Rob Whelan 🌱

I remember this being really hard — at some point I was using noise canceling headphones so I could still think straight, figure out what I hadn't tried or just stay in comfort mode. It was like my brain was screaming *fix this now!" when in fact there was little I could do

Courage & good work

How I read it:
• hey, a successful company with a local, young & relatable founder is moving to ___
• you should start your business at ___, too, it'll be great

What the business does isn't really the point for their target reader

(but it's frustrating that they don't _say_ who they're talking to)

My theory is that managing complex state is hard, abstractions leak... tiny failures can be worked around but they keep adding up

... and *startup* gets super well tested (it's how every test begins!), but then everything after that slowly gets less & less likely to be covered or even thought about

But wait, let me think about this carefully. The user needs to see the tests pass — that is, they need to observe the command-line output indicating a successful test run. If I think of this request as a communication task, it's much easier to complete. To get a sample of the right output, I can gr–

Human / cat relationships get so complex, I think because both parties are actively trying to train the other to behave as they should.

He might have been okay with Geoff, but "Jeff" — no

In fact, good behavior change research expressly finds that helping people to believe that they are "good parents" and that part of being a good parent can include changing their minds about things like health choices is far more effective than shaming and tying messages to "you are a bad parent"

How do we line this up with:
• resilient systems tolerate human (or agent) errors
• if you have not tested your disaster recovery, you have no disaster recovery
• the chaos monkey goes around breaking random things & your service must tolerate that

That POC agent is a distracted new hire; so...

SIGH

There absolutely is — the same dense terminology moat that the experts use to defend their turf achieves that by making it inaccessible

The incentives do not align super well for actually making the world better

I think it's just an extremely difficult thing to do — ask someone deep in a domain to "avoid words of art" and they do not even know what that includes; it's like saying "change your brain" — only people actively making that bridge can cross it
m.xkcd.com/2501/

Partial counterpoint: I resurrected a few old projects of mine that I wasn't maintaining because it's tedious

and "build & run the tests, fix any issues, repeat" is a loop that AI tools can do better & faster than I can do them.

I can upgrade & migrate rather than just barely keep smthg working

It is definitely possible to get paranoid about actually random interactions in social media, but organized harassment does also happen, no?

A very slide with an extremely "sparkly futuristic computing" feel to it. Thereʻs lots of text in a screenshot from a UI, and text on the slide, which is styled but all fully readable and only uses real words. The slide content doesnʻt matter for this example, but itʻs titled "Going beyond AI chat", and is about automatically running agents to debug incidents.

It does say "NOT A JOKE" in all caps, but I have a feeling this is a joke...? Or maybe it was from a veeery early version of the feature?

I donʻt normally use Beautify slide; I donʻt much like that "drenched in AI" look, and figure there are better uses of tokens. But hereʻs what it actually does:

Yep, means testing is so expensive; the people demanding it are *not* making a financial argument

Sounds right to me — and humans also drown in attempts to work with (let alone migrate) massive codebases without good boundaries.

For any larger changes, they search & spot-read to understand the relevant code, write a plan to markdown, then go through the plan step by step — aggressively compressing context as they go.

I don't think there's any particular limit this approach will hit migrating a large codebase.

I haven't tried a large scale migration; but from working regularly with Cursor, Claude Code: they rely heavily on non-LLM tooling to understand structure & search codebases, never just load the whole thing in context — that's be costly & quite slow.

(This is also an advantage of using a toaster as a therapist, so admittedly it's not a big benefit)

I agree, this is BS

though one positive thing about an LLM therapist: no one would get stuck with a terrible one because they felt unable to reject them

It's worth trying different ways from "write a prompt" especially when that means "ask for something I don't understand yet."

If you're in learning mode, that's not going to help you much; but you can 100% list your goals and ask for common options, to plan & explain the very first step, etc.

Many hotels have extra / alternative pillows in a cupboard or on a shelf; fingers crossed

I'll bet all the clever fellas who realized "if I say a deity demands it... so many more people play along than when I demand it!"

didn't ever stop to think that a few years down the road, kids would be immanentizing the eschaton and causing SUCH a ruckus

Two novels on a bookshelf: I tilt my head to the right to read them: "Never Let Me Go" by Kazuo Ishiguro, "I Let You Go" by Clare Mackintosh

I love hiding jokes like this in the house, but no idea if anyone will ever notice

(to be read in the voice of a movie mob boss)

"Hey claude what does 'drop table' do? I got some advice in a bsky thread, curious if it was a good suggestion"

We can burn off tech debt, be really nimble, be SO much better in sync & responsive to what actual people are experiencing, using what we've built

...but not if we drain the reservoir of human expertise the moment AI tools achieve a mediocrity that's sort of the same shape as what we had before.

The companies who jealously guard their domain experts AND use AI to boost execution are going to blow past them so fast their heads will spin, imho.

Folks, we can have ACTUALLY GOOD customer support, not just switch between barely tolerable options.

When I see companies cutting jobs because "with AI we can do the same with fewer people" I instantly think about the quality of their products, customer service, everything.

They're dropping people with years of understanding of their business & customers — to stand still.

As if that were success.

I think even non-math-inclined guitarists are used to the compromise

...you can tune a guitar with harmonics (vibrate a string at 1/2 length, 1/3rd length etc) but it doesn't quite line up, so you gotta fudge it

But it all works out! Reality is messy & imperfect but we can still find much to love

I am expecting the entire extra inch to be forehead