Christian Minich

Christian Minich

@christianminich.com

Dad, data, process, philosophy | Token Alchemist @Vercel | My wife says my headstone will read “sorry I’m late”

3.9K followers2.1K following2.4K posts

Just finished the Hyperion Cantos. Nothing matched the breathtaking poetry of Hyperion, but I enjoyed Rise of Hyperion a lot. And found the conflicting visions of parasitic technology and empathic humanity to be prophetic for the age we live in.

My kingdom for a “humans” internet and an “agents” internet. Human + AI mixed are near unusable by humans. For example, recently I was trying to read comments on one of my pull requests. It was a sea of Code Rabbit, Cursor, Claude, automated security review, agent soup. 1/2

A lot of it was useful, but it also meant that I missed the human comments in the PR because my ape brain was overwhelmed. More and more areas of the internet will look like GitHub, just as agents had their first major successes in coding and coding tooling. I don’t think we’re ready for it. 2/2

New career goal: get hired by every video conferencing software company and fix it. First feature: “Grandparent mode” — hides the hang up, mute, and video options behind a Captcha

Christian Minich reposted

Replying to @emilyriederer.bsky.social

Sidenote: this all illustrates an recurrent fascination of mine -- how often the blocked to good work is never having gotten to see "what good looks like" like if you've never seen a of datamart or an orchestrator, this feels reasonable. And DE has far fewer public examples than DS 12/

I never got a dopamine hit or flow state in coding. Maybe I never dug deep enough, could be a personal failing. But I do get a dopamine hit from a beautifully designed process, a boring and functional architecture. Basically I was AI native before it was cool

I really don’t get the hate for slop UI elements because they’re slop. Text? Deep hate, because it’s an expression of intent from the author (and because standardization is often a bug not a feature). But like isn’t the intent of a red callout box the same everywhere?

Maybe it’s because I’m not a designer and if I was I’d hate it like text but .. isn’t a UI the one place where “average” (in this case: de facto universal) exactly what you want?

I’ve seen people say that agents will make high level programming languages obsolete. Like why should Python et al exist if LLMs could just write Rust, C or even Assembly?

I think it’s interesting that agents write Python all day long in their cycles. Maybe that’s because of training data, maybe that’s because humans made them that way in the process or within the harness. It’s hard to say, but I think it’s an observation in tension with that prediction

Amazing examples of what the level of data engineering is on average at many, many, many organizations

Emily Riederer@emilyriederer.bsky.social

Doing some volunteer data eng work and their infra is... interesting (probably built by a series of random vols over time) 📊 Not pointing fingers, but sharing a running list of antipatterns for your amusement and as I keep track for a future thing I want to write up 🧵

One of my life’s greatest joys is seeing my kids enjoy things I enjoyed as a kid. Recent example: checked out 2 beautiful adaptations from the Epic of Gilgamesh and read them to my 4 year old yesterday. He enjoyed it so much he asked to read it today 🥹

Short, to the point technical reports are both a flex and a sign of deep respect. “I would have written a shorter letter but I didn’t have the time” has never been more true.

Youngest daughter (1.5) is obsessed with Moana and so now are we all. People singing Moana, songs playing on Alexas, the whole thing. And I’m not complaining, I missed these songs

The complaint in data science 15 years ago was “I came here to do machine learning but all I do is data cleaning.” In AI it’s “I came here work with LLMs but all I do is auth and context management.” The most effective people see the shit work as the leverage for huge outcomes

Love the description by Jeremiah Lowin of the change in data pipelines from a DAG of tasks to a DAG of outcomes in the age of AI. Declarative is the programming model humans want but has been incredibly hard to pull off.

With AI, how an outcome is achieved is less important than the outcome + the evals and tests that validate that ourcome, and in that way AI makes declarative models more likely.

Co-sign

Emily Riederer@emilyriederer.bsky.social

Last night I tracked down Jenny Bryan’s absolutely lovely 2019 UseR! keynote on Code Smells (worth the watch, btw!) riffing off of Martin Fowler Confronting more AI slop documents this morning, may is suggested we adopt the “Claude smells”?

Got back from a family beach trip today. On the drive I was in charge of music, and we were listening to “Daddy songs” (meaning: no suggestions). The 4 year old doesn’t accept this and asks if we can listen to a song he knows I like, that he also likes. We do Kid knows selling already.

My kids (oldest are 13) won’t have phones for as long as possible but I do have a MacBook Air that I’ve branded the “family computer” that they can go on whenever. Amazing to me none of them know how to type on a laptop