Fred Hebert

(I realize I posted quickly and it almost looks like a dismissive review. It was useful but also it is a solid book with good advice if you find yourself owning or helping with MySQL databases, would and did recommend!)

I did read the book, it was useful!

It feels like there’s lots of stories behind these two words.

A flowering white sweetclover plant in front of a wooden fence

I’ve sowed white sweetclover for the last two years in the strip of soil next to my house and they’ve never even managed to sprout there.

Yet these fuckers are growing wild right outside my fence, taunting me.

Yes! The issue could be considered “live” but idle.

Alternatives would be: need time to think, let me take a break and tackle something else for now.

For people with full agentic setups: spin up a new workspace and have AIsolve issues in parallel.

In accounting my own time, I would exclude that review time from my own work time, but not commenting and addressing the feedback.

From experience and charitable interpretation, I would say they could overlap when you have to account for the review cycles in issues, meaning you write code, submit it, and while other people (not part of the study) take their own time to provide feedback, you start working on something else.

You know videogame bad guys are software folks because it starts with a tiny issue (your hero) & they keep delegating to straightforward minions but as the problem grows they have to churn ever fancier solutions until final stage where they must fight you themselves but by then you’re way too strong

When I was reading that I sort of instantly mapped it to personal experience going “oh yeah if I know exactly what needs to happen I’ll be far more critical of whatever code is generated, and more likely to give up and take over”

just gotta load up on context, understand the problem and what's already done so you can write a good soluti—

Oh yeah, I also had noticed that not all tasks were completed and not in the same proportion for all devs so there’s something hiding in there too that screams for analysis

Oh it isn’t. I think your comment pointing out its more of a field study than a controlled trial probably is the best criticism I had trouble articulating.

(Also large code bases are known to be more challenging for current AI agents, so this also plays on the possible outcomes)

(Of course ultimately it doesn’t measure actual productivity but whatever they find seems to be based on more realistic/situated observations than what I expected)

While looking for task completion after real review.

This made it sound much much nicer than commit or PR counting in very large populations or through toy problems

Interestingly enough this is why this study felt better to me than many others trying to measure productivity (though it look like a field study): they used real tasks in real repos people were familiar with with more realistic instructions (use as much ai as you see fit), and asked for estimates—

My “just post” to “actual writing” pipeline relies on noticing that once I’ve ranted and left comments about a thing enough times, it’s time to turn it to a blog post I can point at and then move on

Yeah they mention providing training, but unsure what it was.

The questions that came up to me while reading were about whether learning/productivity could be non-linear (stepwise? S-curve?) beyond the 50h mark and what would trigger the improvement it if so.

They reported not finding such effect in most participants (all <50h experience, regardless of time buckets), except one individual with >50h of exp. with the cursor setup who did better than the rest at estimates.

I think it's not clearly defined but from the comments reported it seems to be interpreted as "ended up fixing more/unrelated things than what the initial task specified" by the participants.

I first wrote this about managing to have EKS properly shut down web-fronting nodes talking to an ALB by using their node-termination-handler with ASG lifecycle hooks + SQS queues from their event bus. I re-posted it today due to some IAM issues, and another peer finding weird cases in R53.

I finally came to the realization that this was never supposed to work. It is meant to make me feel stupid. Someone at AWS successfully built a set of features whose sole purpose is to trap engineers into hating themselves, a self-service bullying machine, and they got me.

Throwing the Poor Crumbs Isn’t “Increasing Access” From crypto speculation to Chatbot "therapists" to "tiny homes," our media routinely accepts the premise of manufactured austerity to justify a cruel caste system.

I can’t claim to have come up with the phrase as a whole. I think I first encountered the core idea in this article by Adam Johnson and more or less remixed it with the more recent conversation.

Expressing stigma and inappropriate responses prevents LLMs from safely replacing mental health providers. | Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency

arxiv.org/abs/2504.18412 for the pre-print.

Looks like it ended up in the ACM in June and is full open access now dl.acm.org/doi/10.1145/...

Skimming through a preprint on the risks of using LLMs as therapists, then reading tech commenters who were even less thorough.

Aside from “the poor demand harmful crumbs”, there’s a surprising amount of folks claiming they’re too informed to be harmed—the “not me, I’m built different” defence.

Might be ‘angst’ as defined by the existentialists?

Distancing through Differencing

Probably am at the point where coping with events by going “maybe this will finally make people react” doesn’t work anymore.

“In 1845, 24% of all Irish tenant farms were of 0.4–2 hectares (1–5 acres) in size, while 40% were of 2–6 hectares (5–15 acres). Holdings were so small that no crop other than potatoes would suffice to feed a family.“

What could go wrong?

Computer-vision research powers surveillance technology - Nature An analysis of research&nbsp;papers and citing patents indicates the extensive ties between computer-vision research and surveillance.

New paper hot off the press www.nature.com/articles/s41...

We analysed over 40,000 computer vision papers from CVPR (the longest standing CV conf) & associated patents tracing pathways from research to application. We found that 90% of papers & 86% of downstream patents power surveillance

1/

using this to talk about llms and adaptive behaviors

part of what people mean by "chat isn't intelligent" is that core pieces like "noticing that gassing your homie up is hurting them" are simply not present