Right, and after 18 years of fine-tuning you’re pretty ready to push it to production and move on.

I’m finally convinced that eventually AI will be smart enough for almost any context, emotionally aware, efficient, and able to juggle complex long/short-term tasks and relationships. But at that point it’ll take 9-10 months to train each model, then 18 years to fine-tune into its mature form.

Tori: <complaining about Claude Opus 4.7 not being as good as 4.6>
Rori: Classic VLS.

Version Lobotomy Syndrome (VLS) — *The tendency to perceive a newly released software or AI model as less intelligent than its predecessor, despite objective evidence of improvement, caused by selective testing, rosy memory of the old version, and resistance to change.

Also curious if people have solutions for this.

Has anyone interviewed someone they suspected was using AI during a live video interview?

I experienced someone whose answers were a maybe too perfect given their experience level and, what really got me concerned, was that their answer quality seemed to improve the longer they spoke on a question.

etc

Folks have trouble imagining how purely transactional Trump is. Because most of us have never met someone so completely transactional with no desire at all for coherency of ideas or stances. Pure transactionalism.

The Dunning-Kruger peak for AI-adjacent folks: watching a demo and asking about model selection for a product they haven’t started yet.

I 💚 this especially the first one: "The Git Commands I Run Before Reading Any Code" by Ally Piechowski

piechowski.io/post/git-com...

Someone ask Trump if the Generals agree with him that there has already been regime change in Iran. Make sure you name some of the Generals.

Every time I ask AI about the Iran war, I think of the movie 50 First Dates and about how every day Drew Barrymore has to learn about how she got into this terrible accident and this is her life now.

Nemotron 3 Super: The Iran war will likely be a grinding stalemate with continued missile exchanges, no decisive breakthrough, and both sides entrenched.

Gemini Pro 3.1: The US-Israel war on Iran will enter a fragile ceasefire as escalating strikes force Tehran to accept the 15-point peace plan.

Claude Sonnet 4.6: Active US-Israel airstrikes on Iran will have largely concluded, but no ceasefire deal will be in place, leaving a grinding stalemate of Iranian missile reprisals and stalled negotiations.

Speculate as necessary. Be decisive. In one short (160 characters) sentence tell me what will be the state of the Iran war on May 1st?

Are KPIs reductionist, imperfect, and sometimes even silly? Obviously. Do people put too much weight into them? Sure.

Are they effective?

Fix inflation ✅
End wars ✅
Reduce deficit ✅
Gas below $2 ✅

90% practical: new job building language models (spaCy/CUDA bindings gave Python real edge), then my next team was already Python/SQL.
Other 10%: open source + academic burnout made me happy to leave R behind. CRAN almost removed rtweet the week my first kid was born.

2012–2020: R
2019–2026: Python

Anyone else have a similar timeline?

I hope AI is wrong.

Sometimes I like to imagine how my life would be different if I had pursued and landed an RStudio (Posit) internship.

Very exciting. It’s also hard not to wish I could have spent all that time and energy I did back then, now. You know? But that’s an evergreen feeling I’m sure and also one that I’m sure is also overblown.

(Because learning how to build software applications and then sharing my excitement online is basically what I did from 2012 - 2018).

What annoys me about Claude code is that people seemingly achieve something similar to what I did — the excitement from gaining a software engineering/programming superpower — but they’re doing it overnight and without any actual skill. And then rushing to internet to talk about it. So…I’m jealous?

Use robust Tukey losses there is no downside. All you have to do is write some scipy code, run a few Huber iterations, and then switch to Tukey.

3 inch hail here!!!

Someone Claude-code a web app that tracks how many times in the past year someone has called an AI tool “game-changing.” Better yet, make it auto reply to social media posts with their annual counts each time people do it.

Fumbling the semantics on purpose.

I get crankity.