviously no, but we knew this. Does AI improves the quality of the result (regardless of the quality of the idea): apparently no Does AI improves productivity: again, given this example: no Are these people better, more skilled or else: no

Am I too demanding ? Am I asking too much ?